Search papers, labs, and topics across Lattice.
This paper introduces MCTH (Monte Carlo Tree Hallucination), an innovative framework for biomolecular design that utilizes uncertainty-aware planning to optimize sequence-structure co-design across various modalities, including protein-RNA and protein-DNA interactions. By leveraging pretrained folding and inverse-folding models as black-box operators, MCTH employs Monte Carlo Tree Search to efficiently allocate inference resources while accounting for model confidence and cross-expert consensus. The results demonstrate that MCTH significantly outperforms traditional sampling methods and shows effective transfer capabilities in evaluations with AlphaFold3 and Chai-1, highlighting its versatility and robustness in biomolecular design tasks.
MCTH revolutionizes biomolecular design by enabling adaptive search strategies that outperform conventional methods, even in complex multi-modal scenarios.
Biomolecular design underpins applications from molecular recognition to therapeutics and synthetic biology, yet de novo interaction design remains challenging-especially for DNA/RNA, underexplored non-protein modalities with scarce, heterogeneous complex data and sharper geometric and chemical constraints. We introduce MCTH (Monte Carlo Tree Hallucination), an inference-only framework that casts all-atom sequence-structure co-design as uncertainty-aware planning over hallucinated states from pretrained folding and inverse-folding models, with optional biophysical control within the same decision loop. MCTH treats these models as frozen black-box operators and uses Monte Carlo Tree Search to allocate a fixed inference budget across competing design trajectories, incorporating model confidence and uncertainty, as well as cross-expert consensus/disagreement when multiple predictors are available. Across protein-RNA, protein-DNA, protein-protein, and protein-ligand design, matched-budget experiments show that adaptive search improves over simpler sampling and cycling strategies, while held-out AlphaFold3 and Chai-1 evaluations demonstrate transfer beyond the search-time oracle. MCTH provides a shared planning layer across modalities while allowing task-specific folding, inverse-folding, and biophysical modules, requiring no fine-tuning or backpropagation through component models.
