IND-Ready in One Session: ADMET → PBPK → §2.6 Narrative → Signed Audit Trail

The translation from a compound's SMILES string to an IND-ready pharmacokinetic package involves four sequential steps that today require different specialists: a computational chemist for ADMET, a PK modeler for PBPK, a medical writer for the regulatory narrative, and a quality system for the audit trail. BioMate chains all four steps into a single session, with each step's outputs feeding directly into the next.

Why the chain is sequential — and why that matters

The four steps are causally ordered. PBPK simulation requires the compound's molecular properties (MW, logP, pKa, fu,p) — the same values computed during ADMET screening. The §2.6.1 narrative must cite specific NCA values from the PBPK output (Cmax, AUC, t½, CL/F). The audit manifest must hash every output file from all three preceding steps, including the narrative document itself. None of these steps can proceed without the preceding output, and each step is traditionally owned by a different specialist.

For small biotech teams without a dedicated regulatory bioinformatics function, this handoff structure means the IND preparation process spans weeks of calendar time even when the actual computational work takes hours. BioMate compresses that to a single session.

The four-step chain

SMILES input (3 candidate compounds)
     │
     ▼
predict_admet_properties
├── hERG IC50: Compound 1 <1 µM (PASS), Compound 2 >10 µM (FLAG)
├── BBB penetration: Compound 1 CNS-accessible (log BB = 0.3)
├── Cmax (oral 325 mg): Compound 1 ~ 28–32 µg/mL predicted
├── DILI liability: Compound 3 HIGH (FLAG → excluded)
├── Ames mutagenicity: Compounds 1, 2: NEGATIVE
└── Lipinski Ro5: Compound 1 fully compliant
     │
     │  → admet_results.csv (traffic-light per compound per property)
     │    PROCEED: Compound 1
     ▼
pbpk_simulation (Compound 1, oral 325 mg, 70 kg adult)
├── 11-compartment ODE (SciPy solve_ivp LSODA)
│     blood → liver → kidney → gut → lung →
│     heart → brain → muscle → fat → skin → rest
├── Oral absorption: gut lumen → portal blood (ka from logP + MW)
├── Hepatic first-pass extraction (Qh, CLint estimation)
└── NCA: AUC₀₋₂₄, Cmax, Tmax, t½, CL/F, Vd
     │
     │  → nca_table.csv
     │    pk_curve.png (plasma + tissue overlay)
     ▼
BioMate AI — §2.6.1 pharmacokinetic narrative
├── "Following oral administration of 325 mg, Compound 1
│    reached peak plasma concentrations (Cmax) of approximately
│    30 µg/mL at approximately 1 hour post-dose..."
├── Formatted for CTD module 2.6.4 written summary (ICH M4)
└── 400–600 words · clinical regulatory language
     │
     │  → ind_section_2_6_1.docx
     ▼
run_audit_export
├── SHA256 hash: admet_results.csv  → a3f8d1...
├── SHA256 hash: nca_table.csv      → c92b7e...
├── SHA256 hash: ind_section_2_6_1.docx → f1a4c9...
├── claims_index.json:
│     {"Cmax ~30 µg/mL": "nca_table.csv:row3:col_cmax"}
│     {"t½ ~4.2 h": "nca_table.csv:row3:col_t_half"}
└── manifest.json (timestamp · run ID · all hashes · signed)

Step 1 — ADMET screening

BioMate's ADMET predictor accepts SMILES strings inline in the chat. No file upload, no format conversion. For each compound it predicts:

• hERG IC50: cardiotoxicity flag if <1 µM
• BBB penetration score: logBB estimate (CNS accessibility)
• Cmax prediction: oral absorption estimate from logP, MW, pKa
• DILI liability score: hepatotoxicity probability
• Ames mutagenicity: genotoxicity binary prediction
• Lipinski Rule of 5: oral bioavailability filter

The output is a traffic-light table: green (PROCEED), yellow (MONITOR), red (FLAG). Compounds with red flags in hERG or DILI are excluded from the PBPK step by default, though the user can override with a clinical rationale.

