Computer-implemented systems and methods for compound mutagenicity assessment

Abstract

Provided herein are computer-implemented systems and methods for assessing compound pharmacotoxicity, including, inter alia, generating mutagenicity probability scores and / or profiles for compounds, determining mutagenicity, and / or stratifying mutagenicity risk. The computational assessment may be applied across various Ames testing conditions with sensitivity and / or balanced accuracy, including in new chemical spaces.

Description

71TL-865001-WO PATENT COMPUTER-IMPLEMENTED SYSTEMS AND METHODS FOR COMPOUND MUTAGENICITY ASSESSMENT RELATED APPLICATIONS

[0001] This application claims the benefit under 35 U.S.C. § 119(e) of U.S. Provisional Patent Application Ser. No. 63 / 773,321, filed March 17, 2025, and U.S. Provisional Patent Application Ser. No. 63 / 979,684, filed February 10, 2026, the contents of each of which are incorporated herein by reference in their entirety for all purposes.BACKGROUNDField

[0002] The present disclosure generally relates to computational pharmacotoxicology.Description

[0003] New small-molecule pharmaceuticals must undergo a suite of toxicity testing before entering clinical trials. The U.S. Food and Drug Administration (FDA) and international regulators expect Ames genotoxicity data in regulatory submissions as described in the ICH S2(R1) guideline. The Ames test under Good Laboratory Practice (GLP) conditions exceeds $10,000 per compound. This cost excludes the chemical sample price. Such high expenditure makes routine screening of large compound libraries impractical during early drug discovery. GLP Ames testing typically occurs near the regulatory submission stage after significant investment into the preclinical candidate. Preparing for human clinical trials took approximately three years in the average small-molecule discovery timeline. This process can require $10 million or more per candidate. An Ames test failure at this late stage results in years of wasted effort and millions of dollars in lost capital.

[0004] In response to such bottlenecks, the FDA and regulatory agencies around the world have formalized computational approaches as a pathway to reduce and outright replace traditional preclinical testing. The FDA Modernization Act now provides the legal framework for artificial intelligence (Al)-based computer models to support and replace wet-lab and animal testing for drug evaluation. Accordingly, the FDA established programs like the Artificial Intelligence Program for Toxicology (AI4TOX) to operationalize Al within safety reviews. Under this program the SafetAI initiative focuses on developing Al models for toxicological endpoints. DeepAmes was developed in this initiative to predict Ames mutagenicity within regulatory frameworks. Internationally, the ICH M7(R1) guideline explicitly endorses Quantitative Structure-Activity Relationship (QSAR) methodologies to predict Ames mutagenicity. The First and Second Ames / QSAR International Challenge Projects further propelled regulatoryacceptance. During these projects, government agencies in the U.S., Japan and Italy used standardized benchmarks to improve model performance for Ames prediction.

[0005] The Ames test uses specialized bacterial strains in growth-restrictive conditions to detect if a drug candidate might cause genetic mutations. Mutagenic potential is evidenced by bacterial growth after exposure to either the parent compound or its metabolic byproducts. Effective screening, therefore, depends on high sensitivity to avoid misclassifying a dangerous molecule as safe. Meeting this requirement remains a challenge for Ames QSAR models. Low sensitivity is a critical failure. It allows genotoxic compounds to advance undetected. Such failures are amplified when Al models encounter novel chemical structures distinct from their training data. This challenge in novel chemical space is known as out-of-domain (OOD) generalization. Achieving operational Ames model utility simultaneously requires high sensitivity and OOD generalization so that no genotoxic compound escapes detection when screening novel chemical libraries. Benchmark data from the Second Ames / QSAR International Challenge Projects revealed a significant collapse in sensitivity for current Ames models when encountering novel chemotypes. The project yielded a participant average sensitivity of only 0.46 on the OOD test set. High-profile models like the FDA’s DeepAmes achieved a sensitivity of just 0.47 and MIT’s ChemProp yielded 0.32. These results demonstrate that current methods cannot maintain the sensitivity required for the reliable safety screening of novel drug candidates. Such unreliability creates a pathway for unidentified genotoxic molecules to advance into later stage drug development and potentially waste years and millions of dollars in resources.SUMMARY

