Bayes Pharma.ai Discovery is an open, modular, evidence-traceable workspace that connects target biology, molecular design, virtual perturbation hypotheses, synthesis feasibility, physics validation, and PK/PD translation.
Instead of treating docking, QSAR, HTS, and generative chemistry as separate apps, it turns them into a reviewable decision package with provenance, uncertainty, and scientific challenge built in.
Tool Registry + Research Agent
OpenMM / GROMACS / OpenFE Roadmap
Provenance + Verifier Layer
Synthesis-Aware Design
Virtual Cell Perturbation
PK/PD Translation Handoff
The orchestration layer should understand the objective, select the right tools, challenge the results, and preserve every input, output, assumption, and uncertainty.
Starts with disease biology and development constraints, then assembles a defensible path through target evidence, virtual perturbation, tractability, structures, chemistry, ADMET, and PK/PD translation.
Bayes Pharma.ai stays the orchestrator. Biomni-style biomedical agents, TxGemma, Boltz-2, AiZynthFinder, and BioNeMo-class services should run as sandboxed specialist workers with strict schemas and approval gates.
MD and FEP should use an open, licensed-correct stack with uncertainty, applicability checks, and reviewer-visible assumptions.
Open ligand parameterization path for OpenMM, Amber, and GROMACS-compatible workflows.
Open LigandsProtein and nucleic-acid simulation option for prepared structures and congeneric validation sets.
ProteinsUseful for membrane, lipid, and carbohydrate contexts when system preparation is controlled.
MembranesOPLS4 should be described only as a licensed integration path, not as an included open-source force field.
Live tractability, disease association, safety liability, and known-drug evidence from the public Open Targets Platform GraphQL API (no API key required).
Synthetic-accessibility score (Ertl & Schuffenhauer 2009), functional-group disconnection hints, and a purchasable-precursor similarity check. Full retrosynthesis search (AiZynthFinder) remains roadmap.
Each module should feed the same scientific record: objective, evidence, model limits, generated options, experimental constraints, and next decisions.
Live tractability, disease association, safety liability, and known-drug evidence from the public Open Targets Platform API.
LivePlans the workflow, calls specialist tools, challenges results, and compiles a provenance-linked decision package.
OrchestratorGuided workflow across chemistry, docking, pharmacophores, QSAR, HTS analysis, lead optimization, and ADME/Tox triage.
WorkflowConnect competitor programs, trials, approvals, publications, and development signals to the target hypothesis.
EvidenceFrame the design objective, constraints, target rationale, analog strategy, and review questions before tools run.
AIGenerate analogs against potency, selectivity, ADMET, solubility, novelty, cost, and synthesizability constraints.
AIRank genes, compounds, and combinations by predicted disease-cell-state reversal, pathway plausibility, uncertainty, and wet-lab validation priority.
Focused MVPLive RDKit synthetic-accessibility score, disconnection hints, and building-block similarity triage. Full route search and patent proximity remain roadmap.
Live triageUse experimental structures first, then AlphaFold or Boltz-2-style complex prediction, pocket detection, ensemble docking, and consensus confidence.
Live + RoadmapUse quantum chemistry as a selective supporting check for difficult poses, metals, charge states, and polarization-sensitive decisions.
SpecialistLive RDKit + AlphaFold pre-flight readiness triage today. Full MD and relative binding free-energy execution remain roadmap.
Live triageExploratory VQE/QAOA workflows for research contexts where a reviewer can inspect assumptions, limits, and compute cost.
Research3D pharmacophore extraction, consensus modeling, and database screening.
LivePrimary hit identification, IC50 curve fitting, plate QC, and hit prioritization.
LiveRDKit descriptors, multiple ML algorithms, y-randomization, and applicability domain.
LiveSAR analysis, MPO scoring, liability triage, and analog design.
LiveStructure standardization, molecule profiling, library triage, and candidate prioritization.
LiveUse verified experimental results, dataset versioning, leakage checks, temporal validation, and uncertainty-aware acquisition.
LiveUse foundation models and biomedical agents as contributors to an ensemble. The UI should expose disagreement, confidence, and applicability instead of hiding everything behind one score.
