New machine-learning approach could speed precision drug development

Michael Bronstein, senior author of the study

© Natascha Unkart

A research team in Austria has developed a geometric deep learning method that could significantly accelerate the creation of precision therapies.

The approach, detailed this week in the journal Nature, uses a platform called Molecular Surface Interaction Fingerprinting (MaSIF) to design custom proteins that bind to drug-bound target proteins.

“Proteins are the foundation of life as we know it, and their functionality depends largely on how they interact with other molecules,” said Michael Bronstein, lead on the study and scientific director at AITHYRA, a newly established institute of the Austrian Academy of Sciences, in a news release. “One of the surprising and satisfying outcomes of our study is that a neural network trained on interactions between naturally occurring proteins generalizes very well to protein-ligand complexes never seen before.”

MaSIF works by recognizing the surface shapes of proteins and matching them with compatible “neo-surfaces” formed when small molecules or drugs bind to proteins. The new system enables researchers to engineer custom protein binders that activate only when a specific drug is present, allowing for a level of control that could lead to safer and more precise immunotherapies.

Bruno Correia, whose Laboratory for Immunoengineering and Protein Design at École Polytechnique Fédérale de Lausanne collaborated on the study, said the method “opens a new avenue to precise dosing and control of biological drugs such as those used in oncological immunotherapies.”

In proof-of-concept experiments, the team tested its designed protein binders against three different drug-protein complexes: the hormone progesterone, the FDA-approved leukemia medication Venetoclax and the antibiotic Actinonin. According to the researchers, the binders recognized each target with high affinity, underscoring the potential for MaSIF-driven protein design.

The full study, titled "Targeting protein–ligand neosurfaces with a generalizable deep learning tool," is available in Nature.