Combinatorial growth
The language model supplies plausible next residues, yet search must decide which branches justify structure prediction.
Predicted structure in the search loop
Scientific Design Engine · Peptide design
TrioPepLet protein–peptide structure guide the next sequence decision
TrioPep joins a peptide language model, token-level Monte Carlo tree search, OpenFold3 complex prediction, interface analysis, and fixed-pocket docking in one traceable loop. The result is a ranked peptide evidence package for experimental prioritization.
Locate the candidate before interpreting local contacts
The complete receptor-window view establishes where the FGIGI pocket hypothesis sits within XOD.
Evidence boundary. Predicted complex placement is a structural hypothesis and does not demonstrate inhibition.- Design object
- 3–5 residue peptide
- Target study
- XOD / PDB 1N5X
- Structure model
- OpenFold3 complex
- Evidence state
- Computational prioritization
01 / Abstract and project definition
Structure-Feedback Peptide Design
TrioPep represents peptide design as sequential decision-making. A language model proposes amino-acid residues; Monte Carlo tree search (MCTS) decides where to spend a finite structural-evaluation budget; completed sequences become protein–peptide structure hypotheses; and interface confidence plus pocket compatibility return as feedback to the tree.
The current case study targets xanthine oxidase (XOD). The TEI ligand in Protein Data Bank (PDB) entry 1N5X defines the pocket, while OpenFold3 predicts the complex for each completed short peptide. Component scores remain visible during curation so a high combined reward never hides weak interface, local-confidence, or docking evidence.
02 / Scientific problem
Sequence likelihood cannot answer where a peptide will bind or whether a predicted interface is credible.
Every added residue multiplies the search space, while structural evaluation is much more expensive than residue proposal. A single docking score also cannot represent chain placement, local confidence, aligned-error evidence, and pocket geometry. TrioPep allocates structural evaluations during search and keeps each signal inspectable.
Context-dependent placement
The same sequence can adopt different poses, so evaluation preserves receptor context and the TEI-defined pocket.
Multiple imperfect signals
Complex confidence, local confidence, aligned error, and docking energy answer distinct questions and remain separate during review.
03 / Method overview
A completed sequence becomes a structural hypothesis, and its evidence becomes value for the search tree.
MCTS expands peptide sequences one residue at a time using the language model as a proposal prior. At a valid terminal sequence, OpenFold3 predicts the XOD–peptide complex. Interface predicted TM score, weighted predicted local distance difference test, an aligned-error interaction score, and a normalized Vina term are combined for search while retained separately for curation.
Peptide representation
Each design is an amino-acid token sequence and the language model supplies a probability distribution for the next residue.
Output · Next-residue priorToken-level search
MCTS balances accumulated value and exploration under a fixed evaluation budget.
Output · Terminal peptide sequencesComplex prediction
OpenFold3 predicts an XOD–peptide complex and returns chain-placement and local-confidence evidence.
Output · Predicted complexInterface and pocket review
Chain-pair confidence, aligned-error evidence, weighted local confidence, and fixed-pocket Vina scoring form the interpretable reward stack.
Output · Component score bundle04 / Architecture
Every handoff has an explicit scientific contract
Inputs, operations, evidence, and outputs stay named so the source of each decision can be reviewed.
- Input
Target specification
XOD coordinates, receptor window, TEI-defined pocket, peptide length range, and a finite search budget.
→ - Generate
Peptide language model
Next-residue probabilities keep proposal in amino-acid sequence space.
→ - Search
Monte Carlo tree
Selection, expansion, terminal evaluation, and backpropagation decide which sequence families receive more computation.
→ - Evaluate
Structure-feedback stack
OpenFold3 complex confidence, interface evidence, weighted local confidence, and fixed-pocket docking.
→ - Output
Evidence-bearing shortlist
Sequence, rank, component metrics, predicted complex, pocket pose, and review status.
05 / End-to-end workflow
From scientific input to an evidence-bearing decision
Each stage consumes a bounded object and leaves a reviewable artifact for the next stage.
- 01Target specification
Define
Provide receptor coordinates, chain context, target pocket, peptide length range, and search budget.
- 02Sequence tree
Grow
Use language-model proposals and MCTS to extend a partial peptide one residue at a time.
- 03Complex model
Predict
Run OpenFold3 on each valid terminal sequence to obtain a receptor–peptide structural hypothesis.
- 04Score components
Measure
Calculate interface placement, aligned-error confidence, weighted local confidence, and fixed-pocket docking evidence.