[0006] Disclosed herein, in certain aspects, include methods of assessing compound mutagenicity. In some embodiments, the method comprises: (a) providing a compound or a molecular structure representation thereof; (b) providing (optionally, selecting) one or more Ames testing conditions; (c) combining (optionally, concatenating) an adaptive molecular embedding vector corresponding to the compound, with one or more Ames condition vectors corresponding to the one or more Ames testing conditions, to generate one or more condition-specific embedding vectors; (d) processing the one or more condition-specific embedding vectors through a task-conditioned machine-learning model (optionally, through a multilayer perceptron (MLP)) to generate a mutagenicity probability score for each of the one or more Ames testing conditions, thereby generating a profile of mutagenicity probability scores for the compound; and (e) determining the compound as mutagenic or non-mutagenic, or stratifying a mutagenicity risk of the compound, based on the profile of mutagenicity probability scores.

[0007] Some embodiments of the methods provided herein can comprise generating, using a convolutional neural network, the adaptive molecular embedding vector from a molecularstructure representation of the compound. In some embodiments, the convolutional neural network is a graph convolutional neural network. In some embodiments of the methods, (c) comprises iteratively aggregating structural information (optionally, from neighboring atoms, motifs or fragments) of the compound to generate the adaptive molecular embedding vector. In some embodiments, the (optionally, graph) convolutional neural network comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split. In some embodiments, the (optionally, graph) convolutional neural network comprises a plurality of convolutional layers configured (optionally, trained) to iteratively aggregate structural information (optionally, from neighboring atoms, motifs or fragments) of the compound to generate the adaptive molecular embedding vector. In some embodiments, the structural information comprises one or more mutagenicity-associated molecular features of the compound.

[0008] Some embodiments of the methods provided herein can comprise training the (optionally, graph) convolutional neural network using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions. In some embodiments, the training dataset is partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split. In some embodiments, one or more reference compounds in the training dataset comprise one or more mutagenicity-associated molecular features.

[0009] Some embodiments of the methods provided herein can comprise identifying one or more mutagenicity-associated molecular features in the compound. The method can comprise providing a report that describes the identified one or more mutagenicity-associated molecular features as present or absent in the compound.

[0010] Some embodiments of the methods provided herein can comprise, prior to (c), providing the one or more Ames condition vectors.

[0011] Some embodiments of the methods provided herein can comprise training the task-conditioned machine-learning model (optionally, the multilayer perceptron) using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions. In some embodiments, training the task-conditioned machine-learning model (optionally, the multilayer perceptron) comprises: (i) initializing a plurality of parameters of the task-conditioned machine-learning model (optionally, the multilayer perceptron); (ii) processing the training dataset through the task-conditioned machine-learning model (optionally, the multilayer perceptron) to generate predicted mutagenicity probability scores; (iii) computing a loss function based on differences between the predicted mutagenicity probability scores andknown mutagenicity classifications; (iv) updating the plurality of parameters of the task-conditioned machine-learning model (optionally, the multilayer perceptron) to minimize the loss function; and (v) repeating steps (ii)-(iv). In some embodiments, the training dataset is partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

[0012] Some embodiments of the methods provided herein can comprise optimizing parameters of the task-conditioned machine-learning model (optionally, the multilayer perceptron) until the task-conditioned machine-learning model (optionally, the multilayer perceptron) achieves an improved or substantially maximal value of sensitivity, specificity, accuracy, balanced accuracy, precision, recall or area under a receiver operating characteristic curve (AUROC) (optionally, on a validation dataset). In some embodiments, the task-conditioned machine-learning model (optionally, the multilayer perceptron) comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

[0013] In some embodiments of the methods provided herein, the (optionally, graph) convolutional neural network and the task-conditioned machine-learning model (optionally, the multilayer perceptron) are jointly trained. In some embodiments, the (optionally, graph) convolutional neural network and the task-conditioned machine-learning model (optionally, the multilayer perceptron) are separately trained. In some embodiments, the (optionally, graph) convolutional neural network and the task-conditioned machine-learning model (optionally, the multilayer perceptron) are sequentially or iteratively trained. In some embodiments, the (optionally, graph) convolutional neural network is trained prior to or after training the task-conditioned machine-learning model (optionally, the multilayer perceptron). In some embodiments, the (optionally, graph) convolutional neural network and the task-conditioned machine-learning model (optionally, the multilayer perceptron) are trained using the same dataset. In some embodiments, the (optionally, graph) convolutional neural network and the task-conditioned machine-learning model (optionally, the multilayer perceptron) are trained using different datasets. In some embodiments, the task-conditioned machine-learning model (optionally, the multilayer perceptron) comprises the convolutional neural network (optionally, the graph convolutional neural network).

