EPFL Machine Learning Course, Autumn 2022 - Class Project 1

Team members: Lubor Budaj, Matthew Dupraz, Anton Hosgood

Report

This repository contains the code produced during this project. The aim of the project is to create a model that can classify a protein as either soluble or insoluble based only on its FASTA sequence.

Simply run the script src/run.py to run a given model on the dataset. Hyperparameters and models can be configured in src/config.py.

• data/ - datasets containing FASTA sequences and labels denoting solubility

• src/ - source code used in our pipeline

The rest of the root directory contains notebooks holding analyses and other work carried out.

• src/config.py - (default) configuration of the models and training
• src/data.py - methods for loading and encoding data
• src/models.py - defines general architectures of the models used
• src/scores.py - methods for evaluating model performance
• src/train.py - helper methods for training
• src/run.py - main script for training

A good part of our work is found in the numerous Hupyter notebooks in the root directory:

• Data_Expl.ipynb - contains the initial phases of our exploratory data analysis, as well as our attempts at regression analysis, which ended up giving us better results than the deep learning models
• Model#_Eval.ipynb - evaluating performance of model # with given set of parameters by taking the average over several runs
• Embed_Visualisation.ipynb - visualising the embedding of residues into 2D space that is obtained as a result of training model 3
• CNN_Visualisation.ipynb - visualising the output of the first layer of the CNN on random test sequences
• Regr_Analysis.ipynb - contains an attempt at regression analysis by applying PCA to the one-hot representation of the FASTA sequence

We use Python 3.9.12 and PyTorch to build our deep learning models. Several other libraries are used including NumPy, pandas, scikit-learn.

Matplotlib and seaborn are used for visualisation purposes.