ICH M7 mutagenicity assessment and impurity reports

Human Chemical reasons over predictions from statistical and rule-based models with an AI agent, to help a qualified reviewer reach an ICH M7 class.

Request Demo See the examples

Impurity assessment Agent reasoning

Ethyl 2-(2-aminothiazol-4-yl)acetate

CAS 53266-94-7

Alert model: Primary aromatic amine
Statistical model: 7.7%
Nearest neighbors
Key evidence: Free acid negative in GLP OECD 471

Class 3 → Class 4

Awaiting reviewer sign-off

94.3%

Accuracy on 35 lansoprazole impurities, compared to 70% for other commercial products.

21,501

Compounds with OECD-compliant data in the training set, double that of other vendors. Every structure is visible.

70%+

Of Class 3 predictions downgraded to Class 4, with reasoning that experts agree with.

Standard and workflow

Many compounds with structural alerts are still negative in the Ames test

37.0%

of 54 tested pharmaceutical intermediates with an aromatic amine, amide or N-hydroxylamine alert were Ames positive.

Source: FDA docket FDA-2013-D-0369-0012

46%

of primary aromatic amine predictions were overturned in an FDA analysis of 1,002 drug impurities.

Source: Jayasekara et al., 2021

ICH M7(R2) classes

Class 3

The structure has an alert and no mutagenicity data. The impurity is controlled at the threshold of toxicological concern (TTC), 1.5 μg/day for lifetime exposure.

→ Class 4

The structure has an alert, and the same alert appears in a related compound, in the same chemical environment, that tested negative. The impurity is treated as non-mutagenic.

→ Class 5

The structure has no alert, or there are enough data to show that it is not mutagenic. The impurity is treated as non-mutagenic.

How it works

How Human Chemical assesses an impurity

Individual models

Alert model

An expert rule-based model with 100+ curated alerts, including algorithmic alerts.
Individual models

Statistical model

A 20-member OECD-compliant ensemble model with an applicability domain.
AI agent

Investigates

The agent searches for analogs, checks databases and evaluates the evidence.
Context

Evidence package

Nearest neighbors, chemical environment, mechanisms, database records and literature.
Outcome

ICH M7 class

The qualified reviewer approves the class. The report gives the rationale and full citations, each visible to the user and verified.

Benchmark

94.3% accuracy on a benchmark of 35 impurities

The impurities come from a published study of lansoprazole (Chintalapati et al. 2026): 9 are likely positive and 26 are likely negative. Each system combines its expert rule-based and statistical models, and the combined call is positive when either model is positive.

Accuracy of the combined call

Human Chemical with database lookup

97.1%

1 false positive, 0 false negatives
Human Chemical

94.3%

2 false positives, 0 false negatives
Lhasa (Derek Nexus and Sarah Nexus)

71.4%

9 false positives, 1 false negative
CASE Ultra (GT_Expert and GT1_BMUT)

65.7%

12 false positives, 0 false negatives

Calls on six compounds with a negative reference label

Compound	Derek Nexus	Sarah Nexus	CASE GT_Expert	CASE GT1_BMUT	Human Chemical
2-Mercaptobenzimidazole	Positive	Positive	Out of domain	Positive	Negative
Lansoprazole sulfide	Positive	Negative	Negative	Positive	Negative
Lansoprazole N-oxide	Negative	Negative	Negative	Positive	Negative
Lansoprazole sulfone	Negative	Negative	Negative	Positive	Negative
Lansoprazole sulfone N-oxide	Negative	Negative	Negative	Positive	Negative
4-Dimethylaminopyridine (DMAP)	Negative	Negative	Out of domain	Out of domain	Positive

Green cells are correct calls, orange cells are false positives, and outlined cells are out of domain.

21,501

compounds in the training set, filtered to OECD compliance, with one entry per structure. Twice as many disclosed structures as any other vendor.

