Dataset Viewer
text
stringlengths 15
7.07M
|
|---|
Title: A deep neural network to de-noise single-cell RNA sequencing data | Body:
I.
Introduction
Recent single-cell RNA sequencing (scRNA-seq) technologies provide invaluable insight into gene expression at individual cell levels. However, they often result in many genes displaying zero expression levels. These zero values can arise from two distinct events: (i) biological events, referred to as structural zeros [1], when a gene expresses no RNA at the time of the experiment, and (ii) non-biological events, also known as dropout events [1]–[3], stemming from the insufficiencies in library preparation and amplification processes [4], [5]. Non-biological zeros occur due to various factors in the employed technology. These factors can include mRNA degradation after cell lysis, variability in converting mRNA to cDNA, inconsistencies in amplification efficiency, cell library concentration or sequencing depth, and other batch effects [3].
Unique Molecular Identifiers (UMIs) are a valuable tool in rectifying amplification biases in non-zero gene expression measurements by identifying and eliminating reads from cDNA duplicates generated during amplification [6]. Despite their utility, UMIs cannot rectify sampling zeros and the corresponding cDNA copy numbers for these zeros remain unidentified [7].
In the context of scRNA-seq data, imputation involves inferring missing expression values or non-biological zeros. The aim is to estimate the actual expression values that were not detected due to experimental limitations. Imputation methods are often developed based on the principle that similar cells will likely have similar gene expression profiles [8], [9].
Noise in scRNA-seq data can arise from various sources, including the aforementioned technical variabilities and the inherent stochasticity of gene expression. Denoising methods strive to eliminate or reduce noise in the data by identifying and adjusting for these sources of variability, making the signal more pronounced. For instance, a common denoising approach adjusts the expression values for each gene based on the total amount of RNA detected in each cell, thereby correcting for differences in sequencing depth [10]. Statistical learning and deep learning models, such as autoencoders, are among the most powerful tools for processing complex biological data like scRNA-seq.
Imputation methods for scRNA-seq data are generally built upon probabilistic and mathematical models. They aim to replace missing values, often zeros, with estimates derived from the existing data points and their interrelationships. To infer the missing values, these methods often exploit the structure and characteristics of the data, such as the correlation patterns in the scRNA-seq data, i.e., the similarities among cells and genes.
One example is the Markov affinity-based graph imputation of cells [11] (MAGIC), which is among the most renowned statistical methods tailored for denoising and imputation of scRNA-seq data. This computational approach denoises the cell count matrix and imputes missing gene expression values. MAGIC constructs a graph where nodes represent cells and edges reflect similarities between cells, facilitating information sharing across similar cells via a diffusion process. Another notable method, SAVER [12], uses an empirical Bayes-like approach to estimate the true expression levels, leveraging expression similarities between cells to impute non-biological zeros. scImpute [8] is another method that finds and imputes highly probable dropout values using a clustering approach. This method draws on information from the same gene in similar cells to the imputed cell. A similar approach DrImpute [13], employs k-means clustering [14], [15] to identify similar cells and then imputes dropouts based on the average expression values in those cells.
Deep learning models have emerged as revolutionary tools in the last two decades due to their unique ability to solve complex problems. With the exponential growth of biological data, these models help determine patterns and draw inferences from large, complex datasets. Deep models leverage artificial neural networks with multiple hidden layers to perform downstream tasks, including feature extraction, dimension reduction, prediction, and classification. They offer novel approaches to tackle and interpret the complexity of this information, providing new insights into various areas such as genomics, proteomics, and cell biology [16]–[18]. Applications vary from predicting protein structure and detecting genetic variants to understanding gene regulatory networks and interpreting scRNA-seq data. As these models evolve, their application in biological research is expected to drive significant breakthroughs and help reveal the complex underlying mechanisms.
Notably, self-supervised models like autoencoders play a prominent role in deep learning. Deep autoencoders are explicitly designed to provide efficient data representations. First, the encoder compresses the input into a lower-dimensional representation. Next, the decoder reconstructs the original data from the compressed representation. Training an autoencoder involves minimizing the difference between the input and the output. Incorporating multiple layers in the encoder and the decoder results in deep autoencoders that can model complex data with a high-dimensional feature space. However, a challenge known as the “vanishing gradients” emerges during the backpropagation stage of training, particularly in deeper models. Because of the chain rule and multiplication of the gradients, there is a possibility that when some of the gradients are small, the multiplication would be numerically too small, leading to no significant learning in back-propagation. Techniques such as batch normalization [19] and residual connections [20] fix the vanishing gradients problem [21]. Deep autoencoders are valuable in different applications, including feature extraction, dimension reduction, and denoising.
Regardless of the denoising approach, discrete statistical distributions are needed to model the count data in scRNA-seq. Various count distributions such as negative binomial (NB), zero-inflated NB (ZINB), Poisson, and zero-inflated Poisson (ZIP) have been widely employed for modeling UMIs or gene reads in scRNA-seq [4]. For instance, the deep count autoencoder (DCA) [22], is used to denoise the scRNA-seq data using an NB model for noise. This approach characterizes the reconstruction error by evaluating the probability distribution of the noise model rather than directly reconstructing the input data. The ZINB distribution introduces a zero-inflation parameter that signifies the proportion of additional zeros, distinct from those originating from the NB component. Similarly, the ZIP distribution also incorporates a zero-inflation parameter, specifically targeted at the zeros not accounted for by the Poisson distribution. Employing a zero-inflated model versus a non-zero-inflated model does not depend on the prevalence of zeros in the data. Zero inflation is a statistical consideration that takes into account the inevitable nonbiological zeros. Zero inflation and its biological and technical validity for different sequencing methods are among the highly debated topics.
Here, we introduce ZiPo, a deep model that takes scRNA-seq data as its input and performs two tasks: (a) imputing the missing data using zero-inflated distributions, and (b) providing an embedding of the single-cell gene expression data in a low-dimensional space, offering a nonlinear transformation akin to PCA’s dimensionality reduction capabilities that can be used in downstream tasks like cell-type clustering. We introduce a regularization parameter that not only adjusts the zero-inflation level but also offers a more refined, adaptable, and biologically accurate approach for handling the intricacies of sequencing data, setting our work apart in this domain. This model can also be used for single-nucleus data, but we only use single-cell terminology in this manuscript for simplicity.
II.
Methods
In ZiPo, we first encode the high-dimensional gene expression data of the cells into a low-dimensional space (e.g., with 32 latent variables) using a deep model. Next, we consider these latent variables as unknown covariates of the scRNA-seq data and recover the expression levels using these unknown covariates together with other possibly known covariates that may influence the gene expression patterns observed in individual cells. Such covariates might include tissue or cell types, sequencing depth, spatial location, age, sex, etc. [23].
We assume a zero-inflated distribution for gene expression levels. The two most common distributions for scRNA-seq data are zero-inflated Poisson (ZIP) and zero-inflated negative binomial (ZINB). The results in this manuscript are based on ZIP, but they can readily apply to ZINB with minimal adjustments. We prioritized ZIP because the straightforward Poisson distribution aligns well with UMI-based sequencing data. This approach eliminates the need for ZINB’s additional parameter that might account for potentially greater dispersion in the model.
The model takes each cell’s measured gene expression vector as its input. The encoder consists of dense linear layers for reducing the dimension followed by sigmoid activation. The model is then followed by parallel layers from the encoded variables (unknown covariates) to predict:
the library size,
the zero-inflation rate, and
the parameters of the desired distribution.
A schematic representation of the model is provided in Fig. 1.
A.
Advantages of predicting the library size
We separate the library size (i.e., the sum of all gene expressions for each cell) from the other predictions because what we want to model in the zero-inflated distribution is the relative expression levels of genes in a cell, not the absolute values, which depend on less relevant variables like the sequencing quality, assay efficiency, etc. To normalize gene expression data, we can adjust raw count data using count per million (CPM) [24] to account for differences in library size or sequencing depth between samples, allowing for fair comparisons of gene expression levels across samples. In this method, the raw counts for each gene in a sample are divided by the total number of reads (or counts) in that sample. Predicting the library size offers at least two distinct advantages over calculating it via adding all gene expressions for each cell: (1) it can help reduce technical noise in the data, and (2) it retains potentially valuable information about the library size in the latent variables that can be used in subsequent analyses. Notably, the library size could convey information relevant to the cell size [25].
B.
Sparsifying of the weights
We expect each feature in the second layer to represent a pathway or an eigengene, incorporating only a small subset of the genes [26], [27]. Therefore, most of the weights of the first encoding layer should be zero or near zero, forming a sparse weight matrix. The sparsity of this layer can be enforced by adding a term to the loss function that acts similarly to an L1 penalty. Not all activation functions can produce the intended sparsity as discussed in detail in section II-D.
C.
The model structure in mathematical notation
Throughout the manuscript, we use the subscript c for any variable to indicate that it is calculated for the cell c. The input matrix is denoted as X = {xcg}, where xcg is the expression of gene g in cell c. Xc denotes the full expression vector of cell c over all available genes. For t encoding layers, the cellular embedding for cell c is formulated as:
Uc=Gt⊕…⊕G2⊕G1(Xc),
where Uc is the embedding for cell c. Each Gi is composed of a linear layer, an activation function (e.g., sigmoid), and a batch normalization layer, which helps resolve the vanishing gradients problem and accelerates the training [19], [28]. We add known covariates of the cells, Uc′, to the embedding vector Uc, yielding the complete encoded representation of the cell,
Vc=concat(Uc,Uc′).
Subsequently, optional layers are used to predict the following variables:
log(lc)=liblog(Vc),
logit(zc)=N(Vc),
log(μc)=Mlog(Vc),
θc(j)=Θj(Vc);1≤j≤p,
where l is the library size, μ is the mean parameter for the distribution, z is the zero-inflation probability, θ represents additional relevant parameters of the distribution, and p is the number of these parameters. In the case of the Poisson distribution, we only have the mean parameter, μ. For the NB distribution, we need an extra parameter representing the difference between the mean and the dispersion of gene expressions. For N, we use a dense linear layer. For Mlog, we use a dense linear layer followed by logSoftmax activation, forcing the sum of the average gene expressions for each cell to be equal to 1.
Given this configuration, the following linear relationships are established:
zc=sig(AzVc+Bz)
μ≈exp(AμVc+Bμ).
To model the library size, we use a series of dense linear layers followed by batch normalization and an activation function. We can readily employ deep models to predict distribution parameters, resulting in nonlinear rather than linear Vc functions in the above equations.
D.
The loss function
The zero-inflated distribution is given by:
Pr(x∣z,l,μ,θ)=zδ0(x)+(1−z)expf(x;m,θ),m=lμ,
where f represents the logarithm of the distribution for non-zero-inflated expression levels with mean m and possibly other parameters θ and δ0 is the atomic distribution on zero. Note that m = lμ is the expected expression level according to our definitions.
We denote the rate of zero inflation by r, i.e., r = logit(z) and the likelihood of our samples, given the distribution parameters by ℒ. The first component of the loss function is the log-likelihood of the zero-inflated distribution:
logℒ=logPr(X∣z,μ,θ)=log∑xcgPr(xcg∣z,μ,θ)=∑xcg=0log(z+(1−z)expf(0;lμ,θ))+∑xcq>0(log(1−z)+f(x;lμ,θ))=∑log(1−z)+∑xcg=0log(z1−z+expf(0;lμ,θ))+∑xcg>0f(x;lμ,θ)=∑logsig(−r)+∑xcg=0log(expr+expf(0;lμ,θ))+∑xcg>0f(x;lμ,θ)=∑logsig(−r)+∑xcg=0r−logsig(r−f(0;lμ,θ))+∑xcg>0f(x;lμ,θ).
A typical approach employed in the DCA method involves using truncated log functions by adding a tiny offset quantity to the argument of the log function to avoid infinity (or very large) numeric values when the argument is zero (or near zero). Although this approach might be appropriate in specific conditions, it is inaccurate when we have no a priori knowledge of the input argument to decide on a proper value for the offset. To prioritize precision in our calculations, instead of relying on truncated log functions, we use the composite function log sig [29], which stabilizes the numerical schema without truncation.
The ZiPo model is regularized in two ways. We apply L2 regularization to the predicted zero probabilities and other distribution parameters. The regularization hyperparameter in ZiPo, αz, controls the level of zero inflation. An excessively large αz prevents the model from using significant zero inflation. Setting αz to zero will yield a model with strong zero inflation that overestimates the expression rates. We also impose a regularization term on the weights of the first layer to sparsify them. We use this regularization term to shrink weights appearing to be less important. We then prune those relatively small weights after ensuring that removing them does not affect the model performance.
We use NLL for “negative log-likelihood” and MSE for “mean square error” throughout the manuscript.
E.
Which activation function and weight regularization approach should be used?
In deep neural networks, activation functions such as ReLU [30] or GELU [31] are commonly employed to tackle the vanishing gradients issue. However, for the encoding segment of ZiPo, we favor the sigmoid activation function. This choice is influenced by the specific type of regularization applied to the weights (parameters) of the first encoding layer.
A limitation of ReLU arises from its linear properties. That is, the model can decrease the absolute weights in the initial layer while compensating for this by increasing the weights of the second layer, without affecting the final output of the second layer and the final performance of the model. Therefore, simple L1 or L2 regularization penalties would not significantly change the model’s performance or output. Some variations of the above activation functions like ReLU6, defined as min(max(0,x),6), could potentially resolve this problem. However, we prefer the nonlinearity of the sigmoid function over the piecewise linearity of ReLU6.
Similarly, the optimizer can lower the weights of the first encoding layer but compensate for them by increasing the batch normalization weights in the following layer. To protect the first layer weights, we use a scale-invariant loss function defined by dividing L1 by L2:
−E[logℒc]+αzE[zcg2]+αwmeanimeang|Wig|(meangWig2)12+E[αθ⋅θ2],
where Wigs are the weights of the first linear layer and α’s are regularization hyperparameters determining the strength of the corresponding penalty terms (i.e., αz controls the zeroinflation rate and αw controls the weights in the first layer and αθ controls the parameters of the distribution).
If the features of the initial layer are sufficiently interpretable, the model should perform well because those features are directly related to the genes. Employing similar regularization across all layers is possible. However, introducing an additional hyperparameter for each layer complicates the model and the training process.
F.
Datasets
We have used three datasets to illustrate ZiPo performance (Table I). The ROSMAP Alzheimer’s disease single-nucleus RNA-seq (snRNA-seq) dataset was generated by Mathys et al. [32], and is available from https://www.synapse.org/ with synapse ID: syn18485175. This dataset includes 80,660 single nuclei derived from the dorsolateral prefrontal cortex of 48 postmortem human brains. The liver [33] and myeloma [34] scRNA-seq datasets are accessible through GEO with accession numbers GSE125188 and GSE117156, respectively. In the liver dataset, 70,706 human CD45+ cells from the paired liver perfusion, spleen, and human peripheral blood mononuclear cells (PBMC) were profiled. The myeloma dataset includes 51,840 cells from 29 diagnosed patients and 11 control donors.
While the distributions of library sizes (i.e., the sum of gene expression per cell) were similar across these three datasets, the total gene expression (i.e., the sum of expression levels per gene) varied (Fig. 2). We used the liver dataset to conduct hyperparameter analysis. Then, we assessed the performance of the optimized architecture in all three datasets.
G.
Model parameters
Here, we systematically investigate the hyperparameter space to identify reasonable values and provide insight for ranges that can be extensively explored using an automatic hyperparameter optimization software such as optuna package [35]. Specifically, we use two middle layers with 16 and 4 neurons to predict the library size. The Adam optimizer [36] is used with the default linear layer initialization of weights in pytorch [37]. The desired learning rate is 0.001. However, due to the relatively large library sizes, the optimizer needs large learning rates to converge quickly. Therefore, to accelerate the training process, we set the initial learning rate at 0.0212 and decrease it by a factor of 0.9 to reach 0.001 at epoch 30. If the learning process halts, i.e., no improvement of at least 0.1% in validation loss, the learning rate will be reduced automatically after 15 epochs using the class ReduceLROnPlateau of pytorch-ignite [38]. The learning process stops early if the validation loss does not improve over 30 epochs.
In our implementation, users can either manually specify the model hyperparameters or allow our program to search for the optimum settings in a predefined sampling space using the Optuna package.
H.
Model comparison
We compared the trained models based on the average loss they obtained using 12 random seeds. The error bars in the figures in the Results section represent λ times of standard error of the mean, where λ=Φ−1(0.99)/2≈1.64, where Φ is the cumulative distribution function for the normal distribution. With this choice, assuming a one-sided t test (i.e., normality of distributions and same variance), non-overlapping bars indicate that we can reject the null hypothesis with 99% confidence.
A.
Advantages of predicting the library size
We separate the library size (i.e., the sum of all gene expressions for each cell) from the other predictions because what we want to model in the zero-inflated distribution is the relative expression levels of genes in a cell, not the absolute values, which depend on less relevant variables like the sequencing quality, assay efficiency, etc. To normalize gene expression data, we can adjust raw count data using count per million (CPM) [24] to account for differences in library size or sequencing depth between samples, allowing for fair comparisons of gene expression levels across samples. In this method, the raw counts for each gene in a sample are divided by the total number of reads (or counts) in that sample. Predicting the library size offers at least two distinct advantages over calculating it via adding all gene expressions for each cell: (1) it can help reduce technical noise in the data, and (2) it retains potentially valuable information about the library size in the latent variables that can be used in subsequent analyses. Notably, the library size could convey information relevant to the cell size [25].
B.
Sparsifying of the weights
We expect each feature in the second layer to represent a pathway or an eigengene, incorporating only a small subset of the genes [26], [27]. Therefore, most of the weights of the first encoding layer should be zero or near zero, forming a sparse weight matrix. The sparsity of this layer can be enforced by adding a term to the loss function that acts similarly to an L1 penalty. Not all activation functions can produce the intended sparsity as discussed in detail in section II-D.
C.
The model structure in mathematical notation
Throughout the manuscript, we use the subscript c for any variable to indicate that it is calculated for the cell c. The input matrix is denoted as X = {xcg}, where xcg is the expression of gene g in cell c. Xc denotes the full expression vector of cell c over all available genes. For t encoding layers, the cellular embedding for cell c is formulated as:
Uc=Gt⊕…⊕G2⊕G1(Xc),
where Uc is the embedding for cell c. Each Gi is composed of a linear layer, an activation function (e.g., sigmoid), and a batch normalization layer, which helps resolve the vanishing gradients problem and accelerates the training [19], [28]. We add known covariates of the cells, Uc′, to the embedding vector Uc, yielding the complete encoded representation of the cell,
Vc=concat(Uc,Uc′).
Subsequently, optional layers are used to predict the following variables:
log(lc)=liblog(Vc),
logit(zc)=N(Vc),
log(μc)=Mlog(Vc),
θc(j)=Θj(Vc);1≤j≤p,
where l is the library size, μ is the mean parameter for the distribution, z is the zero-inflation probability, θ represents additional relevant parameters of the distribution, and p is the number of these parameters. In the case of the Poisson distribution, we only have the mean parameter, μ. For the NB distribution, we need an extra parameter representing the difference between the mean and the dispersion of gene expressions. For N, we use a dense linear layer. For Mlog, we use a dense linear layer followed by logSoftmax activation, forcing the sum of the average gene expressions for each cell to be equal to 1.
Given this configuration, the following linear relationships are established:
zc=sig(AzVc+Bz)
μ≈exp(AμVc+Bμ).
To model the library size, we use a series of dense linear layers followed by batch normalization and an activation function. We can readily employ deep models to predict distribution parameters, resulting in nonlinear rather than linear Vc functions in the above equations.
D.
The loss function
The zero-inflated distribution is given by:
Pr(x∣z,l,μ,θ)=zδ0(x)+(1−z)expf(x;m,θ),m=lμ,
where f represents the logarithm of the distribution for non-zero-inflated expression levels with mean m and possibly other parameters θ and δ0 is the atomic distribution on zero. Note that m = lμ is the expected expression level according to our definitions.
