Reuse a chemical prior
A fragment policy should retain broad chemical knowledge while a new pocket supplies its own evaluation context.
Pocket-conditioned molecular generation
Scientific Design Engine · Small-molecule design
TrioMol2One frozen generator, target-specific evidence, every pocket
TrioMol2 adapts a fragment-based molecular generator at inference time through an explicit memory of receptor states, poses, contacts, and labels. Molecular proposals, state selection, scoring, and feedback remain inspectable across both pocket-scale studies and an all-target campaign.
The evaluated pocket sits inside the full receptor state
A professional molecular rendering places the selected pocket and accepted-memory ligand evidence within the receptor model used by the computational study.
Evidence boundary. Computational receptor and pose context; no affinity or inhibition is established.- Design object
- Pocket-conditioned small molecule
- Adaptation
- Inference-time target memory
- Campaign scale
- 26,562 target pockets
- Evidence state
- Computational prioritization
01 / Abstract and project definition
All-Target Molecular Generation
TrioMol2 separates a transferable chemical proposal model from a target-specific adaptation state. The Fragment-based Generative Pre-trained Transformer (FRAGPT) remains frozen while physical target evidence guides Monte Carlo tree search (MCTS), receptor-state selection, docking, contact evaluation, and synthesis-aware ranking.
The project asks whether one generator can serve many targets without target-by-target retraining. Its answer is an evolving target memory: each accepted candidate keeps the receptor state, pose, score decomposition, and search trajectory that produced its rank. Updating the memory changes the target evidence available to later rounds while leaving global generator weights unchanged.
02 / Scientific problem
Target specificity is often hidden inside retrained weights or reduced to one static structure.
A reusable generator must respond to distinct pocket geometries, receptor states, contact patterns, and chemistry constraints. Retraining for every target makes adaptation expensive and difficult to audit. Treating one receptor conformation as ground truth also hides uncertainty in state selection. TrioMol2 exposes these choices as evidence carried through search.
Select physical evidence explicitly
Receptor states, accepted poses, residue contacts, and optional labels remain addressable inputs to candidate evaluation.
Keep the search path reviewable
Every ranked molecule retains the fragments, selected state, physical scores, synthesis evidence, and feedback history behind the decision.
03 / Method overview
A frozen fragment policy proposes chemistry while physics-informed memory supplies the target condition.
For a partial molecule, retrieval selects relevant memory entries. MCTS allocates search among fragment continuations, reaction-template priors favor defensible transformations, and an explicit reward combines docking, contact agreement, affinity priors, pose-quality penalties, and synthesis evidence. Accepted evidence can be added to memory for the next round.
Frozen proposal model
FRAGPT remains in evaluation mode and proposes chemically local fragment continuations without target-specific weight updates.
Output · Partial molecular trajectoriesPhysics-informed target memory
Entries couple receptor states with pose geometry, contact evidence, and optional activity labels so target adaptation remains inspectable.
Output · Retrieved target contextTree search and chemical priors
MCTS balances exploitation and exploration while reaction-template priors shape plausible branch expansion.
Output · Completed candidate poolDecomposed evaluation
State-aware docking, contacts, affinity priors, pose diagnostics, and synthesis terms are recorded separately before ranking.
Output · Ranked evidence packets04 / 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
Pocket and memory state
A target pocket, one or more receptor states, geometric or contact evidence, and an initial memory state.
→ - Proposal
Fragment trajectory
The frozen generator supplies next-fragment probabilities for each partial molecule.
→ - Evaluation
State-aware physical score
The selected receptor state, docked pose, contact agreement, pose quality, and synthesis evidence remain separate.
→ - Output
Auditable molecule
A ranked structure with its receptor state, pose, score decomposition, and complete search trajectory.
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 the pocket
Load receptor coordinates, pocket geometry, contact context, and any prior target labels.
- 02Target memory
Initialize memory
Register the receptor state and any accepted ligand, pose, or contact evidence available for the task.
- 03Search tree
Expand fragments
Use the frozen fragment model and MCTS to explore chemically informed molecular trajectories.
- 04Evaluation packet
Evaluate states
Dock completed candidates, select a receptor state, and preserve physical and synthesis score components.
