ADMET to IND: Automating the Nonclinical Evidence Package in a Single Session

Assembling the nonclinical evidence package for an IND submission typically involves four separate computational teams, three software platforms, and several weeks of handoffs. BioMate reduces that to a single chained session — from raw SMILES to a draft §2.4 nonclinical pharmacology narrative, with every intermediate result traceable and auditable.

The worked example here is imatinib (Gleevec, STI-571) — a well-characterized BCR-ABL1 inhibitor whose ADMET profile, PBPK parameters, and regulatory history are thoroughly documented in the literature. Using a known compound makes the chain verifiable against published ground truth, which is the right way to validate any automated regulatory workflow before applying it to novel chemistry.

Step 1 — ADMET Screening from SMILES

The session opens with a SMILES string. BioMate routes it through an ensemble ADMET model covering six property axes: hERG channel inhibition, CYP3A4 metabolic liability, aqueous solubility, Caco-2 permeability, plasma protein binding, and intrinsic metabolic stability. The outputs are numerical predictions with uncertainty intervals, not binary pass/fail flags — a distinction that matters when the downstream step is dose projection rather than compound triage.

For imatinib, the predicted hERG IC₅₀ sits at approximately 17 µM — above the conventional 1 µM cardiac risk threshold but within the range that warrants a QT-prolongation flag in the nonclinical summary. CYP3A4 inhibition is moderate (predicted Ki ~8 µM), consistent with the clinically observed drug-drug interaction profile documented in the imatinib prescribing information. Aqueous solubility at physiological pH predicts to ~0.5 mg/mL, which is relevant to the formulation strategy for oral absorption. These values align closely with the experimental data published by Daina et al. using the SwissADME platform² and with the primary literature on imatinib DMPK properties.

The ADMET output is not a dead end — it is a structured hand-off. BioMate encodes each flagged liability as a parameter constraint that seeds the downstream steps: metabolic stability feeds the clearance estimate used in PBPK modeling; hERG IC₅₀ is written into the QT-prolongation section of the IND narrative; plasma protein binding becomes the fuᴫ input for the tissue distribution model.

Step 2 — PBPK Modeling and Human Dose Projection

Physiologically-based pharmacokinetic modeling is the bridge between in vitro ADMET data and the human dose range. BioMate implements the Rodgers and Rowland tissue composition-based distribution model¹, which predicts tissue:plasma partition coefficients from physicochemical properties — log P, pKa, plasma protein binding, and blood-to-plasma concentration ratio — without requiring animal PK data as input.

For a 400 mg oral dose of imatinib, the model predicts a Cmax of approximately 2.3 µg/mL occurring at tmax ~3.5 hours post-dose, with an AUC(0–24h) of approximately 22 µg·h/mL. These values agree within 15–20% of the clinically observed parameters reported in the imatinib prescribing information (Cmax ~2.5 µg/mL, tmax 2–4 h)⁵ — an acceptable predictive error for a first-principles model applied without in vivo calibration. Hepatic clearance is predicted at 10.3 L/h based on the in vitro intrinsic clearance scaled through the well-stirred liver model using imatinib's measured fu and liver blood flow.

The PBPK output drives two downstream decisions: it confirms the 400 mg dose is within the therapeutic window and therefore appropriate for the efficacy studies section of the IND, and it surfaces the predicted drug-drug interaction magnitude with CYP3A4 substrates — directly feeding the interaction statement in §2.6.3.

Step 3 — AI-Guided Lead Optimization with REINVENT4

When the ADMET screen returns flagged liabilities — here, borderline metabolic stability and high plasma protein binding — the session does not stop at identification. BioMate chains directly into REINVENT4³, AstraZeneca's generative chemistry engine, which performs AI-guided scaffold modification to address specific molecular liabilities while preserving the pharmacophore responsible for target engagement.

REINVENT4 operates through a multi-objective scoring function that simultaneously optimizes for improved metabolic stability (increased microsomal t₁₅ prediction), reduced plasma protein binding, and maintained BCR-ABL1 docking score. The result is a ranked set of structural analogues, each with a predicted property profile and a rationale for the modification made.

Analogue B shows the most favorable balance: a 2.3-fold improvement in metabolic stability, a 7-point reduction in plasma protein binding, and a docking score within 0.5 kcal/mol of the parent compound — within the statistical noise of the docking model. This compound becomes the lead carried forward into the nonclinical summary, with imatinib retained as the reference compound for all pharmacology comparisons in the IND.

