Clarification about FAISS index content: SwissProt vs UniRef50

[zoeyzyq]

Hi and thanks a lot for releasing ProTrek and the pretrained FAISS indices!

I am currently using the following FAISS index for retrieval:

faiss_index/SwissProt/ProTrek_650M_UniRef50/sequence/sequence.index

In the README, it is mentioned that sequence embeddings for SwissProt are stored in this FAISS index. However, the directory name contains UniRef50, which makes me a bit confused about what the index actually contains.

Could you please help clarify the following points?

1. Does sequence.index contain embeddings for SwissProt entries only, or for UniRef50 (cluster representatives), or some combination of both?
2. For the corresponding ids.tsv in the same folder:
   • Do these IDs correspond to SwissProt accessions, UniRef50 cluster IDs, or something else?
   • Is there any additional mapping file I should use to interpret the IDs (e.g., mapping UniRef50 clusters back to SwissProt)?

Thanks for your help!

[LTEnjoy]

Hi,

The UniRef50 marker here only represents the version of our ProTrek model, i.e. it is pre-trained on the combination dataset of UniRef50 and Swiss-Prot. So the sequence.index file only contains embeddings of all SwissProt proteins.

Likewise, the ids.tsv only contains corresponding sequences and IDs of SwissProt entries. The ith line in ids.tsv corresponds to the ith embedding in sequence.index file.

[zoeyzyq]

Thanks for your response, it clears up my confusion :)

I was wondering if you have any precomputed embeddings available for UniRef90 or UniRef100 as well, or if there are any plans to release them in the future?

[LTEnjoy]

We didn't precompute embeddings for UniRef90 and UniRef100 and probably will not release them due to the computational cost. Thank you for your understanding :)