- 05Candidate ledger
Rank and review
Deduplicate structures and review reward, docking, contacts, pose quality, and synthesis evidence together.
- 06Updated memory
Feed back evidence
Add accepted poses, contacts, scores, or labels to the target memory for a later search round.
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.
Human-proteome pocket campaign
One fixed protocol generated ten molecules for every pocket in a curated human-proteome atlas. The coverage audit accounts for every pocket packet and every planned output.
- Pockets
- 26,562
- UniProt targets
- 9,919
- Receptor structures
- 11,260
- Generated molecules
- 265,620
AML1–ETO9a pocket study
Ten search rounds and eight seeds per round produced a merged computational record used to inspect reward trajectories, receptor context, contacts, candidate chemistry, and closed-loop memory behavior.
- Merged evaluations
- 640
- Unique top candidates
- 176
- Filtered candidates
- 61
- Curated set
- 20
OpenBind retrieval and backend checks
A historical diagnostic separates evaluator reliability, receptor-state aggregation, and retrieval quality. The design keeps an engineering success signal visible beside an unresolved scientific retrieval gap.
- Successful Vina calls
- 994 / 1,000
- Timeouts
- 0
- State-specific held-out rows
- 7 / 52
- Retrieval conclusion
- Open
07 / Results and evidence
Read the signal with its evidence boundary
Quantitative summaries, structural views, and scientific interpretation remain separate layers.
Pocket coverage
26,562 of 26,562 pocket packets accounted for
Generated molecules
Ten molecules under the same campaign budget for every pocket
AML1 curated candidates
Selected from 61 filtered and 176 unique top candidates
Historical evaluator success
Backend reliability diagnostic; scientific retrieval remains open
All-target execution reached the planned atlas scope
The coverage audit reports 26,562 completed pockets, 9,919 UniProt targets, 11,260 receptor structures, and 265,620 generated molecules. The same pocket-conditioned budget was applied throughout the atlas.
Boundary. Coverage establishes execution breadth. It does not establish binding, potency, selectivity, or biological activity.The AML1 record links scores to a physical pocket hypothesis
The study preserves the evaluated receptor state, accepted-memory pose, residue-level contact record, search trajectory, and curated chemistry. The round-six leader retained 100% validity and Lipinski pass rate in its recorded quality summary, with 98.4% uniqueness and 96.9% diversity.
Boundary. Docking, contacts, and molecular quality metrics support computational triage and require prospective testing.Backend reliability and retrieval quality diverge
The historical packet recorded 994 successful Vina calls out of 1,000 with no timeouts or segmentation faults. At the same time, evaluated real retrieval methods did not exceed the shuffled cross-cluster control, leaving target-state retrieval as an open research problem.
Boundary. Engineering reliability cannot substitute for evidence that the retrieved state is biologically correct.
The accepted memory preserves a residue-level environment
The pocket close-up connects the selected ligand pose to the recorded residue neighborhood and nearest heavy-atom separations.
Evidence boundary. Distance annotations are computational geometry and do not assign experimentally confirmed interaction types.
The curated pool retains multiple chemical scaffolds
Six leading structures show how the final pool is read through molecular identity, source round, combined score, and contact evidence.
Evidence boundary. These structures are computationally prioritized and have not been experimentally tested in this study.Reading key
Scientific abbreviations
- FRAGPT
- Fragment-based Generative Pre-trained Transformer used as the frozen molecular proposal model.
- MCTS
- Monte Carlo tree search, which allocates evaluations across competing molecular branches.
- Vina
- A molecular docking scoring engine used here for computational pose and score evaluation.
08 / Limitations and provenance
What the current evidence can establish
Public values are limited to reviewed computational snapshots already present in the TrioMol2 project. No unpublished biological outcome is inferred.
Physical scores remain approximations
Docking, contact rules, affinity priors, and synthesis models simplify receptor flexibility, solvation, entropy, kinetics, and experimental feasibility.
Target-level evidence is uneven
The AML1 study is a single-pocket analysis. The OpenBind packet is target-level, and weak retrieval performance limits a broader generalization claim.
Wet-lab validation remains required
The current record does not establish biochemical potency, cellular activity, selectivity, pharmacology, or efficacy for generated candidates.