Target-disease evidence, tractability, safety, known drugs, and disease biology
Evidence GraphComplex structure and affinity signal for rapid triage, not a final FEP replacement
Structure/AffinityRetrosynthesis routes, purchasable precursors, policy control, and route confidence
SynthesisTherapeutic prediction specialist for toxicity, BBB, reaction, and explanation tasks
EnsembleOptional sandboxed biomedical agent for omics, CRISPR, literature, and experiment planning
SandboxedGPU-scale protein, genomic, and fine-tuning infrastructure when workload justifies it
InfrastructureThe priority is not more isolated models. It is target evidence, structure consensus, synthesis feasibility, provenance, and validated learning loops.
Bayes Pharma.ai Agent OS already exposes typed JSON tool schemas per domain (discovery, generics, clinical, rescue) with risk levels and live introspection.
LiveLive tractability, disease association, safety liability, and known-drug evidence from the public Open Targets Platform API.
LiveLive focused MVP: TB-infected macrophage cell-state reversal ranking with model agreement and uncertainty. See the Decision Stack grid above.
LiveUse complex prediction and affinity as one ranking signal beside docking, QSAR, and experimental context.
NextLive grid-based (LIGSITE-style) pocket detection auto-fills the Vina docking box in AlphaDock OS; AlphaFold/Enhanced ML/Vina consensus runs in Compare mode.
LiveLive RDKit synthetic-accessibility score, disconnection hints, and route-risk rating above. Full route search and cost/precursor sourcing remain roadmap.
Live triageLive hash-chained provenance dossier in AlphaDock OS: every run is recorded and exportable as tamper-evident JSON.
LiveAdd therapeutic-property specialists while displaying disagreement with QSAR and experimental domain limits.
NextRun OpenMM or GROMACS only after protein prep, protonation review, ligand parameters, and QC gates.
NextUse relative free energy on prepared congeneric series with MBAR uncertainty and cycle-closure checks.
NextClose the loop with verified assay results, temporal validation, leakage prevention, and acquisition strategy.
NextKeep AlphaGenome/Evo-style variant-to-target workflows separate from small-molecule docking.
LaterAdopt GPU-scale biomolecular model infrastructure after workloads need fine-tuning or distributed inference.
LaterUse dynamics and relative free-energy calculations only when the system is prepared, the ligand series is appropriate, and the uncertainty can be reviewed.
Real RDKit structural checks plus AlphaFold model metadata. Full system building (solvation, ion placement) and production OpenMM/GROMACS/OpenFE runs remain roadmap.
Established docking suites and AI discovery platforms are strong in their lanes. Bayes Pharma.ai should win by connecting target evidence, chemistry, developability, PK/PD translation, and audit-ready provenance.
| Decision Question | Standalone Docking Suite | AI Discovery Platform | Bayes Pharma.ai Discovery |
|---|---|---|---|
| Where should the program start? | Often after target selection | Often target or molecule first | Disease-to-target evidence graph before design |
| What supports the pose? | Docking and physics workflows | Model confidence and generated hypotheses | PDB/AlphaFold/Boltz-style hierarchy plus Vina/GNINA consensus and pose QC |
| Can the compound be made? | Usually a separate chemistry workflow | Sometimes implicit in generation | Synthesis route, precursor access, complexity, novelty, and cost shown beside potency |
| Will it translate? | Typically outside the discovery workbench | Often weakly connected | PBPK, PK/PD, developability, formulation, and clinical pharmacology handoff |
| Can reviewers reconstruct the decision? | Project-dependent reporting | Variable explainability | Every input, model version, assumption, failed option, and uncertainty preserved |
The strongest version of Bayes Pharma.ai does not pretend every model is right. It records evidence, rejected paths, failed runs, model disagreement, and what must be verified experimentally.
Store molecular inputs, structures, parameters, model versions, prompts, generated alternatives, rejected candidates, and reviewer notes.
A separate reviewer layer checks whether the chosen tools fit the objective, whether assumptions are overextended, and where independent evidence disagrees.
Use fast screens first, then escalate only defensible candidates to docking ensembles, MD, RBFE, or specialist quantum chemistry.
Active learning should be driven by verified experimental results, not simple thumbs-up feedback. The loop must protect against leakage, version datasets, and prove prospective improvement.
Bring target evidence, molecular design, synthesis feasibility, physics validation, and PK/PD translation into one traceable Bayes Pharma.ai workflow.