We denote the rate of zero inflation by r, i.e., r = logit(z) and the likelihood of our samples, given the distribution parameters by ℒ. The first component of the loss function is the log-likelihood of the zero-inflated distribution:
logℒ=logPr(X∣z,μ,θ)=log∑xcgPr(xcg∣z,μ,θ)=∑xcg=0log(z+(1−z)expf(0;lμ,θ))+∑xcq>0(log(1−z)+f(x;lμ,θ))=∑log(1−z)+∑xcg=0log(z1−z+expf(0;lμ,θ))+∑xcg>0f(x;lμ,θ)=∑logsig(−r)+∑xcg=0log(expr+expf(0;lμ,θ))+∑xcg>0f(x;lμ,θ)=∑logsig(−r)+∑xcg=0r−logsig(r−f(0;lμ,θ))+∑xcg>0f(x;lμ,θ).
A typical approach employed in the DCA method involves using truncated log functions by adding a tiny offset quantity to the argument of the log function to avoid infinity (or very large) numeric values when the argument is zero (or near zero). Although this approach might be appropriate in specific conditions, it is inaccurate when we have no a priori knowledge of the input argument to decide on a proper value for the offset. To prioritize precision in our calculations, instead of relying on truncated log functions, we use the composite function log sig [29], which stabilizes the numerical schema without truncation.
The ZiPo model is regularized in two ways. We apply L2 regularization to the predicted zero probabilities and other distribution parameters. The regularization hyperparameter in ZiPo, αz, controls the level of zero inflation. An excessively large αz prevents the model from using significant zero inflation. Setting αz to zero will yield a model with strong zero inflation that overestimates the expression rates. We also impose a regularization term on the weights of the first layer to sparsify them. We use this regularization term to shrink weights appearing to be less important. We then prune those relatively small weights after ensuring that removing them does not affect the model performance.
We use NLL for “negative log-likelihood” and MSE for “mean square error” throughout the manuscript.
E.
Which activation function and weight regularization approach should be used?
In deep neural networks, activation functions such as ReLU [30] or GELU [31] are commonly employed to tackle the vanishing gradients issue. However, for the encoding segment of ZiPo, we favor the sigmoid activation function. This choice is influenced by the specific type of regularization applied to the weights (parameters) of the first encoding layer.
A limitation of ReLU arises from its linear properties. That is, the model can decrease the absolute weights in the initial layer while compensating for this by increasing the weights of the second layer, without affecting the final output of the second layer and the final performance of the model. Therefore, simple L1 or L2 regularization penalties would not significantly change the model’s performance or output. Some variations of the above activation functions like ReLU6, defined as min(max(0,x),6), could potentially resolve this problem. However, we prefer the nonlinearity of the sigmoid function over the piecewise linearity of ReLU6.
Similarly, the optimizer can lower the weights of the first encoding layer but compensate for them by increasing the batch normalization weights in the following layer. To protect the first layer weights, we use a scale-invariant loss function defined by dividing L1 by L2:
−E[logℒc]+αzE[zcg2]+αwmeanimeang|Wig|(meangWig2)12+E[αθ⋅θ2],
where Wigs are the weights of the first linear layer and α’s are regularization hyperparameters determining the strength of the corresponding penalty terms (i.e., αz controls the zeroinflation rate and αw controls the weights in the first layer and αθ controls the parameters of the distribution).
If the features of the initial layer are sufficiently interpretable, the model should perform well because those features are directly related to the genes. Employing similar regularization across all layers is possible. However, introducing an additional hyperparameter for each layer complicates the model and the training process.
F.
Datasets
We have used three datasets to illustrate ZiPo performance (Table I). The ROSMAP Alzheimer’s disease single-nucleus RNA-seq (snRNA-seq) dataset was generated by Mathys et al. [32], and is available from https://www.synapse.org/ with synapse ID: syn18485175. This dataset includes 80,660 single nuclei derived from the dorsolateral prefrontal cortex of 48 postmortem human brains. The liver [33] and myeloma [34] scRNA-seq datasets are accessible through GEO with accession numbers GSE125188 and GSE117156, respectively. In the liver dataset, 70,706 human CD45+ cells from the paired liver perfusion, spleen, and human peripheral blood mononuclear cells (PBMC) were profiled. The myeloma dataset includes 51,840 cells from 29 diagnosed patients and 11 control donors.
While the distributions of library sizes (i.e., the sum of gene expression per cell) were similar across these three datasets, the total gene expression (i.e., the sum of expression levels per gene) varied (Fig. 2). We used the liver dataset to conduct hyperparameter analysis. Then, we assessed the performance of the optimized architecture in all three datasets.
G.
Model parameters
Here, we systematically investigate the hyperparameter space to identify reasonable values and provide insight for ranges that can be extensively explored using an automatic hyperparameter optimization software such as optuna package [35]. Specifically, we use two middle layers with 16 and 4 neurons to predict the library size. The Adam optimizer [36] is used with the default linear layer initialization of weights in pytorch [37]. The desired learning rate is 0.001. However, due to the relatively large library sizes, the optimizer needs large learning rates to converge quickly. Therefore, to accelerate the training process, we set the initial learning rate at 0.0212 and decrease it by a factor of 0.9 to reach 0.001 at epoch 30. If the learning process halts, i.e., no improvement of at least 0.1% in validation loss, the learning rate will be reduced automatically after 15 epochs using the class ReduceLROnPlateau of pytorch-ignite [38]. The learning process stops early if the validation loss does not improve over 30 epochs.
In our implementation, users can either manually specify the model hyperparameters or allow our program to search for the optimum settings in a predefined sampling space using the Optuna package.
H.
Model comparison
We compared the trained models based on the average loss they obtained using 12 random seeds. The error bars in the figures in the Results section represent λ times of standard error of the mean, where λ=Φ−1(0.99)/2≈1.64, where Φ is the cumulative distribution function for the normal distribution. With this choice, assuming a one-sided t test (i.e., normality of distributions and same variance), non-overlapping bars indicate that we can reject the null hypothesis with 99% confidence.
III.
Results
Our systematic investigation of the hyperparameter space starts with seeking an optimum design for the encoder and decoder, using the ROSMAP dataset.
A.
A deep enough encoder with residuals leads to better performance.
Almost half of the model variables are embedded in the first layer of the encoder. The first layer also influences the size of the other layers. Therefore, it is important to optimize the first layer. To focus on the encoder, we used a simple linear structure for the decoder, as discussed in II-C.
1)
The first layer’s size:
In a 3-layer model with 64 latent variables, we assessed the effect of the first layer size, s1, on model training time and performance (Fig. 3). The middle layer size was set to be the geometric mean of the first layer size and the number of the latent variables (i.e., s1*64). In Fig. 3 and throughout this paper, we use the s1–s2–···–sp convention to describe the number of neurons used in the structure of a model with p layers, e.g., 8192–512–64 is a model with 8192, 512, and 64 neurons in the first, second, and third layers, respectively.
With increasing s1, the average loss was reduced until s1 = 4096. The model became relatively unstable, and loss obtained a wider range with s1 = 8096 (Fig. 3a). The s1 = 4096 model had the lowest average loss, and according to error bars, was significantly better than the models with a smaller first layer, except the models with s1 = 2048. Because we are interested in smaller networks, for the rest of our experiments, we chose s1 = 2048, a smaller model than s1 = 4096, but not significantly worse in terms of the likelihood. The rationale was that smaller models require less memory and computation for training and have a better chance of interpretability.
2)
The number of latent variables:
In a 3-layer architecture with a first layer of 2048 neurons and the size of the middle layer determined by the geometric mean as described, the likelihood improved as we increased the number of latent variables on the third layer up to 128 (Fig. 4a). Subsequent increases in the number of latent variables caused no significant improvement. As a smaller quantity of latent variables is computationally beneficial (Fig. 4b), we chose 128 latent variables for our model.
3)
Deeper encoders with residuals perform better:
We investigated the effect of the encoder depth and the incorporation of residual connections from the encoder layers to the latent layer on models featuring 2048 neurons in the first layer.
Adding residual connections to the models with 3 or more layers lowered their NLL loss (Fig. 5). The deepest model with the full residual connections was the optimal model. Because adding the middle layers did not significantly affect the model size (i.e., most weights are in the first layer), we continued to use this configuration for the rest of our experiments.
B.
A simple decoder is better than a deep one.
Using the same optimal encoder structure, (i.e., 2048 neurons in the first layer, 128 latents, 1024, 512, and 256 neurons in the middle layers, respectively, equipped with full residual connections from all encoder layers to the latent layer), we trained the model with different decoder structures (Fig. 6). More complex decoders needed more epochs for training and resulted in larger NLL losses. Therefore, we chose the simple linear dense decoder described in the Methods section.
C.
Batch normalization is essential for convergence.
The effect of batch normalization on training depended on the activation function. With the GELU activation function, removing the batch-normalization layer destabilized the training process, and the optimizer did not converge. With the sigmoid activation function, removing the batch-normalization layer did not destabilize the training process, and the optimizer converged, but got stuck in local minima (Fig. 7). Either way, batch normalization was essential for training.
D.
Controlling the zero-inflation rate improves library size estimation and MSE of reconstruction.
With no regularization on zero inflation, the model overestimated their rates, leading to a much larger MSE in 3 out of 8 tests with different seeds (Fig. 8c). Also, the mean of zero probabilities was very large (Fig. 8d). With a very large zero-inflation regularization parameter (αz = 108 in Fig. 8d), the model had almost no zero inflation, the training needed more epochs, and the NLL loss was larger than the models with smaller regularization. Apart from these extreme cases, the models had stable reconstruction regarding NLL loss and MSE rates for a vast range of regularization parameters (i.e., 0.01 ≤ αz ≤ 10, 000). However, smaller regularization parameters generally led to smaller NLL losses and larger mean zero probabilities. In the following experiments, we set αz = 0.1, which appeared to be a reasonable value for the mean of zero-inflation probabilities.
E.
Sparsifying the weights through weight regularization
With αz = 0.1, we investigated the effect of sparsifying the weights in the first layer using the weight regularization parameter αw. Note that the relative weights are important for each neuron on the first layer. One can multiply all the weights of a neuron by a constant and compensate for it by changing the batch normalization coefficient.
We normalized the weights of each neuron in the first layer by dividing all weights by the maximum weight. Then, we prune the first layer weights by setting those smaller than a threshold δ to zero. We investigated the effect of this sparsification on the model performance.
With no regularization (i.e., αw = 0), more than 90% of the weights were above 0.01 (Fig. 10a). With a large regularization parameter, say αw ≥ 1, many weights were very small, making the model sensitive to pruning (Fig. 10, c and d). In the middle range (i.e., 0.01 ≤ αw ≤ 0.1), the sensitivity of MSE of rates to the pruning seemed reasonable (Fig. 10c), we had even better NLL losses compared to no regularization (Fig. 9a), and the weight loss stayed almost the same (Fig. 9d).
F.
ZINB compared to ZIP
Our experiments showed that due to the greater complexity of the NB distribution compared to the Poisson distribution, we would need to apply another regularization (although very small) on the predicted rates of the model to train it. Therefore, we added 0.1*E[(lcμc−∑gxcg)2] to the loss function. In models with ZINB, training stopped earlier than in models with ZIP. In terms of the MSE of the rates, the models with ZIP and ZINB were not significantly different (Fig. 11). However, the models with ZIP performed significantly better regarding NLL, and their first layer weights were much sparser (Fig. 12) than those with ZINB.
G.
ZiPo compared to DCA
Due to technical difficulties, we could not use the DCA package to compare the corresponding method with ZiPo. Instead, we leveraged our implementation, which is in a sense a generalization of DCA, to simulate DCA using three approaches: (1) the recommended parameters by DCA authors in our settings, (2) the Poisson distribution (zipDCA), and (3) the settings that are similar to our optimized ZiPo model (tunedDCA). For the base settings, we used a 128–64–128 structure for the number of neurons in the model, applied logarithmic transformation, and normalized the input matrix by dividing the expression of each gene by the library size of each cell.
In terms of NLL, ZiPo, and DCA performed similarly and better than other models (Fig. 13a). The simplest model, i.e., DCA, was trained faster than the others but had the largest MSE for the rates (Fig. 13b). Poisson distribution would decrease the rates’ MSE, but the NLL was not as low as ZiPo.
H.
Automatic hyperparameter optimization
We used the “Optuna” package to optimize ZiPo’s hyperparameters. Specifically, we searched five different encoder structures, 2048–1024–512–256–128, 1024–512–256–128, 1024–512–256–128–64, 512–256–128, and 512–256–128–64 with full residual connections. For the decoder, we used a simple linear model For the regularization parameter of zero inflation and the weights, we searched the range 0.01–1 in the logarithmic space. After 96 iterations, the hyper-optimizer chose the deepest model, which resulted in a relatively reasonable value for the MSE of expression rates (Fig. 14). For the best model, the regularization parameter of zero inflation was 0.10594359 and the regularization parameter of weights was 0.14182903, both of which were near our choice of 0.1.
I.
Three additional datasets
Using the same hyperparameters optimized on the liver dataset, we trained the ZiPo model on the three datasets mentioned in II-F. The ZINB distribution generally seemed more powerful due to its larger number of parameters (Fig. 15). ZINB requires fewer epochs, which means faster convergence of the optimizer, most probably because it has more parameters and can easily fit the input distribution. However, ZIP acted even better regarding NLL in cases like the Liver dataset.
A.
A deep enough encoder with residuals leads to better performance.
Almost half of the model variables are embedded in the first layer of the encoder. The first layer also influences the size of the other layers. Therefore, it is important to optimize the first layer. To focus on the encoder, we used a simple linear structure for the decoder, as discussed in II-C.
1)
The first layer’s size:
In a 3-layer model with 64 latent variables, we assessed the effect of the first layer size, s1, on model training time and performance (Fig. 3). The middle layer size was set to be the geometric mean of the first layer size and the number of the latent variables (i.e., s1*64). In Fig. 3 and throughout this paper, we use the s1–s2–···–sp convention to describe the number of neurons used in the structure of a model with p layers, e.g., 8192–512–64 is a model with 8192, 512, and 64 neurons in the first, second, and third layers, respectively.
With increasing s1, the average loss was reduced until s1 = 4096. The model became relatively unstable, and loss obtained a wider range with s1 = 8096 (Fig. 3a). The s1 = 4096 model had the lowest average loss, and according to error bars, was significantly better than the models with a smaller first layer, except the models with s1 = 2048. Because we are interested in smaller networks, for the rest of our experiments, we chose s1 = 2048, a smaller model than s1 = 4096, but not significantly worse in terms of the likelihood. The rationale was that smaller models require less memory and computation for training and have a better chance of interpretability.
2)
The number of latent variables:
In a 3-layer architecture with a first layer of 2048 neurons and the size of the middle layer determined by the geometric mean as described, the likelihood improved as we increased the number of latent variables on the third layer up to 128 (Fig. 4a). Subsequent increases in the number of latent variables caused no significant improvement. As a smaller quantity of latent variables is computationally beneficial (Fig. 4b), we chose 128 latent variables for our model.
3)
Deeper encoders with residuals perform better:
We investigated the effect of the encoder depth and the incorporation of residual connections from the encoder layers to the latent layer on models featuring 2048 neurons in the first layer.
Adding residual connections to the models with 3 or more layers lowered their NLL loss (Fig. 5). The deepest model with the full residual connections was the optimal model. Because adding the middle layers did not significantly affect the model size (i.e., most weights are in the first layer), we continued to use this configuration for the rest of our experiments.
1)
The first layer’s size:
In a 3-layer model with 64 latent variables, we assessed the effect of the first layer size, s1, on model training time and performance (Fig. 3). The middle layer size was set to be the geometric mean of the first layer size and the number of the latent variables (i.e., s1*64). In Fig. 3 and throughout this paper, we use the s1–s2–···–sp convention to describe the number of neurons used in the structure of a model with p layers, e.g., 8192–512–64 is a model with 8192, 512, and 64 neurons in the first, second, and third layers, respectively.
With increasing s1, the average loss was reduced until s1 = 4096. The model became relatively unstable, and loss obtained a wider range with s1 = 8096 (Fig. 3a). The s1 = 4096 model had the lowest average loss, and according to error bars, was significantly better than the models with a smaller first layer, except the models with s1 = 2048. Because we are interested in smaller networks, for the rest of our experiments, we chose s1 = 2048, a smaller model than s1 = 4096, but not significantly worse in terms of the likelihood. The rationale was that smaller models require less memory and computation for training and have a better chance of interpretability.
2)
The number of latent variables:
In a 3-layer architecture with a first layer of 2048 neurons and the size of the middle layer determined by the geometric mean as described, the likelihood improved as we increased the number of latent variables on the third layer up to 128 (Fig. 4a). Subsequent increases in the number of latent variables caused no significant improvement. As a smaller quantity of latent variables is computationally beneficial (Fig. 4b), we chose 128 latent variables for our model.
3)
Deeper encoders with residuals perform better:
We investigated the effect of the encoder depth and the incorporation of residual connections from the encoder layers to the latent layer on models featuring 2048 neurons in the first layer.
Adding residual connections to the models with 3 or more layers lowered their NLL loss (Fig. 5). The deepest model with the full residual connections was the optimal model. Because adding the middle layers did not significantly affect the model size (i.e., most weights are in the first layer), we continued to use this configuration for the rest of our experiments.
B.
A simple decoder is better than a deep one.
Using the same optimal encoder structure, (i.e., 2048 neurons in the first layer, 128 latents, 1024, 512, and 256 neurons in the middle layers, respectively, equipped with full residual connections from all encoder layers to the latent layer), we trained the model with different decoder structures (Fig. 6). More complex decoders needed more epochs for training and resulted in larger NLL losses. Therefore, we chose the simple linear dense decoder described in the Methods section.
C.
Batch normalization is essential for convergence.
The effect of batch normalization on training depended on the activation function. With the GELU activation function, removing the batch-normalization layer destabilized the training process, and the optimizer did not converge. With the sigmoid activation function, removing the batch-normalization layer did not destabilize the training process, and the optimizer converged, but got stuck in local minima (Fig. 7). Either way, batch normalization was essential for training.
D.
Controlling the zero-inflation rate improves library size estimation and MSE of reconstruction.
With no regularization on zero inflation, the model overestimated their rates, leading to a much larger MSE in 3 out of 8 tests with different seeds (Fig. 8c). Also, the mean of zero probabilities was very large (Fig. 8d). With a very large zero-inflation regularization parameter (αz = 108 in Fig. 8d), the model had almost no zero inflation, the training needed more epochs, and the NLL loss was larger than the models with smaller regularization. Apart from these extreme cases, the models had stable reconstruction regarding NLL loss and MSE rates for a vast range of regularization parameters (i.e., 0.01 ≤ αz ≤ 10, 000). However, smaller regularization parameters generally led to smaller NLL losses and larger mean zero probabilities. In the following experiments, we set αz = 0.1, which appeared to be a reasonable value for the mean of zero-inflation probabilities.
E.
Sparsifying the weights through weight regularization
With αz = 0.1, we investigated the effect of sparsifying the weights in the first layer using the weight regularization parameter αw. Note that the relative weights are important for each neuron on the first layer. One can multiply all the weights of a neuron by a constant and compensate for it by changing the batch normalization coefficient.
We normalized the weights of each neuron in the first layer by dividing all weights by the maximum weight. Then, we prune the first layer weights by setting those smaller than a threshold δ to zero. We investigated the effect of this sparsification on the model performance.
With no regularization (i.e., αw = 0), more than 90% of the weights were above 0.01 (Fig. 10a). With a large regularization parameter, say αw ≥ 1, many weights were very small, making the model sensitive to pruning (Fig. 10, c and d). In the middle range (i.e., 0.01 ≤ αw ≤ 0.1), the sensitivity of MSE of rates to the pruning seemed reasonable (Fig. 10c), we had even better NLL losses compared to no regularization (Fig. 9a), and the weight loss stayed almost the same (Fig. 9d).