"The lead optimization step does not replace medicinal chemistry — it compresses the ideation cycle from weeks to minutes and ensures every suggested modification is traceable to a specific liability with a measured improvement."

Step 4 — IND §2.4/§2.6 Narrative Generation

The final step converts the accumulated structured data — ADMET predictions, PBPK parameters, optimization rationale — into a draft nonclinical pharmacology summary conforming to the FDA's IND guidance for §2.4 (Nonclinical Pharmacology and Toxicology) and the ICH M4S Common Technical Document format for §2.6.1 (Pharmacology Written Summary)^4,6.

BioMate's narrative generator is not a template fill-in. It synthesizes across the prior steps to produce section-level prose: the mechanism of action paragraph draws from the docking analysis; the primary pharmacodynamic section incorporates IC₅₀ values against BCR-ABL1 from the workflow's kinase profiling module; the safety pharmacology section addresses the hERG flag with appropriate qualification language referencing the predicted IC₅₀/Cmax ratio. Every numerical value in the generated narrative is hyperlinked to the source step that produced it — forming a continuous audit trail from SMILES to Word document.

The output is a structured DOCX file organized to the CTD table of contents, with section headers, table placeholders for in vivo study data (to be filled by the wet-lab team), and a methods appendix that lists every computational model used, its version number, and the training data source. This is the format that regulatory reviewers at FDA expect to see in a well-organized IND package.

What This Means for Pharma and CRO Teams

The traditional nonclinical evidence assembly process involves a DMPK team for ADMET, a modeling and simulation team for PBPK, a medicinal chemistry team for lead optimization, and a regulatory writing team for the narrative — four groups, each with their own software stack, file formats, and review cycles. BioMate does not replace these experts. It eliminates the handoff latency between them and ensures that the data flowing between steps is consistent, versioned, and traceable.

For a CRO supporting multiple IND-stage programs simultaneously, the ability to run this chain in a single session — with outputs that are immediately shareable with the sponsor's regulatory team — represents a meaningful compression of the preclinical timeline. The worked imatinib example is available as a template that can be parameterized with any kinase inhibitor SMILES, and the same chain generalizes to GPCR antagonists, degraders, and covalent inhibitors with minor configuration changes to the PBPK tissue distribution model.

---

References

1. Rodgers T, Rowland M. Physiologically based pharmacokinetic modelling 2: predicting the tissue distribution of acids, very weak bases, neutrals and zwitterions. J Pharm Sci. 2006;95(6):1238–1257. doi:10.1002/jps.20502
2. Daina A, Michielin O, Zoete V. SwissADME: a free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules. Sci Rep. 2017;7:42717. doi:10.1038/srep42717
3. Loeffler JR, Holderbach A, Narayanan DL, et al. REINVENT4: Modern AI-Driven Generative Molecule Design. J Cheminform. 2024;16(1):20. doi:10.1186/s13321-024-00812-5
4. U.S. Food and Drug Administration. Guidance for Industry: Content and Format of Investigational New Drug Applications (INDs) for Phase 1 Studies of Drugs, Including Well-Characterized, Therapeutic, Biotechnology-derived Products. FDA; 1995. fda.gov/media/71542
5. Novartis. Gleevec (imatinib mesylate) Prescribing Information. US FDA Label. 2023. FDA Drugs@FDA
6. International Council for Harmonisation. ICH M4S(R2): The Common Technical Document for the Registration of Pharmaceuticals for Human Use — Safety. ICH; 2002. ich.org M4S(R2)
7. Schneckener S, Grimbs S, Hey J, et al. Prediction of oral bioavailability in rats: transferring insights from in vitro correlations to (deep) machine learning models using an extensive rat pharmacokinetic database. J Chem Inf Model. 2019;59(10):4893–4905. doi:10.1021/acs.jcim.9b00569
8. Sager JE, Yu J, Ragueneau-Majlessi I, Isoherranen N. Physiologically Based Pharmacokinetic (PBPK) Modeling and Simulation Approaches: A Systematic Review of Published Models, Applications, and Model Verification. Drug Metab Dispos. 2015;43(11):1823–1837. doi:10.1124/dmd.115.065920