DrugSynthAI
User Manual

New to DrugSynthAI? This section gives you everything you need to understand the platform and see your first drug candidate in under five minutes. No prior drug discovery knowledge required.

DrugSynthAI is designed for three primary user groups, each engaging with the platform at different levels of technical depth.

Drug discovery is among the most expensive and failure-prone processes in science. The industry average for bringing a drug to approval is 10–15 years? and $2.6 billion? — and 95% of candidates still fail in clinical trials.

For rare diseases, the problem is even more acute: patient populations are too small to attract commercial investment, academic labs lack the computational infrastructure to run systematic screening, and most targets have never been computationally validated.

DrugSynthAI consists of two layers: the platform? (governance, AI optimization, audit) and the domain applicationDomain ApplicationA therapeutic-area-specific configuration running on DrugSynthAI. A therapeutic-area-specific configuration running on DrugSynthAI. (disease-specific targets, scoring weights, constraints). The mitochondrial therapeutics program is the first domain applicationDomain ApplicationA therapeutic-area-specific configuration running on DrugSynthAI. A therapeutic-area-specific configuration running on DrugSynthAI. built on the platform.

Two-Layer Architecture

S00–S10 Stage Gate Pipeline

Every campaignCampaignA complete drug discovery run from S00 (init) through S10 (IVVP packaging). Each campaign has a unique CMP-{8hex} ID and produces a governed artifact bundle. flows through 11 stages, each with defined entry criteria, exit criteria, and kill conditions. No candidate advances without a valid StageDecisionRecordStageDecisionRecordA cryptographically signed record that documents every stage transition in the pipeline. No agent can advance to the next stage without an SDR approved by the RunController. (SDR).

7-Component Reward Function

Stage S09–S10 score every candidate using a weighted composite of seven components: binding affinity?, ADMET Tier A rate?, rescue mechanism compatibility?, structural novelty?, synthetic accessibility?, network perturbationNetwork PerturbationA measure of how strongly a drug candidate disrupts the disease-relevant signaling network. Higher perturbation = greater therapeutic potential. score, and ΔΨm mitochondrial accumulation?. Weights are calibrated for mitochondrial therapeutics and are IP-protected.

Click any page card to expand its full reference guide.

Walk through a complete drug discovery program from disease selection to synthesis priority list. This example targets Alzheimer's disease.

All tiers include full governance audit trail. Patent Pending. For Enterprise pricing: jyborges@bu.edu — FxMEDUS LLC, Boston, MA

Patient-Specific Generative Therapeutics extends DrugSynthAI from population-level drug discovery to individualized therapeutic design. Available in the Precision Medicine tier and above.

Platform Outputs

Every report, export, and data package the platform generates. Click any item for a detailed explanation of its contents and use case.

These endpoints return real-time data and are consumed by the platform UI. They can also be accessed programmatically for integration with external tools.

DrugSynthAI is the reference implementation of AIDD-GOV (AI Drug Discovery Governance), the first open governance standard for AI-driven drug discovery pipelines. The standard defines machine-readable schemas for every governance artifact the platform produces.

What AIDD-GOV Requires

Blockchain Anchoring (Roadmap)

AIDD-GOV governance reports contain SHA-256 checksums for every StageDecisionRecord. These checksums form a verifiable chain of provenance from campaign initialization through final candidate nomination. A planned integration will anchor these checksum chains to a public blockchain, providing tamper-proof, third-party-verifiable evidence that no governance record was modified after issuance.

How it works: At campaign completion, the platform computes a Merkle root over all SDR checksums in the campaign ledger. This single hash is written to a public blockchain (target: Ethereum or Polygon). Anyone with the governance report can independently verify that every SDR matches the anchored Merkle root — without revealing any proprietary compound data or reward weights.

What this enables: A regulator, CRO partner, or disease foundation can verify that the governance trail presented to them is identical to the trail that existed at campaign completion. No trust in the platform operator is required — the blockchain provides the proof.

Regulatory Alignment

AIDD-GOV aligns with FDA Draft Guidance on AI/ML in Drug Development (2023), ICH Q8(R2) Quality-by-Design principles, and 21 CFR Part 11 electronic records requirements. The open specification provides a framework for demonstrating algorithmic accountability when AI methods are used in drug candidate nomination.

