Leveraging generative hallucination and biophysics-informed modeling for unified biomolecular sequence-structure co-design

Details

Designing novel biomolecules that interact with specific targets — proteins, RNA, DNA, small molecules — is central to drug discovery and synthetic biology. Most approaches require either expensive wet lab iteration or computationally intensive fine-tuning of large generative models. MCTH takes a different path: it treats pretrained folding models (like AlphaFold) and inverse-folding models as frozen black-box operators, then uses Monte Carlo Tree Search to allocate a fixed inference budget across competing design trajectories. The key insight is that MCTS can incorporate model confidence, uncertainty, and cross-expert consensus/disagreement to guide search without backpropagating through any component model. This makes MCTH broadly applicable: the same planning layer works across protein-RNA, protein-DNA, protein-protein, and protein-ligand design tasks, with task-specific modules plugged in per modality. The paper focuses especially on DNA and RNA — underexplored non-protein modalities with scarce complex data and sharper geometric constraints. Held-out evaluations using AlphaFold3 and Chai-1 demonstrate that gains transfer beyond the models used during search. The code appears to be available. DNA/RNA targeting is particularly relevant to RNA therapeutics, which has seen major investment since the success of mRNA vaccines.