README.md

metadata

license: mit
task_categories:
  - feature-extraction
tags:
  - embeddings
size_categories:
  - 100K<n<1M

Want to analyze some proteins, but lack embeddings? Want to perform vector similarity search? Want a context of known proteins embeddings? Look no further!

This repository is a dataset of Reviewed Swiss-Prot Proteins. Each protein I compute the embeddings for ESM2 (6 layer model) and ProteinCLIP.

Specs

See the data viewer for all information. Most of the metadata on each protein and the sequences themself come from Reviewed Swiss-Prot Proteins.

Important columns

• accession: the Uniprot Accession Number that identifies each protein uniquely
• embedding: 128 dimensional vectors computed from the ProteinCLIP models (first from ESM 6 layer model last layer).

Examples

Yeah this dataset is easy to use! See some quick examples!

Example 1

Upload all the embeddings to Nomic Atlas to create https://atlas.nomic.ai/data/donnybertucci/swissprot-proteinclip/map

Example 2

Similarity search with cosine similarity https://github.com/xnought/DS569k-viewer (live site https://ocular.cc.gatech.edu/DS569k/)

Credit

All credit for the data goes to https://www.uniprot.org/ and https://www.expasy.org/resources/uniprotkb-swiss-prot and to the original authors of each protein. I directly took the data from them.

Large pieces of code were copied from https://github.com/wukevin/proteinclip to embed both ESM and ProteinCLIP. Without their pretrained models and code, I could not have produced the embeddings.

And credit to Fair ESM for the pretrained ESM2 models https://github.com/facebookresearch/esm.