Protein function prediction for CAFA 6

A deep learning solution for CAFA 6, a Kaggle competition on predicting protein function. Given a protein sequence, the model predicts the Gene Ontology terms that annotate it.

Each protein is first turned into an embedding by the ESM-2 protein language model, in its three-billion-parameter version. A multi-layer perceptron is then trained on those embeddings as a multi-label classifier over Gene Ontology terms, with class weights to offset the imbalance between common and rare terms.

Training and validation sets are split by sequence clustering with MMseqs2, and the decision threshold for each term is tuned separately.

Built with: Python, PyTorch, Transformers (ESM-2 embeddings), MMseqs2.

Code: github.com/Enhso/cafa