Joseph Szymborski

This paper identifies data leakage in pre-trained pLLMs used for protein-protein interaction inference and shows their limitations in generalization.

CLASP integrates protein sequence, structure, and text-based descriptions using geometric deep learning and contrastive learning to create embeddings that improve zero-shot classification and clustering.

INTREPPPID improves cross-species protein-protein interaction prediction using orthologue-informed neural networks and outperforms existing methods on evaluation datasets.

RAPPPID improves protein-protein interaction prediction by using regularised AWD-LSTM twin networks, outperforming existing methods on unseen proteins.