In-silico toxicology & chemical safety assessment (QSAR, read-across for REACH/CLP & cosmetics)

Under REACH, cosmetics law and ICH M7, companies must show that substances are safe, yet new animal testing is restricted or banned. Regulatory toxicologists therefore use (Q)SAR models, read-across from similar substances and other new approach methodologies (NAMs) to fill data gaps. AI speeds up analogue search, model prediction, evidence gathering and dossier drafting, while the expert remains accountable for the weight-of-evidence conclusion.

Problem & value

• Today: a registrant with a data gap for an endpoint (e.g. skin sensitisation, mutagenicity, aquatic toxicity) either runs a test, or builds an adaptation under REACH Annex XI using (Q)SAR or read-across. Toxicologists search ECHA's registered-substance data, pick analogues by hand, document the rationale in Word and defend it against ECHA compliance checks. Read-across justifications are a frequent reason for dossier rejection or follow-up requests [2].
• Cosmetics: the EU bans animal testing for finished products and ingredients (Regulation 1223/2009), so the safety report (CPSR) of the Safety Assessor relies on NAMs, existing data and in-silico evidence [4].
• Pharma: ICH M7 expects mutagenic impurities to be assessed with two complementary (Q)SAR methodologies (one expert rule-based, one statistical) before any Ames test is considered [5].
• AI value: faster screening of large substance portfolios, systematic analogue search, consistent documentation, early detection of structural alerts in R&D (safe-and-sustainable-by-design), fewer animal tests. No sourced percentage savings are given here; measure your own baseline.

Variants

• REACH/CLP registrants (chemical industry, downstream users): gap-filling for registration, classification proposals, SVHC and substitution screening.
• Cosmetics and personal-care manufacturers: ingredient assessment, next-generation risk assessment (NGRA) workflows combining in-silico, in-vitro and exposure modelling.
• Pharma/biotech: ICH M7 impurity assessment, genotoxicity and DILI/hERG liability flags (overlaps SCI-002).
• Agro, food contact, biocides: residue, metabolite and packaging-migrant assessments.
• Regulators and public bodies: screening and prioritisation (ECHA, national authorities, EFSA).
• Product development: early toxicity flags for new molecules and materials (SCI-003, SCI-014).
• Private use: not applicable; consumer apps only provide label look-ups (not an assessment).

How it works

1. substance identity (CAS/EC/SMILES, impurities, metabolites)
2. structure curation & standardisation (salts, tautomers, UVCBs handled separately)
3. endpoint prediction (expert-rule alerts + statistical QSAR + GNN/foundation models)
4. applicability-domain & reliability check (similarity, confidence, OECD validation principles)
5. analogue search & read-across (structural + metabolic + mechanistic similarity, data matrix)
6. weight-of-evidence synthesis (in-silico + in-vitro + existing in-vivo + exposure)
7. LLM-assisted documentation (QMRF/QPRF, IUCLID text, literature extraction)
8. expert review and sign-off
9. dossier / safety report

• Models: rule-based alerts (e.g. Derek-type knowledge bases), statistical QSAR (random forest, gradient boosting on fingerprints), graph neural networks (Chemprop-style message passing) and pretrained molecular foundation models. Models must meet the five OECD principles for QSAR validation (defined endpoint, unambiguous algorithm, defined applicability domain, performance measures, mechanistic interpretation if possible) [3].
• Read-across: ECHA's Read-Across Assessment Framework (RAAF) lists the scenarios and assessment options an expert must satisfy; AI supports analogue search and data-matrix building, not the justification itself [2].
• LLM role: extracting study data from dossiers and literature, drafting rationale text, checking completeness against RAAF. Numbers always come from the database, never generated.
• Human-in-the-loop: the toxicologist approves every prediction used regulatorily, overrides domain-outliers and signs the report.

Data preparation

Needed are substance identities (CAS/EC numbers, structures as SMILES, impurity profiles), existing toxicity study results, supplier safety data sheets, dossiers (IUCLID) and exposure data. In an SMB they are scattered across supplier PDFs, SDS folders, Excel lists, the ERP material master and the regulatory affairs person's mailbox.

Typical gaps: missing or wrong CAS numbers, no structures for mixtures, polymers and UVCBs, scanned or outdated SDS, duplicate substance records, study results without test guideline or species, and unclear confidentiality of supplier data.

Preparation steps: - Build a substance master with identifiers and structures; curate and standardise structures (salts, tautomers) with RDKit (PREP-011, PREP-012). - Extract study data from SDS and reports into a table, using OCR and parsing where needed (PREP-006, DOC-014). - Pull ERP material data (PREP-005) and flag substances where no structure exists. - Create a validation set of substances with known outcomes (PREP-018) and record the source of every value. - Classify supplier data for confidentiality before it enters shared tools (PREP-024).

