HighPlay

Cyclic peptides occupy a useful middle ground between small molecules and biologics: the ring constrains conformation, which sharpens affinity and slows proteolysis while keeping polarity low enough for membrane permeability. De novo design is the hard part. Template modification stays trapped near natural scaffolds, combinatorial libraries cannot cover a space of roughly 10^13 decamers, and structure-based design presumes a solved target structure and a known binding site. HighPlay, developed by the Duan group at Macao Polytechnic University with Zhejiang University of Technology, Shenzhen Highslab Therapeutics, and clinical collaborators at Shanghai Jiao Tong University, removes all three assumptions: it takes a target protein sequence and nothing else.

The method treats binder design as a game. A cyclic peptide sequence is a board state, a single-residue substitution is a move, and the reward is how confidently HighFold — the group's earlier cyclic-peptide structure predictor, built on AlphaFold 2 — folds the resulting peptide-target complex. Monte Carlo tree search explores mutation paths, guided by a transformer policy-value network trained on the search's own episodes.

This follows the pattern of design tools that wrap a frozen structure predictor in an optimization loop, as BindCraft does over AlphaFold 2 hallucination and BoltzDesign1 does over an all-atom co-folding model. HighPlay's distinction is the search strategy: tree search with a learned prior accumulates experience across moves, where hallucination and random-mutation directed evolution treat each proposal independently.

#Key Features

• Target sequence is the only input: No binding pocket, template backbone, or length constraint is supplied, so it applies to targets with no solved complex and no characterized interface.
• Learned search prior: A transformer encoder with policy and value heads reads the L x 20 state matrix and scores candidate mutations, focusing tree expansion on high-potential regions rather than mutating at random.
• Folding confidence as the objective: Peptide pLDDT drives the reward, and it correlates with Rosetta interface energy per unit buried surface, so optimizing confidence also improves modeled interface quality.
• Cyclization built into initialization: Starting sequences carry two cysteines separated by more than two-thirds of the length, so every candidate is a disulfide-closed macrocycle from the first move.
• Experimentally confirmed output: TEAD4 designs were synthesized and measured by surface plasmon resonance rather than resting on simulation alone.

#Technical Details

The environment state is an L x 20 binary matrix over positions and amino acid types, with a flattened action space of the same size; each action rewrites one row. MCTS selects children by the largest sum of action value Q and an exploration term U weighted by the network's prior and visit counts. Episodes terminate when a move lowers the reward or reproduces a visited sequence, and a new starting sequence is drawn after thirty episodes without improvement. Episode data accumulates in a replay buffer used to train the policy-value network against a combined value and policy loss.

Designs were run against three targets (PDB 1SSC, 3R7G, 6SEO), 13 runs each, with peptide lengths from 8 to 20 residues and 300 search iterations per run. For 1SSC, 8 of 13 runs reached the pLDDT > 70 threshold, with a best of 90.15 at length 14. Candidates passing that threshold and a Rosetta dG_separated/dSASA x 100 below -1.5 advanced to 100 ns Amber molecular dynamics with the ff19SB force field, where backbone RMSD converged around 3.0 Å and MM-GBSA binding energies all fell below -35 kcal/mol, reaching -128.45 kcal/mol for the best 3R7G binder. Ten synthesized TEAD4 binders all bound in surface plasmon resonance, with dissociation constants between 3.78 and 7.62 µM against 1.46 µM for the natural YAP(61-99) control.

#Applications

HighPlay suits targets where structure-based design stalls — interfaces with no obvious pocket, or proteins whose complexes have never been crystallized — because it needs only the sequence and finds a binding site as part of the search. The practical output is a short ranked list per target, filtered by folding confidence and Rosetta interface score before synthesis, with molecular dynamics as a second gate. TEAD4 is the worked example: a Hippo-pathway transcription factor of oncology interest, approached with no prior binder information.

#Impact

Published in the Journal of Medicinal Chemistry after its preprint, the work is separated from purely in-silico design reports by its wet-lab confirmation: every one of ten synthesized cyclic peptides bound at micromolar affinity, though none beat the natural YAP control, so these are starting points rather than optimized leads. The framework inherits its ceiling from the structure predictor it queries, and the released code is the search procedure — the AlphaFold parameters it scores with are downloaded separately. The repository's README states MIT terms but ships no LICENSE file, and the preprint carries a CC BY-NC license, so reuse terms are unsettled. Within the group's own work it sits alongside High-PepBinder as a second route to target-specific peptide binders, one search-based and one generative.