Project in progress for master's thesis at University of Chemistry and Technology.
Problem and data come from TDC.
The goal of the project is to train a model capable of predicting an antibody's developability index based on its protein sequence.
The data directory contains files necessary to run training. The chen directory contains the SAbDab / Chen et al. dataset along with all its data representations and train-test splits. The tap directory contains the TAP dataset and its representations.
Evaluation metrics and hyperparameters of trained models are stored in the evaluations directory. Actual trained models are not stored here. However, the best LSTM models are stored in the models directory. Fine-tuned ProteinBERT and Sapiens models are not present in this repository.
The first set of jupyter notebooks (in notebooks/data_processing) serve to download both datasets and create all respective data representations. Seven different data representations are created. These are then saved to the data directory.
Notebook 04 also produces the train-test splits based on clustering.
Six different ML models (k-NN, logistic regression, random forest, gradient boosting, SVM and multilayer perceptron) were trained for the classification task. Additionally, LSTM networks and two pre-trained transformer networks (ProteinBERT and Sapiens) were trained.
The first four notebooks and notebook 06 in the directory notebooks/experiments provide functionality for training ML models. In practice, these were trained using scripts in the bin directory, in particular 06_high_level_training.py. test_on_tap.py was then used for testing these models.
LSTM models were trained using notebook 05 in notebooks/experiments and ProteinBERT was fine-tuned using notebook 09.
All data and scripts for fine-tuning the Sapiens model are in fairseq directory. The notebook sapiens_fine_tune.ipynb also contains fine-tuning functionality, but serves mainly for evaluation.
Data exploration and evaluation notebooks are in notebooks/reports along with various plots.