DrugSynthAI includes eight governance-exclusive features that are structurally impossible for competitors to replicate. Each feature requires the governed pipeline infrastructure (immutable StageDecisionRecords, append-only audit trails, pre-declared constraints) as a prerequisite. Platforms that operate as black boxes cannot build these capabilities without first retrofitting their entire architecture.

1. Compound Provenance Waterfall

What it proves: Any compound is traceable from gene selection to nomination in one visual.

Select any drug candidate and see its complete governed history: which gene was selected and why, which binding pocket was analyzed, which fragment was chosen, how the analog was expanded, what the docking result was, how ADMET profiling classified it, whether it passed novelty screening, how RL optimization ranked it, and whether it was nominated for experimental validation. Every checkpoint shows the corresponding StageDecisionRecord with its SHA-256 integrity checksum.

Endpoint: GET /api/provenance/{compound_id}

Access: Drug Candidates page → select compound → Provenance button

2. AIDD-GOV Self-Assessment Report

What it proves: Machine-readable compliance proof, third-party verifiable.

Generates a structured compliance report mapping each of the 10 AIDD-GOV schemas to its implementation artifact in the platform. A regulator, CRO partner, or disease foundation can independently verify that the platform implements every required schema at the claimed conformance level. The report is exportable as JSON for automated compliance checking.

Endpoint: GET /api/aidd-gov/self-assessment

Access: Dashboard → AIDD-GOV Compliance card → click for full report

3. Regulatory Readiness Score

What it proves: Computed percentage of the FDA pre-IND package already generated.

Maps platform artifacts to FDA pre-IND submission requirements and computes a readiness percentage. For each requirement (target identification, lead characterization, ADMET summary, synthesis feasibility, novelty assessment, clinical protocol), the score shows whether the governed pipeline has already produced the necessary documentation. Items requiring wet lab validation (GLP toxicology, clinical pharmacology) are marked as pending with clear next steps.

Endpoint: GET /api/regulatory/readiness/{campaign_id}

Access: Dashboard → Regulatory Readiness gauge → click for full FDA checklist

4. Campaign Replay (DVR)

What it proves: Step-by-step playback of a governed campaign.

Replays an entire campaign execution from S00 initialization through S10 nomination, showing stage-by-stage timing, inputs, outputs, and SDR decisions. Functions like a DVR for drug discovery: pause at any stage, inspect the governance decision, examine the data that was evaluated, and see the exact criteria that were applied. The RL convergence trace animates as the replay progresses, showing how candidate scores improved across optimization iterations.

Endpoint: GET /api/replay/{campaign_id}

Access: Dashboard → Campaign card → Replay button

5. Governance Diff Engine

What it proves: Side-by-side comparison of two campaign governance trails.

Compares two campaigns stage by stage: how target selection differed, whether constraint policies changed, how RL convergence rates compared, and where scoring outcomes diverged. Color-coded diffs highlight improvements (green), regressions (red), and unchanged parameters (gray). Enables systematic analysis of how different configurations affect pipeline outcomes while maintaining governance traceability for both campaigns.

Endpoint: GET /api/governance/diff?campaign_a={id}&campaign_b={id}

Access: Analytics → Governance Analytics → Campaign Diff selector

6. Kill Switch Dashboard

What it proves: Real-time visualization of safety monitoring.

Displays all kill switch conditions defined for a campaign with their current values, thresholds, and headroom. Each kill switch (reward collapse, constraint violation rate, convergence failure, toxicity rate, resource exhaustion) shows a gauge meter indicating how far the current state is from the trigger threshold. Evaluation counts and trigger history are displayed for full audit transparency. A green shield badge confirms when all switches are in safe state.

Endpoint: GET /api/kill-switches/{campaign_id}

Access: Dashboard → Kill Switch Status card → click for full panel

7. Constraint Sensitivity Analysis

What it proves: Constraints are meaningful, not arbitrary.

For each hard constraint in the constraint policy, simulates what happens if the threshold is tightened by 10%, 25%, and 50%. Shows how many candidates would survive at each threshold level. This proves that constraints are calibrated to the actual compound library rather than set at arbitrary default values. A constraint where tightening by 50% eliminates 40% of candidates is meaningfully engaged with the data. A constraint where even 50% tightening changes nothing indicates headroom in the current library.