Tools: OECD QSAR Toolbox, IUCLID, RDKit, OpenRefine. Plan about 30 % of the total effort for data work, as the models are ready-made; correct identity data matters most. See GUIDE-003.

Implementation paths

Which path for whom

• SMB: Path C, because the free OECD QSAR Toolbox, VEGA and Toxtree run on a workstation (HW1) and keep confidential structures in house, which commercial web tools would not.
• Enterprise: Path A, because commercial platforms such as Leadscope or Derek Nexus come with documented validation and vendor support that regulators and auditors accept.
• Developer: Path C, because RDKit, Chemprop and open QSAR models allow validated in-house models with a defined applicability domain.
• Public: Path C, because authorities and institutes need transparent, reproducible open models on sovereign infrastructure and must justify procurement.

Path A — Buy (off-the-shelf)

• Free regulatory tooling: OECD QSAR Toolbox (profilers, analogue search, read-across workflow; developed with ECHA) and ECHA's IUCLID for dossier building [1][6].
• Commercial platforms: Instem Leadscope (Model Applier, expert-alert and statistical models, ICH M7 workflows), Lhasa Limited (UK, Derek Nexus and Sarah Nexus) [7], Simulations Plus ADMET Predictor, Certara, Optibrium. EU/EEA-based: Molecular Networks / altAMES (Germany, CHEMOTION/ChemTunes), ToxMinds (EU), Toxometris.ai, Chemaxon (Hungary) for chemistry tooling. Check: OECD-principle documentation, QMRF availability, audit trail, 21 CFR Part 11 / GxP validation support.
• Choose A when: the assessment must be defensible to ECHA/SCCS/FDA and the team has no ML staff.

Path B — Build on cloud APIs

• Use cloud LLMs (EU region offerings of Azure OpenAI, Google Vertex, Mistral, Anthropic via EU-capable clouds) only for text tasks: literature mining, structured extraction from study reports, drafting justifications from verified data tables.
• Data access: ECHA's public registered-substances data, US EPA CompTox Dashboard APIs (ToxCast/Tox21 bioactivity) [8], PubChem.
• Orchestration: Python with RDKit, workflow engine (n8n, Airflow) and a database of curated study records. Prediction itself should stay in validated models, not in a general LLM.

Path C — Self-host (open source)

• Cheminformatics: RDKit, Mordred, scikit-learn, XGBoost; GNNs: Chemprop, DGL-LifeSci, PyTorch Geometric.
• Open QSAR tools: VEGA (Istituto Mario Negri, Italy; many validated models with applicability-domain scoring), OPERA (US NIEHS/EPA), Toxtree (JRC), OECD QSAR Toolbox (free desktop software), ProTox and ADMETlab as public web tools (not for confidential structures).
• LLM: a 30-70B open-weight model (Qwen, Llama, Mistral families) via vLLM for private extraction from confidential dossiers.
• Hardware: HW1 for classic QSAR; HW3 only if a local LLM is used. Ops effort is mainly model validation and data curation.

Path D — Custom training

Worthwhile when the company holds proprietary toxicity data in a chemical space public models cover poorly (e.g. surfactants, polymers, UVCBs, organometallics). Need hundreds to thousands of high-quality labelled results per endpoint, strict train/validation splits (scaffold or cluster split), applicability-domain definition and a documented validation report following OECD guidance. Consider federated approaches across companies for sensitive data.

State of the art (2026)

• Mature: Ames mutagenicity, skin sensitisation alerts, acute aquatic toxicity and basic physicochemical properties are covered by validated tools; the ICH M7 two-method approach is routine in pharma.
• Advancing: graph neural networks and pretrained chemical foundation models improve accuracy on benchmarks such as Tox21 and ToxCast, and multi-task learning helps sparse endpoints. Generative AI for analogue retrieval and dossier drafting is becoming standard in vendor roadmaps.
• Read-across and NAMs: combining in-silico with in-vitro bioactivity, toxicogenomics and PBPK models in next-generation risk assessment is the direction of travel in both ECHA and cosmetics practice [2][4]. The EU is also moving towards a roadmap to phase out animal testing in chemical safety assessments.
• Limits and failure modes: complex endpoints (repeated-dose, reproductive and developmental toxicity, carcinogenicity) remain poorly predicted; statistical models degrade outside the applicability domain; training data contain noise and publication bias; mixtures, UVCBs, nanomaterials, polymers and metabolites are weakly covered; LLMs may hallucinate study values or citations; regulators judge the justification, not the score.

