Transfer learning to annotate the protein universe

This is the source code for experiments in “Transfer learning: the key to functionally annotate the protein universe,” by L. A. Bugnon, E. Fenoy, A. Edera, J. Raad, G. Stegmayer and D.H. Milone. Patterns, Cell Press, 2023. https://doi.org/10.1016/j.patter.2023.100691

Installation

First, download the repository or clone it with

git clone https://github.com/sinc-lab/transfer-learning-pfam.git

We recommend using a python virtual environment such as conda or venv. For example, with conda:

conda create -n tlpfam python
conda activate tlpfam

Now on this environment it is better to install pytorch first (it may depend on your hardware setup, follow these recommendations)

pip install torch

Move to the repository folder and install the required packages

pip install -r requirements.txt

Run a quick verification demo

Since embedding sequences takes some time, precomputed embeddings are provided. You can download them for dev and test partitions from here. Copy and unzip all the files on root folder (transfer-learning-pfam/). This contains the sequences and labels as well.

To check if everything is ok, a fast converging model can be trained using only the dev partition. Go to transfer-learning-pfam/ directory and run

python train.py --train data/Clustered_data/dev/ --dev data/Clustered_data/dev/

To evaluate all models in the folder, run

python test.py

Reproducing results

To reproduce the results of the paper, compute the embeddings for train sequences with the following

python compute_embeddings.py -i data/Clustered_data/train/

Then, train a TL-ProtCNN-ensemble of 5 models with the full dataset

python train.py --train data/Clustered_data/train/ --dev data/Clustered_data/dev/ -n 5

For reference, one model takes about a day to converge using a RTX A5000 GPU.

Using the test script you can see the results for each model individually, and the ensemble with all the trained models found in the folders named as results_*

python test.py