CeNT Seminar (Wed, March 11, 2026): Beyond AlphaFold3: From Physical Chemistry Principles to Machine Learned Physics in Molecular Design

09 03 2026

The Centre of New Technologies, University of Warsaw invites to a seminar by:

prof. Chia-en A Chang

Department of Chemistry, University of California, Riverside

Title: Beyond AlphaFold3: From Physical Chemistry Principles to Machine Learned Physics in Molecular Design
Date: Wed Mar 11th, 2026
Time: 11:00 (Central European Time)
Host: prof. Joanna Trylska
The seminar will be held in the 00.142 auditorium, Banacha 2c

Abstract:

The presentation will discuss our recent advances in protein variants and cyclic-peptide design in the AI era. While powerful tools such as AlphaFold3 and ProteinMPNN are widely used, what can we do when these data-driven AI methods based on statistical correlations fail? Here we will introduce 1) using residue correlation networks to design high-affinity protein-based binders; 2) teaching AI to learn physics to understand molecular motions and assist cyclic-peptide design. Using all-atom molecular dynamics (MD) simulations, we identify the essential network of residue interactions and dihedral angle correlations critical in protein–protein recognition and guide residue substitution to enhance protein-protein binding. Using ubiquitin (Ub), a central player in multiple cellular functions, and MERS coronaviral papain-like protease (PLpro), an antiviral drug target, our designed ubiquitin variant (UbV) hosting 3 mutated residues achieved a ~3,500-fold increase in functional inhibition relative to wild-type Ub, demonstrating the effectiveness of this physical-chemistry-based approach to design high-affinity protein binders for cell biology research and future therapeutics [1]. We will then introduce our Internal Coordinate Net (ICoN) [2], an autoencoder-based deep learning model that utilizes atomistic bond-angle-torsion coordinates as features and force-field based lost function to learn the physical principles of conformational changes from MD simulation data and to sample novel conformations in the latent space. The talk will highlight how coordinated torsion rotations drive conformational transition pathways in macro-cyclic peptides — systems with highly concerted atomic motions and complex energy landscapes that challenge traditional conformational sampling and design strategies. We will show how ICoN improves accuracy, interpretability, and computational efficiency of deep learning models to bring scientific insights and assist macrocycle design.

References:

[1] Hung, T. I., Hsieh, Y-J., Lu, W-L., Wu, K-P., and Chang, C-E. A., What Strengthens Protein-Protein Interactions: Analysis and Applications of Residue Correlation Networks. J. of Molecular Biology , 435, 168337, 2023.

[2] Ruzmetov T, Hung T. I., Jonnalagedda S. P., Chen S. H., Fasihianifard P., Guo Z., Bhanu B., Chang C-E. A., Sampling Conformational Ensembles of Highly Dynamic Proteins via Generative Deep Learning. Journal of Chemical Information and Modeling, 65(5), 2487–2502, 2025