ProHiFlo generates proteins 4x faster with 58.9% success rate
New AI generates custom proteins from scratch, outperforming RFDiffusion by 17%.
ProHiFlo, developed by Chuanzhen Wang, Meade Cleti, and Pete Jano, is a new hierarchical flow matching framework for generating novel proteins from scratch. Unlike prior diffusion or flow matching methods that operate at a single resolution and lack functional constraints, ProHiFlo uses three key innovations: coarse-to-fine generation that first models backbone geometry then refines to all-atom coordinates, reducing computational costs; functional guidance that leverages pretrained predictors to steer generation toward desired properties without retraining; and an adaptive SE(3)-equivariant architecture for efficient multi-scale processing.
In experiments, ProHiFlo achieves state-of-the-art performance on unconditional generation, motif scaffolding, and functional design. Notably, on enzyme active site scaffolding, it attains a 58.9% success rate compared to 41.2% for RFDiffusion, while requiring 4 fewer sampling steps. This breakthrough could accelerate therapeutic design, enzyme engineering, and synthetic biology by enabling targeted protein creation with higher efficiency and accuracy.
- Coarse-to-fine generation reduces computational cost while maintaining all-atom accuracy.
- Functional guidance uses pretrained predictors to steer protein properties without retraining.
- 58.9% success on enzyme scaffolding vs 41.2% for RFDiffusion, with 4 fewer sampling steps.
Why It Matters
ProHiFlo could accelerate drug discovery and enzyme engineering by generating custom proteins faster and more accurately.