Mutational Effect Transfer Learning (METL) is a framework for pretraining and finetuning biophysics-informed protein language models.

This repository contains a wrapper meant to facilitate the ease of use of METL models. Usage of this wrapper will be provided below. Models are hosted on Zenodo and will be downloaded by this wrapper when used.

METL is discussed in the paper in further detail. The GitHub repo contains more documentation and includes scripts for training and predicting with METL. Google Colab notebooks for finetuning and predicting on publicly available METL models are available as well here.

Use the code below to get started with the model.

Running METL requires the following packages:

transformers==4.42.4
numpy>=1.23.2
networkx>=2.6.3
scipy>=1.9.1
biopandas>=0.2.7

In order to run the example, a PDB file for the GB1 protein structure must be installed. It is provided here and in raw format here.

After installing those packages and downloading the above file, you may run METL with the following code example (assuming the downloaded file is in the same place as the script):

Biophysics-based protein language models for protein engineering.
Sam Gelman, Bryce Johnson, Chase R Freschlin, Arnav Sharma, Sameer D'Costa, John Peters, Anthony Gitter⁺, Philip A Romero⁺.
Nature Methods 22, 2025.
⁺ denotes equal contribution.

For questions and comments about METL, the best way to reach out is through opening a GitHub issue in the METL repository.