ICH M7 Workflow Guide
From SMILES to expert-review-ready mutagenicity table — what ICH M7 (R2) requires and how an MCP-supported workflow delivers it in under 90 seconds per batch, without compliance compromises.
1. What ICH M7 (R2) Requires
ICH M7 (R2) is the international guideline for assessing and controlling DNA-reactive mutagenic impurities in pharmaceuticals. It defines five classes (1: known mutagens with cancer evidence, 2: known mutagens without cancer data, 3: alerting structures with insufficient data, 4: alerting structures with negative test data, 5: no alerts). Each synthetic impurity must be assessed using two complementary QSAR methods: an expert-rule-based system (e.g., Derek Nexus) plus a statistical system (e.g., Sarah Nexus, Leadscope, Bayesian-based models).
2. Classic Workflow (Excel + Manual Statistics)
Typical med-chem team workflow without MCP: manually extract SMILES from synthesis notes, paste into two separate web interfaces, export outputs as CSV, merge in Excel, walk through ICH M7 decision tree manually for classification, check domain applicability, prepare expert review table. Effort: 4 hours for 50 structures. Reproducibility after 18 months: weak (tool versions change, input strings drift, reconstructing results becomes detective work).
3. MCP Workflow with covatox_assess_ichm7_batch
Using the covatox tool covatox_assess_ichm7_batch, the agent passes a list of SMILES strings as input. The tool runs two deterministic QSAR models (rule-based + statistical), combines them per ICH M7 classes 1–5, documents domain applicability per prediction, and returns a structured JSON table. Output is version-pinned: same inputs always yield same outputs, reproducible at any time.
4. Example Call
# MCP call via Claude Desktop or any MCP client
{
"tool": "covatox_assess_ichm7_batch",
"input": {
"smiles_list": ["CC(=O)Nc1ccc(O)cc1", "Nc1ccc(N)cc1", ...],
"include_domain_check": true
}
}5. Audit Trail
Every call is logged with: tool version, model version (rule-based + statistical), input hash (SHA-256 over the SMILES list), output hash, timestamp (UTC), authenticated user. During an EMA or FDA inspection, any individual assessment can be re-instantiated years later — the machine produced exactly this output at this time with these model versions.
6. What Stays Human
Class-3 substances (alerts present, insufficient data) still require expert review by a qualified toxicologist. Class-1 (known mutagens) is flagged but not auto-blocked — the discovery decision stays with the med-chem lead. Class-5 (no alert) auto-clears with audit trail. The machine offloads, the human decides.
7. GAMP 5 Software Category 4 Argumentation
covatox_assess_ichm7_batch falls into GAMP 5 Software Category 4 (Configured Product), not Category 5 (Custom Software). Meaning: validation effort focuses on URS, FS, configuration, risk assessment, and user acceptance testing — not source-code reviews and code-coverage metrics. CovaSyn provides a prepared validation pack in the Enterprise subscription with URS templates, risk assessment matrix, and IQ/OQ/PQ templates that the QA lead adapts and signs. Validation effort typically drops from 4–8 weeks to 1–2 weeks.
8. Caveats
The tool does not replace a qualified toxicologist. It makes the pre-work auditable and reduces manual effort. The final ICH M7 classification and submission strategy remain the responsibility of QA and reg affairs. Domain applicability flags must be taken seriously — structures outside the trained model space require explicit expert assessment.