F.
ZINB compared to ZIP
Our experiments showed that due to the greater complexity of the NB distribution compared to the Poisson distribution, we would need to apply another regularization (although very small) on the predicted rates of the model to train it. Therefore, we added 0.1*E[(lcμc−∑gxcg)2] to the loss function. In models with ZINB, training stopped earlier than in models with ZIP. In terms of the MSE of the rates, the models with ZIP and ZINB were not significantly different (Fig. 11). However, the models with ZIP performed significantly better regarding NLL, and their first layer weights were much sparser (Fig. 12) than those with ZINB.
G.
ZiPo compared to DCA
Due to technical difficulties, we could not use the DCA package to compare the corresponding method with ZiPo. Instead, we leveraged our implementation, which is in a sense a generalization of DCA, to simulate DCA using three approaches: (1) the recommended parameters by DCA authors in our settings, (2) the Poisson distribution (zipDCA), and (3) the settings that are similar to our optimized ZiPo model (tunedDCA). For the base settings, we used a 128–64–128 structure for the number of neurons in the model, applied logarithmic transformation, and normalized the input matrix by dividing the expression of each gene by the library size of each cell.
In terms of NLL, ZiPo, and DCA performed similarly and better than other models (Fig. 13a). The simplest model, i.e., DCA, was trained faster than the others but had the largest MSE for the rates (Fig. 13b). Poisson distribution would decrease the rates’ MSE, but the NLL was not as low as ZiPo.
H.
Automatic hyperparameter optimization
We used the “Optuna” package to optimize ZiPo’s hyperparameters. Specifically, we searched five different encoder structures, 2048–1024–512–256–128, 1024–512–256–128, 1024–512–256–128–64, 512–256–128, and 512–256–128–64 with full residual connections. For the decoder, we used a simple linear model For the regularization parameter of zero inflation and the weights, we searched the range 0.01–1 in the logarithmic space. After 96 iterations, the hyper-optimizer chose the deepest model, which resulted in a relatively reasonable value for the MSE of expression rates (Fig. 14). For the best model, the regularization parameter of zero inflation was 0.10594359 and the regularization parameter of weights was 0.14182903, both of which were near our choice of 0.1.
I.
Three additional datasets
Using the same hyperparameters optimized on the liver dataset, we trained the ZiPo model on the three datasets mentioned in II-F. The ZINB distribution generally seemed more powerful due to its larger number of parameters (Fig. 15). ZINB requires fewer epochs, which means faster convergence of the optimizer, most probably because it has more parameters and can easily fit the input distribution. However, ZIP acted even better regarding NLL in cases like the Liver dataset.
IV.
Discussion
We have described a novel deep model, ZiPo, for rate estimation and library size prediction in scRNA-seq data. ZiPo presents several unique features described below. Most prominently, the present model features an adjustable zero-inflation rate.
a)
Handling big data with deep models and batches:
A significant advantage of using deep models is their capability to handle massive datasets in batches. In contrast, other methods would be impractical as they would need to load the whole dataset in the memory. Also, similar to other deep models, ZiPo can be partially trained with one dataset and then fine-tuned using other datasets. This usually cannot be done easily with other methods.
b)
Software quality:
To ensure the longevity of our code and its maintenance in the long term, we have used popular software packages and modern programming techniques in implementing ZiPo. Using features like configuration files, we have provided a flexible platform for software developers and end-users.
c)
ZIP overall surpasses ZINB:
With an appropriate architecture, zero-inflation models using ZIP distribution perform well. As shown here, the models with ZINB distribution can achieve smaller losses, hence better performances than those using ZIP. However, due to the need for extra regularization in the models with ZINB, they are relatively larger with more trainable weights, their training is more complicated and time-consuming, and they are harder to interpret than those with ZIP.
d)
Latents for downstream analysis:
To assess the validity of the latent variables, we tested ZiPo for clustering the cells in the ROSMAP dataset. We compared the results with the predefined cell types. The tSNE plots [39] based on the latent variables from the two ZIP and ZINB models showed that these models generally performed well except for a few outliers (Fig. 16).
e)
Future work:
The latent variables can be investigated more thoroughly and used in downstream analyses including regressing eigengenes [40] or classifying senescent cells [41].
a)
Handling big data with deep models and batches:
A significant advantage of using deep models is their capability to handle massive datasets in batches. In contrast, other methods would be impractical as they would need to load the whole dataset in the memory. Also, similar to other deep models, ZiPo can be partially trained with one dataset and then fine-tuned using other datasets. This usually cannot be done easily with other methods.
b)
Software quality:
To ensure the longevity of our code and its maintenance in the long term, we have used popular software packages and modern programming techniques in implementing ZiPo. Using features like configuration files, we have provided a flexible platform for software developers and end-users.
c)
ZIP overall surpasses ZINB:
With an appropriate architecture, zero-inflation models using ZIP distribution perform well. As shown here, the models with ZINB distribution can achieve smaller losses, hence better performances than those using ZIP. However, due to the need for extra regularization in the models with ZINB, they are relatively larger with more trainable weights, their training is more complicated and time-consuming, and they are harder to interpret than those with ZIP.
d)
Latents for downstream analysis:
To assess the validity of the latent variables, we tested ZiPo for clustering the cells in the ROSMAP dataset. We compared the results with the predefined cell types. The tSNE plots [39] based on the latent variables from the two ZIP and ZINB models showed that these models generally performed well except for a few outliers (Fig. 16).
e)
Future work:
The latent variables can be investigated more thoroughly and used in downstream analyses including regressing eigengenes [40] or classifying senescent cells [41]. |
Title: Tamoxifen metabolites treatment promotes ERα+ transition to triple negative phenotype | Body:
Introduction
Breast cancer (BC) is the most common malignancy in women worldwide. In 2020, BC had an approximate incidence of 2 million cases and generated an estimated 685,000 deaths (Global Cancer Observatory, 2022) associated with metastatic tumors in vital organs and tissues. Breast tumors can be classified by molecular subtypes, depending on the hormonal receptor expression [1], the main subtypes are Luminal A with Estrogen Receptor alpha (ERα)/Progesterone Receptor (PR) expression and no expression of human epidermal growth factor receptor type 2 (HER2); Luminal B with ERα/PR and HER2 expression; HER2 positive with HER2 overexpression and no expression of ERα/PR, and triple-negative (TNBC) with no hormonal receptors expression [2].
Approximately 80% of breast cancer tumors are ERα+, and for these cases, the neoadjuvant election treatment for premenopausal women is the selective estrogen receptor modulators (SERMS), mainly tamoxifen. Tamoxifen is a prodrug whose plasmatic primary active metabolites, 4-hydroxytamoxifen (4OH-Tam) and 4-hydroxy-N-desmethyl-tamoxifen (endoxifen), are generated by cytochrome P450 isoenzymes families such as CYP3A4 and CYP2D6 [3]. 4OH-Tam and Endoxifen have similar potency, possibly associated with the hydroxyl group at the C4 position in their structure [4]; although, endoxifen is the most abundant plasmatic metabolite in tamoxifen-treated patients [3]. Tamoxifen and its metabolites function as ERα antagonists in breast tissue, interfering with growth and proliferation pathways in cancer cells [4].
Tamoxifen treatment has been associated with a decrease in recurrence rate by 41% and a mortality rate of up to 31% [5]; however, nearly one-third of the patients relapse due to chemoresistance [6]. Several mechanisms responsible for tamoxifen resistance have been proposed in chemoresistant tumoral ERα+ cells; one of them corresponds to overexpression of sulfotransferase 1A1 (SULT1A), a phase II metabolism enzyme that inactivates 4OH-Tam and endoxifen by sulfation mechanism [7]. Other chemoresistance mechanisms are related to the ERα; for instance, loss or down-regulation of ERα expression by mutations, epigenetic modifications (ESR1), or post-translational alterations [8]. Changes in other hormonal receptors expression could occur as chemoresistant mechanism and/or associated with disease progression [9].
Changes in hormonal receptors expression among primary and metastatic tumors are frequently reported in clinical studies [10,11]. A meta-analysis that included 39 studies described a 19.3% change in ERα status, a 30.9% change in PR status, and a 10.3% change in HER2 status in metastatic tumors; the change from positive to negative status was the most frequent in this study [12]. Other meta-analysis evaluating hormonal status in BC brain metastasis in comparison with primary tumors reported a 42.6% discordance in any receptor status and individually a 17% in ERα status, 23% in PR status, and a 12% in HER2 status [13]. Although the causes of this phenomenon must be multifactorial, the effect of dyslipidemia in patients that undergo this complication has not been studied in detail.
High low-density lipoproteins (LDL) serum levels and low high density lipoprotein (HDL) serum levels or a high c-LDL/c-HDL ratio, commonly associated to an increased cardiovascular risk [14], has been also widely associated with an increased risk of developing BC cancer and BC cancer progression [15–19]. Focusing in LDL, it has been demonstrated in vitro that LDL cholesterol promotes proliferation and migration in ERα negative cell lines, while its effect on ERα+ cell lines remains debatable [19,20]. In animal models, it has been described a relationship between high LDL and total cholesterol levels and cancer progression [19]. In vivo, plasma LDL-C levels above 117 mg/dL have been positively associated with higher histological grade, higher proliferative rate and more advanced clinical stage [21]; on the other hand, tumors overexpressing LDLR, which allow them to capture more cholesterol, correlates to a more aggressive behavior and show a greater metastasis potential [22]. Although, the molecular mechanisms triggered by LDL high concentrations have not been described in detail in several contexts.
Metabolic adaptations are commonly related to cancer progression. It is common for tumoral cells to have increased free radical synthesis and, in consequence, augmented oxidative stress due to their accelerated metabolic activity [23]. Nuclear factor erythroid-2-related factor 2 (Nrf2), a leucine zipper transcription factor described as the redox homeostasis master regulator, plays an essential role in managing oxidative stress in tumoral cells [23,24] in addition to increasing metastatic capacity [25] and potentially chemoresistance. In turn, this phenotype could be triggered by high LDL concentration.
Therefore, this study aimed to evaluate the effect of LDL on tamoxifen resistance generation in a cellular ERα+ breast cancer model. For the first time, the induced chemoresistance was reached through treatment with primary plasmatic tamoxifen-derived metabolites (4-OH Tam and endoxifen), promoting the transition to a TNBC phenotype-like. However, contrary to what was expected, this phenotype was not observed under co-incubation of metabolites with a high LDL concentration (75 µg/ml). Elucidating LDL role in ERα BC tamoxifen chemoresistance acquisition and the molecular mechanisms involved is critical to stablish the most suitable disease management and to propose new therapeutic targets.
Materials and methods
Materials
Cell culture reagents were purchased from Thermo-Fisher (Carlsbad, CA, U.S.A.), tissue culture plates and other plastic materials were obtained from Corning Inc. Tamoxifen metabolites 4OH-Tam and endoxifen were obtained from Santa Cruz Biotechnology with a purity greater than 98%. Tunicamycin (Tum) was obtained from Sigma-Aldrich. Buffers and Dimethylthiazol-2-yl-2,5-diphenyltetrazolium bromide (MTT) were obtained from Merck (Darmstadt, Germany). Monoclonal antibodies anti-Fas (1:250), anti-LDH-A (1:400), anti-PCNA (1:2500), anti-Nrf2 (1:300), anti-SNAI-1 (1:250) anti-β-actin (1:300), anti-FAK (1:300), Anti-p-FAK (1:400), anti-Ki67 (1:900), anti-HER2 (Neu) (1:200), anti-ER alpha (1:250), anti-PR (1:400), anti-HO-1 (1:500), anti-SULT1A1 (1:400), and anti-NQO1 (1:500) were obtained from Santa Cruz Biotechnology. LDLR (1:800) from Abcam. Anti-mouse secondary antibodies coupled to Horseradish peroxidase from Thermo-Fisher. Chemiluminescence detection was performed using Immobilon Western kit (Millipore, MA, U.S.A.) and X-ray film for some blots, for others Bio-Rad ChemiDoc XRS+ was used and a digital image was obtained.
Cell culture
MCF-7 cells obtained from American Type Culture Collection (ATCC), were grown in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% fetal bovine serum (FBS), 50 U/ml penicillin and 50 µg/ml streptomycin and insulin (0.57 UI/ml). Cells were maintained at 37°C with 90% humidity and 5% CO2.
Chemoresistance generation
Through the development of this protocol, two chemoresistant variants were generated. MCF-7Var-H: the treatment was performed with equimolar 1 µM concentration of tamoxifen metabolites 4OH-Tam and endoxifen in culture media for 30 days. Later, the concentration was increased to 2 µM for both metabolites in culture media for 15 days. Cells were maintained under 50 nM concentration of both metabolites during cellular expansion and recovery period. MCF-7Var-I: treatment was followed as referred to above for MCF-7Var-H plus 75 µg/ml of LDL for 45 days. Cells were maintained in 50 nM concentration of both metabolites and 7.5 µg/ml of LDL for expansion. Chemoresistance acquisition in both variants was confirmed through MTT assay. In each experimental panel, parental MCF-7 cells were used as a control.
LDL isolation and fluorescent labeling
Human plasma samples were obtained from a healthy donor who signed an informed consent. The protocol was designed and carried out according to the Declaration of Helsinki and registered in the Research Ethics Committee of Facultad de Medicina Mexicali (FMM/CEI-FMM/006/2022-1). Briefly, plasma density was initially adjusted to density 1.019 g/ml by adding KBr and then centrifuged at 345,000 g for 160 min at 4°C. Later, fraction <1.019 g/ml was discarded, and the remaining plasma was adjusted to 1.053 g/mL density. It was centrifuged at 377,000 g for 200 min. Finally, <1.053 g/ml fraction was recovered, dialyzed against 150 mM NaCl/EDTA 0.024 mM, and filtered through 0.45 μm. In the LDL fraction, protein concentration was measured with the bicinchoninic acid assay (BCA), and the determination of LDL-cholesterol was carried out (Spinreact). A defined volume of buffer (NaCl 150 mM, EDTA 0.024 mM) corresponding to a specific dose was evaluated as a control for the experimentation. The quality of the isolates was evaluated by characterization of apolipoproteins apoB-100 and apoA1 (Supplementary Figure S1).
Labeling of LDL fraction was carried out with the dilC18 probe (D3911), which is incorporated into the phospholipid monolayer, through incubation of 10 µl of the probe (2 mg/ml) for each 1 mg of protein-LDL for 18 h at 37 °C, obtaining dil-LDL. The solution density was adjusted to a 1.053 g/ml and centrifuged at 377,000 g for 200 min to recover fluorescent LDL. The fraction was recovered and dialyzed against PBS. The experiments evaluated a range of 0–75 µg/ml.
dil-LDL cytometer assays
Before internalization experiments, MCF-7 and cell variant cultures at 90% confluence were incubated in a FBS-free culture medium. Two hours fasting later, cells were treated under diferent dil-LDL concentrations (0–75 µg/ml) for 24 h. Later, cells were washed twice with PBS and recovered in a 200 µl volume. Cellular characterization was performed in a Beckman-Coulter cytometer Cytoflex. Employing the PC7-A channel to record the dil-LDL fluorescence based on a previous work, 30,000 events were registered [26].
Cell viability assay
Cell viability was assessed using an MTT reduction assay. Cells were seeded into 96-well plates with a density of 20,000 cells/well and allowed to grow to 75% confluence. The DMEM-supplemented medium was replaced with Opti-MEM and incubated at 37°C for 2 h; later, several treatments were added (0.125–16 µM of 4-OH Tam or Endoxifen) and incubated for 72 h. The next step was the addition of 30 µl MTT (2.1 g/ml) in Opti-MEM and incubation at 37°C for 4 h. Formazan crystals generated by mitochondrial activity were dissolved with a lysis buffer (20% SDS, 50% N, N-dimethyl-formamide, pH 3.7). A lecture on absorbances at 590 nm was performed 12 h later. The IC50 of each metabolite was calculated with GraphPad Prism 8.
Wound-healing assay
Cells were placed in a 6-well plate under a density of 400, 000 cells/well and allowed to grow to 90–95% confluence, forming a cellular monolayer. The DMEM-supplemented medium was replaced with the DMEM experimental medium, and cells were incubated at 37°C for 2 h. Later, cells were treated with mitomycin C for 2 h to prevent cell proliferation. The monolayer was scratched with a pipette tip, washed with PBS to remove floating cells, and treatments were added. Cells were allowed to migrate for 72 h at 37°C. Each well was fixed with paraformaldehyde, stained with Blue Coomassie, and photographed using a Motic Images Plus 3.0 camera coupled to an optical inverted microscope. Results were processed using ImageJ software, and statistical analysis was calculated using GraphPad Prism 8.
Cellular lysates and Western blot analysis
Cells were seeded into 6-well plates with a density of 400,000 cells/well and allowed to reach 90–95% confluence. Next, cells were incubated under the specific treatments. Cells were washed with PBS and lysed with a protein lysis buffer supplemented with protease and phosphatase inhibitors. Sample proteins were quantified by the BCA assay. Samples (25 µg/lane) were analyzed by SDS-PAGE on 8–12% gels and further transferred to PVDF membranes. Later, membranes were blocked in TBS/0.1% tween (TBST)/5% low-fat milk at 37°C for 1 h and incubated at 4°C overnight with the corresponding primary antibody. Next, membranes were washed with TBST and incubated for 120 min at 37°C with the corresponding horseradish peroxidase-conjugated (HRP) secondary antibodies. Membranes were washed with TBST, and the HRP activity was detected with the Immobilon Western Kit (Millipore, MA, U.S.A.). Analysis of immunoblot films was made with the ImageJ software (NIH, Maryland, DC, U.S.A.).
Optical microscopy
Cells were photographed using a Motic Images Plus 3.0 camera coupled to an optical inverted microscope.
Zymography analysis
Variants and MCF-7 cells were seeded at a density of 200,000 cells/ml. Cellular cultures were incubated under 90% confluency under the specific schemes, and the extracellular medium was collected. Volumes of 40 µl non-heated conditioned medium samples were mixed with 5× sample buffer (0.313 M Tris pH 6.8, 10% SDS, 50% glycerol, and 0.05% bromophenol blue) and loaded on 8% polyacrylamide gels copolymerized with gelatin (1% w/v). Gels were rinsed twice with 2.5% Triton X-100 and then incubated in development buffer (50 mM Tris-HCl pH 7.4, 10 mM CaCl2, and 0.02% NaN3) for 40 h at 37 °C. Gels were fixed and stained with 0.25% Coomassie Brilliant Blue G-250 in 10% acetic acid and 30% methanol. Proteolytic activity was detected as clear bands against the background stain. In these experiments, β-actin as a reference for the amount of protein was used as a loading control.
Nrf2 localization
We performed the analysis of Nrf2 cellular localization. Briefly, cells were seeded on slides placed in untreated 6-well dishes (400,000 cells/well) and allowed to reach 60–70% confluence. Cells were incubated under several treatments for 24 h at 37°C. Later, cells were washed with filtered PBS and fixed with 2% paraformaldehyde/PBS. Next, cells were treated with a 0.01% triton solution for 5 min and blocked with 2% bovine albumin/PBS for 30 min. Cells were incubated with anti-Nrf2 antibody overnight at 4°C. Later, cells were washed with PBS and incubated with a FITC-coupled secondary antibody. Subsequently, counterstaining was carried out with Propidium Iodide. Finally, the cells were observed using the green and red filter in an EPI-fluorescence microscope (Axio VertA.1, Zeiss, Göttingen, Germany).
GEPIA analysis
We used the GEPIA database (http://gepia.cancer-pku.cn/about.html) to evaluate the Nrf2 role in breast cancer. We analyzed the mRNA expression levels of NFE2L2 (Nrf2 coding gene, Gene ID: 4780) and their relationship with overall survival in BC patients using GEPIA which collects information from the TCGA (The Cancer Genome Atlas) [27] and GTEx (Genotype-Tissue-Expression) [28].
