"By removing structural dependence, we unlock the full potential of AI for drug discovery"

Gero, a Singapore-based biotech company focused on aging and chronic diseases, launched ProtoBind-Diff, an AI model that generates small molecules for specific protein targets—using only the amino acid sequence, without needing 3D structural data.

- Ankit Kakhar

This is a major shift from traditional drug discovery tools, which rely heavily on high-quality structural information that often doesn’t exist for many disease-relevant proteins. It can generate drug-like compounds for targets that have been historically difficult or impossible to hit by opening up new possibilities for oncology, infectious disease, immunology, and aging-related conditions. In an interaction with BioSpectrum Asia, Dr Peter Fedichev, CEO, Gero, Singapore shared his views on how ProtoBind-Diff performs, tackles challenges and play a role in fast-tracking treatments during future pandemics. Edited excerpts;

How does ProtoBind-Diff tackle the challenge of targeting proteins without 3D structural data?

ProtoBind-Diff was designed from the ground up to overcome a fundamental bottleneck in drug discovery: the limited availability of high-quality 3D structural data for protein—ligand complexes. While structure-based approaches like docking or AlphaFold-guided generative models rely on resolved protein structures or predicted pockets, these are not available or reliable for a significant fraction of biologically relevant targets—especially novel, disordered, or poorly characterized proteins. This scarcity of structural data restricts the druggable target space, especially in challenging therapeutic areas such as cancer, aging, and neurodegeneration. ProtoBind-Diff sidesteps this dependency entirely by using only the linear amino acid sequence of a protein as its input. The model is a masked diffusion language model that learns the joint distribution between protein sequences and chemically valid small molecules that bind to them. This allows it to operate effectively in sequence space, bypassing the need for any structural input.

What makes ProtoBind-Diff perform better than models like Pocket2Mol on harder targets?