Endpoint: GET /api/sensitivity/{campaign_id}

Access: Analytics → Governance Analytics → Constraint Sensitivity chart

8. Cross-Campaign Meta-Analysis

What it proves: Aggregate learning across campaigns without exposing compound structures.

Aggregates statistical patterns across all governed campaigns: ADMET score distributions, molecular weight histograms, binding affinity ranges, RL convergence rates, constraint violation counts, and governance compliance rates. No individual compound data is exposed. This enables the platform to demonstrate learning effects (do later campaigns converge faster?) and quality trends (are ADMET profiles improving?) while preserving compound confidentiality for each campaign commissioner.

Endpoint: GET /api/meta-analysis

Access: Dashboard → Meta-Analysis card → Analytics → Cross-Campaign section

Why Competitors Cannot Build These

When you select a disease in the Discovery Lab, DrugSynthAI automatically fetches epidemiological and clinical evidence from four public databases in parallel, grounding your drug discovery campaign in real patient data before a single compound is generated.

Live Public Sources

Planned Registry Sources (Pending Data Sharing)

Privacy and Data Safety

All evidence displayed in the Real-World Evidence panel is aggregate, population-level data only. No patient-identifiable information is fetched, stored, or displayed. ClinVar variant data is de-identified by definition. Orphanet and GARD data are fully public.

Evidence Panel (Discovery Lab)

When you select a disease in Step 1 of the Discovery Lab, the RWE panel automatically appears below the disease grid showing: prevalence class, total pathogenic variant count across your selected genes (ClinVar), HPO phenotype count, and inheritance pattern. The top 12 HPO phenotype terms are displayed as clickable chips. All four sources query in parallel — typical response time is under 3 seconds.

GOV-FM: Governance Foundation Model

GOV-FM is a self-improving governance quality scoring system built into DrugSynthAI. Every completed campaign generates a GovernanceCampaignRecord — a structured snapshot of governance metadata. The record is scored across 5 dimensions, gaps are identified, and recommendations are stored as governance memory for the next campaign. The loop compounds over time.

What GOV-FM Evaluates (5 Dimensions)

Risk Thresholds

How Scoring Works — Phase 1 (Rule-Based)

GOV-FM Phase 1 uses a deterministic rule engine. Each dimension is scored 0–1 using conditional logic derived from the AIDD Governance Standards v1.0 and 21 CFR Part 11. Gaps are identified as specific violations; recommendations are actionable remediation steps. The composite score is the weighted average of all 5 dimension scores.

Phase 2 (planned): LLM-based semantic scoring of governance narratives, cross-campaign pattern learning, and automated gap prioritization using campaign outcome data.

The Self-Improving Loop

Governance Memory

The file data/gov_fm_training/governance memory persists across campaigns and contains:

• last_campaign — ID of the most recently scored campaign
• last_score — composite governance score (0.0–1.0)
• trend — FIRST_CAMPAIGN | IMPROVING | STABLE | DECLINING
• active_recommendations — actionable steps from the last scoring run
• total_campaigns_scored — how many campaigns have been scored to date

Improvement Trend Tracking

Data Privacy

GOvernanceCampaignRecord contains only governance metadata — never compound SMILES, molecular weights, patient data, or optimization reward weights. Training data is safe for institutional sharing under standard data use agreements.

Training Data Location

All training data accumulates at data/gov_fm_training/:

• {campaign_id}.yaml — GovernanceCampaignRecord (training input)
• {campaign_id} score data — GovernanceScore (training label)
• governance memory — active recommendations for next campaign

Dashboard Integration

The GOV-FM Score card on the Dashboard displays the composite score, risk level, trend indicator, 5-dimension progress bars, detected gaps, and the top recommendation for any selected campaign. Select a campaign from the dropdown to load its score in real time via GET /api/governance/score/{campaign_id}.

DrugSynthAI closes the gap between in silico optimization and the first physical molecule. After campaign completion, any nominated compound can be routed through the synthesis planning pipeline to produce a complete CRO dispatch package — ready to send to a contract research organization for synthesis and biological testing.

ASKCOS Retrosynthesis (MIT)

ASKCOS routes include: step-by-step retrosynthetic disconnections, reaction template scores (success probability), literature precedent counts per reaction, commercially available starting materials (eMolecules 7M+ compounds, ZINC15 230M+ compounds), and estimated synthesis difficulty (LOW / MODERATE / HIGH / VERY_HIGH). Availability rate and price estimates are included in every synthesis route response.