Molecular docking
For molecular docking experimentation, the atomic coordinates of protein SULT1A1 (PDB ID: 4GRA) with a resolution of 2.56 Å were evaluated. The structures of tamoxifen (CID:10540-29-1), 4OH-Tam (CID: 112093-28-4) and endoxifen (CID:11002528-0) were obtained from the PubChem Database. The protein structure was prepared, waters and small molecules were removed. The ligands and protein were 3D-protonated and energy minimization was carried out by employing Molecular Operating Environment software (MOE) using default parameters (Placement: Triangle Matcher, Rescoring 1: London G, AMBER99 forcefield). Each ligand was generated up to different conformations and protein was visualized with ligand interactions implemented in MOE. Estradiol was used as a natural ligand (CHEBI: 23965), obtained from ChEBI Database.
Statistical analysis
Data are expressed as mean± SD. The IC50 and ANOVA probes were calculated in GraphPad Prism 8.
Materials
Cell culture reagents were purchased from Thermo-Fisher (Carlsbad, CA, U.S.A.), tissue culture plates and other plastic materials were obtained from Corning Inc. Tamoxifen metabolites 4OH-Tam and endoxifen were obtained from Santa Cruz Biotechnology with a purity greater than 98%. Tunicamycin (Tum) was obtained from Sigma-Aldrich. Buffers and Dimethylthiazol-2-yl-2,5-diphenyltetrazolium bromide (MTT) were obtained from Merck (Darmstadt, Germany). Monoclonal antibodies anti-Fas (1:250), anti-LDH-A (1:400), anti-PCNA (1:2500), anti-Nrf2 (1:300), anti-SNAI-1 (1:250) anti-β-actin (1:300), anti-FAK (1:300), Anti-p-FAK (1:400), anti-Ki67 (1:900), anti-HER2 (Neu) (1:200), anti-ER alpha (1:250), anti-PR (1:400), anti-HO-1 (1:500), anti-SULT1A1 (1:400), and anti-NQO1 (1:500) were obtained from Santa Cruz Biotechnology. LDLR (1:800) from Abcam. Anti-mouse secondary antibodies coupled to Horseradish peroxidase from Thermo-Fisher. Chemiluminescence detection was performed using Immobilon Western kit (Millipore, MA, U.S.A.) and X-ray film for some blots, for others Bio-Rad ChemiDoc XRS+ was used and a digital image was obtained.
Cell culture
MCF-7 cells obtained from American Type Culture Collection (ATCC), were grown in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% fetal bovine serum (FBS), 50 U/ml penicillin and 50 µg/ml streptomycin and insulin (0.57 UI/ml). Cells were maintained at 37°C with 90% humidity and 5% CO2.
Chemoresistance generation
Through the development of this protocol, two chemoresistant variants were generated. MCF-7Var-H: the treatment was performed with equimolar 1 µM concentration of tamoxifen metabolites 4OH-Tam and endoxifen in culture media for 30 days. Later, the concentration was increased to 2 µM for both metabolites in culture media for 15 days. Cells were maintained under 50 nM concentration of both metabolites during cellular expansion and recovery period. MCF-7Var-I: treatment was followed as referred to above for MCF-7Var-H plus 75 µg/ml of LDL for 45 days. Cells were maintained in 50 nM concentration of both metabolites and 7.5 µg/ml of LDL for expansion. Chemoresistance acquisition in both variants was confirmed through MTT assay. In each experimental panel, parental MCF-7 cells were used as a control.
LDL isolation and fluorescent labeling
Human plasma samples were obtained from a healthy donor who signed an informed consent. The protocol was designed and carried out according to the Declaration of Helsinki and registered in the Research Ethics Committee of Facultad de Medicina Mexicali (FMM/CEI-FMM/006/2022-1). Briefly, plasma density was initially adjusted to density 1.019 g/ml by adding KBr and then centrifuged at 345,000 g for 160 min at 4°C. Later, fraction <1.019 g/ml was discarded, and the remaining plasma was adjusted to 1.053 g/mL density. It was centrifuged at 377,000 g for 200 min. Finally, <1.053 g/ml fraction was recovered, dialyzed against 150 mM NaCl/EDTA 0.024 mM, and filtered through 0.45 μm. In the LDL fraction, protein concentration was measured with the bicinchoninic acid assay (BCA), and the determination of LDL-cholesterol was carried out (Spinreact). A defined volume of buffer (NaCl 150 mM, EDTA 0.024 mM) corresponding to a specific dose was evaluated as a control for the experimentation. The quality of the isolates was evaluated by characterization of apolipoproteins apoB-100 and apoA1 (Supplementary Figure S1).
Labeling of LDL fraction was carried out with the dilC18 probe (D3911), which is incorporated into the phospholipid monolayer, through incubation of 10 µl of the probe (2 mg/ml) for each 1 mg of protein-LDL for 18 h at 37 °C, obtaining dil-LDL. The solution density was adjusted to a 1.053 g/ml and centrifuged at 377,000 g for 200 min to recover fluorescent LDL. The fraction was recovered and dialyzed against PBS. The experiments evaluated a range of 0–75 µg/ml.
dil-LDL cytometer assays
Before internalization experiments, MCF-7 and cell variant cultures at 90% confluence were incubated in a FBS-free culture medium. Two hours fasting later, cells were treated under diferent dil-LDL concentrations (0–75 µg/ml) for 24 h. Later, cells were washed twice with PBS and recovered in a 200 µl volume. Cellular characterization was performed in a Beckman-Coulter cytometer Cytoflex. Employing the PC7-A channel to record the dil-LDL fluorescence based on a previous work, 30,000 events were registered [26].
Cell viability assay
Cell viability was assessed using an MTT reduction assay. Cells were seeded into 96-well plates with a density of 20,000 cells/well and allowed to grow to 75% confluence. The DMEM-supplemented medium was replaced with Opti-MEM and incubated at 37°C for 2 h; later, several treatments were added (0.125–16 µM of 4-OH Tam or Endoxifen) and incubated for 72 h. The next step was the addition of 30 µl MTT (2.1 g/ml) in Opti-MEM and incubation at 37°C for 4 h. Formazan crystals generated by mitochondrial activity were dissolved with a lysis buffer (20% SDS, 50% N, N-dimethyl-formamide, pH 3.7). A lecture on absorbances at 590 nm was performed 12 h later. The IC50 of each metabolite was calculated with GraphPad Prism 8.
Wound-healing assay
Cells were placed in a 6-well plate under a density of 400, 000 cells/well and allowed to grow to 90–95% confluence, forming a cellular monolayer. The DMEM-supplemented medium was replaced with the DMEM experimental medium, and cells were incubated at 37°C for 2 h. Later, cells were treated with mitomycin C for 2 h to prevent cell proliferation. The monolayer was scratched with a pipette tip, washed with PBS to remove floating cells, and treatments were added. Cells were allowed to migrate for 72 h at 37°C. Each well was fixed with paraformaldehyde, stained with Blue Coomassie, and photographed using a Motic Images Plus 3.0 camera coupled to an optical inverted microscope. Results were processed using ImageJ software, and statistical analysis was calculated using GraphPad Prism 8.
Cellular lysates and Western blot analysis
Cells were seeded into 6-well plates with a density of 400,000 cells/well and allowed to reach 90–95% confluence. Next, cells were incubated under the specific treatments. Cells were washed with PBS and lysed with a protein lysis buffer supplemented with protease and phosphatase inhibitors. Sample proteins were quantified by the BCA assay. Samples (25 µg/lane) were analyzed by SDS-PAGE on 8–12% gels and further transferred to PVDF membranes. Later, membranes were blocked in TBS/0.1% tween (TBST)/5% low-fat milk at 37°C for 1 h and incubated at 4°C overnight with the corresponding primary antibody. Next, membranes were washed with TBST and incubated for 120 min at 37°C with the corresponding horseradish peroxidase-conjugated (HRP) secondary antibodies. Membranes were washed with TBST, and the HRP activity was detected with the Immobilon Western Kit (Millipore, MA, U.S.A.). Analysis of immunoblot films was made with the ImageJ software (NIH, Maryland, DC, U.S.A.).
Optical microscopy
Cells were photographed using a Motic Images Plus 3.0 camera coupled to an optical inverted microscope.
Zymography analysis
Variants and MCF-7 cells were seeded at a density of 200,000 cells/ml. Cellular cultures were incubated under 90% confluency under the specific schemes, and the extracellular medium was collected. Volumes of 40 µl non-heated conditioned medium samples were mixed with 5× sample buffer (0.313 M Tris pH 6.8, 10% SDS, 50% glycerol, and 0.05% bromophenol blue) and loaded on 8% polyacrylamide gels copolymerized with gelatin (1% w/v). Gels were rinsed twice with 2.5% Triton X-100 and then incubated in development buffer (50 mM Tris-HCl pH 7.4, 10 mM CaCl2, and 0.02% NaN3) for 40 h at 37 °C. Gels were fixed and stained with 0.25% Coomassie Brilliant Blue G-250 in 10% acetic acid and 30% methanol. Proteolytic activity was detected as clear bands against the background stain. In these experiments, β-actin as a reference for the amount of protein was used as a loading control.
Nrf2 localization
We performed the analysis of Nrf2 cellular localization. Briefly, cells were seeded on slides placed in untreated 6-well dishes (400,000 cells/well) and allowed to reach 60–70% confluence. Cells were incubated under several treatments for 24 h at 37°C. Later, cells were washed with filtered PBS and fixed with 2% paraformaldehyde/PBS. Next, cells were treated with a 0.01% triton solution for 5 min and blocked with 2% bovine albumin/PBS for 30 min. Cells were incubated with anti-Nrf2 antibody overnight at 4°C. Later, cells were washed with PBS and incubated with a FITC-coupled secondary antibody. Subsequently, counterstaining was carried out with Propidium Iodide. Finally, the cells were observed using the green and red filter in an EPI-fluorescence microscope (Axio VertA.1, Zeiss, Göttingen, Germany).
GEPIA analysis
We used the GEPIA database (http://gepia.cancer-pku.cn/about.html) to evaluate the Nrf2 role in breast cancer. We analyzed the mRNA expression levels of NFE2L2 (Nrf2 coding gene, Gene ID: 4780) and their relationship with overall survival in BC patients using GEPIA which collects information from the TCGA (The Cancer Genome Atlas) [27] and GTEx (Genotype-Tissue-Expression) [28].
Molecular docking
For molecular docking experimentation, the atomic coordinates of protein SULT1A1 (PDB ID: 4GRA) with a resolution of 2.56 Å were evaluated. The structures of tamoxifen (CID:10540-29-1), 4OH-Tam (CID: 112093-28-4) and endoxifen (CID:11002528-0) were obtained from the PubChem Database. The protein structure was prepared, waters and small molecules were removed. The ligands and protein were 3D-protonated and energy minimization was carried out by employing Molecular Operating Environment software (MOE) using default parameters (Placement: Triangle Matcher, Rescoring 1: London G, AMBER99 forcefield). Each ligand was generated up to different conformations and protein was visualized with ligand interactions implemented in MOE. Estradiol was used as a natural ligand (CHEBI: 23965), obtained from ChEBI Database.
Statistical analysis
Data are expressed as mean± SD. The IC50 and ANOVA probes were calculated in GraphPad Prism 8.
Results
ERα+ cells internalize LDL
We focused on the LDL endocytosis capability of the ERα+ cell model (MCF-7 cells). LDL isolation was performed by the KBr ultracentrifugation method (Figure 1A) and validated through the evaluation of protein targets apoA-1 and apoB-100 on LDL enriched fraction, but also in VLDL, HDL enriched fractions and in plasma-free fraction. Results verified the appropriate enrichment of LDL in the desired fraction (Supplementary Figure S1). The yields in the LDL fraction were characterized by protein determination (0.76 ± 0.13 µg/µl) and LDL-cholesterol quantification (1.190 ± 0.39 mg/dL). LDL was labeled with a fluorescent dil-C18 probe, obtaining dil-LDL particles based on previous reports [29,30] (Section 2.4). Then, MCF-7 cells were treated for 24 h with dil-LDL (0–75 µg/ml) and evaluated by flow cytometry, with results suggesting MCF-7 cell internalization (Figure 1B).
Figure 1
LDL internalization and chemoresistance induction in MCF-7 cells
(A) Scheme corresponding to LDL isolation by KBr method. (B) Evaluation of dil-LDL internalization on parental MCF-7 cells (0, 7.5, 75 µg/ml) through flow-cytometry. (C) Scheme of chemoresistance protocol for generation of MCF-7Var-H and MCF-7Var-I cell variants. (A) and (C) were created in BioRender.com.
For chemoresistance induction, for the first time, MCF-7 cultures were incubated under an equimolar ratio of 4OH-Tam and endoxifen, the primary plasmatic tamoxifen-derived metabolites. In this regard, we generated two cellular variants (Figure 1C); for the variant denominated MCF-7Var-H, we treated MCF7 cells with 1 µM of 4OH-Tam and 1 µM Endoxifen for a 30-day period; the metabolites concentration was increased to 2 µM and treatment was continued for 15 days. Later, cell cultures remained in a proliferation and recovery period for 45 days with 50 nM of metabolites (Figure 1C). For MCF-7Var-I, cells were treated with 1 µM of 4-OH Tam and 1 µM Endoxifen for a 30-day period and concomitant 75 µg/ml concentration of LDL. Metabolite concentration was increased to 2 µM, and treatment was continued for 15 days; LDL concentration remained unchanged. After treatment was finished, cells remained in a proliferation and recovery period for 45 days with 50 nM of metabolites plus 7.5 µg/ml of LDL (Figure 1C).
Chemoresistance induction by tamoxifen’s metabolites
Later, we determined the chemoresistance through half-inhibitory concentration (IC50) for 4OH-Tam and endoxifen on cellular variants regarding parental cells, using the MTT assay. A range of 0.125–8 μM metabolite concentration was employed and treatments were maintained for 72 h (Figure 2B,C). The IC50 values of 4OH-Tam (7.64 µM) and endoxifen (7.85 µM) were higher for MCF-7Var-H than for the parental cells, whose values were 6.16 and 5.33 µM, respectively (Figure 2B,C).
Figure 2
Chemoresistance evaluation in MCF-7 derived cellular variants
(A) Tamoxifen metabolites chemical structures and in vivo synthesis representation by CYP isoenzymes, structures were generated with ChemDraw. Cell viability for 4OH-Tam (B) and Endoxifen (C) evaluated by MTT assay in a metabolite range of 0.25–8 µM. Treatments were performed during 72 h. Inserts showed values of inhibitory concentration 50 (IC50) in cellular variants.
On the other hand, MCF-7Var-I showed IC50 values of 12.36 and 5.7 µM for 4OH-Tam and endoxifen, respectively (Figure 2B,C), a different behavior from the observed one in MCF-7Var-H. For the first time, we reported the chemoresistance acquisition under equimolar concentrations of the main pharmacological active tamoxifen-derived metabolites and the influence of high LDL concentrations (75 µg/ml).
Important hormonal receptor down-regulation associated with chemoresistance in MCF7VarH cells
Several clinical studies have described the association between chemoresistance acquisition and the down-regulation of hormonal receptors in breast tumors [11,31]. Then, hormonal receptor status in chemoresistant cellular variants was evaluated. We analyzed the expression of ERα, HER2, and progesterone receptor (PR). A significant down-regulation in the three hormonal receptor expression levels in the MCF7VarH was registered (Figure 3A). However, MCF7VarI, obtained by co-incubation of MCF-7 cells with tamoxifen’s metabolites and LDL, did not display these drastic changes (Figure 3A). Considering that the 4-OH Tam and endoxifen major pharmacological effects are related to ERα inhibition, we decided to evaluate the receptor levels in cellular variants under treatment with tamoxifen metabolites (8 µM), a standard concentration according to IC50 values. Interestingly, down-regulation of ERα remains in MCF7VarH cells even under 4OH-Tam and endoxifen treatment, and its effect on MCF-7 and MCF7VarI appears to be not statistically significant (Figure 3B,C).
Figure 3
Chemoresistance is associated with a transition of ER+ to TNBC like phenotype
(A) Evaluation of hormonal receptor expression, ERα, PR and HER2 by western blot. β-actin was used as a loading control in parental, MCF-7Var-H, and MCF-7Var-I cells. (B) ERα expression under 4OH-Tam and endoxifen treatment (8 µM) in three cellular variants. (C) Densitometry analysis of ERα; results are reported as mean ± SD (n = 3); ***P<0.001 regarding respective control. (D) Expression of PCNA in variants MCF-7, MCF-7Var-H, and MCF-7Var-I cells. (E) Semiquantitative analysis of PCNA in three cellular variants. Optical microscopy images corresponding to MCF-7 (F), MCF-7Var-H (G), and MCF-7Var-I cells (H), images were taken at low cellular density with 40× objective.
As a complementary characterization, proliferation cell markers were evaluated. Expression levels of PCNA related to cell division were higher in MCF-7Var-H and MCF-7Var-I cells with regard to MCF-7 parental (Figure 3D,E). Likewise, we observed an important morphological change in MCF-7Var-H, these cells acquired a mesenchymal-like morphology distinct to epithelial morphology of MCF-7 and MCF-7Var-I cells (Figure 3F–H). The down-regulation of hormonal receptors, PCNA levels, and the mesenchymal morphology led us to suggest the induction of TNBC-like phenotype in MCF-7Var-H variant.
MCF-7Var-H cells showed enhanced migratory capability in vitro and augmented metalloproteinase secretion
TNBC tumors are associated with aggressive behavior and a poorer prognosis than other molecular subtypes. Then, we decided to evaluate the metastatic potential in resistant cellular variants, mainly in MCF-7Var-H, to determine the potential acquisition of TNBC-like features. First, we assessed migration capability through wound healing assay (Figure 4A), and we found that MCF-7Var-H showed high migratory capability to the wound area under FBS 10% stimuli. Meanwhile, neither MCF-7 nor MCF-7Var-I showed migratory capacity under the evaluated conditions (Figure 4A,B). We identified in wound healing assay a mesenchymal morphology in MCF-7Var-H migrated cells, mainly on the migration fronts (Figure 4A), probably associated with rearrangements in their cytoskeleton.
Figure 4
Characterization of cell migration and invasion capability in chemoresistant cellular variants
(A) Representative images of the wound-healing assay in MCF-7, MCF-7Var-H, and MCF-7Var-I cells under FBS stimuli (10%) and respective controls. In an amplified image of cells, lamellipodia and filopodia were observed in MCF-7VarH migrating cells, indicated by a red arrow. (B) Wound-healing results expressed as percentage respect to controls in parental, MCF-7Var-H, and MCF-7Var-I cells. Mean values are presented (n=3, mean ± SD); **P<0.005 respect to MCF-7 control. (C) Under tamoxifen-derived metabolites treatment (8 µM), evaluation of FAK activation (p-FAKY397) by Western blot. β-Actin was used as a loading control. (D) Representative image of MMP-9 activity in extracellular media of cellular variants, under 4-OHTam and endoxifen treatment (8 µM). (E) Statistic evaluation of MMP-9 activity in extracellular media of three biological replicates, ***P<0.0001 with respect to controls.
In aggressive tumors, the Focal Adhesion Kinase (FAK) pathway is often hyperactivated, promoting stromal remodeling and inducing tissue stiffness, cell proliferation, survival, and chemoresistance [32]. Then, we evaluated FAK expression and activation (p-FAKY397); data suggest FAK activation (p-FAKY397) in MCF-7Var-H (Figure 4C), wherein the data corroborate a potential greater malignancy in MCF-7Var-H compared to MCF-7Var-I and parental cells. Even more, our previous evidence showed that FAK pathway activation was associated with chemoresistance in TNBC cells [26]
In a complementary way, we evaluated the extracellular matrix degradation capability of the chemoresistant variants and parental cells, a phenomenon associated with metastasis processes [33]. In the first focus, matrix metallopeptidase 9 (MMP-9) activity was evaluated in the extracellular media of cellular variants. In an important way, we observed a high activity of MMP-9-mediated collagen degradation only in MCF-7Var-H cells (Figure 4D,E); this phenomenon was not affected by 4OH-Tam or endoxifen treatments (8 µM) (Figure 4D), a concentration close to the IC50 for both metabolites. This phenomenon suggests a high invasive capability in the MCF-7Var-H despite pharmacological treatment. Even more, it has been described that metabolic adaptations can occur, for instance, in targets involved in lipoprotein metabolism, associated with an increase in tumor aggressiveness [34].
