School of Biomedical Sciences cordially invites you to join the following seminar:

Speaker: Dr. Gregor Popowicz, Group Leader of Institute for Structural Biology, Helmholtz Zentrum München, CSO & Co-founder of Khumbu.AI GmbH, Munich
Talk Title: Reading your drug target wishes - a structure-aware AI model guiding drug discovery

Date: 2 September 2025 (Tuesday)
Time: 4:00 pm – 5:00 pm
Venue: Lecture Theatre 1, G/F, William M.W. Mong Block, 21 Sassoon Road 
Host: Professor Clive Chung

Biography

I have dedicated my career to exploring the intersection of physics and the health sciences. After completing a Master's degree in Medical Physics, I joined the Max Planck Institute for Biochemistry to pursue a PhD in structural biology under the supervision of Nobel Laureate Professor Robert Huber. My doctoral research rapidly transitioned from fundamental structural biology to its applied potential in drug discovery. Initially, I contributed to several oncology-focused drug discovery projects, but I soon redirected my efforts toward rare and neglected diseases. Since 2012, I have led a major anti-parasitic drug discovery initiative targeting Chagas disease, in collaboration with the Drugs for Neglected Diseases initiative (DNDi). With the rise of modern machine learning methods, I became increasingly involved in developing AI tools capable of understanding and analyzing structural biology data. This work led me to join a startup company to further advance and translate our AI-driven technologies. My focus remains on deeply integrative, structure-based drug discovery, with a particular emphasis on rare and neglected diseases. I am strongly convinced that the convergence of experimental and computational approaches holds the key to unlocking long-needed acceleration in the drug discovery process.

Abstract

Applying machine learning (ML) in structural biology has already transformed the field. Not only because of improved accuracy but also the much better accessibility of structural models to life scientists. Meanwhile, ML models gain accuracy in predicting not only single-domain structures but also multimeric complexes of diverse molecules. Moreover, reverse folding models are able to generate realistic structures with pre-programmed properties without biological templates. Yet, the progress of structure-based drug discovery seems to lag behind the overall wave of ML developments. This can be attributed to much noise in the training data and poor coverage of nearly infinite chemical combinatorial space by known experimental structures.

We present a set of ML models designed to circumvent these limitations. The approach, called Target Preference Mapping, treats each interface between a biomolecule and a drug as a large set of small microenvironments. They are then used to predict biomolecule preference for specific ligand chemistry. Chemical connectivity and binding energy are not used for the training to avoid overtraining of the model on the known complexes. We show that this approach allows us to predict drug optimization and perform virtual screening at unprecedented speed and accuracy. A similar approach was recently tested for Protein-protein interactions, we show some preliminary observations on these interactions as well.

ALL ARE WELCOME.