AmesNet

Every new small-molecule drug needs Ames mutagenicity data before it can enter human trials, and a GLP-compliant Ames test costs upward of $10,000 per compound. That expense pushes the assay late in development, where a positive result can invalidate years of work and more than $10 million of investment. Regulators have endorsed in silico screening as the remedy — the ICH M7(R1) guideline accepts QSAR predictions of mutagenicity. The catch is that these models fail precisely where screening matters: in the Second Ames/QSAR International Challenge, participating models averaged a sensitivity of 0.46 on chemically novel compounds, meaning most true mutagens escaped detection.

AmesNet, developed by Model Medicines, attacks that gap by changing what the model is asked to predict. Prior Ames models are unconditioned: they map a molecule to one mutagenicity call, averaging over the eight Salmonella tester strains and over the presence or absence of the S9 liver fraction that mimics human metabolism. AmesNet instead treats the assay context as an explicit second input, so one trained network emits a strain-specific, activation-specific prediction for each of the 16 (strain, ±S9) conditions. The authors call this Task-Conditioned Learning.

The molecular half of the network is an adaptation of ChemPrint, the geometric graph convolutional encoder inside Model Medicines' GALILEO discovery platform; all AmesNet parameters are learned end-to-end from random initialization rather than imported from a pretrained checkpoint. The work was peer reviewed in Chemical Research in Toxicology, and the model runs inside GALILEO as a commercial screening service. Its single-endpoint focus sets it apart from broader safety-profiling models such as Tox21mer and OpenADMET.

#Key Features

• Explicit assay-state conditioning: Nine one-hot features encode the Salmonella strain and whether S9 metabolic activation was present, entering the network on a dedicated channel alongside the molecular embedding.
• One model, sixteen assay conditions: A single set of weights covers all strain and ±S9 combinations, replacing the sixteen task-specific models that single-task baselines require.
• Sensitivity without a discrimination penalty: On the held-out out-of-domain split, sensitivity of 0.73 and balanced accuracy of 0.81 improve simultaneously over every baseline, with non-overlapping bootstrap intervals on balanced accuracy.
• Chemically interpretable recovery: The compounds it rescues from false-negative status concentrate in context-dependent alert classes — planar aromatic intercalators, which report in frameshift strains, and aromatic amines, which require S9 activation.
• Portable conditioning framework: Injecting the same channel into ChemProp lifts its sensitivity from 0.54 to 0.65, isolating the contribution of the paradigm from that of the encoder.

#Technical Details

AmesNet is a dual-branch network. The molecular branch encodes each compound as a geometric graph with atoms as nodes and bonds as edges, producing a 512-dimensional embedding; the context branch carries the nine one-hot assay features. The two vectors are concatenated and passed to a multilayer perceptron. Training used a public compilation of the ISSSTY, OASIS, EFSAP, and MHLW Ames datasets covering four base-pair substitution strains (TA100, TA102, TA104, TA1535) and four frameshift strains (TA97, TA98, TA1537, TA1538). After removing conflicting and duplicate (compound, strain, S9) triplets, eliminating cross-split leakage, and restricting molecular weight to 100–1000 Da, 40,129 training and validation points and 4,528 out-of-domain test points remained. All training ran on a single NVIDIA A100 40GB GPU.

Metrics are sample-size-weighted across the 16 tasks with 95% confidence intervals from a within-task stratified bootstrap (n = 1,000). AmesNet reached sensitivity 0.73 (0.68–0.77) and balanced accuracy 0.81 (0.79–0.83). Baselines on the same split, as sensitivity/balanced accuracy: single-task MLP 0.50/0.72, ChemProp 0.54/0.74, GROVER 0.55/0.75, DeepAmes 0.67/0.75, ungrouped multitask MLP 0.57/0.73, grouped multitask MLP 0.59/0.75. The headline 46% relative sensitivity gain is measured against the single-task MLP; against the re-implemented DeepAmes the margin is 0.06 absolute, or 9%.

#Applications

The intended use is early triage: scoring large compound libraries for mutagenic risk while chemical series are still cheap to abandon, rather than discovering the liability after a candidate has been nominated. Because predictions are strain- and activation-resolved, a medicinal chemist can see whether a flag comes from a direct-acting frameshift mutagen or from a metabolite generated under S9, which points to different mitigation strategies. Model Medicines exposes AmesNet through GALILEO and describes extending the conditioning framework to further ADMET endpoints.

#Impact

AmesNet argues that the sensitivity collapse plaguing Ames QSAR models is an artifact of the modeling paradigm rather than a limit of the data: conditioning on assay state recovers mutagens whose signal exists only within a particular strain or activation context. The encoder-swap experiment strengthens the claim, showing the lift transfers to a message-passing baseline though not to the GROVER graph transformer. The caveats are real — training and evaluation used a single public dataset, and the model addresses mutagenicity alone, not a full preclinical safety profile. The published repository contains the baselines, data splits, and bootstrap analysis, but neither AmesNet's weights nor the ChemPrint encoder is released; the model is available only as a commercial service.