Inverse folding from MD simulations

Summary

The combination of MD simulations with inverse folding was explored by (1) in redesigning a nanobody to have slower dissociation from its antigen.

Details

Brotzakis et al. use a Bayesian framework to link exploration of the conformational space by MD ($p(struct|seq)$) with exploration of sequence space by inverse folding ($p(seq|struct)$). For a single sequence, comparison of the latter to the former would be as follows. (1)

$$p(seq|struct)=\frac{p(struct|seq)p(seq)}{p(struct)}$$

They rearrange this formula by only considering the conformational space of mutants relative to the starting sequence:

$$p(struct|se{q}_{mut})=p(struct|se{q}_{wt})\frac{p(se{q}_{mut}|struct)p(se{q}_{wt})}{p(se{q}_{wt}|struct)p(se{q}_{mut})}$$

See also

• Protein language models and PLM-based structure prediction generalize to de novo designed proteins
• Inverse folding can generate more stable sequences when jointly run alongside a protein folding model
• Inverse folding NNs are better predictors of equilibrium dynamics than protein folding NNs

1.

Brotzakis ZF, Vendruscolo M, Skretas G. Design of Protein Sequences with Precisely Tuned Kinetic Properties. openRxiv; 2025. Available from: https://doi.org/10.1101/2025.02.13.638027