Metabolic evaluation and the LDL role in chemoresistance acquisition
We evaluated LDL internalization in cellular variants after 7.5 µg/ml dil-LDL treatment (24 h). Results suggest a slightly higher capability for LDL endocytosis in MCF-7Var-H (97.12%) cells in comparison with MCF-7 (80.46%) (Figure 5A). Indeed, the lowest levels of LDL internalization were recorded in variant MCF-7Var-I (7.82%) (Figure 5B). In an attempt to explain this behavior, we determined the LDLR expression. We observed a correspondence of LDL internalization capability and LDLR expression in chemoresistant variants (Figure 5C,D).
Figure 5
Characterization of lipid metabolism in chemoresistant variants
LDL internalization in MCF-7Var-H (A) and MCF-7Var-I (B) regarding MCF-7 parental cells. Cell cultures were dil-LDL (7.5 µg/l) treated for 24 h. C) Expression of the protein targets LDLR and FAS in MCF-7, MCF-7Var-H and MCF-7Var-I cells. Densitometry analysis of LDLR (D) and FAS (E), mean values are presented (n=3, mean ± SD) and expressed as % of control, *P<0.01. β-Actin was used as a loading control.
A down-regulation trend in Fatty Acid Synthase (FAS) was registered in MCF-7Var-H as well as an increase in the MCF-7Var-I, suggesting a potential different rearrangement in lipid metabolism in both variants (Figure 5C,E). Other metabolic changes related to chemoresistance could take place; it is common for chemoresistant tumoral cells to have an enhanced antioxidant and detoxification capacity [23]; this response may be related to the activation of specific transcriptional factors.
Overexpression of transcription factor Nrf2 could be associated with phenotypic changes
Nrf2, a transcription factor involved in the management of cellular redox conditions [35], has been associated with chemoresistance and metastasis [36,37], and could be related to changes in hormone receptor status (personal communication). Critically, we found a Nrf2 overexpression in MCF-7Var-H cells (Figure 6A) and this behavior was not influenced by tamoxifen-derived metabolites treatment (Figure 6A). Likewise, our data suggest an overexpression of a lower molecular weight (45 kDa) Nrf2 isoform in MCF-7Var-H, and this phenomenon is also not influenced by metabolites treatment (Figure 6A). Lower molecular weight Nrf2 bands have been previously reported in vitro [38–40] and in vivo in pulmonary tumors [39]; however, in breast cancer cells have not been reported.
Figure 6
Overexpression of Nrf2 could be associated with chemoresistance induced through treatment with tamoxifen derived metabolites in MCF-7Var-H
(A) Western blot of Nrf2 expression in variants MCF-7, MCF-7Var-H, and MCF-7Var-I under tamoxifen derived metabolites treatment (8 μM). β-Actin was used as a loading control. (B) Localization of Nrf2 in MCF-7, MCF-7Var-H, and MCF-7Var-I cells. Images showed nuclei stained with propidium iodide (PI) (red), FITC-Nrf2 (green) and the merge. Scale bar corresponds to 50 µm. (C) Characterization of the expression of HO-1, NQO1 and SULT1A1 in MCF-7, MCF-7Var-H, and MCF-7Var-I cells under the tamoxifen metabolites treatment (8 µM). β-Actin was used as a loading control. (D) Overall survival plot for low and high expression of NFE2L2 (Nrf2 coding gene) in BC patients from GEPIA.
In a complementary way, we characterized the Nrf2 cellular localization; we employed the monoclonal Nrf2 antibody and FITC coupled secondary antibody, and nuclear colocalization was determined by the use of propidium iodide. Results suggest nuclear and cytoplasmic Nrf2 localization in MCF-7Var-H cells, compared with MCF-7Var-I and parental cells (Figure 6B) wherein its localization is mostly in cytoplasm. Several controls were used in these experiments (Section 2.11). Therefore, results correspond with data of protein expression (Figure 6A).
In order to broaden the impact of Nrf2 on chemoresistance development, we characterized protein targets dependent on Nrf2 activity. In this case, we found a differential expression of Nrf2 most representative targets such as heme oxygenase-1 (HO-1) and NAD(P)H dehydrogenase [quinone] 1 (NQO1) (Figure 6C). In MCF-7Var-H cells, HO-1 was overexpressed while NQO1 was down-regulated compared with MCF-7 cells. For its part, in MCF-7Var-I HO-1 and NQO1 expression levels remain similar to those in MCF-7, despite tamoxifen-derived metabolites treatment (8 µM). This phenomenon suggests the presence of a specific antioxidant profile adapted to cell survival in chemoresistant MCF-7Var-H cells (Figure 6C).
Subsequently, we evaluated the expression of SULT1A1, an enzyme whose expression is influenced by Nrf2 activity and has been related to 4OH-Tam and endoxifen inactivation by a sulfate group addition reaction [41]. Contrary to what was expected, a SULT1A1 down-regulation was observed in MCF-7Var-H cells (Figure 6C; Supplementary Figure S3), this response is not influenced by metabolites treatment (8 µM) in any of the variants. In complementary way, we performed a molecular docking characterization to evaluate SULT1A1 affinity for tamoxifen metabolites in comparison with estradiol, one of its known target molecules, founding a lower affinity (Supplementary Figure S4). Data suggest Nrf2 regulates the expression of necessary genes in the MCF-7Var-H cells based on a specific gene regulation condition, suggesting a specific adaptive cellular response.
Considering the potential implications of Nrf2 overexpression, we assessed the role of Nrf2 in breast cancer patient’s prognosis databases. Therefore, we evaluated the role of Nrf2 expression levels in breast cancer patients using the GEPIA database (http://gepia.cancer-pku.cn/about.html). The data showed a marked relationship between high expression of Nrf2 and a lower overall survival (Figure 6D).
The activation of the Nrf2 pathway may be related to chemoresistance development and augmented metastasis-related processes in the MCF-7Var-H cells, even more so with the transition toward a TNBC phenotype. In this regard, we registered an increment in Nrf2 expression dependent on the concentration of Tunicamycin (Tuni) (0–4 µg/ml) in the cellular variants (Supplementary Figure S5A,C), a compound that inhibits protein N-glycosylation in the endoplasmic reticulum (ER) [42], triggering the unfolded protein response (UPR) and activating Nrf2 [43]. Tuni concentrations as low as 0.25 μg/ml could upregulate Nrf2. Subsequently, in the evaluation of hormonal receptors ERα and HER2 expression in the cellular variants, we observed an inverse relationship among Tuni dose and hormonal receptor expression (Supplementary Figure S5B,D).
ERα+ cells internalize LDL
We focused on the LDL endocytosis capability of the ERα+ cell model (MCF-7 cells). LDL isolation was performed by the KBr ultracentrifugation method (Figure 1A) and validated through the evaluation of protein targets apoA-1 and apoB-100 on LDL enriched fraction, but also in VLDL, HDL enriched fractions and in plasma-free fraction. Results verified the appropriate enrichment of LDL in the desired fraction (Supplementary Figure S1). The yields in the LDL fraction were characterized by protein determination (0.76 ± 0.13 µg/µl) and LDL-cholesterol quantification (1.190 ± 0.39 mg/dL). LDL was labeled with a fluorescent dil-C18 probe, obtaining dil-LDL particles based on previous reports [29,30] (Section 2.4). Then, MCF-7 cells were treated for 24 h with dil-LDL (0–75 µg/ml) and evaluated by flow cytometry, with results suggesting MCF-7 cell internalization (Figure 1B).
Figure 1
LDL internalization and chemoresistance induction in MCF-7 cells
(A) Scheme corresponding to LDL isolation by KBr method. (B) Evaluation of dil-LDL internalization on parental MCF-7 cells (0, 7.5, 75 µg/ml) through flow-cytometry. (C) Scheme of chemoresistance protocol for generation of MCF-7Var-H and MCF-7Var-I cell variants. (A) and (C) were created in BioRender.com.
For chemoresistance induction, for the first time, MCF-7 cultures were incubated under an equimolar ratio of 4OH-Tam and endoxifen, the primary plasmatic tamoxifen-derived metabolites. In this regard, we generated two cellular variants (Figure 1C); for the variant denominated MCF-7Var-H, we treated MCF7 cells with 1 µM of 4OH-Tam and 1 µM Endoxifen for a 30-day period; the metabolites concentration was increased to 2 µM and treatment was continued for 15 days. Later, cell cultures remained in a proliferation and recovery period for 45 days with 50 nM of metabolites (Figure 1C). For MCF-7Var-I, cells were treated with 1 µM of 4-OH Tam and 1 µM Endoxifen for a 30-day period and concomitant 75 µg/ml concentration of LDL. Metabolite concentration was increased to 2 µM, and treatment was continued for 15 days; LDL concentration remained unchanged. After treatment was finished, cells remained in a proliferation and recovery period for 45 days with 50 nM of metabolites plus 7.5 µg/ml of LDL (Figure 1C).
Chemoresistance induction by tamoxifen’s metabolites
Later, we determined the chemoresistance through half-inhibitory concentration (IC50) for 4OH-Tam and endoxifen on cellular variants regarding parental cells, using the MTT assay. A range of 0.125–8 μM metabolite concentration was employed and treatments were maintained for 72 h (Figure 2B,C). The IC50 values of 4OH-Tam (7.64 µM) and endoxifen (7.85 µM) were higher for MCF-7Var-H than for the parental cells, whose values were 6.16 and 5.33 µM, respectively (Figure 2B,C).
Figure 2
Chemoresistance evaluation in MCF-7 derived cellular variants
(A) Tamoxifen metabolites chemical structures and in vivo synthesis representation by CYP isoenzymes, structures were generated with ChemDraw. Cell viability for 4OH-Tam (B) and Endoxifen (C) evaluated by MTT assay in a metabolite range of 0.25–8 µM. Treatments were performed during 72 h. Inserts showed values of inhibitory concentration 50 (IC50) in cellular variants.
On the other hand, MCF-7Var-I showed IC50 values of 12.36 and 5.7 µM for 4OH-Tam and endoxifen, respectively (Figure 2B,C), a different behavior from the observed one in MCF-7Var-H. For the first time, we reported the chemoresistance acquisition under equimolar concentrations of the main pharmacological active tamoxifen-derived metabolites and the influence of high LDL concentrations (75 µg/ml).
Important hormonal receptor down-regulation associated with chemoresistance in MCF7VarH cells
Several clinical studies have described the association between chemoresistance acquisition and the down-regulation of hormonal receptors in breast tumors [11,31]. Then, hormonal receptor status in chemoresistant cellular variants was evaluated. We analyzed the expression of ERα, HER2, and progesterone receptor (PR). A significant down-regulation in the three hormonal receptor expression levels in the MCF7VarH was registered (Figure 3A). However, MCF7VarI, obtained by co-incubation of MCF-7 cells with tamoxifen’s metabolites and LDL, did not display these drastic changes (Figure 3A). Considering that the 4-OH Tam and endoxifen major pharmacological effects are related to ERα inhibition, we decided to evaluate the receptor levels in cellular variants under treatment with tamoxifen metabolites (8 µM), a standard concentration according to IC50 values. Interestingly, down-regulation of ERα remains in MCF7VarH cells even under 4OH-Tam and endoxifen treatment, and its effect on MCF-7 and MCF7VarI appears to be not statistically significant (Figure 3B,C).
Figure 3
Chemoresistance is associated with a transition of ER+ to TNBC like phenotype
(A) Evaluation of hormonal receptor expression, ERα, PR and HER2 by western blot. β-actin was used as a loading control in parental, MCF-7Var-H, and MCF-7Var-I cells. (B) ERα expression under 4OH-Tam and endoxifen treatment (8 µM) in three cellular variants. (C) Densitometry analysis of ERα; results are reported as mean ± SD (n = 3); ***P<0.001 regarding respective control. (D) Expression of PCNA in variants MCF-7, MCF-7Var-H, and MCF-7Var-I cells. (E) Semiquantitative analysis of PCNA in three cellular variants. Optical microscopy images corresponding to MCF-7 (F), MCF-7Var-H (G), and MCF-7Var-I cells (H), images were taken at low cellular density with 40× objective.
As a complementary characterization, proliferation cell markers were evaluated. Expression levels of PCNA related to cell division were higher in MCF-7Var-H and MCF-7Var-I cells with regard to MCF-7 parental (Figure 3D,E). Likewise, we observed an important morphological change in MCF-7Var-H, these cells acquired a mesenchymal-like morphology distinct to epithelial morphology of MCF-7 and MCF-7Var-I cells (Figure 3F–H). The down-regulation of hormonal receptors, PCNA levels, and the mesenchymal morphology led us to suggest the induction of TNBC-like phenotype in MCF-7Var-H variant.
MCF-7Var-H cells showed enhanced migratory capability in vitro and augmented metalloproteinase secretion
TNBC tumors are associated with aggressive behavior and a poorer prognosis than other molecular subtypes. Then, we decided to evaluate the metastatic potential in resistant cellular variants, mainly in MCF-7Var-H, to determine the potential acquisition of TNBC-like features. First, we assessed migration capability through wound healing assay (Figure 4A), and we found that MCF-7Var-H showed high migratory capability to the wound area under FBS 10% stimuli. Meanwhile, neither MCF-7 nor MCF-7Var-I showed migratory capacity under the evaluated conditions (Figure 4A,B). We identified in wound healing assay a mesenchymal morphology in MCF-7Var-H migrated cells, mainly on the migration fronts (Figure 4A), probably associated with rearrangements in their cytoskeleton.
Figure 4
Characterization of cell migration and invasion capability in chemoresistant cellular variants
(A) Representative images of the wound-healing assay in MCF-7, MCF-7Var-H, and MCF-7Var-I cells under FBS stimuli (10%) and respective controls. In an amplified image of cells, lamellipodia and filopodia were observed in MCF-7VarH migrating cells, indicated by a red arrow. (B) Wound-healing results expressed as percentage respect to controls in parental, MCF-7Var-H, and MCF-7Var-I cells. Mean values are presented (n=3, mean ± SD); **P<0.005 respect to MCF-7 control. (C) Under tamoxifen-derived metabolites treatment (8 µM), evaluation of FAK activation (p-FAKY397) by Western blot. β-Actin was used as a loading control. (D) Representative image of MMP-9 activity in extracellular media of cellular variants, under 4-OHTam and endoxifen treatment (8 µM). (E) Statistic evaluation of MMP-9 activity in extracellular media of three biological replicates, ***P<0.0001 with respect to controls.
In aggressive tumors, the Focal Adhesion Kinase (FAK) pathway is often hyperactivated, promoting stromal remodeling and inducing tissue stiffness, cell proliferation, survival, and chemoresistance [32]. Then, we evaluated FAK expression and activation (p-FAKY397); data suggest FAK activation (p-FAKY397) in MCF-7Var-H (Figure 4C), wherein the data corroborate a potential greater malignancy in MCF-7Var-H compared to MCF-7Var-I and parental cells. Even more, our previous evidence showed that FAK pathway activation was associated with chemoresistance in TNBC cells [26]
In a complementary way, we evaluated the extracellular matrix degradation capability of the chemoresistant variants and parental cells, a phenomenon associated with metastasis processes [33]. In the first focus, matrix metallopeptidase 9 (MMP-9) activity was evaluated in the extracellular media of cellular variants. In an important way, we observed a high activity of MMP-9-mediated collagen degradation only in MCF-7Var-H cells (Figure 4D,E); this phenomenon was not affected by 4OH-Tam or endoxifen treatments (8 µM) (Figure 4D), a concentration close to the IC50 for both metabolites. This phenomenon suggests a high invasive capability in the MCF-7Var-H despite pharmacological treatment. Even more, it has been described that metabolic adaptations can occur, for instance, in targets involved in lipoprotein metabolism, associated with an increase in tumor aggressiveness [34].
Metabolic evaluation and the LDL role in chemoresistance acquisition
We evaluated LDL internalization in cellular variants after 7.5 µg/ml dil-LDL treatment (24 h). Results suggest a slightly higher capability for LDL endocytosis in MCF-7Var-H (97.12%) cells in comparison with MCF-7 (80.46%) (Figure 5A). Indeed, the lowest levels of LDL internalization were recorded in variant MCF-7Var-I (7.82%) (Figure 5B). In an attempt to explain this behavior, we determined the LDLR expression. We observed a correspondence of LDL internalization capability and LDLR expression in chemoresistant variants (Figure 5C,D).
Figure 5
Characterization of lipid metabolism in chemoresistant variants
LDL internalization in MCF-7Var-H (A) and MCF-7Var-I (B) regarding MCF-7 parental cells. Cell cultures were dil-LDL (7.5 µg/l) treated for 24 h. C) Expression of the protein targets LDLR and FAS in MCF-7, MCF-7Var-H and MCF-7Var-I cells. Densitometry analysis of LDLR (D) and FAS (E), mean values are presented (n=3, mean ± SD) and expressed as % of control, *P<0.01. β-Actin was used as a loading control.
A down-regulation trend in Fatty Acid Synthase (FAS) was registered in MCF-7Var-H as well as an increase in the MCF-7Var-I, suggesting a potential different rearrangement in lipid metabolism in both variants (Figure 5C,E). Other metabolic changes related to chemoresistance could take place; it is common for chemoresistant tumoral cells to have an enhanced antioxidant and detoxification capacity [23]; this response may be related to the activation of specific transcriptional factors.
Overexpression of transcription factor Nrf2 could be associated with phenotypic changes
Nrf2, a transcription factor involved in the management of cellular redox conditions [35], has been associated with chemoresistance and metastasis [36,37], and could be related to changes in hormone receptor status (personal communication). Critically, we found a Nrf2 overexpression in MCF-7Var-H cells (Figure 6A) and this behavior was not influenced by tamoxifen-derived metabolites treatment (Figure 6A). Likewise, our data suggest an overexpression of a lower molecular weight (45 kDa) Nrf2 isoform in MCF-7Var-H, and this phenomenon is also not influenced by metabolites treatment (Figure 6A). Lower molecular weight Nrf2 bands have been previously reported in vitro [38–40] and in vivo in pulmonary tumors [39]; however, in breast cancer cells have not been reported.
Figure 6
Overexpression of Nrf2 could be associated with chemoresistance induced through treatment with tamoxifen derived metabolites in MCF-7Var-H
(A) Western blot of Nrf2 expression in variants MCF-7, MCF-7Var-H, and MCF-7Var-I under tamoxifen derived metabolites treatment (8 μM). β-Actin was used as a loading control. (B) Localization of Nrf2 in MCF-7, MCF-7Var-H, and MCF-7Var-I cells. Images showed nuclei stained with propidium iodide (PI) (red), FITC-Nrf2 (green) and the merge. Scale bar corresponds to 50 µm. (C) Characterization of the expression of HO-1, NQO1 and SULT1A1 in MCF-7, MCF-7Var-H, and MCF-7Var-I cells under the tamoxifen metabolites treatment (8 µM). β-Actin was used as a loading control. (D) Overall survival plot for low and high expression of NFE2L2 (Nrf2 coding gene) in BC patients from GEPIA.
In a complementary way, we characterized the Nrf2 cellular localization; we employed the monoclonal Nrf2 antibody and FITC coupled secondary antibody, and nuclear colocalization was determined by the use of propidium iodide. Results suggest nuclear and cytoplasmic Nrf2 localization in MCF-7Var-H cells, compared with MCF-7Var-I and parental cells (Figure 6B) wherein its localization is mostly in cytoplasm. Several controls were used in these experiments (Section 2.11). Therefore, results correspond with data of protein expression (Figure 6A).