6,028 positive (28.0%) 15,473 negative (72.0%)

Sources include EURL ECVAM, NTP and CEBS, Hansen, NIHS, EFSA, Mendeley and ISSSTY, EPA CompTox, and JPMA (Hakura 2021). Every compound structure is visible in the system.

The agent at work

The agent adds context to each model score

1 · Model score

Descriptors and fingerprints
Probability, alert and class label, with uncertainty
Applicability domain, which does not consider mechanism

2 · Context checks

Applicability domain: Nearest neighbors, scaffold distance and label reliability.
Mechanism: Reactivity, metabolism, transport and target biology.
Evidence: Literature, public assays, analogs and regulatory dossiers.
Hypotheses: Positive and negative controls, counter-hypotheses and falsification.
Skeptic review: Bias checks, impossible values and human judgment.

3 · Decision

Trust: The structure is in the domain, the mechanism is coherent and the evidence agrees.
Reject: The structure is out of domain, or an assay and the literature contradict the result.
Test: The agent recommends the experiment that most reduces uncertainty.

Worked example

Downgrading a prediction

1-(2-fluoro-3-hydroxy-6-nitrophenyl)propan-2-one, an ICH M7 impurity assessment

Class 3

The alert model flags an aromatic nitro alert of high severity. The statistical model predicts a probability of 0.77, which is mutagenic, so the impurity is Class 3.
Class 4

Does experimental Ames data already exist? The agent finds analogs with negative Ames results that match the local environment, a nitro group on a nitrophenol scaffold.
Is the nitro alert mechanistically relevant here? The -OH group para to the nitro group raises the LUMO energy. The calculated log P is low at 1.57 [1]. A CH₂ group separates the carbonyl from the ring. The fluorine is meta to the nitro group, not ortho or para.
The agent assigns Class 4 and recommends an Ames test.
Class 5

Bristol Myers Squibb reported that a GLP Ames test (OECD 471) was negative, which confirms Class 5 [2].*

*Shared with permission

[1] Debnath et al., Environ Mol Mutagen 1992;19:53-70.

[2] Seal S et al., AI Agents Bring Context to Computational Toxicology: QSAR Predictions to Auditable Recommendations. Chem Res Toxicol (invited).

Examples

Six Class 3 calls the agent reasoned to Class 4

Case 1 · Read-across to a negative comparator

Class 3 → Class 4

Ethyl 2-(2-aminothiazol-4-yl)acetate

CAS 53266-94-7 · Primary aromatic amine alert

The alert is correct for 2-aminothiazoles in general. The free acid has the same alert in the same chemical environment and is negative in a full GLP Ames study.

Free acid Ames negative Methyl ester Ames negative

Alert model: Positive; The alert is a primary aromatic amine of high severity.
Statistical model: 7.7%; The probability of a positive Ames result is 7.7%, so the call is negative.
Nearest neighbors: 9 of 10; Nine of the ten closest neighbors are Ames negative, including the free acid and the methyl ester.
Key evidence: OECD 471; The free acid is negative in a GLP study, with and without S9.

Case 2 · Different chemical environment

Class 3 → Class 4

2-Amino-5-(trifluoromethyl)pyridine

CAS 74784-70-6 · Primary aromatic amine alert

The alert is in a different chemical environment from other anilines. The ring nitrogen and the CF₃ group make the reactive nitrenium ion less likely to form.

2-Aminopyridine Ames negative 4-(Trifluoromethyl)aniline Ames positive

Alert model: Positive; The alert is a primary aromatic amine of high severity.
Statistical model: 6.2%; The probability of a positive Ames result is 6.2%, so the call is negative.
Nearest neighbors: Pyridines; The nearest negatives are pyridines, and the nearest positive is an aniline.
Key evidence: NTP 1987; 2-Aminopyridine is negative with rat and hamster S9, at up to 10,000 μg/plate.

Case 3 · Deactivated amine with a remote substituent

Class 3 → Class 4

4-Amino-2-methylbenzoic acid

CAS 2486-75-1 · Primary aromatic amine alert

The para carboxyl group deactivates the amine, as in PABA, and the 2-methyl group is remote from it. o-Toluidine lacks the carboxyl group.

