abyssal-pytorch

Implementation of Abyssal, a deep neural network trained with a new "mega" dataset to predict the impact of an amino acid variant on protein stability.

Installation

$ pip install abyssal-pytorch

Usage

import torch
from abyssal_pytorch import Abyssal

model = Abyssal()

wt_emb = torch.randn(1, 1280)    # ESM2 embedding for the wildtype amino acid.
mut_emb = torch.randn(1, 1280)   # ESM2 embedding for the mutated amino acid.

out = model(wt_emb, mut_emb)  # (1, 1) predicted ddG.

Training from scratch

Please refer to note/data-preprocessing.ipynb notebook for data preprocessing. Running this notebook will generate mega.train.csv, mega.val.csv and mega.test.csv containing metadata for model training and evaluation.

For faster training, extract ESM2 embeddings and save them to the disk. It takes ~5 hours.

$ python prefetch_embeddings.py -i data/mega.train.csv -o data/embeddings
$ python prefetch_embeddings.py -i data/mega.val.csv -o data/embeddings
$ python prefetch_embeddings.py -i data/mega.test.csv -o data/embeddings

This will save embedding vectors in .pt format in data/embeddings directory.

You can now train model.

$ python -m abyssal_pytorch.train \
    --train data/mega.train.csv \
    --val data/mega.val.csv \
    --test data/mega.test.csv \
    --emb-dir data/embeddings

Reproduction status

Unfortunately, I could not exactly reproduce the results based on the model and training specification in the current version of Abyssal preprint. Below is the results from my best effort so far. Any ideas to reproduce the model performance would be appreciated!

Metric	Target	Reproduced
Pearson's r	0.85+-0.00	0.7661
Spearman's r	0.81+-0.01	0.7846
MSE, kcal/mol	0.89	0.6003
Accuracy (?)	0.79	?
PCC(f-r)	-0.98	-0.9611
delta	-0.01	-0.0330

Citations