In order to broaden the impact of Nrf2 on chemoresistance development, we characterized protein targets dependent on Nrf2 activity. In this case, we found a differential expression of Nrf2 most representative targets such as heme oxygenase-1 (HO-1) and NAD(P)H dehydrogenase [quinone] 1 (NQO1) (Figure 6C). In MCF-7Var-H cells, HO-1 was overexpressed while NQO1 was down-regulated compared with MCF-7 cells. For its part, in MCF-7Var-I HO-1 and NQO1 expression levels remain similar to those in MCF-7, despite tamoxifen-derived metabolites treatment (8 µM). This phenomenon suggests the presence of a specific antioxidant profile adapted to cell survival in chemoresistant MCF-7Var-H cells (Figure 6C).
Subsequently, we evaluated the expression of SULT1A1, an enzyme whose expression is influenced by Nrf2 activity and has been related to 4OH-Tam and endoxifen inactivation by a sulfate group addition reaction [41]. Contrary to what was expected, a SULT1A1 down-regulation was observed in MCF-7Var-H cells (Figure 6C; Supplementary Figure S3), this response is not influenced by metabolites treatment (8 µM) in any of the variants. In complementary way, we performed a molecular docking characterization to evaluate SULT1A1 affinity for tamoxifen metabolites in comparison with estradiol, one of its known target molecules, founding a lower affinity (Supplementary Figure S4). Data suggest Nrf2 regulates the expression of necessary genes in the MCF-7Var-H cells based on a specific gene regulation condition, suggesting a specific adaptive cellular response.
Considering the potential implications of Nrf2 overexpression, we assessed the role of Nrf2 in breast cancer patient’s prognosis databases. Therefore, we evaluated the role of Nrf2 expression levels in breast cancer patients using the GEPIA database (http://gepia.cancer-pku.cn/about.html). The data showed a marked relationship between high expression of Nrf2 and a lower overall survival (Figure 6D).
The activation of the Nrf2 pathway may be related to chemoresistance development and augmented metastasis-related processes in the MCF-7Var-H cells, even more so with the transition toward a TNBC phenotype. In this regard, we registered an increment in Nrf2 expression dependent on the concentration of Tunicamycin (Tuni) (0–4 µg/ml) in the cellular variants (Supplementary Figure S5A,C), a compound that inhibits protein N-glycosylation in the endoplasmic reticulum (ER) [42], triggering the unfolded protein response (UPR) and activating Nrf2 [43]. Tuni concentrations as low as 0.25 μg/ml could upregulate Nrf2. Subsequently, in the evaluation of hormonal receptors ERα and HER2 expression in the cellular variants, we observed an inverse relationship among Tuni dose and hormonal receptor expression (Supplementary Figure S5B,D).
Discussion
High LDL levels have been associated with BC development and progression in many in vitro [18], in vivo [19], and clinical studies [44]; however, the potential LDL impact on chemoresistance in ER+ BC cells has not been described clearly. We developed two chemoresistant cellular variants by co-incubation of ER+ MCF7 cells with endoxifen and 4-OH tam. Although several protocols have been described for tamoxifen resistance generation in MCF-7 cells [45] it is the first time that chemoresistance is achieved under an equimolar treatment scheme of the two primary plasmatic tamoxifen’s active metabolites. MCF-7Var-I was obtained by maintaining high LDL concentrations (75 µg/ml) in culture media during chemoresistance acquisition period, and MCF-7Var-H was obtained with metabolites treatment only (Figure 1C, Section 2.3). This strategy was designed to evaluate the LDL effect on chemoresistance development, simulating the mechanisms triggered in ER+ breast cancer patients with high LDL plasma levels during tamoxifen treatment.
We determined chemoresistance acquisition by MTT cell viability assays in MCF-7Var-H and MCF-7Var-I by comparing their IC50 metabolites values with those of MCF-7 parental cells (Figure 2B,C). Among differences observed in cellular variants, one of the most noticeable was that MCF-7Var-H showed a significant knockdown in the hormonal receptors ERα and HER2. In contrast, although not significant, MCF-7Var-I showed a slight increase in the expression of ERα and HER2 compared to MCF-7 cells (Figure 3A–C).
MCF-7Var-H apparently showed diminished estradiol (E2)-ERα mediated signaling. Furthermore, another target of E2 is the G protein-coupled receptor 30 (GPR30) [46]; our data suggest a knockdown in GPR30 expression on MCF-7Var-H (personal communication), providing evidence that in these chemoresistance cells the response mediated by E2 is down-regulated.
Moreover, MCF-7Var-H showed enhanced proliferation corroborated by PCNA (Figure 3D,E) and Ki67 (Supplementary Figure S2) by Western blot compared with MCF-7Var-I and MCF-7. Another visible difference was the morphology, MCF-7Var-H acquired a mesenchymal-like phenotype, while MCF-7Var-I conserved an epithelial one (Figure 3F–H). Those results led us to propose the similarity of MCF-7Var-H variant with TNBC cells, which are commonly related to a basal phenotype and have a greater metastatic capability compared with ER+ BC cells [47]. To corroborate this proposal, we evaluated some processes associated with metastasis in vitro, such as migration capability and MMP-9 activity in extracellular media. As expected, MCF-7Var-H had an increased migratory capability compared with MCF-7Var-I and MCF-7 (Figure 4A,B). Indeed, the mesenchymal morphology of MCF-7Var-H cells at the wound edge was evident. In addition, only MCF-7Var-H metalloproteinase secretion was detected in zymography, which was not affected by treatment with Tamoxifen metabolites (Figure 4D,E). Therefore, MCF-7Var-H acquired a TNBC like phenotype, while MCF-7VarI maintained its luminal cell behavior and characteristics (Figure 7).
Figure 7
Phenotypic differences between MCF-7Var-H and MCF-7Var-I
MCF-7Var-H (left side) generated by treating the MCF-7 cell line with equimolar concentrations of 4-OH Tam and Endoxifen showed decreased hormonal receptors, increased migration capacity, and increased MMP-9 secretion, features similar to those observed in TNBC. Our proposed mechanism involves an increased Nrf2 expression and activity favored by metabolites ROS generation. We suggest that Nrf2 activity promotes proteins related to migration and invasion processes transcription. MCF-7Var-I (right side) generated by treating the MCF-7 cell line with equimolar concentrations of 4-OH Tam and Endoxifen in addition to 75 mg/dl of LDL cholesterol maintained its luminal ER+ phenotype related to an epithelial morphology and a low migration capability. We proposed that high intracellular concentrations of cholesterol, generated by LDL treatment and the effect of tamoxifen metabolites in LDL intracellular traffic, favor Nrf2 interaction with cav-1 in lipid rafts, preventing its nuclear accumulation and transcriptional activity. Dashed lines refer to proposed mechanisms; solid lines refer to observed phenomena in variants. Triple negative breast cancer (TNBC), low-density lipoprotein receptor (LDL), human epidermal growth factor receptor type 2 (HER2), estrogen receptor α (ERα), cholesterol (CHOL), fatty acid synthase (FAS), metalloproteinase-9 (MMP-9), reactive oxygen species (ROS), antioxidant response element (ARE). Created with BioRender.com
In another instance, we registered significant changes in lipid metabolism targets in both variants. MCF-7VarI showed an augmented expression of FAS which catalyzes palmitic acid syntheses using acetyl-CoA, malonyl-CoA and NADPH as cofactor [48], suggesting active lipogenic metabolism in this variant. MCF-7VarI also exhibited a significant low dil-LDL internalization of only 7.82%, that could be related to a negative feedback mechanism due to exposure to high LDL concentrations during chemoresistance acquisition. For its part, MCF-7VarH expressed statistically significant higher LDLR expression levels than MCF-7 and showed a corresponding enhanced LDL internalization after treatment with 7.5µg/ml dil-LDL (Figure 5). Previously, increased LDL internalization and neutral lipid storage were described in the TNBC cell line MDA-MB-231 compared with MCF-7 [49]. This evidence suggests again an ER+-TNBC transition in MCF-7VarH cells.
Furthermore, it has been described that SERMS such as tamoxifen and its metabolites can alter lipidic metabolism in cells independently of their action on ERα [50]. SERMS can favor the activation of SREBP2 and, in turn, inhibit the intracellular traffic of LDL-cholesterol, increasing cellular cholesterol accumulation [51]. Cholesterol accumulation facilitates the generation of oxysterols such as 27-hydroxy-cholesterol [34], metabolite related to ERα cell proliferation in low E2 environments [34,52]. E2-deprived MCF-7 cells exert an increased response to oxysterols in contrast with MCF-7 native cells, which favors REα mediated transcription [53]. To explain the phenotypic differences observed in MCF-7Var-H and MCF-7Var-I, mainly in relation to hormonal receptors expression, a proposal generated so far is that LDL overload achieved in MCF-7VarI may have favored 27-hydroxy-cholesterol intracellular generation and ERα activation, allowing latency of this receptor.
Other significant changes were described in both variants, such as increased expression of the Nrf2 master regulator of redox homeostasis, but in a significant way in MCF-7VarH (Figure 6A). We also observed increased nuclear localization in this cell variant (Figure 6B). It has been widely described that Nrf2 overexpression in different tumor types contributes to pro-oncogenic processes and chemoresistance [54] by promoting the transcription of antioxidant (NQO1), detoxifying (SULT1A1), and anti-apoptotic (BCL-2) enzymes with ARE element [55]. It has been described that Tamoxifen treatment can promote ROS generation, favoring Nrf2 activation and chemoresistance acquisition to therapy and other therapies in a concomitant way [56]. Interestingly, although both variants were generated by tamoxifen metabolites treatment, an increased expression of Nrf2 and augmented nuclear localization was observed predominately in MCF-7VarH. Inactive Nrf2 is mainly localized in cell membranes interacting with caveolin-1 (Cav-1) in lipid rafts. Lipid raft formation favored by high cholesterol concentrations, such as in MCF-7Var-I, could hinder Nrf2 nuclear translocation [57] and explain the observed behavior among variants (Figure 7).
Furthermore, differential expression of Nrf2-dependent genes was observed in both variants. Although the canonical Nrf2 pathway involves Nrf2-sMaf complex formation, other transcription factors could be influencing Nrf2 differential downstream genes expression among variants. It has been described that Nrf2 interacts physically and forms a complex with Activating Transcription Factor 4 (ATF4) promoting specifically HO-1 transcription, while Nrf2-ATF3 can act as a repressor of Glutathione S-transferase (GST) expression [58]. Even more, observed metabolic adaptation in variants could be related to Nrf2 action as well. It has been described that increased Nrf2 signaling is associated with suppression of lipogenesis related genes like FAS to avoid NADPH expenditure, necessary for detoxification reactions [59], which corresponds with what we observed in MCF-7Var-H.
Finally, Nrf2 could be implicated in chemoresistance adaptations in other ways that have not yet been described. Notably, under the subtle induction of ER stress by Tuni and FBS, which leads to an indirect Nrf2 activation, data suggest an inverse relationship between Tuni dose and hormonal receptor expression. This relationship has to be verified with more analysis.
Elucidating the role of LDL in tamoxifen chemoresistance development and the mechanism involved bring us closer to propose new strategies and/or pharmacological targets to prevent or counteract chemoresistance, avoiding complications in ER+ tumors treatment. Nrf2 inhibition is a promising therapeutic approach for Nrf2-dependent cancers and Nrf2-inhibitors are actively being developed in our group (personal communication).
Conclusions
We evaluated the LDL involvement in chemoresistance acquisition to the two mainly active tamoxifen metabolites, endoxifen and 4-OHTam, by the development of two cellular variants derived from an ERα+ (MCF-7) cell line. We observed a completely different behavior when comparing variant MCF-7VarH generated by tamoxifen metabolites treatment and variant MCF-7VarI generated by co-incubation of tamoxifen metabolites and high LDL concentrations (75 µg/ml). Variants showed differences in the IC50 to metabolites, morphology, expression of hormone receptors, metabolic characteristics, and processes associated with metastasis. MCF-7VarH showed similar features to those presented by triple-negative cells and increased antioxidant response, while MCF-7VarI only showed subtle changes in its phenotype compared with MCF-7 cells. A better understanding of the metabolic changes displayed in these cellular variants allows us to propose new pharmacological targets, such as the master factor of the antioxidant response Nrf2, which can be modulated with current pharmacological treatments and potentially allow recurrence-free recovery in patients.
Supplementary Material
Supplementary Figures S1-S9 |
Title: Anticancer Potential of Tocopherols-Containing Plants and Semi-Synthetic Tocopherols | Body:
1. Introduction
Bioactive compounds are substances with biological activity that directly affect living organisms. The effects of bioactive chemicals can differ based on the substance, dosage, and bioavailability [1]. However, bioactive compounds consumed at sufficient levels can provide medicinal advantages, including avoiding and curing several illnesses [2]. Recently, the increase in illnesses, including heart attack, obesity, and carcinoma, has led to the development of research on herbal plants as drugs in the pharmaceutical industry sector for cancer prevention. Bioactive compounds are found in fruits, vegetables, and grains [3,4]. They belong to a very heterogeneous class of compounds (polyphenolic compounds, carotenoids, phytosterols, and tocopherols) with different chemical structures (hydrophilic or lipophilic) [3,5].
Tocopherols are various forms of vitamin E, which is a group of bioactive, fat-soluble compounds composed of tocopherols and tocotrienols as two main classes, essential for human nutrition [6]. Alpha-tocopherol is the most biologically active among beta-, gamma- and delta-tocopherol in human. According to epidemiological studies, vitamin E could potentially reduce cancer and heart attacks, increase the immune system’s efficiency, and protect against various age-degenerative conditions (cataracts, spinal disorders, and arthritis) [7]. The antioxidant vitamin E supplements, have been used for its antioxidant qualities to reduce the risk of developing cancer [7]. Tocopherols have different variants, namely four tocopherol groups: α, β, γ, and δ. The difference in tocopherol groups lies in their function and ability to inhibit cells [8]. Tocopherol derivatives can be synthesized through various chemical modifications, which have been shown to enhance their anti-cancer properties, making them promising candidates for further pharmacological research [9]. Plants can be the main source of the body’s antioxidant vitamin E since tocopherols can be exclusively generated by photosynthesis in organisms like plants.
Exploring medicinal plants for anticancer agents is important as these plants have been used in traditional medicine for centuries and may offer novel compounds that could enhance cancer treatment [10]. Plants may produce a range of chemical compounds with biochemical and medicinal properties and are ideal for research as basic therapeutic ingredients, including anticancer [11]. The potential of plants being employed as therapeutic agents is partly because of their accumulation of bioactive chemicals [12]. Some herbal medicinal plants that can be used as inhibitors of cancer cell growth, are Typhonium flagelliforme [13,14], Moringa oleifera [15,16], Zingiber officinale [17,18,19], Curcuma longa [20,21], Withania somnifera [22,23], and Pisum fulsum [24,25].
In some herbal medicinal plants, tocopherol content is found in all plant parts, including leaves, tubers, stems, roots, and seeds [26]. Plants that have tocopherol content, as showcased in Table 1, include Bulbine anguistifolia Poelln (Asphodelaceae family), Moringa oleifera (Moringaceae family), Kigelia pinnata (Bignoniaceae family), Cissus assamica (Vitaceae family), Asclepias curassavica L. (Apocynaceae family), and Typhonium flagelliforme (Araceae family) [27,28,29,30,31,32,33].
Despite the promising potential of tocopherols as an anticancer agent, research on tocopherol-containing plants has been relatively limited compared to other bioactive compounds such as polyphenols and carotenoids [34,35,36]. This research gap can be attributed to several factors, including the complexity of isolating and studying fat-soluble compounds like tocopherol, the historical focus on synthetic vitamin E in medical research, and challenges in standardizing tocopherol content across different plant sources. However, the unique properties of plant-based tocopherols, including their natural co-occurrence with other beneficial plant compounds and potential synergistic effects, make them worthy of increased research attention. Furthermore, the growing interest in natural product-based pharmaceuticals underscores the timely nature of focused research on tocopherol-containing plants for anticancer applications.
This review discusses herbal medicinal plants that contain potential anticancer tocopherols and may be developed as anticancer drugs in further research. When testing anticancer activity using unpurified extracts, it is important to consider that the observed effects may not be solely due to the compound present in the highest concentration. The interactions between various components in the extract, including potential synergistic or antagonistic effects, can significantly influence the overall bioactivity. This review selects medicinal plants from Scopus publications from 2016–2024 because Scopus provides a comprehensive collection of multidisciplinary research, making it suitable for our focus on medicinal plants and tocopherols. Scopus was chosen for its extensive coverage of peer-reviewed literature and its robust content curation process overseen by the Content Selection and Advisory Board [37,38,39]. It appears that reviews on herbal medicinal plants containing tocopherol for anticancer purpose are not available. Furthermore, it explores the emerging field of semi-synthetic tocopherol derivatives, potentially leading to improved therapeutic outcomes in cancer treatment. The objectives of this review are as follows: (1) To select publications containing tocopherols with anticancer activities. (2) To analyze the various forms and concentrations of tocopherols in the selected medicinal plants or extracts and their contribution to anticancer activities. (3) To identify factors influencing the anticancer activities of tocopherols in the selected plants, including common traits among species, plant parts utilized, tocopherol forms and concentrations, the presence of other phytochemicals, stability, degradation, and analytical methods employed.
2. Tocopherol Amounts in Selected Herbal Medicinal Plant Species
Anticancer tocopherol variants in herbal medicinal plants and their biological activity are shown in Table 1, with a total of 15 publications derived from the literature search (see Section 5 for literature search and plant selection method). Table 1 presents the results of diverse studies aimed at extracting tocopherol variants from various medicinal plant species, focusing on inhibiting cancer cells. The chemical structure forms of tocopherol compounds presented in Table 1 are illustrated in Figure 1. This figure showcases the structural variants of tocopherols, which are essential for understanding their biochemical properties and potential anticancer activities. By highlighting the different forms of tocopherols, this visual representation emphasizes the diversity of these bioactive compounds and their implications for medicinal applications.
plants-13-02994-t001_Table 1
Table 1
Compilation of existing studies on anticancer tocopherol variants in herbal medicinal plants.
No.
Plant Species
Plant Source
Biological Activity
Total Tocopherol Content
Structure Number
IC50 & Cancer Cell Lines
Reference
1.
Bulbine anguistifolia Poelln (Asphodelaceae family)
Roots extracts
Exhibit significant cytotoxic effects against cancer cells, indicating potential as an anticancer agent.
α-tocopherol succinate *: 4.16%
10
- IC50 of acetone extract = 28.21 μg/mL, COX-2- LC50 of dichloromethane extract (50% lethal concentration) = 22.46 μg/mL- MCF-7 cells and Caco-2 cell line
[28]
2.
Punica granatum L. (Punicaceae family)
Fruits extracts
Red pomegranate exhibited the highest cytotoxic activity among the three varieties, followed closely by easy red and white pomegranates, which displayed comparable effects.
Total tocols (α-, γ-, δ-tocopherol, and β-, γ-tocotrienol):- 32.1 ± 0.2 μg (white pomegranate) - 69.4 ± 1.0 μg (pink pomegranate) - 110.0 ± 0.2 μg (red pomegranate)
α-tocopherol: 1γ-tocopherol: 2δ- tocopherol: 3β-tocotrienol: 4γ- tocotrienol: 5
- IC50 > 200 μg/mL - MCF-7 cells (breast), HCT116 cells (colorectal), PC-3 cells (prostate), A549 cells (lung)
[40]
3.
Calligonum comosum (Polygonaceae family)
Leaves extracts
Significant anticancer for MCF-7 and MDA 231 cancerous breast cells. Has a specific apoptotic effect on cancer cells and shows minimal toxicity against non-cancerous breast cells (MCF-12).
DL-α-tocopherol: 14.75%
6
- IC50 = 269 μg/mL, MCF-7 cells- IC50 = 258 μg/mL- MDA 231 cells, MCF-7 cells
[41]
4.
