Bayes Pharma.ai Discovery Target-to-Lead Copilot
Evidence-to-Candidate Discovery

Target-to-lead
decisions, not
isolated scores.

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

5
Decision Layers
13+
Specialist Tools
QC
Verifier Agent Layer
PK
PBPK / PK-PD Handoff
Target Evidence Graph Research Agent Generative Chemistry Structure & Affinity Studio Synthesis Feasibility Assay-Driven Active Learning Virtual Perturbation Lab Open Physics Validation Specialist Model Ensemble
Research Agent Architecture

From scientific objective to reviewable decision dossier

The orchestration layer should understand the objective, select the right tools, challenge the results, and preserve every input, output, assumption, and uncertainty.

Target-to-Lead Copilot

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.

Disease Evidence Target Score Cell-State Shift Structure Hierarchy Consensus Docking Decision Dossier
Launch the live discovery orchestrator

Specialist Worker Model

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.

JSON Tool Schemas MCP-Ready Services Sandboxed Execution QC Verifier Audit Trail
View the live tool registry
Validated Physics Stack

Use physics as evidence, not decoration

MD and FEP should use an open, licensed-correct stack with uncertainty, applicability checks, and reviewer-visible assumptions.

OpenFF Sage / GAFF

Open ligand parameterization path for OpenMM, Amber, and GROMACS-compatible workflows.

Open Ligands

AMBER

Protein and nucleic-acid simulation option for prepared structures and congeneric validation sets.

Proteins

CHARMM

Useful for membrane, lipid, and carbohydrate contexts when system preparation is controlled.

Membranes

OPLS4 should be described only as a licensed integration path, not as an included open-source force field.

Live · Open Targets Platform

Target Evidence Studio

Live tractability, disease association, safety liability, and known-drug evidence from the public Open Targets Platform GraphQL API (no API key required).

Look up a target

Evidence result

Enter a target symbol and fetch evidence.
Live · RDKit

Synthesis Intelligence

Synthetic-accessibility score (Ertl & Schuffenhauer 2009), functional-group disconnection hints, and a purchasable-precursor similarity check. Full retrosynthesis search (AiZynthFinder) remains roadmap.

Triage a molecule

Triage result

Enter a SMILES string and triage synthesis.
Decision Stack

The evidence-to-lead workspace

Each module should feed the same scientific record: objective, evidence, model limits, generated options, experimental constraints, and next decisions.

Target Evidence Studio

Live tractability, disease association, safety liability, and known-drug evidence from the public Open Targets Platform API.

Live

Bayes Pharma.ai Research Agent

Plans the workflow, calls specialist tools, challenges results, and compiles a provenance-linked decision package.

Orchestrator

Discovery Suite

Guided workflow across chemistry, docking, pharmacophores, QSAR, HTS analysis, lead optimization, and ADME/Tox triage.

Workflow

Opportunity Intelligence

Connect competitor programs, trials, approvals, publications, and development signals to the target hypothesis.

Evidence

Research Agent Briefing

Frame the design objective, constraints, target rationale, analog strategy, and review questions before tools run.

AI

Generative Chemistry

Generate analogs against potency, selectivity, ADMET, solubility, novelty, cost, and synthesizability constraints.

AI

Virtual Perturbation Lab

Rank genes, compounds, and combinations by predicted disease-cell-state reversal, pathway plausibility, uncertainty, and wet-lab validation priority.

Focused MVP

Synthesis Intelligence

Live RDKit synthetic-accessibility score, disconnection hints, and building-block similarity triage. Full route search and patent proximity remain roadmap.

Live triage

Structure & Affinity Studio

Use experimental structures first, then AlphaFold or Boltz-2-style complex prediction, pocket detection, ensemble docking, and consensus confidence.

Live + Roadmap

QM-Aware Pose Review

Use quantum chemistry as a selective supporting check for difficult poses, metals, charge states, and polarization-sensitive decisions.

Specialist

MD & FEP Studio

Live RDKit + AlphaFold pre-flight readiness triage today. Full MD and relative binding free-energy execution remain roadmap.

Live triage

Advanced Quantum Methods

Exploratory VQE/QAOA workflows for research contexts where a reviewer can inspect assumptions, limits, and compute cost.

Research

Pharmacophore Modeling

3D pharmacophore extraction, consensus modeling, and database screening.

Live

HTS Analyzer

Primary hit identification, IC50 curve fitting, plate QC, and hit prioritization.

Live

QSAR Builder

RDKit descriptors, multiple ML algorithms, y-randomization, and applicability domain.

Live

Lead Optimization

SAR analysis, MPO scoring, liability triage, and analog design.

Live

Cheminformatics Workspace

Structure standardization, molecule profiling, library triage, and candidate prioritization.

Live

Assay-Driven Active Learning

Use verified experimental results, dataset versioning, leakage checks, temporal validation, and uncertainty-aware acquisition.

Live
Specialist Workers

Models as auditable evidence sources

Use 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.