PABA Ames negative o-Toluidine IARC Group 1

Alert model: Positive; The alert is a primary aromatic amine on a carbocyclic ring.
Statistical model: 19.7%; The probability of a positive Ames result is 19.7%, so the call is negative.
Nearest neighbors: 10 of 10; All ten closest neighbors are Ames negative, including amino and hydroxy-amino benzoic acids.
Key evidence: NTP; 4-Aminobenzoic acid (PABA) has the same alert and is Ames negative. IARC places it in Group 3.

Case 4 · Endogenous scaffold with one data gap

Class 3 → Class 4

1H-Imidazole-4-carbaldehyde

CAS 3034-50-2 · Heteroaryl carbaldehyde alert

The imidazole ring is electron-rich, and the compound is related to histidine pathway metabolites. Every tested imidazole neighbor is negative.

Urocanic acid Ames negative 4-Methylimidazole Ames negative

Alert model: Positive; The alert is a heteroaryl carbaldehyde of medium severity.
Statistical model: 12.6%; The probability of a positive Ames result is 12.6%, so the call is negative, with moderate confidence.
Nearest neighbors: Negative; The closest tested neighbors are urocanic acid, 4-methylimidazole and histamine, and all three are Ames negative.
Key evidence: Imidazole; Imidazole and its metabolites are Ames negative.

Case 5 · Michael acceptor that does not react

Class 3 → Class 4

4-(2-Methylprop-2-enoylamino)benzoic acid

CAS 15286-99-4 · Michael acceptor alert

The α-methyl group sterically hinders conjugate addition, and methacrylamide, with the same alert, is negative in OECD 471 and OECD 473 studies.

Methacrylamide Ames negative Acedoben Ames negative

Alert model: Positive; The alert is a Michael acceptor of medium severity. A second alert, for acyl aromatic amines, was set aside.
Statistical model: 8.4%; The probability of a positive Ames result is 8.4%, so the call is negative.
Nearest neighbors: Negative; The closest neighbors are predominantly negative, including acedoben.
Key evidence: OECD 471; Methacrylamide is negative at up to 5,000 μg/plate, and it is negative in OECD 473.

Case 6 · Electron-donating group on a negative sub-class

Class 3 → Class 4

3-Methoxythiophene-2-carbaldehyde

CAS 35134-07-7 · Heteroaryl carbaldehyde alert

The methoxy group donates electrons and lowers the reactivity of the aldehyde. Thiophene carbaldehydes with the same alert in the same environment are negative in published Ames data.

3-Thiophenecarboxaldehyde Ames negative Vanillin Ames negative

Alert model: Positive; The alert is a heteroaryl carbaldehyde of medium severity.
Statistical model: 11.8%; The probability of a positive Ames result is 11.8%, so the call is negative.
Nearest neighbors: 10 of 10; All ten closest neighbors are Ames negative, including the carboxylic acid and vanillin.
Key evidence: Lee 1994; Thiophene carbaldehydes are negative in strains TA98 and TA100.

Next steps

From prediction to submission

01

Regulatory engagement

Our pharma partners are using the system for inclusion in regulatory submissions under ICH M7.
02

Validation

We run pilots with pharmaceutical sponsors and risk-assessment partners.
03

Documentation

Documentation in QSAR Model Reporting Format (QMRF) is available for users.
04

Experimental confirmation

Partners often confirm with OECD 471 studies on compounds to provide confidence in models.

A qualified toxicologist signs off on each Class 3 and Class 4 call before regulatory use, and experimental Ames data overrule (Q)SAR predictions under ICH M7.

Built to the highest standards of IP security. We never use your inputs, outputs, or uploaded content to train our AI models. SOC 2 Type II compliant.

Talk to us about ICH M7

See the system on your own impurities, or request the QMRF documentation.

Request Demo