- 05Updated tree values
Backpropagate
Return the combined reward to every ancestor of the evaluated branch.
- 06Ranked evidence set
Curate
Deduplicate and review sequence composition, pocket occupancy, pose geometry, confidence, and docking side by side.
06 / Experimental design
Questions, protocols, and comparison boundaries remain attached
Every study below describes what was varied, what was measured, and where interpretation must stop.
XOD short-peptide design
The case study uses Protein Data Bank entry 1N5X, receptor window 760–1280, and the TEI ligand to define the pocket. Candidate lengths span three to five residues.
- Evaluated winner rows
- 377
- Unique sequences
- 365
- Curated set
- 20
- Candidate length
- 3–5 aa
Matched evaluation allocation
Across 203 tasks, TrioPep and comparator generators receive the same budget of 500 candidate evaluations per task. TrioPep allocates evaluations during search; comparator outputs are externally selected after official sampling.
- Tasks
- 203
- Budget per task
- 500
- TrioPep selection
- Top 1 in search
- Comparator selection
- Top 1 of 500
Five reported XOD-inhibitory peptides
Five literature peptides with reported half-maximal inhibitory concentration values are rescored by the project reward. They remain outside the de novo ranking and provide a descriptive directional check only.
- Literature peptides
- 5
- Pearson correlation
- −0.867
- Spearman correlation
- −0.821
- Interpretation
- Descriptive
07 / Results and evidence
Read the signal with its evidence boundary
Quantitative summaries, structural views, and scientific interpretation remain separate layers.
Evaluated XOD winner rows
Every retained winner has a predicted structure record
Unique peptide sequences
Computational campaign set before manual curation
FGIGI residue-pair contacts
Coordinate-derived pairs within 4 Å across 25 XOD residues
Curated candidates
Component scores and structural context reviewed together
FGIGI is the current curated lead, with visible trade-offs
FGIGI has the highest curation score in the current set at 0.9020 and the strongest recorded Vina value at −10.70. Its combined reward is 0.9094, interface placement is 0.5426, aligned-error confidence is 0.7294, and weighted local confidence is 82.20.
Boundary. These values describe model and search outputs. They do not demonstrate XOD inhibition.The predicted pose exposes a residue-level interface hypothesis
The OpenFold3 FGIGI pose contains 41 peptide–receptor residue-pair contacts within 4 Å across 25 XOD residues. The five peptide positions contribute 9, 7, 12, 5, and 8 recorded contacts respectively.
Boundary. Geometric proximity identifies a predicted neighborhood and does not assign interaction type or biochemical causality.Reported activity context has the expected direction in a very small sample
Across five literature peptides, higher TrioPep reward is associated with lower reported half-maximal inhibitory concentration: Pearson r −0.867 and Spearman ρ −0.821.
Boundary. The sample is descriptive, has no significance claim, and does not validate FGIGI or the de novo ranking.
Pocket placement connects the search score to physical context
The close-up asks whether FGIGI occupies the intended TEI-defined site rather than an unrelated receptor surface.
Evidence boundary. The view supports prioritization and requires prospective structural and activity validation.
Five peptide residues distribute 41 predicted contacts
Dot position identifies the XOD residue and peptide position; dot size reflects minimum heavy-atom separation.
Evidence boundary. Coordinate-derived proximity does not establish a bond type or biochemical inhibition.
Four sequences occupy the same pocket frame differently
The shared receptor frame enables direct visual comparison of four predicted peptide placements.
Evidence boundary. Pose similarity and overlap are computational observations, not measured activity.Reading key
Scientific abbreviations
- XOD
- Xanthine oxidase, the target used by the current pocket-specific case study.
- ipTM
- Interface predicted TM score, a model confidence signal for relative chain placement.
- ipSAE
- Interaction prediction score derived from predicted aligned errors across the interface.
- pLDDT
- Predicted local distance difference test, used here as a local structural-confidence measure.
08 / Limitations and provenance
What the current evidence can establish
Model outputs, literature measurements, and generated-candidate hypotheses remain separate throughout the page.
Predicted structures are hypotheses
OpenFold3 confidence and interface geometry describe model behavior; alternative conformations, receptor flexibility, and induced-fit effects remain possible.
The reward is a search objective
The weighted total allocates computation. Similar totals can arise from different component strengths, so each component remains visible.
Activity evidence is limited
The five-peptide literature association is descriptive, and the generated candidates lack prospective inhibition, selectivity, stability, and cellular validation.