Research

Multi-level Studies of Protein-Protein Interactions

Systems Biology

In biology, proteins typically function in coordinated teams through various pathways. Predicting protein-protein interactions (PPIs) on a genome-wide scale is crucial for understanding cellular functions, signaling pathways, and the mechanisms of diseases. Recent advances in deep learning and AI have shown remarkable potential in predicting PPI structures with accuracy. These approaches harness vast datasets and sophisticated algorithms to infer interactions, offering valuable insights into the molecular basis of biological processes. However, a major challenge lies in managing the immense computational demands of genome-wide predictions. Achieving a balance between computational efficiency and predictive accuracy—particularly for dynamic and modified protein regions—remains a critical area for future developments.

Building on our previous works, PrePPI-AF and ZEPPI, we are leveraging state-of-the-art deep learning models, complemented with traditional sequence analyses, to re-score the 30 billion PPI interface models generated by PrePPI-AF. Our goal is to construct a comprehensive structural database of the human interactome, the first of its kind. Using this methodology, we are particularly focused on predicting the Host-Pathogen Interactome (HPI). We hope to unravel the molecular basis that underpin infectious diseases, enabling the development of targeted therapies, vaccines, and strategies to combat existing and emerging pathogens.

Protein Dynamics and Protein Design

Moreover, with millions of predicted PPIs, we aim to explore the complex networks these proteins form, with the goal of identifying novel pathways and gaining a more comprehensive understanding of protein functions. Once specific target PPIs are identified, we are further interested in studying their binding dynamics and energetics, with the goal of designing optimized binders and small molecule inhibitors.

Currently, we are seeking 1 postdoc and 1-2 graduate students to join us in this exciting direction. We are always enthusiastic to collaborate with experimental groups.

Predicting Protein Binding Affinity and Their Changes

Calculating protein binding free energy in silico remains a challenging task. Molecular dynamics-based methods often struggle with the complexity introduced by protein conformational flexibility, in addition to the computation demands required for accurate simulations. Recently, machine learning approaches have emerged as promising alternatives, either by fitting structural data to simple regression models or by developing deep learning models based solely on sequence information.

Our goal is to create a general and widely applicable machine learning model to predict binding affinities. Specifically, we are building a transferable, attention-based deep learning model that incorporates both structural features of protein complexes and sequence-derived data. We aim to extend this model to predict changes in binding free energy due to mutations, interactions with small molecules, or post-translational modifications, providing a versatile tool for a range of biological applications.

We are currently seeking 1 postdoc and 1-2 graduate students for this exciting direction.