Project part of the 2021 Copenhagen Protein Biohackathon.
We participated in the challenge Predicting multi mutant miniprotein stability
Our team name is MARS (Miniprotein stAbility fRom Sequence) 🌕
We built a convolutional Variational Autoencoder architecture using PyTorch to tackle the task of predicting miniprotein stability from sequence using the dataset by Rocklin et.al. Science 2017.
The idea is to learn a lower dimensional embedding of the miniprotein sequence space in the latent dimension from which the original sequence can reconstructed or from which new sequences can be sampled. Additionally, a prediction task is used to predict the stability score of the miniproteins from the learned embedding in the latent space.
This architecture has the advantage that one can sample sequences around a sequence of interest for which one knows or has predicted that it has high stability.
A sketch of the architecture is as follows:
The results for the single and multi mutant dataset are as follows:
Each sequence is encoded using one hot encoding with 3 extra entries to denote secondary structure
- Single mutants Rp 0.72 Spearman 0.73 p<0.00001
- Multi mutants Rp 0.47 Spearman 0.35 p<0.00001
Single mutants Rp 0.80 Spearman 0.81 p<0.00001
Interactive image
Multi mutants Rp 0.53 Spearman 0.39 p<0.00001
Interactive image
Create conda environment
conda env create -f environment.yml
activate environment
conda activate biohackathon
Install bioembeddings
pip3 install -U pip > /dev/null
pip3 install -U "bio-embeddings[all] @ git+https://github.com/sacdallago/bio_embeddings.git" > /dev/null
Run model
python embeddings_protbert_single.py
Examples were run using PyTorch 1.8.1, CUDA 10.1 on a machine equipped with 2x GTX 2070, 377 GB RAM and 2x Intel Xeon Gold 5120 CPU @ 2.20GHz