Step 2 — PBPK simulation

For each compound that passes ADMET, BioMate runs an 11-compartment physiologically-based pharmacokinetic model. The compartments are blood, liver, kidney, gut, lung, heart, brain, muscle, fat, skin, and rest-of-body, with blood flows matched to the International Commission on Radiological Protection (ICRP) reference person.

Molecular parameters (volume of distribution, protein binding fu,p, intrinsic clearance CLint) are estimated from SMILES-derived physicochemical properties using Obach 1999 prediction equations, or taken from user-provided values when available. The ODE system is integrated with SciPy solve_ivp LSODA for stiff systems.

Output is a full NCA report:

The aspirin retrospective serves as the canonical validation case. Published pharmacokinetic data from Needs & Brooks 1985 (Clin Pharmacokinet 10:164–177) reports Cmax of 25–35 µg/mL across healthy volunteers after 325 mg oral dose. BioMate's PBPK prediction falls within this range for all five NCA parameters, without access to any measured data — the model is fully prospective from molecular properties.

Step 3 — §2.6.1 pharmacokinetic narrative

After the PBPK run completes, BioMate AI generates the §2.6.1 Summary of Pharmacokinetics narrative from the NCA output. The narrative is formatted for CTD module 2.6.4, following ICH M4 structure guidance for the written summary of pharmacokinetics studies. The language is clinical and regulatory-grade — not a technical report summary, but the actual section text that would appear in the IND filing.

"Following oral administration of 325 mg, the compound reached peak plasma concentrations of approximately 30 µg/mL at approximately 1 hour post-dose. The area under the plasma concentration-time curve from 0 to 24 hours (AUC₀₋₂₄) was approximately 210 µg·h/mL. The apparent terminal half-life was approximately 4.2 hours, consistent with the expected pharmacokinetic profile for this compound class..."

The narrative cites specific values from the PBPK output, with each claim traceable to a row in nca_table.csv via the claims_index.json generated in the next step.

Step 4 — Audit trail export

run_audit_export produces three files:

• manifest.json: SHA256 hash of every output file (admet_results.csv, nca_table.csv, pk_curve.png, ind_section_2_6_1.docx), plus run ID, timestamp, platform version, and model version. This is the chain of custody document.
• claims_index.json: Maps every quantitative claim in the narrative to its source data. For example: {"Cmax ~30 µg/mL": "nca_table.csv:row3:col_cmax"}. An auditor can open nca_table.csv, go to row 3, column cmax, and verify the number without any interpretation.
• cite_as.txt: Auto-generated citation strings for each output file (e.g., "BioMate PBPK simulation, run ID abc123, 2026-07-20, version 2.4.1") for use in IND section references.

From a CRO perspective

The audit manifest makes computational pharmacokinetic data IND-citable without requiring a dedicated regulatory bioinformatics function. For small CRO clients — particularly those running early Phase 1 INDs with limited regulatory infrastructure — the ability to generate a complete, traceable PK package from SMILES to signed manifest in a single session changes the economics of the regulatory preparation step entirely.

Hesperos, whose organ-on-chip technology produces CTD-quality PK data from microfluidic tissue models, provides the wet-lab data layer. BioMate provides the computational layer and the audit infrastructure. Together they produce a complete IND pharmacokinetics package from microphysiological system data — the same two-page PK section that would otherwise require a CRO pharmacokineticist, a medical writer, and a QA review cycle.

References: ICH M4 Common Technical Document (CTD); Needs CJ & Brooks PM, Clin Pharmacokinet 1985;10(2):164-177; FDA Guidance: Computerized Systems Used in Clinical Investigations (2007); Rowland M & Tozer TN, Clinical Pharmacokinetics and Pharmacodynamics, 4th ed.