Open Targets

Target-disease evidence, tractability, safety, known drugs, and disease biology

Evidence Graph

Boltz-2

Complex structure and affinity signal for rapid triage, not a final FEP replacement

Structure/Affinity

AiZynthFinder

Retrosynthesis routes, purchasable precursors, policy control, and route confidence

Synthesis

TxGemma

Therapeutic prediction specialist for toxicity, BBB, reaction, and explanation tasks

Ensemble

Biomni Worker

Optional sandboxed biomedical agent for omics, CRISPR, literature, and experiment planning

Sandboxed

BioNeMo

GPU-scale protein, genomic, and fine-tuning infrastructure when workload justifies it

Infrastructure
Product Roadmap

Build the defensible sequence first

The priority is not more isolated models. It is target evidence, structure consensus, synthesis feasibility, provenance, and validated learning loops.

Tool Registry & Schemas

Bayes Pharma.ai Agent OS already exposes typed JSON tool schemas per domain (discovery, generics, clinical, rescue) with risk levels and live introspection.

Live

Target Evidence Graph

Live tractability, disease association, safety liability, and known-drug evidence from the public Open Targets Platform API.

Live

Virtual Perturbation Lab

Live focused MVP: TB-infected macrophage cell-state reversal ranking with model agreement and uncertainty. See the Decision Stack grid above.

Live

Boltz-2 Worker

Use complex prediction and affinity as one ranking signal beside docking, QSAR, and experimental context.

Next

Pocket + Consensus Docking

Live grid-based (LIGSITE-style) pocket detection auto-fills the Vina docking box in AlphaDock OS; AlphaFold/Enhanced ML/Vina consensus runs in Compare mode.

Live

Synthesis Planning

Live RDKit synthetic-accessibility score, disconnection hints, and route-risk rating above. Full route search and cost/precursor sourcing remain roadmap.

Live triage

Provenance & Benchmarks

Live hash-chained provenance dossier in AlphaDock OS: every run is recorded and exportable as tamper-evident JSON.

Live

TxGemma Ensemble

Add therapeutic-property specialists while displaying disagreement with QSAR and experimental domain limits.

Next

MD Preparation

Run OpenMM or GROMACS only after protein prep, protonation review, ligand parameters, and QC gates.

Next

OpenFE RBFE

Use relative free energy on prepared congeneric series with MBAR uncertainty and cycle-closure checks.

Next

Assay Learning Loop

Close the loop with verified assay results, temporal validation, leakage prevention, and acquisition strategy.

Next

Genomic Target Intelligence

Keep AlphaGenome/Evo-style variant-to-target workflows separate from small-molecule docking.

Later

BioNeMo Infrastructure

Adopt GPU-scale biomolecular model infrastructure after workloads need fine-tuning or distributed inference.

Later
Physics Validation

MD & FEP Studio

Use dynamics and relative free-energy calculations only when the system is prepared, the ligand series is appropriate, and the uncertainty can be reviewed.

Readiness Triage Inputs

Real RDKit structural checks plus AlphaFold model metadata. Full system building (solvation, ion placement) and production OpenMM/GROMACS/OpenFE runs remain roadmap.

Readiness Report

Run the readiness triage to see blocking issues, advisory notes, and passed checks.
Differentiation

Designed around evidence, not just scores

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 selectionOften target or molecule firstDisease-to-target evidence graph before design
What supports the pose?Docking and physics workflowsModel confidence and generated hypothesesPDB/AlphaFold/Boltz-style hierarchy plus Vina/GNINA consensus and pose QC
Can the compound be made?Usually a separate chemistry workflowSometimes implicit in generationSynthesis route, precursor access, complexity, novelty, and cost shown beside potency
Will it translate?Typically outside the discovery workbenchOften weakly connectedPBPK, PK/PD, developability, formulation, and clinical pharmacology handoff
Can reviewers reconstruct the decision?Project-dependent reportingVariable explainabilityEvery input, model version, assumption, failed option, and uncertainty preserved
Positioning is intentionally capability-based. The page avoids unverifiable "feature-for-feature" and cost claims, and focuses on the Bayes Pharma.ai lane: evidence-linked discovery to human dosing.
Decision Memory

Every claim needs a review trail

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.

Provenance Ledger

Store molecular inputs, structures, parameters, model versions, prompts, generated alternatives, rejected candidates, and reviewer notes.

  • Dataset and model version tracking
  • Assumptions visible in reports
  • Negative results preserved
View the live provenance ledger →

Verifier Agent

A separate reviewer layer checks whether the chosen tools fit the objective, whether assumptions are overextended, and where independent evidence disagrees.

  • Applicability-domain checks
  • Cross-model disagreement flags
  • Experimental validation prompts

Tiered Compute Gates

Use fast screens first, then escalate only defensible candidates to docking ensembles, MD, RBFE, or specialist quantum chemistry.

  • Vina/GNINA for early triage
  • MD/RBFE for prepared shortlists
  • Cost and runtime approval gates
CLOSED-LOOP LEARNING

Models improve from experiments, not casual clicks

Active Learning Discipline

Closed-loop assay intelligence

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.

  • Verified assay uploads drive model updates
  • Temporal validation separates past from future
  • Acquisition balances potency, uncertainty, diversity, ADMET, cost, and synthesis
  • Negative and failed results remain part of the evidence base
Evidence First Scientific models update only from governed data
QCVersioned assays
Split
Temporal validation
No Leak
Dataset controls

Build a reviewable target-to-lead package

Bring target evidence, molecular design, synthesis feasibility, physics validation, and PK/PD translation into one traceable Bayes Pharma.ai workflow.

Start molecule triage Explore Discovery Suite Review architecture
Built for auditable scientific decisions, not black-box score chasing.