CRO Brief Generator

The CRO Brief Generator produces a complete dispatch package from campaign data. A single API call generates everything a contract research organization needs to quote and execute.

Assay Protocol Library

The library (registries/assay protocol library) contains pre-defined protocols for 5 mitochondrial disease targets, plus three standard panels applied to all compounds.

How to Use

API Endpoints Phase A

Phase B closes the gap between optimized compounds and real patients. Every gene click in the Discovery Lab now fires the Patient Intelligence panel, connecting genotype data to candidate compounds, population estimates, and companion diagnostic specifications.

Genotype-to-Compound Matching

Given a patient's genetic variant, the platform automatically selects the correct rescue compound using a 4-tier matching hierarchy.

Population Estimation

For each variant, the platform estimates the global patient population using gnomAD allele frequencies, Orphanet prevalence data, and Hardy-Weinberg equilibrium calculations. The pre-computed registry covers 14 variants representing 54,204 patients globally.

Companion Diagnostic Specification

Each campaign generates a 3-tier companion diagnostic panel specification, defining the genetic test needed to identify eligible patients.

Discovery Lab Integration

When you click a gene in the Discovery Lab, the Patient Intelligence panel fires automatically, displaying: matched compound candidates with rescue direction, estimated patient population for the selected variant, and the recommended diagnostic tier. This data flows directly into the CRO Brief and foundation proposal generator.

API Endpoints Phase B

Molecular Dynamics & Free Energy Perturbation

Phase C adds physics-based validation to every shortlisted compound. Molecular dynamics (MD) simulates how each compound behaves in a solvated protein environment over time, while free energy perturbation (FEP) estimates binding affinity as ΔΔG = ΔG_complex − ΔG_solvent.

MD Simulation Pipeline

Stability Scoring

Each compound receives a stability score (0.0–1.0) based on three trajectory metrics:

Classification: STABLE (RMSD < 0.20 nm, H-bond > 60%) · FLEXIBLE (RMSD < 0.35 nm) · UNSTABLE (RMSD ≥ 0.35 nm).

Free Energy Perturbation

FEP uses alchemical λ-windows (12 equidistant steps, λ = 0 → 1) to perturb each compound between its bound (complex) and unbound (solvent) states. ΔΔG is computed using a descriptor-based proxy with BAR/TI scaffolding for full alchemical activation:

• FAVORABLE: ΔΔG < −1.0 kcal/mol (predicted binder)
• NEUTRAL: −1.0 ≤ ΔΔG ≤ +1.0 kcal/mol
• UNFAVORABLE: ΔΔG > +1.0 kcal/mol (predicted non-binder)

GPU Compute Platforms

The platform is auto-detected at runtime (CUDA > OpenCL > CPU > Reference). For cloud GPU runs, use the Export Colab button in the MD panel to download a pre-wired Colab notebook configured for T4/A100.

MD Panel — Discovery Lab

The MD Validation panel appears in the Discovery Lab (right column). For each shortlisted compound:

1. Click Run MD to submit a preparation + simulation job.
2. Results appear as a stability badge: STABLE / FLEXIBLE / UNSTABLE with RMSD, H-bond occupancy, and score.
3. Click FEP Compare to rank shortlisted compounds by ΔΔG. The table sorts with FAVORABLE (green) on top.
4. Click Export Colab to download the Jupyter notebook for the selected compound. Open in Google Colab and select Runtime → T4 GPU for production-quality MD.

API Endpoints

Dashboard MD Validation Badge

The Campaign Dashboard shows a MD Validation badge summarizing the count of STABLE / FLEXIBLE / UNSTABLE compounds for the active campaign. Click the badge to expand the full MD results table with per-compound RMSD, stability score, and FEP ΔΔG. Compounds rated STABLE + FAVORABLE are highlighted as Stage Gate ready.

Delivery Engineering

Phase D closes the gap between optimized compounds and the target compartment. Nine of 25 mitochondrial targets sit inside the mitochondrial matrix behind a double membrane and a −180 mV electrochemical gradient. Every compound must GET THERE. Phase D solves the delivery problem computationally before synthesis.

