ICH M7(R2) compliant · OECD QMRF v2.1
Ames mutagenicity. Done properly.
Two complementary (Q)SAR methodologies, automatic ICH M7 classification of every enumerated impurity, metabolite activation reasoning, and read-across. In one call.
Two methodologies, one call
Statistical and expert. Both fire on every prediction.
ICH M7(R2) Section 6.1 requires two complementary (Q)SAR methodologies. The Rasayan engine runs both as a single API call and returns a weight-of-evidence narrative, not two disjoint outputs you have to reconcile.
Deep neural network ensemble trained on 23,009 curated compounds. Probability-calibrated output.
62 curated structural alerts including 3 ICH M7 Cohort-of-Concern overrides. Mechanistic reasoning per alert.
Per-impurity screen
Don't just classify the parent. Classify every impurity.
ICH M7(R2) requires assessment of every impurity above the Threshold of Toxicological Concern, not just the active substance. The Rasayan engine enumerates structural-derivation impurities from the parent — from FDA GSRS, pharmacopeia, chemistry-derived oxidation pathways, and the curated nitrosamine library — and assigns ICH M7 Class 1 to 5 to each. Cohort-of-Concern flags are raised automatically.
Example: running the engine on Losartan returns the parent (Class 5, Ames negative, corroborated by the FDA label) plus an enumerated impurity panel. N-nitroso impurities, where present in the candidate set, trigger explicit Cohort-of-Concern handling per ICH M7(R2).
“Any presence of such impurities in drug products is not acceptable.”— FDA, press announcement on the ARB (angiotensin II receptor blocker) drug products investigation, which identified nitrosamine impurities (NDMA, NDEA, NMBA) in certain lots of valsartan, losartan, and irbesartan.
Metabolite activation
Reactive intermediates surface with the CYP isoform.
Phase-I CYP450 metabolites are generated for every query and scored against the same Ames pipeline. Reactive intermediates from epoxidation, hydroxylation, and N-oxidation are surfaced with the CYP isoform responsible. When the parent is a marketed drug, the CYPs cited in the FDA label are cross-referenced against the predicted metabolites.
Read-across
Curated experimental neighbours, not just a similarity score.
Tanimoto-based search over a chemical-structure index of approximately 120 million compounds, cross-referenced against the curated Ames database, regulatory bioactivity records, and the carcinogenicity potency database. The closest experimentally-tested analogs are returned per call with strain-level outcomes and source attribution.
The integrated platform
Ames sits next to everything else you need.
Same chemical standardisation, same applicability-domain methodology, one audit trail for the whole impurity-safety question.
Open access
Free on OCSR.ai. Full report on the Rasayan Engine.
The Ames prediction itself is openly available on OCSR.ai. Paste a SMILES, draw a structure, or upload a photograph and you see the model call, the active alerts, and an applicability-domain check. The integrated per-impurity screen, FDA-label integration, metabolite activation analysis, and the QMRF citation chain are in the Rasayan Engine.
Validation
The numbers, as measured.
Blind hold-out (N = 1,000), classification cut-off 0.40
Y-scrambling control (same architecture, random labels)
FAQ
Regulatory questions, with citations.
Is the Rasayan Ames Engine approved or endorsed by FDA, EMA, or any other regulator?
No. No QSAR system is. FDA, EMA, PMDA, MHRA, and Health Canada do not formally approve or endorse any specific (Q)SAR software. The regulatory standard under ICH M7(R2) is methodological compliance: two complementary methodologies, OECD validation principles, and a QMRF document on regulator request. The Rasayan Ames Engine delivers all three.
But aren't Derek Nexus and CASE Ultra approved by FDA?
They are not. Lhasa (Derek Nexus, Sarah Nexus) and MultiCASE (CASE Ultra) hold Research Collaboration Agreements with FDA CDER. An RCA is a data-sharing contract: FDA provides non-clinical toxicology datasets, the vendor improves their software. The FDA's published disclaimer states participation in the RCA "should not be interpreted as a direct or indirect endorsement of any Lhasa product or service." The same applies to every vendor in an RCA.
Do you comply with ICH M7(R2)?
Yes. ICH M7(R2) Section 6.1 requires two complementary (Q)SAR methodologies: one expert rule-based and one statistical. The Rasayan Ames Engine implements both (RS_BMUT statistical + RS_Expert structural-alert). Both are documented in OECD QMRF v2.1, comply with the five OECD validation principles, and include the Cohort-of-Concern overrides mandated by ICH M7(R2).
What is a QMRF and can we see yours?
The (Q)SAR Model Reporting Format is the OECD-defined template for documenting a QSAR model's endpoint, algorithm, applicability domain, validation, and mechanistic interpretation per the five OECD validation principles. Current standard is QMRF v2.1, published as Annex I of OECD ENV/CBC/MONO(2023)32. Rasayan publishes both QMRFs (RS_BMUT and RS_Expert) under v2.1, linked at the top of this page.
What validation evidence do you publish?
Locked external blind hold-out (N = 1,000; 332 positive, 668 negative), set aside before any training: AUC-ROC 0.9329 at classification cut-off 0.40. Y-scrambling control on the same architecture under randomly permuted labels collapses to AUC 0.4539, confirming the production performance reflects genuine structure-activity learning. Full validation tables across goodness-of-fit, internal cross-validation, Y-scrambling, and external hold-out are in Sections 6 and 7 of the RS_BMUT QMRF.
What is the Cohort of Concern?
ICH M7(R2) defines three structural classes that drive classification independently of general toxicophore reasoning: N-nitroso compounds, aflatoxin-like structures, and alkyl-azoxy / cycasin-type compounds. RS_Expert implements explicit overrides: when any matching pattern is detected, ICH M7 classification escalates and the compound is flagged in the per-prediction report.
Citations
Can I try the model for free?
Yes. The Ames prediction is built into OCSR.ai, our free chemistry platform. Paste any SMILES and see the model call, active structural alerts, and applicability-domain check at no cost. For the full report and the per-impurity screen, sign in to the Rasayan Engine.
Citations
Can a regulator audit the underlying data and methodology?
Yes, under non-disclosure arrangements on regulatory request. QMRFs publicly disclose model type, algorithm family, descriptor families, training-set provenance, validation methodology, applicability-domain methodology, and validation statistics. Training-set composition, model weights, hyperparameters, SMARTS pattern strings, and per-alert PPV thresholds are available under NDA. Same disclosure level used by all commercial systems.
How is this priced?
Per-prediction or annual subscription. Volume discounts for CDMOs and CROs. The Ames endpoint is part of a multi-endpoint platform licence. Contact sales for the current price book.
Citations