PhD defence: A Hitchhiker’s Journey through Machine - Learning for Structural Biology of Antibodies

Protein–protein interactions (PPIs) are fundamental to biological processes, and understanding their three-dimensional structures is essential for uncovering cellular mechanisms and advancing applications such as drug discovery and protein engineering. This thesis investigates the integration of machine learning into computational structural biology, focusing on antibody–antigen interactions, which remain challenging due to the high variability of antibody binding regions and weak evolutionary signals.

The thesis first reviews recent advances in AI-driven structural biology and outlines the key challenges in antibody modelling. It then presents DeepRank-GNN-esm, a graph neural network that replaces computationally expensive evolutionary profiles with ESM-2 protein language model embeddings, improving docking model scoring while broadening applicability to systems with limited evolutionary information.

To address the difficulty of predicting the flexible CDR-H3 loop, a workflow combining AlphaFlow-generated antibody ensembles with HADDOCK3 docking is introduced, demonstrating that increased conformational diversity significantly improves docking success. The thesis also explores sequence-specific fine-tuning of protein language models, showing improved performance in antibody structure prediction, mutation effect prediction, and binding affinity estimation with substantially lower computational cost than training antibody-specific models.

Finally, DeepRank-Ab, a geometric deep learning scoring function tailored to antibody–antigen interfaces, is presented. Trained on over 2.3 million docking models, it consistently outperforms existing methods, including AlphaFold3 and HADDOCK, in identifying near-native complexes. Together, these contributions demonstrate how machine learning can enhance antibody modelling and support the computational design of therapeutic antibodies.