IC Colloquium: Can we Design Any Biomolecule? From Binders to Assemblies
Par : Hannes Stärk - MIT
IC Faculty candidate
Abstract
The vision is to build machine learning tools capable of designing any biomolecule: binders as starting points for therapeutics, therapeutics with high selectivity and low immunogenicity; assemblies as delivery vehicles for drugs; enzymes engineered to perform endosomal escape or degrade toxins. Realizing this vision demands innovation at the level of core inference algorithms and machine learning methodology alike. As a central example, I will present BoltzGen, a framework for designing protein and peptide binders against target biomolecules of any molecular modality. BoltzGen discovers binders against targets with no existing bound structures, demonstrated across eight diverse design campaigns with functional and affinity readouts spanning 26 targets. The binder modalities range from nanobodies to disulfide-bonded peptides; the targets, from intrinsically disordered proteins to small molecules. Notably, we identified nanomolar-affinity protein and nanobody binders for six of ten novel targets bearing low structural similarity to any protein with a known bound partner, a regime where prior methods falter. Several successful designs are now advancing into mouse models and human cell lines to assess their efficacy against a range of cancers. The talk closes by charting research trajectories toward the broader ambition: the generative design of any biomolecule from binders to assemblies.
Bio
Hannes Stark is a PhD Student at MIT advised by Regina Barzilay and Tommi Jaakkola. His works, including DiffDock, Dirichlet Flow matching, ProtComposer, Boltz-2, and BoltzGen, develop new generative models and inference algorithms to understand and design biomolecules.
More information
IC Faculty candidate
Abstract
The vision is to build machine learning tools capable of designing any biomolecule: binders as starting points for therapeutics, therapeutics with high selectivity and low immunogenicity; assemblies as delivery vehicles for drugs; enzymes engineered to perform endosomal escape or degrade toxins. Realizing this vision demands innovation at the level of core inference algorithms and machine learning methodology alike. As a central example, I will present BoltzGen, a framework for designing protein and peptide binders against target biomolecules of any molecular modality. BoltzGen discovers binders against targets with no existing bound structures, demonstrated across eight diverse design campaigns with functional and affinity readouts spanning 26 targets. The binder modalities range from nanobodies to disulfide-bonded peptides; the targets, from intrinsically disordered proteins to small molecules. Notably, we identified nanomolar-affinity protein and nanobody binders for six of ten novel targets bearing low structural similarity to any protein with a known bound partner, a regime where prior methods falter. Several successful designs are now advancing into mouse models and human cell lines to assess their efficacy against a range of cancers. The talk closes by charting research trajectories toward the broader ambition: the generative design of any biomolecule from binders to assemblies.
Bio
Hannes Stark is a PhD Student at MIT advised by Regina Barzilay and Tommi Jaakkola. His works, including DiffDock, Dirichlet Flow matching, ProtComposer, Boltz-2, and BoltzGen, develop new generative models and inference algorithms to understand and design biomolecules.
More information
Practical information
- General public
- Free
Contact
- Host: Charlotte Bunne