Moringa oleifera (Moringaceae family)
Seed extract
Significant in vitro anticancer effects have been observed, demonstrating potent inhibitory activity against target proteins associated with breast and prostate cancer.
α-tocopherol: - 13.76 ± 0.89 mg/100 mL (crude oil)- 12.32 ± 0.43 mg/100 mL (degummed oil)
1
- No IC50 value mentioned- Breast and prostate cancer cells based on molecular docking
[29]
5.
Kigelia pinnata (Bignoniaceae family)
Fruits extracts
Significant cytotoxic effects against liver cancer (Huh-7), pancreatic cancer (PANC-1), colorectal cancer (Colo-205 and HT-29), gastric carcinoma (SNU-16), colorectal adenocarcinoma (SW620), and colon carcinoma (HCT116) cells. It has high free radical scavenging activity and iron ion reduction ability.
Not mentioned
-
- IC50 = 6.79 μg/mL, SW620 (colorectal adenocarcinoma)- IC50 = 8.69 μg/mL, SNU-16 (gastric carcinoma)- IC50 = 10.34 μg/mL, PANC-1 (pancreatic cancer)- IC50 = 65.55 μg/mL, Huh-7 (liver cancer)- IC50 = 52.92 μg/mL, Colo-205 (colorectal cancer)- IC50 = 91.32 μg/mL, HT-29 (colorectal cancer)
[33]
6.
Haloxylon salicornicum (Amaranthaceae family)
Whole of plant bodies extracts
Anticancer (liver cancer) activity through apoptosis (AO/EB staining) and cell cycle arrest (flow cytommetry)
Not mentioned
-
- Antioxidant IC50 = 4120 μg/mL- Liver cancer cells (HCAM)
[42]
7.
Prunus persica Linn. (Rosaceae family) & Malpighia glabra Linn. (Malpighiaceae family)
Leaves extracts
Anticancer activities on colon or HCT-116 and breast or MCF-7 cancer cells
α-tocopherol: 0.139 mg/100 g
1
- IC50 = P. persica cv. Desert red: 249.5 μg/mL, Florida prince: >1000 μg/mL, Swelling: 617 μg/mL, M.glabra: 490 μg/mL, breast cancer cells (MCF-7) - IC50 = P. persica cv. Desert red: 617 μg/mL, Florida prince: 617 μg/mL, Swelling: 490 μg/mL, M. glabra: 302 μg/mL- Colon cancer cells (HCT-116)
[43]
8.
Petroselium crispum L. & Anethum graveolens L. (Apiaceae family)
Shoot tissues extracts
Anticancer activities in HepG2 (carcinoma), Colo205 (colon carcinoma), 293 (embryonic kidney adenocarcinoma), and T24P (urinary bladder carcinoma) cell lines.
- P. crispum L. = 0.211 ± 0.09 μmole CO2/mole air, eCO2
- A. graveolens L. = 0.338 ± 0.06 μmole CO2/mole air, eCO2
1
- No IC50 value mentioned- MCF-7 (breast), HepG2 (liver), Colo205 (colorectal), and T24P (bladder) cancer cells
[44]
9.
Cissus assamica (Vitaceae family)
Stems purified extracts
Anticancer activity with cytotoxic effects on non-small cell lung carcinoma cells (NCI-H226) and colon cancer cells (HCT-116)
α-tocopherol: 200 mg tocopherol trimer IVa: 24.3 mg tocopherol trimer IVb: 28.1 mg 1,2-bis-(5- γ- tocopheril)- ethane:15.4 mg α-tocospiro B: 6.1 mg
α-tocopherol: 1tocopherol trimer IVa: 7tocopherol trimer IVb: 8α-tocospiro B: 9
Betulinic acid:- IC50 = 2.0 μM, non-small cell lung carcinoma cancer cells (NCI-H226) - IC50 = 1.6 μM, colorectal cancer cells (HCT-116) Epi-glut-5(6)- en-ol compound:- IC50 = 9.1 μM, non-small cell lung carcinoma cancer cells (NCI-H226) - IC50 = 6.0 μM, colorectal cancer cells (HCT-116)
[30]
10.
Asclepias curassavica L. (Apocynaceae family)
Whole of plant bodies extracts
Exhibits anticancer activities.
- vitamin E (tocopherol): 3.95%- γ -tocopherol: 0.79%
γ-tocopherol: 2
No IC50 value and cancer cell specificity mentioned
[31]
11.
Rubus fairholmianus (Rosaceae family)
Root methanolic column subfraction (RFM) extract
Promotes apoptosis in MCF-7 cancer cells, characterized by decreased ATP, increased LDH, increased apoptosis, and Caspase 3/7 activity
Not mentioned
-
- No IC50 value mentioned- MCF-7 (breast) cancer cells
[45]
12.
Michelia nilagirica (Magnoliaceae family)
Bark extract
α -glucosidase inhibitory activity associated with blood sugar management as well as anticancer activity in HepG2 cells
2.43%
1
- IC50 = 303.26 ± 2.30 mg/mL- HepG2 cancer cells
[46]
13.
Typhonium flagelliforme Lodd. (Araceae family)
Shoots and tubers extracts
Not investigated
Vitamin E (α-tocopherol): 0.46%
1
Not investigated
[32]
14.
Epilobium Species (E. hirsutum L., E. parviflorum Schreb., E. palustre L., E. dodonaei Vill., and E. angustifolium L.) (Onagraceae family)
Aerial parts and leaves extracts
Anticancer activities on LNCap cell line through decreasing cellular viability.
- α-tocopherol: 9435.35 ± 398.51 ng/g d.w.- δ-tocopherol: 572.76 ± 33.14 ng/g d.w.
α-tocopherol: 1δ-tocopherol: 3
- E. hirsutum L. leaves (EtOH 50%), IC50 = 6.10 μg/mL- E. hirsutum L. leaves (EtOH 30%), IC50 = 6.69 μg/mL- E. hirsutum L. aerial parts (EtOH 30%), IC50 = 9.10 μg/mL- E. parviflorum aerial parts (EtOH 30%), IC50 = 8.49 μg/mL- E. palustre aerial parts (EtOH 30%), IC50 = 5.84 μg/mL- E. dodonaei aerial parts (EtOH 30%), IC50 = 6.85 μg/mL- E. angustifolium leaves (EtOH 30%), IC50 = 7.05 μg/mL- E. angustifolium aerial parts (EtOH 30%), IC50 = 10.56 μg/mL- LNCap cancer cells
[47]
15.
Rubus sp. (Rosaceae family)
Fruits purified extracts
Anticancer against HeLa cells with an increase in apoptosis reaching 67%
430 μg/g FW
1
- IC50 = 35 μg/mL- HeLa cells
[48]
Note: This table summarizes key findings from selected studies (Scopus 2016–2024) on tocopherol-containing herbal medicinal plants and their potential anticancer properties. Compounds marked (*) are semi-synthetic. Structure numbers in the “Structure Number” column refer to Figure 1.
Table 1 also include studies employing gas chromatography-mass spectrometry (GC-MS), liquid chromatographic-mass spectrometry (LC-MS), and high-performance liquid chromatography (HPLC) analysis to determine the amount/content of tocopherol, and using 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide assay (MTT assay) to analyze the toxicity of test samples toward cancer cells.
In their study, Raletsena and Mongalo [28] identified 56 compounds in the crude acetone extract of Bulbine anguistifolia Poelln. Among these, the semi-synthetic derivative α-tocopheryl succinate was quantified at 4.16%. The study used gas chromatography time-of-flight mass spectrometry (GC-ToF-MS) analysis of the acetone extract to quantify the compound content. GC-ToF-MS is an analytical technique that combines the separation capabilities of gas chromatography (GC) with the susceptible and high-resolution detection of time-of-flight mass spectrometry (ToF-MS). Another case was with Kumar et al. [41], who used the GC-MS analysis method to obtain a more significant number of tocopherol variants. They managed to identify DL-α-Tocopherol as much as 14.75% in Calligonum comosum leaves. Many studies have used GC-MS analysis to measure high tocopherol content (Table 1).
Cortez-Trejo et al. [40] detailed the total content of tocols from the fruit of Punica granatum L. Tocols, including α-, δ-, γ-tocopherol, and β-, γ-tocotrienol, had a total amount of 32.1 ± 0.2 μg in white pomegranate, 69.4 ± 1.0 μg in pink pomegranate, and 110.0 ± 0.2 μg in red pomegranate. Red pomegranate has a higher nutrient and phytochemical content than pink and white pomegranate. Moringa oleifera, the seed part, has α-tocopherol content with α-tocopherol type of 13.76 ± 0.89 mg/100 mL in crude oil and 12.32 ± 0.43 mg/100 mL in oil that has undergone a degumming process [29]. Prunus persica Linn. from the Malpighiaceae family and Malpighia glabra Linn. from the Rosaceae family have high amounts of vitamin E (tocopherol) and vitamin C based on GC-MS analysis [43]. Vitamin E content (α-tocopherol) in P. persica cv. Florida prince reached 0.139 mg/100 g, and vitamin C in acerola (M. glabra) reached 34 mg/100 g.
As mentioned, most research in obtaining tocopherol variants is through GC-MS analysis. After knowing the variation of tocopherol content in various medicinal plant species through GC-MS analysis, it is essential to understand the factors that can affect the increase of this compound content. One method proven effective in increasing tocopherol content in plants is elevated CO2 (eCO2) treatment, as reported by Saleh et al. [44]. Elevated CO2 can stimulate metabolic pathways that produce secondary compounds, including antioxidants such as tocopherols. The eCO2 treatment can increase the accumulation of phenolic compounds and flavonoids, often associated with increased tocopherol content [44]. The total tocopherol content in Petroselium crispum L. before eCO2 treatment was 0.211 ± 0.09 μmole CO2/mole air, and after eCO2 treatment increased to 0.496 ± 0.02 μmole CO2/mole air, eCO2. The same treatment was also given to Anethum graveolens L. with total tocopherol content from 0.338 ± 0.06 μmole CO2/mole air to 0.421 ± 0.02 CO2/mole air, eCO2 [44]. The eCO2 treatment in the study could produce a significant increase in β- and γ-tocopherol levels in P. crispum yet not for A. graveolens, and δ-and α-tocopherol only increased for A. graveolens. However, total tocopherol levels increased in both plants. Other studies have also shown that increasing atmospheric CO2 concentrations can stimulate the production of (tocopherol) compounds in soybean (Glycine max) plants by increasing photosynthetic activity and protecting cell membranes from oxidative stress [49].
In addition to eCO2, other studies have explored various plant species for their unique tocopherol content and potential as bioactive agents. For example, Chan et al. [30] reported five types of tocopherols isolated from Cissus assamica species which showed significant potential in health studies. The five types of tocopherols had a total amount of tocopherols of 273.9 mg [30]. α-tocopherol was 200 mg, tocopherol trimer IVa was 24.3 mg, tocopherol trimer IVb was 28.1 mg, 1,2-bis-(5-γ-tocopheril)-ethane was 15.4 mg, and α-tocospiro B was 6.1 mg [42]. Among the five types of tocopherols, the first natural isolation of 1,2-bis-(5-γ-tocopheril)-ethane from C. assamica was identified. In addition, GC-MS analysis of Asclepias curassavica L. plants revealed that the vitamin E content, in the form of γ-tocopherol, reached 3.95% [31].
Furthermore, analytical techniques such as LC-MS and GC-MS not only help identify tocopherol content but also other compounds related to bioactive properties. In plants such as Rubus fairholmianus root methanolic column subfraction (RFM), α-tocopherol content and other chemicals lead to its biological function [45]. LC-MS analysis of RFM (at dominant LC-MS peaks at 469.71, 779, and 893) showed mass equivalents of α-tocopherol and flavonol glycosidic compounds. GC-MS analysis and Fourier transform infra-red (FT-IR) spectroscopy stated that 28 compounds were successfully detected from the extract of ethyl acetate of Michelia nilagirica stem bark, one of which was vitamin E (tocopherols) as much as 2.43% [46]. Likewise, the plant Typhonium flagelliforme Lodd. (family: Araceae) has an amount of vitamin E (α-tocopherol) as much as 0.46% in the tuber [32]. The α-tocopherol content is a new type of tocopherol detected in rodent tubers.
LC-MS analysis revealed that Epilobium species (E. parviflorum Schreb., E. hirsutum L., E. dodonaei Vill., E. angustifolium L., and E. palustre L.) contained α-tocopherol as much as 9435.35 ± 398.51 ng/g d.w and δ-tocopherol as much as 572.76 ± 33.14 ng/g d.w [47]. These amounts are the highest concentrations of α-tocopherol and δ-tocopherol. In addition to the species discussed, several other herbal medicinal plants exhibit high tocopherol content, although the total amount has not been specifically reported. These examples include Kigelia pinnata (fruit parts), Haloxylon salicornicum (throughout the plant), Rubus sp. (fruit parts), and Typhonium flagelliforme Lodd. (leaf and tuber parts), which are essential to explore further [32,42,48,50].
It is important to note that not all tocopherols have the same structure and function. Each type of tocopherol, such as α-, β-, γ-, and δ-tocopherol, has a unique role in plants and different anticancer activities [51]. In plants, α-tocopherol protects chloroplasts from damage caused by reactive oxygen species (ROS) during photosynthesis [52]. Meanwhile, β-tocopherol is less common in nature and is usually present in lower plant concentrations. γ-tocopherol plays a role in protecting plant cells from environmental stress, and δ-tocopherol is very effective in detoxifying harmful substances and protecting lipids in cell membranes from oxidation [53].
Apart from tocopherols, there are derived compounds such as tocopherol trimers IVa, IVb, 1,2-bis-(5-γ-tocopheril)-ethane, and α-tocospiro B, each with specific roles in responding to stress and maintaining cell membrane stability. Recent research has increasingly focused on understanding how these compounds are formed and accumulated under different environmental conditions, such as stress and elevated CO2 levels, to boost tocopherol content for health benefits and medical uses [54,55]. Tocopherols have beneficial biological effects by influencing gene expression, signal transmission, and altering cell function through protein-membrane interactions [51,56]. The variants of tocopherols in different medicinal plants can significantly impact their therapeutic properties, especially anticancer activity.
3. Herbal Medicinal Plants Containing Tocopherol with Anticancer Potential
Through a comprehensive understanding of tocopherol levels across a variety of plants and the factors that impact their biosynthesis, we can evaluate the potential therapeutic value of these compounds, particularly in the context of cancer treatment. Tocopherol have been shown to have potent anticancer activity (Table 1). All compounds presented in Table 1 are natural, except for those marked with an asterisk, which are semi-synthetic. For example, Raletsena and Mongalo [28] reported that the semi-synthetic tocopherol content of B. anguistifolia Poelln roots has anticancer potential against soybean lipoxygenase enzyme 15-LOX in addition to Cyclooxygenase-1 (COX-1) and COX-2. The acetone extract of this plant had a minimum inhibitory concentration (MIC) of 0.05 mg/mL for several pathogens. It showed antiproliferative effects against MCF-7 cells with LC50 concentrations of 25.33 and 22.46 μg/mL for acetone and dichloromethane extracts, respectively. The IC50 concentration of dichloromethane extract for 2,2′-Azino-bis(3-ethylbenzothiazoline-6-sulfonic acid) (ABTS) was 12.52 μg/mL, while the IC50 concentration of acetone extract against 2,2-Diphenyl-1-picrylhydrazyl (DPPH) was 2.88 μg/mL.
Kumar et al. [41] observed that the ethanol extract from C. comosum (EECC) demonstrates significant antiproliferative effects against MCF-7 (IC50 approximately 269 μg/mL) and MDA-MB-231 (IC50 approximately 258 μg/mL) breast cancer cells. Treatment with EECC induced notable apoptosis in both types of cancer cells, suggesting its potential as an effective anticancer agent. Additionally, Cortez-Trejo et al. [40] reported that pomegranate extract exhibits anticancer potential by inhibiting cancer cell growth. Moreover, the presence of tocopherol and phenolic compounds in pomegranate contributes to maintaining cell integrity, preventing genetic mutations, and reducing inflammation, which is a known risk factor for various cancer types [40]. Furthermore, molecular docking studies have shown that M. oleifera seed oil displays promising anticancer potential compared to degummed oil [29].
The methanolic concentrate of another species, the ripe fruit of K. penata, had significantly greater cytotoxic activities compared to the ethyl acetate extract and even compared to the drug doxorubicin, a commonly used chemotherapeutic agent [33]. Interestingly, the study showed that this extract exhibited substantial cytotoxic activity on numerous types of cancer cells, suchHuh-7 (liver cancer), PANC-1 (pancreatic cancer), Colo-205, HT-29, SW620, and HCT116 (all colorectal cancers), as well as SNU-16 (gastric carcinoma). Another study also showed that P. crispum L. and A. graveolens L. plants have anticancer potential in cell lines HepG2 (carcinoma), Colo205 (colon carcinoma), 293 (embryonic kidney adenocarcinoma), and T24P (urinary bladder carcinoma) [44]. This aligns with the results obtained from M. glabra plant extract, which showed anticancer activity against colon cancer cells with effectiveness equivalent to standard doxorubicin (0.1 μg/mL) [43]. These findings underscore the importance of further exploration of the M. glabra plant as a potential cancer drug.
However, not all plant extracts show similar anticancer activity. For example, H. salicornicum only has cytotoxic effects on liver cancer cells (HCAM), using microwave-assisted extraction (MAE) with an ethanol solvent [42]. The study showed significant toxicity activity at a dosage of 1000 μg/mL, with a cell death rate of 42.35%. This suggests that the anticancer activity of H. salicornicum may be more limited compared to other plants tested. On the other hand, tocopherols in C. assamica plants showed promising cytotoxicity potential against non-small cell lung carcinoma (NCI-H226) and colon cancer (HCT-116) with IC50 values in the range of 1.6 to 9.1 μM [30]. The significant decrease in cell viability at specific concentrations suggests that tocopherol can induce apoptosis, adding to the evidence that tocopherol plays a role in inhibiting cancer cell proliferation. For example, the plant R. fairholmianus showed a 67.73% decrease in cell viability at a concentration of 20 μg/mL, which induced apoptosis [45]. In addition, Rubus sp. has anticancer activity against HeLa cells, with an increase in apoptosis reaching 67% [48]. The PI3K/PTEN/AKT/mTOR pathway is a key signaling route involved in regulating cell growth, proliferation, and metabolism. Similar findings were also found in M. nilagirica and Epilobium [46,47].
Extraction of T. flagelliforme Lodd. has exhibited significant potential in mitigating breast cancer cell proliferation, attributed to its substantial tocopherol content [13,14,57]. Tocopherol, recognized for its potent anticancer properties, mitigates oxidative stress in cancer cells, thereby precipitating apoptosis and impeding cancer cell proliferation by safeguarding cellular membrane structures against oxidative harm [58]. These properties hold significance as oxidative stress contributes to the onset and progression of cancer. The foregoing studies affirm the antioxidative function of tocopherols in herbal medicinal plants, effectively curtailing cancer cell proliferation and instigating apoptosis.
Accordingly, these findings underscore the promising potential of plants containing tocopherol and methanolic extracts as prospective anticancer agents. Therefore, the utilization of tocopherol-containing plants as anticancer agents offers a potential natural approach to cancer therapy and could be the basis for developing plant-based anticancer drugs in the future.
4. Semi-Synthetic Tocopherols with Anticancer Potential
4.1. Semi-Synthetic Tocopherol Derivatives
Recent studies have significantly expanded our understanding of tocopherols and their derivatives in cancer research. A comprehensive review by Baj et al. [9] highlighted several promising semi-synthetic tocopherol derivatives. α-tocopherol succinate (α-TOS) has shown remarkable potential by inducing apoptosis in breast, prostate, and colon cancer cell lines through the disruption of mitochondrial function and generation of reactive oxygen species [9,59]. Another noteworthy derivative is tocopheryl polyethylene glycol succinate (TPGS), which enhances the bioavailability and efficacy of anticancer drugs. TPGS acts as a P-glycoprotein inhibitor, potentially overcoming multidrug resistance in cancer cells [9,60]. The development of redox-silent vitamin E analogs, such as α-tocopheryl ether-linked acetic acid (α-TEA) and α-tocopheroloxybutyric acid (α-TOS), represents another significant advancement. These compounds induce apoptosis independent of their antioxidant properties and have demonstrated effectiveness against breast, prostate, and ovarian cancer cells [9,61]. This research highlights the potential of modified tocopherols to enhance anticancer efficacy beyond the capabilities of their natural counterparts.
