MCTH

Hallucination-based design hands a structure predictor an input it was never meant to fold — an all-X sequence, a half-specified complex — and keeps the plausible structure it invents, then asks an inverse-folding model to write a sequence for that structure. Cycling the two yields designable molecules without training anything, but the cycle is a fixed schedule marching down a single trajectory: it spends the same expensive structure-prediction call on a branch that has stalled as on one that is still improving.

MCTH (Monte Carlo Tree Hallucination) replaces the schedule with a search. Each node of its tree is a concrete sequence–structure state — the current designable sequence plus the all-atom complex it folds into under fixed target conditioning — and each edge is one inverse-fold-then-refold step. Monte Carlo Tree Search decides which leaf receives the next folding call, through a UCT-style rule the authors call CU-PUCT that combines the action value with the folding model's confidence, the entropy of the inverse-folding proposal, agreement among multiple folding experts where more than one is available, and an optional biophysical penalty. The component models are never touched: Boltz-2, AlphaFold 2, ProteinMPNN, NA-MPNN and LigandMPNN are called as frozen black boxes through forward inference alone, with no gradients, no fine-tuning and no access to their parameters.

What that buys is one planner across protein–RNA, protein–ssDNA, protein–protein and protein–ligand design: only the inverse-folding backend and the biophysical term change with the modality, while the state representation, selection rule and backup stay fixed. It comes from Xuefeng Liu and colleagues at the University of Florida, the University of Chicago, the Toyota Technological Institute at Chicago, Mayo Clinic and Argonne National Laboratory, and shares its planning-over-frozen-experts lineage with tree-search design methods such as HighPlay and with gradient-based hallucination methods such as BoltzDesign1.

#Key Features

• Adaptive compute allocation: A fixed budget of folding calls is distributed across competing design trajectories rather than consumed by one. Matched-budget ablations against greedy cycling and best-of-N sampling hold everything else identical, isolating the planning strategy.
• Uncertainty-aware selection: Inverse-folding entropy and folding confidence enter the selection rule as separate terms, so the search distinguishes ambiguity about which sequence to write from ambiguity about how the complex folds.
• Biophysics inside the decision loop: For DNA aptamers a nearest-neighbour self-folding free energy defines a favourable window; candidates that are too self-stable or too unstructured are down-weighted during selection and repaired by mutations restricted to non-interface positions.
• Structural guidance at inference: Predicted non-interface base pairs are injected into Boltz-2's potential-energy steering as a soft contact prior, preserving an aptamer's pre-organized secondary structure away from the binding site while leaving interface bases free.
• Multiple folding experts: Candidates can be scored by two predictors at once, with their consensus sharpening the action prior and their disagreement acting as an explicit exploration bonus.

#Technical Details

The reference configuration budgets 100 Boltz-2 folding calls per target, expanding up to four children per node to a maximum tree depth of five over 50 search iterations, with early stopping after ten iterations without improvement; each target runs on a single A100 80GB GPU, taking roughly an hour for aptamers and nine hours for protein binders. On seven protein–RNA aptamer targets MCTH reaches a mean best ipTM of 0.946 against 0.923 and 0.894 for matched-budget baselines, and 0.884 against 0.824 and 0.794 when the final designs are re-scored by held-out AlphaFold 3. On the 12-target CaoData binder benchmark it reaches 0.864 mean best ipTM. Adding AlphaFold 2 as a second search-time expert lifts the mean score under held-out Chai-1 from 0.18 to 0.57, and at ten designs per target 37.3% of its binders pass a held-out quality filter against 20.0% for RFdiffusion.

#Applications

MCTH suits groups that already run released structure predictors and want better designs out of them without training anything: aptamer design against a fixed protein target, where RNA and single-stranded DNA are poorly served by protein-centric pipelines, de novo protein binder design, and sequence design around a specified ligand pocket. Because the planner consumes only sequences, structures and confidence scores, a laboratory can substitute whichever folding or inverse-folding checkpoint it is licensed to run.

#Impact

MCTH argues that the gains available at inference time in biomolecular design are a scheduling problem as much as a modelling one, and that a planner kept separate from its predictors inherits every improvement to them for free. Two caveats bound the claims. Evaluation is entirely computational: ipTM under Boltz-2, AlphaFold 3 and Chai-1 plus a Rosetta interface check, with no wet-lab binding data. And nothing MCTH designs can exceed what its components can already fold and redesign. The preprint carries no code release; its authors state that code and configuration files will be published on acceptance.