Shared resources & combinations

• Chemistry stack (RDKit + GNN + tabular-ml): shared with SCI-002 (drug discovery), SCI-003 (materials property prediction) and SCI-014 (formulation development, where toxicity is a design constraint).
• Document pipeline (ocr-docparse + llm-general): DOC-023 (safety data sheet extraction and REACH/CLP substance registers) feeds substance identity and hazard data; SCI-018 supplies analytical identity and impurity profiles.
• Knowledge graph and vector search over study records are reusable for regulatory RAG. Hardware: a single GPU workstation (HW2) serves most self-hosted needs.

Effort, cost & KPIs

• Path A: weeks to onboard; licence-based pricing, free OECD Toolbox needs training time.
• Path B: 2-4 months for extraction and drafting pipeline plus validation.
• Path C: 3-6 months including model validation and documentation; ongoing curation effort.
• Path D: 6-12 months, dependent on data availability and validation burden.
• KPIs: share of data gaps closed by accepted adaptations, ECHA compliance-check outcomes, analogue-search time per substance, concordance of predictions with later experimental data, animal tests avoided, time per dossier/CPSR.

People & change

Management sets the risk appetite: AI results are screening evidence, not a substitute for the expert assessment. The department lead (regulatory affairs or product safety) owns the process, and the toxicologists who sign the safety report are the main users. IT provides workstations or a server; data owners (purchasing, R&D, quality) supply identities and SDS; legal or compliance checks REACH/CLP and cosmetics obligations; suppliers must agree where their data is used.

Works council: normally no. The tools process substances, not employee behaviour, so §87(1) no. 6 BetrVG is rarely triggered unless a platform logs user activity; the council is still to be informed under §90(1) no. 3 when work processes change (GUIDE-004). Public bodies inform the staff council.

DPO: no personal data beyond user accounts; a short confirmation is enough, plus an AVV for any cloud service (GUIDE-005).

AI literacy: support the team with training on applicability domains, OECD validation principles and reading prediction confidence (GUIDE-006). Start with one endpoint (for example skin sensitisation) for a pilot group of 2-3 assessors and compare against known results (GUIDE-009). Typical concern: regulators will not accept it; address it by documenting QMRF/QPRF and keeping the toxicologist as decision maker (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: typically minimal risk, since a toxicity prediction tool is not an Annex III use case and not a safety component of a regulated product by itself; AI literacy (Art. 4) applies. Reassess if the output is embedded in a medical device or product-safety function [9].
• Chemicals law: REACH Annex XI (adaptations), RAAF, CLP classification, Cosmetics Regulation 1223/2009 and SCCS Notes of Guidance; ICH M7. Responsibility stays with the registrant or Safety Assessor.
• GDPR: generally no personal data; study reports may contain staff or contract-lab names and CBI in dossiers.
• Confidential business information: do not submit proprietary structures to public web tools or non-EU LLM endpoints without contractual protection.
• Risks: false negatives leading to unsafe substances; over-trust in black-box scores; bias from data-poor chemical classes. Mitigations: applicability-domain gating, two-method consensus, documented expert review, conservative classification under uncertainty, post-hoc monitoring against new test data.

Getting started

1. Pick two endpoints with recurring data gaps (e.g. Ames, skin sensitisation) and 20-50 substances with known results.
2. Install the OECD QSAR Toolbox and VEGA; benchmark their predictions against your known results.
3. Add a commercial expert-rule tool if you need ICH M7 or regulatory-grade reports.
4. Build a small data matrix and a RAAF-based read-across template; test LLM-assisted extraction with human checks.
5. Write a validation SOP (applicability domain, sign-off, versioning) before using results in a submission.

Sources

1. OECD — QSAR Toolbox — https://qsartoolbox.org/
2. ECHA — Read-across assessment framework (RAAF) — https://echa.europa.eu/support/registration/how-to-avoid-unnecessary-testing-on-animals/grouping-of-substances-and-read-across
3. OECD — Guidance document on the validation of (Q)SAR models (Principles) — https://www.oecd.org/en/publications/guidance-document-on-the-validation-of-quantitative-structure-activity-relationship-q-sar-models_9789264085442-en.html
4. European Commission — Cosmetics Regulation (EC) 1223/2009 and animal testing ban — https://single-market-economy.ec.europa.eu/sectors/cosmetics_en
5. ICH — M7(R2) Assessment and control of DNA reactive (mutagenic) impurities — https://www.ich.org/page/multidisciplinary-guidelines
6. ECHA — IUCLID — https://iuclid6.echa.europa.eu/
7. Lhasa Limited — Derek Nexus — https://www.lhasalimited.org/products/derek-nexus.htm
8. US EPA — CompTox Chemicals Dashboard — https://comptox.epa.gov/dashboard/
9. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
10. VEGA platform (Istituto Mario Negri) — https://www.vegahub.eu/