Decision Tree: Target Location → Delivery Strategy

TPP+ Conjugate Designer

Triphenylphosphonium (TPP+) cations accumulate 100–1000× in the mitochondrial matrix driven by the electrochemical gradient (ΔΨm = −180 mV). The Nernst equation gives:

Accumulation = 10^{(z·F·ΔΨ / 2.303RT)} ≈ 850× at 37°C

TPP+ adds ~263 Da and +4.5 logP units. Alkyl-C6-ester is the default recommendation: optimal membrane partitioning, matrix esterase cleavage at pH 8.0 (t½ ~2 hours), releasing the unmodified parent compound.

Prodrug Designer — 7 Strategies

Formulation Specifications — 5 Routes

LNP composition: ionizable lipid 50% (ALC-0315 or DLin-MC3-DMA) + DSPC 10% + cholesterol 38.5% + PEG-DMG 1.5%. Near-neutral zeta potential (−5 mV) minimizes non-specific protein adsorption while maintaining BBB transcytosis.

CRO Brief Integration

The CRO Brief generator now includes a Delivery & Formulation section per compound. Instead of:

"Synthesize [COMPOUND] (MW 380, NDUFV1 stabilizer)"

The brief now reads:

"Synthesize [COMPOUND]-TPP-C6 (TPP+-alkyl-C6-ester conjugate, MW 690).
Linker: hexyl chain with ester cleavage site.
Predicted matrix accumulation: 850× at ΔΨm = −180 mV.
Ester hydrolysis t½: 2 hours (matrix esterases, pH 8.0).
Formulation: sterile injectable, 10 mg/mL in normal saline.
Excipients: NaCl 0.9%, Polysorbate 80 0.1%.
Sterilization: 0.22 μm filtration + aseptic fill.
Container: Type I borosilicate glass vial."

Delivery Design API

Dashboard Delivery Design Badge

The Campaign Dashboard Delivery Design badge shows the distribution across primary candidates:

• Oral: 5 compounds — cytoplasmic/OMM targets, no mitochondrial targeting required
• Intranasal: 1 compound — CNS-penetrant target requiring BBB bypass
• IV + TPP+: 2 compounds — matrix-targeted FAD/FMN binding sites
• IV + Prodrug: 1 compound — methyl ester strategy for TPSA reduction
• IV minimal: 1 compound — low TPSA, partial passive matrix uptake

Phase E adds a regulatory document engine that transforms campaign data into FDA-formatted document drafts. A researcher clicks Generate Full IND Package for any compound and receives a complete draft package — Pre-IND meeting request, CMC modules, and nonclinical pharmacology summary — in seconds. Regulatory consultant review time: weeks, not months.

FDA IND Structure (21 CFR 312.23)

Pre-IND Meeting Request (21 CFR 312.82)

The Pre-IND generator produces three deliverables per compound:

1. Meeting request letter — FDA Type B format with sponsor details, indication, patient population, and orphan drug eligibility
2. Briefing document — 9-section structured document: executive summary, product information, disease background, nonclinical summary, proposed program, CMC summary, clinical plan, regulatory strategy, proposed questions
3. Proposed questions — 7–9 FDA-formatted questions covering nonclinical adequacy, CMC, Phase I design, safety pharmacology, adaptive trial use, and orphan designation

CMC Documentation (ICH CTD Module 3.2.S/3.2.P)

SMILES redaction policy: All regulatory documents display [REDACTED — disclosed under NDA only] in place of compound structures. The patent-pending IP is never exposed in document exports.

Nonclinical Pharmacology (Modules 2.4 + 2.6)

The nonclinical summary engine drafts:

• Primary pharmacology — target binding assay design, functional assay selection (target-specific: Complex I OCR for NDUFV1, mitophagy flux for PINK1, ARE-luciferase for NFE2L2…), in silico docking ΔG and composite score
• Secondary pharmacology — hERG patch clamp, CYP inhibition (3A4/2D6/2C9), Eurofins SafetyScreen44, kinase selectivity
• Safety pharmacology (ICH S7A) — cardiovascular, CNS (modified Irwin), respiratory
• Proposed toxicology program (5 studies) — single-dose MTD, 14-day rat, 14-day dog, Ames, in vivo micronucleus
• Proposed PK program (4 studies) — rat/dog single-dose PK, in vitro metabolic stability, plasma protein binding

Regulatory Designation Eligibility

API Endpoints