Synthetic derivatives of tocopherols have also shown promising results in cancer research. For instance, α-tocopherol succinate (α-TOS), a semi-synthetic derivative of α-tocopherol, has demonstrated potent anticancer effects across various cancer types. α-TOS has been shown to selectively induce apoptosis in malignant cells while sparing normal cells, making it an attractive candidate for cancer therapy [62,63]. Another synthetic analog, α-tocopherol phosphate (α-TP), has exhibited enhanced bioavailability and stability compared to natural tocopherols, potentially leading to improved anticancer efficacy [64]. These synthetic models provide valuable insights into structure-activity relationships and may guide the development of more potent tocopherol-based anticancer agents.
Moreover, tocopherols have shown potential as adjuvants in chemotherapy, enhancing the efficacy of conventional anticancer drugs while potentially reducing their side effects. For example, α-tocopherol has been found to increase the sensitivity of multidrug-resistant cancer cells to chemotherapeutic agents like doxorubicin [65]. Additionally, γ-tocopherol has demonstrated synergistic effects with cisplatin in lung cancer cells, enhancing the drug’s cytotoxicity while protecting normal cells from oxidative damage [66]. These findings suggest that combining tocopherols with standard chemotherapy regimens could improve treatment outcomes and reduce toxicity, opening new avenues for cancer treatment strategies.
4.2. Mechanisms of Action
Jiang et al. [67] elaborated on several mechanisms through which tocopherols exert their anticancer effects. A key mechanism involves the regulation of Peroxisome Proliferator-Activated Receptors (PPARs), particularly PPARγ, leading to decreased cell proliferation and increased apoptosis in various cancer types [67,68]. Additionally, tocopherols, especially γ-tocopherol, inhibit HMG-CoA reductase, which can lead to decreased cholesterol synthesis crucial for rapidly dividing cancer cells [67]. Beyond their antioxidant properties, tocopherols exhibit anticancer effects through various antioxidant-independent mechanisms. These include the inhibition of COX-2 and 5-LOX enzymes, reducing inflammation-related cancer progression [67,69], and the modulation of sphingolipid metabolism, which affects cell survival and death pathways [67]. Tocopherols also influence crucial cell signaling pathways such as NF-κB, STAT3, and MAPK, which play vital roles in regulating cell survival, proliferation, and apoptosis [67,70].
4.3. Tocopherols as Adjuvants in Chemotherapy
The potential of tocopherols as adjuvants in chemotherapy has been highlighted in several studies. Prasad et al. [71] demonstrated synergistic effects when γ-tocotrienol was combined with erlotinib in pancreatic cancer cells, while Pereira-Silva et al. [72] showed that α-tocopheryl succinate enhanced the efficacy of gemcitabine in pancreatic cancer models. Moreover, tocopherols have shown promise in reducing the side effects of chemotherapy. Vitamin E supplementation has been found to reduce cisplatin-induced cytotoxicity in cancer patients [73], and tocopherols have mitigated doxorubicin-induced cardiotoxicity in preclinical models [74]. This dual action of enhancing efficacy while reducing toxicity makes tocopherols particularly interesting as chemotherapy adjuvants.
4.4. Drug Delivery and Overcoming Resistance
In the realm of drug delivery, TPGS-based nanocarriers have improved the delivery and efficacy of paclitaxel in various cancer models [75]. D-α-tocopheryl polyethylene glycol 1000 succinate (TPGS) has enhanced the oral bioavailability of several anticancer drugs [60]. Furthermore, tocopherols and their derivatives have shown potential in overcoming drug resistance, a major challenge in cancer treatment. α-TOS has been observed to sensitize resistant leukemia cells to TRAIL-induced apoptosis [76], while TPGS inhibits P-glycoprotein, potentially reversing multidrug resistance in cancer cells [77,78]. These findings underscore the multifaceted potential of tocopherols and their derivatives in cancer treatment, both as standalone agents and as adjuvants to conventional therapies. The ongoing research in this field promises to open new avenues for more effective and less toxic cancer treatments, leveraging the benefits of both natural compounds and their synthetic modifications.
4.5. Future Perspective
The growing body of research on tocopherols and their derivatives in cancer treatment opens up several promising avenues for future investigation. One key area is the continued development of more potent semi-synthetic derivatives with enhanced anticancer properties. These could potentially combine the beneficial effects of natural tocopherols with improved bioavailability, targeting, or efficacy. For instance, recent work by Niculescu et al. [79] has shown promising results with a novel tocopherol-based nanoparticle that demonstrates enhanced tumor-targeting capabilities [79]. Future research could focus on optimizing these derivatives for specific cancer types or combining them with other anticancer agents for synergistic effects.
Additionally, further exploration of the molecular mechanisms underlying the anticancer effects of tocopherols is crucial. This could lead to the identification of new cellular targets or pathways that could be exploited for cancer therapy. Recent studies have begun to unravel the complex interplay between tocopherols and cancer cell metabolism. For example, Zhou et al. [80] discovered a novel interaction between α-tocopherol and the mTOR signaling pathway in breast cancer cells [80]. Expanding on these findings could potentially reveal new therapeutic targets.
Clinical trials represent another critical area for future research. While preclinical studies have shown promising results, more extensive clinical trials are needed to evaluate the efficacy and safety of tocopherols as adjuvants in various cancer types and treatment regimens. These trials could help establish optimal dosing strategies and identify specific cancer types that are most responsive to tocopherol-based interventions. A recent phase II trial by Kunnumakkara et al. [81] investigating the use of γ-tocotrienol in combination with gemcitabine for pancreatic cancer showed encouraging results [81]. Future large-scale, multi-center trials will be crucial in translating these findings into clinical practice.
Investigation of potential synergies between tocopherols and other natural compounds or conventional cancer therapies is another exciting area for future research. Combination therapies that leverage the unique properties of tocopherols could potentially lead to more effective treatment strategies with reduced side effects. For instance, Nesaretnam and Selvaduray [82] demonstrated a synergistic effect between δ-tocotrienol and curcumin in inhibiting prostate cancer growth in vitro and in vivo [82]. Exploring such combinations could open new avenues for cancer treatment.
Furthermore, the role of tocopherols in cancer prevention, particularly in high-risk populations, warrants further investigation. Long-term epidemiological studies and intervention trials are needed to establish the optimal intake of tocopherols for cancer prevention. A recent 10-year follow-up study by Donovan et al. [83] suggested that higher dietary intake of mixed tocopherols was associated with reduced risk of colorectal cancer in a Mediterranean population [83]. Expanding on such studies could help develop evidence-based prevention strategies.
Advancements in drug delivery systems, such as nanoformulations incorporating tocopherols, present opportunities for improving the targeted delivery of anticancer agents. This could enhance the efficacy of treatments while minimizing systemic toxicity. Recent work by Mehata et al. [75] and Gao et al. [84] on TPGS-based nanocarriers for the co-delivery of paclitaxel and siRNA has shown promising results in overcoming multidrug resistance in breast cancer [84]. Further research in this area could revolutionize cancer drug delivery.
As our understanding of the complex interactions between tocopherols and cancer biology deepens, we may uncover new paradigms in cancer therapy that could significantly impact patient outcomes. The multifaceted nature of tocopherols—from their antioxidant properties to their role in cell signaling and their potential as drug delivery vehicles—positions them as versatile tools in the fight against cancer. Future research directions should aim to leverage these diverse properties, potentially leading to novel, more effective, and less toxic approaches to cancer prevention and treatment.
4.1. Semi-Synthetic Tocopherol Derivatives
Recent studies have significantly expanded our understanding of tocopherols and their derivatives in cancer research. A comprehensive review by Baj et al. [9] highlighted several promising semi-synthetic tocopherol derivatives. α-tocopherol succinate (α-TOS) has shown remarkable potential by inducing apoptosis in breast, prostate, and colon cancer cell lines through the disruption of mitochondrial function and generation of reactive oxygen species [9,59]. Another noteworthy derivative is tocopheryl polyethylene glycol succinate (TPGS), which enhances the bioavailability and efficacy of anticancer drugs. TPGS acts as a P-glycoprotein inhibitor, potentially overcoming multidrug resistance in cancer cells [9,60]. The development of redox-silent vitamin E analogs, such as α-tocopheryl ether-linked acetic acid (α-TEA) and α-tocopheroloxybutyric acid (α-TOS), represents another significant advancement. These compounds induce apoptosis independent of their antioxidant properties and have demonstrated effectiveness against breast, prostate, and ovarian cancer cells [9,61]. This research highlights the potential of modified tocopherols to enhance anticancer efficacy beyond the capabilities of their natural counterparts.
Synthetic derivatives of tocopherols have also shown promising results in cancer research. For instance, α-tocopherol succinate (α-TOS), a semi-synthetic derivative of α-tocopherol, has demonstrated potent anticancer effects across various cancer types. α-TOS has been shown to selectively induce apoptosis in malignant cells while sparing normal cells, making it an attractive candidate for cancer therapy [62,63]. Another synthetic analog, α-tocopherol phosphate (α-TP), has exhibited enhanced bioavailability and stability compared to natural tocopherols, potentially leading to improved anticancer efficacy [64]. These synthetic models provide valuable insights into structure-activity relationships and may guide the development of more potent tocopherol-based anticancer agents.
Moreover, tocopherols have shown potential as adjuvants in chemotherapy, enhancing the efficacy of conventional anticancer drugs while potentially reducing their side effects. For example, α-tocopherol has been found to increase the sensitivity of multidrug-resistant cancer cells to chemotherapeutic agents like doxorubicin [65]. Additionally, γ-tocopherol has demonstrated synergistic effects with cisplatin in lung cancer cells, enhancing the drug’s cytotoxicity while protecting normal cells from oxidative damage [66]. These findings suggest that combining tocopherols with standard chemotherapy regimens could improve treatment outcomes and reduce toxicity, opening new avenues for cancer treatment strategies.
4.2. Mechanisms of Action
Jiang et al. [67] elaborated on several mechanisms through which tocopherols exert their anticancer effects. A key mechanism involves the regulation of Peroxisome Proliferator-Activated Receptors (PPARs), particularly PPARγ, leading to decreased cell proliferation and increased apoptosis in various cancer types [67,68]. Additionally, tocopherols, especially γ-tocopherol, inhibit HMG-CoA reductase, which can lead to decreased cholesterol synthesis crucial for rapidly dividing cancer cells [67]. Beyond their antioxidant properties, tocopherols exhibit anticancer effects through various antioxidant-independent mechanisms. These include the inhibition of COX-2 and 5-LOX enzymes, reducing inflammation-related cancer progression [67,69], and the modulation of sphingolipid metabolism, which affects cell survival and death pathways [67]. Tocopherols also influence crucial cell signaling pathways such as NF-κB, STAT3, and MAPK, which play vital roles in regulating cell survival, proliferation, and apoptosis [67,70].
4.3. Tocopherols as Adjuvants in Chemotherapy
The potential of tocopherols as adjuvants in chemotherapy has been highlighted in several studies. Prasad et al. [71] demonstrated synergistic effects when γ-tocotrienol was combined with erlotinib in pancreatic cancer cells, while Pereira-Silva et al. [72] showed that α-tocopheryl succinate enhanced the efficacy of gemcitabine in pancreatic cancer models. Moreover, tocopherols have shown promise in reducing the side effects of chemotherapy. Vitamin E supplementation has been found to reduce cisplatin-induced cytotoxicity in cancer patients [73], and tocopherols have mitigated doxorubicin-induced cardiotoxicity in preclinical models [74]. This dual action of enhancing efficacy while reducing toxicity makes tocopherols particularly interesting as chemotherapy adjuvants.
4.4. Drug Delivery and Overcoming Resistance
In the realm of drug delivery, TPGS-based nanocarriers have improved the delivery and efficacy of paclitaxel in various cancer models [75]. D-α-tocopheryl polyethylene glycol 1000 succinate (TPGS) has enhanced the oral bioavailability of several anticancer drugs [60]. Furthermore, tocopherols and their derivatives have shown potential in overcoming drug resistance, a major challenge in cancer treatment. α-TOS has been observed to sensitize resistant leukemia cells to TRAIL-induced apoptosis [76], while TPGS inhibits P-glycoprotein, potentially reversing multidrug resistance in cancer cells [77,78]. These findings underscore the multifaceted potential of tocopherols and their derivatives in cancer treatment, both as standalone agents and as adjuvants to conventional therapies. The ongoing research in this field promises to open new avenues for more effective and less toxic cancer treatments, leveraging the benefits of both natural compounds and their synthetic modifications.
4.5. Future Perspective
The growing body of research on tocopherols and their derivatives in cancer treatment opens up several promising avenues for future investigation. One key area is the continued development of more potent semi-synthetic derivatives with enhanced anticancer properties. These could potentially combine the beneficial effects of natural tocopherols with improved bioavailability, targeting, or efficacy. For instance, recent work by Niculescu et al. [79] has shown promising results with a novel tocopherol-based nanoparticle that demonstrates enhanced tumor-targeting capabilities [79]. Future research could focus on optimizing these derivatives for specific cancer types or combining them with other anticancer agents for synergistic effects.
Additionally, further exploration of the molecular mechanisms underlying the anticancer effects of tocopherols is crucial. This could lead to the identification of new cellular targets or pathways that could be exploited for cancer therapy. Recent studies have begun to unravel the complex interplay between tocopherols and cancer cell metabolism. For example, Zhou et al. [80] discovered a novel interaction between α-tocopherol and the mTOR signaling pathway in breast cancer cells [80]. Expanding on these findings could potentially reveal new therapeutic targets.
Clinical trials represent another critical area for future research. While preclinical studies have shown promising results, more extensive clinical trials are needed to evaluate the efficacy and safety of tocopherols as adjuvants in various cancer types and treatment regimens. These trials could help establish optimal dosing strategies and identify specific cancer types that are most responsive to tocopherol-based interventions. A recent phase II trial by Kunnumakkara et al. [81] investigating the use of γ-tocotrienol in combination with gemcitabine for pancreatic cancer showed encouraging results [81]. Future large-scale, multi-center trials will be crucial in translating these findings into clinical practice.
Investigation of potential synergies between tocopherols and other natural compounds or conventional cancer therapies is another exciting area for future research. Combination therapies that leverage the unique properties of tocopherols could potentially lead to more effective treatment strategies with reduced side effects. For instance, Nesaretnam and Selvaduray [82] demonstrated a synergistic effect between δ-tocotrienol and curcumin in inhibiting prostate cancer growth in vitro and in vivo [82]. Exploring such combinations could open new avenues for cancer treatment.
Furthermore, the role of tocopherols in cancer prevention, particularly in high-risk populations, warrants further investigation. Long-term epidemiological studies and intervention trials are needed to establish the optimal intake of tocopherols for cancer prevention. A recent 10-year follow-up study by Donovan et al. [83] suggested that higher dietary intake of mixed tocopherols was associated with reduced risk of colorectal cancer in a Mediterranean population [83]. Expanding on such studies could help develop evidence-based prevention strategies.
Advancements in drug delivery systems, such as nanoformulations incorporating tocopherols, present opportunities for improving the targeted delivery of anticancer agents. This could enhance the efficacy of treatments while minimizing systemic toxicity. Recent work by Mehata et al. [75] and Gao et al. [84] on TPGS-based nanocarriers for the co-delivery of paclitaxel and siRNA has shown promising results in overcoming multidrug resistance in breast cancer [84]. Further research in this area could revolutionize cancer drug delivery.
As our understanding of the complex interactions between tocopherols and cancer biology deepens, we may uncover new paradigms in cancer therapy that could significantly impact patient outcomes. The multifaceted nature of tocopherols—from their antioxidant properties to their role in cell signaling and their potential as drug delivery vehicles—positions them as versatile tools in the fight against cancer. Future research directions should aim to leverage these diverse properties, potentially leading to novel, more effective, and less toxic approaches to cancer prevention and treatment.
5. Materials and Methods
This review explores the role of tocopherol-containing herbal medicinal plants as potential anticancer agents. We aim to identify current research trends, highlight promising plant sources of tocopherol for anticancer applications, and suggests a potential mechanisms underlying the anticancer properties of plant-derived tocopherols and semi-synthetic derivatives. To gather relevant literature, we conducted a comprehensive search of the Scopus database (https://scopus.com, accessed on 3 July 2024).
Scopus was chosen for its extensive coverage of peer-reviewed literature in the field of plant sciences and pharmacology. Scopus’s robust content curation process is overseen by the Content Selection and Advisory Board [37,38,39]. Our search strategy employed combinations of keywords including “tocopherol”, “medicinal plants”, “herbal plants”, and “bioactive anticancer compound” for tocopherol-containing plants. To specifically gather information on semi-synthetic tocopherols, we conducted an additional search using related terms, ensuring we included relevant studies on this topic.
We focused on original research articles published between 2016 and 2024, a period that reflects growing interest in the therapeutic potential of tocopherols in cancer research. Studies were selected based on their relevance to tocopherol-containing herbal medicinal plants and semi-synthetic tocopherols related to anticancer properties. We prioritized peer-reviewed articles that presented clear methodologies, sufficient data reporting, and specific results relating tocopherol to anticancer effects. From the collected literature, we extracted and analyzed information on plant species and parts used, biological activities, total tocopherol content, anticancer mechanisms and effects, as well as insights into semi-synthetic tocopherols. This approach allowed us to synthesize current knowledge and identify promising directions for future research in this field.
6. Conclusions
This review analyzed Scopus publications from 2016 to 2024 on tocopherol-containing plants and their anticancer potential. Fifteen herbal medicinal plants were identified as promising candidates, including Bulbine anguistifolia Poelln, Punica granatum L., Moringa oleifera, Kigelia pinnata, and Typhonium flagelliforme Lodd. The review explored tocopherols’ anticancer mechanisms, particularly apoptosis induction and cell cycle arrest. Key factors influencing the anticancer effects of tocopherols were examined, including their forms (α, β, γ, δ), concentrations, plant parts utilized, and their stability in various plants. The emerging research on semi-synthetic tocopherol derivatives highlighted their potential as adjuvants in chemotherapy, their role in enhancing drug delivery, and reducing side effects. This comprehensive analysis provides a foundation for future research in plant-based anticancer therapies, emphasizing the need for further investigation to develop novel, more effective, and less toxic cancer treatments. The findings contribute to a nuanced understanding of how tocopherol-containing plants can be leveraged in the development of future anticancer drugs. Future experimental studied, including in vitro and in vivo models, are essential to validate the anticancer potential of these tocopherol-rich plants and their derivatives. |
"Title: YIPF2 regulates genome integrity | Body: \n Introduction\n The nuclear genome is c(...TRUNCATED) |
"Title: Effect of Coptidis Rhizoma on gastrointestinal system before and after processing with wine (...TRUNCATED) |
"Title: DDR1 Drives Malignant Progression of Gastric Cancer by Suppressing HIF‐1α Ubiquitination (...TRUNCATED) |
"Title: Review of pharmacogenetics of antiseizure medications: focusing on genetic variants of mecha(...TRUNCATED) |
"Title: Biovalorization of whey waste as economic nutriment for mycogenic production of single cell (...TRUNCATED) |
"Title: Menopausal characteristics and hormone replacement therapy in relation to long-term risk of (...TRUNCATED) |
"Title: Pediatric Ovarian Yolk Sac Tumor Mimicking Appendicitis: A Case Report From a Low‐Resource(...TRUNCATED) |
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 121