Promera: a unified model for biomolecular structure prediction, filtering, and design

Area of Science:

• Computational Biology
• Structural Biology
• Protein Engineering

Background:

• Generative models are crucial for biomolecular structure modeling and design.
• Current models struggle with accurate filtering of designed binders and lack controllable design features.

Purpose of the Study:

• Introduce Promera, a unified generative model for all-atom structure prediction, enhanced filtering, and controllable biomolecular design.
• Evaluate Promera's performance in binder filtering and compare its co-folding capabilities against existing models.

Main Methods:

• Developed Promera, integrating structure prediction with confidence metrics for binder/non-binder discrimination.
• Employed Promera for controllable binder generation by predicting masked protein sequences with specified constraints (epitope, paratope, template).
• Assessed Promera's co-folding performance and binder design success rates using in silico evaluations.

Main Results:

• Promera's confidence metrics show superior accuracy in filtering binders for miniproteins and nanobodies.
• Promera's co-folding performance exceeds popular open-source models for therapeutically relevant protein categories.
• Designed nanobodies achieve in silico success rates comparable to backpropagation-based methods when filtered by Promera's co-folding confidence.

Conclusions:

• Promera offers a unified approach to biomolecular structure prediction and design with improved filtering and control.
• Demonstrated Promera's versatility through in silico applications like epitope targeting and GPCR stabilization.
• Proposed a scaling law for co-folding models to guide future performance enhancements in generative biomolecular modeling.