Request for info
Hi! I was looking for a dataset of prtein AA sequences and their foldseek 3di sequences. This looks a lot like that, but could you tll me a little bit about hwo it came to be (what filters/seelctions/protein dataset etc). Just making sure it is what I think it is before trainign a model on it...
Hi @MichelNivard ,
Just updated the readme with more information. This dataset is not mine, it is from the ProstT5 project. Their paper details the data compilation and processing:
ProstT5 data set
Our translation from 1D amino acid sequences to 1D structure sequences (3Di tokens) began with a recently published (36) clustered version of the AlphaFold Protein Structure Database (AFDB (33)). This dataset was created by two-step clustering and one step of quality filtering.
MMseqs2 (35) clustered 214 million (M) UniprotKB (34) protein sequences from AFDB such that no pair had over 50% pairwise sequence identity (PIDE) at 90% sequence overlap. For each of the 52M resulting clusters, the protein with the highest predicted local distance difference test (pLDDT) score (32) was selected as the representative.
Foldseek (1) clustered the 52M representatives further into 18.8M clusters enforcing a pairwise minimal E-value of 10−2 at 90% sequence overlap for each Foldseek (structural) alignment. From those 18.8M, 2.8M clusters contained two or more members (16M were singletons, i.e. no other protein could be aligned using the procedure above). To avoid bias towards exotic outliers and to increase the training set, we expanded each cluster, maximally, by its 20 most diverse members using HHBlits (37). This expansion increased from 2.8M clusters to 18.6M proteins leading to a total set size of 34.6M proteins when combined with the singletons.
Finally, we added three filtering steps: remove (a) low-quality structure predictions (pLDDT < 70), (b) short proteins (length < 30) and (c) proteins with highly repetitive 3Di-strings (>95% of assigned to single 3Di token). The final training set contained 17M proteins (4.3M singletons with respect to the original 16M). As we are translating into both directions, i.e. from 3Di to amino acids (AA) and vice versa, this corresponded to 34M training samples. From those, we randomly split off 1.2k proteins for validation and 1.2k for final testing while ensuring that clusters were moved to either of the sets such that all members of one cluster always end up in the same split. After keeping only representatives to avoid bias towards clusters, we ended up with 474 proteins for validation and final testing each.
Hope this is helpful.
Best,
Logan