ECE PhD Thesis Defense: Hao Yu

ECE PhD Thesis Defense: Hao Yu

Title: AI and Physics-Based Methods for Molecular Recognition in Proteins

Presenter: Hao Yu

Advisor: Professor Sandor Vajda

Chair: Professor Enrico BEllotti

Committee: Professor Sandor Vajda, Professor Ioannis Paschalidis, Professor David Castañón, Professor Kayhan Batmanghelich

Google Scholar Link: https://scholar.google.com/citations?user=pDuTVbEAAAAJ&hl=en

Abstract: Understanding how proteins recognize and bind other molecules is central to both basic biology and drug discovery. Despite advances brought by crystallography and computational docking, predicting molecular interactions at atomic resolution across diverse molecular systems remains challenging. The recent emergence of deep learning-based protein structure prediction - culminating in AlphaFold2 and AlphaFold3 - has dramatically expanded what is computationally possible. However, important limitations remain. Antibody-antigen docking succeeds for only a fraction of targets, epitope prediction methods lack rigorous evaluation benchmarks, and deep learning-based co-folding models exhibit systematic biases inherited from their training data.

This thesis advances the understanding of molecular recognition through three studies that combine AI-driven prediction with physics-based modeling. First, we show that combining physics-based docking by ClusPro with AlphaFold-Multimer refinement achieves a 50% success rate in top-10 predictions on a benchmark of 52 complexes. We find that the improvement arises primarily from increased structural diversity among candidate complexes, which enables more successful sampling of near-native binding modes. Second, we benchmark six epitope prediction methods across six input conditions that reflect real-world constraints on antibody availability. We demonstrate that antibody-agnostic predictors face a ceiling that cannot be overcome through algorithmic improvements alone, and that AlphaFold3 combined with the ClusPro AbEMap scoring framework advances the state of the art for antibody-specific epitope prediction. Third, we find that AlphaFold3’s predictions of ligand-induced conformational change are determined primarily by the apo-to-holo distribution of related structures in its training data rather than by the underlying physical chemistry of the interaction. Furthermore, AlphaFold3 predictions of non-binding ligand-protein interactions produce distributions indistinguishable from those of binders.

Together, these results demonstrate that advances in molecular recognition require the integration of data-driven learning with physics-based reasoning. They also highlight specific limitations of current co-folding models and suggest directions for developing more physically grounded architectures.

When 1:00 pm - 3:00 pm on 23 June 2026
Building PHO 339