"Bridge-IF: Learning Inverse Protein Folding with Markov Bridges"

A novel diffusion model that understands the language of protein structures

Bridge-IF, proposed in this paper, connects protein structures to sequences using Markov bridges for better protein design

https://arxiv.org/abs/2411.02120

🎯 Original Problem:

Inverse protein folding aims to design protein sequences that fold into desired backbone structures. Current methods use discriminative approaches which face two key issues: error accumulation during sequence generation and inability to handle one-to-many mapping where multiple sequences can fold into same structure.

-----

🔧 Solution in this Paper:

Bridge-IF, a generative diffusion bridge model that:

→ Uses an expressive structure encoder to create informative prior sequences from input protein structures

→ Employs a Markov bridge to progressively refine sequences through multiple steps

→ Integrates pre-trained Protein Language Models with structural conditions

→ Introduces AdaLN-Bias and Structural adapter components for better structural information integration

-----

💡 Key Insights:

→ First use of Markov bridge formulation enables better handling of discrete sequences

→ Novel reparameterization perspective simplifies loss function for more effective training

→ Structural conditions can be effectively integrated into pre-trained models while maintaining compatibility

→ Progressive refinement from structure-aware prior works better than random noise prior

-----

📊 Results:

→ Achieves state-of-the-art 58.59% sequence recovery on CATH benchmark

→ Outperforms previous methods in both perplexity (3.83) and recovery metrics

→ Shows superior TM-score (0.81) indicating better foldability of generated sequences

→ Requires only 25 diffusion steps compared to 500 in previous approaches