README.md

---
language:
- en
tags:
- biology
- ESM
- language-model
- protein
- VEP
pretty_name: VESM scores
size_categories:
- 100M<n<1B
---

# Proteome-wide VESM variant effect scores 

This repository provides precomputed **proteome-wide (UniProt, hg19, and hg38) variant-effect prediction scores** using the latest VESM models developed in the paper ["Compressing the collective knowledge of ESM into a single protein language model"](vesm_arxiv) by Tuan Dinh, Seon-Kyeong Jang, Noah Zaitlen and Vasilis Ntranos.
- **Models:** VESM, VESM1, VESM2, VESM3, VESM++ (available at https://huggingface.co/ntranoslab/vesm).

```VESM1, VESM2 and VESM3 are individual protein language models based on ESM1b, ESM2-650m and ESM3. VESM is the average of VESM1 and VESM2 (sequence-only) and VESM++ is the ensemble of all three. ```

Please see the corresponding GitHub repo (https://github.com/ntranoslab/vesm) for more details.

## License  <a name="license"></a>
The predictions of VESM, VESM1, and VESM2 are distributed under the MIT License. 
The VESM3 and VESM++ models are built with ESM3-Open (EvolutionaryScale), which is available under a [non-commercial license agreement](https://www.evolutionaryscale.ai/policies/cambrian-open-license-agreement).