[0014] In some embodiments of the methods provided herein, (d) comprises applying an ensemble of neural network models. In some embodiments, each neural network model in the ensemble processes the condition-specific embedding vectors to generate a set of mutagenicity probability scores, and the mutagenicity probability scores are averaged (optionally, with respect to each Ames testing condition) across the ensemble to generate the profile of mutagenicity probability scores for the compound. In some embodiments, the ensemble comprises at least five,six, seven, eight, nine, or ten neural network models.

[0015] In some embodiments of the methods provided herein, (a) comprises providing a molecular structure representation of the compound. In some embodiments, the method is performed on a plurality of compounds or molecular structure representations thereof. In some embodiments, the plurality of compounds comprises at least 100, 200, 300, 400, 500, 600, 700, 800, 900, or 1,000 compounds (optionally, small molecules). In some embodiments, (a) occurs prior to (b). In some embodiments, (a) occurs subsequent to (b).

[0016] Some embodiments of the methods provided herein can comprise verifying a (optionally, mutagenic or non-mutagenic) determination. The method can comprise testing the compound under at least one of the profiled Ames testing conditions. The method can comprise testing the compound under each of the profiled Ames testing conditions.

[0017] In some embodiments, the method provided herein determines or stratifies the plurality of compounds at a sensitivity of at least 0.70, 0.75, or 0.80. In some embodiments, the method determines or stratifies the plurality of compounds at an accuracy (optionally, a balanced accuracy) of at least 0.70, 0.75, or 0.80. In some embodiments, the method provided herein determines or stratifies the plurality of compounds at a sensitivity of at least 0.70, 0.75, or 0.80 and an accuracy (optionally, a balanced accuracy) of at least 0.70, 0.75, or 0.80. In some embodiments, the method determines or stratifies the plurality of compounds at a specificity of at least 0.70, 0.75, or 0.80.

[0018] Some embodiments of the methods provided herein can comprise providing a report that describes the determination or stratification of the compound made in (e). In some embodiments, the method can comprise, displaying on a user interface connected to an output device, a report that describes the determination or stratification of the compound made in (e).

[0019] Disclosed herein, in certain aspects, include systems for assessing compound mutagenicity. In some embodiments, the system comprises: one or more processors, and memory storing a convolutional neural network (optionally, a graph convolutional neural network) and a plurality of lines of instructions. In some embodiments, the system comprises: one or more processors, and memory storing a task-conditioned machine-learning model (optionally, a multilayer perceptron (MLP)) and a plurality of lines of instructions. In some embodiments, the system comprises: one or more processors, and memory storing a convolutional neural network (optionally, a graph convolutional neural network), a task-conditioned machine-learning model (optionally, a multilayer perceptron)), and a plurality of lines of instructions.

[0020] In some embodiments of the systems provided herein, the one or more processors are configured to execute the plurality of lines of instructions to perform the methods disclosed herein. In some embodiments, the one or more processors are configured to execute theplurality of lines of instructions to: (a) receive a compound or a molecular structure representation thereof; (b) receiving one or more Ames testing conditions; (c) combine (optionally, concatenate) an adaptive molecular embedding vector corresponding to the compound, with one or more Ames condition vectors corresponding to the one or more Ames testing conditions, to generate one or more condition-specific embedding vectors; (d) process the one or more condition-specific embedding vectors through the task-conditioned machine-learning model (optionally, the multilayer perceptron) to generate a mutagenicity probability score for each of the one or more Ames testing conditions, thereby generating a profile of mutagenicity probability scores for the compound; and (e) determine the compound as mutagenic or non-mutagenic, or stratify a mutagenicity risk of the compound, based on the profile of mutagenicity probability scores.

[0021] In some embodiments of the systems provided herein, the one or more processors are configured to execute the plurality of lines of instructions to generate, using the (optionally, graph) convolutional neural network, the adaptive molecular embedding vector from a molecular structure representation of the compound. In some embodiments, the (optionally, graph) convolutional neural network is configured to iteratively aggregate structural information (optionally, from neighboring atoms, motifs or fragments) of the compound to generate the adaptive molecular embedding vector. In some embodiments, the (optionally, graph) convolutional neural network comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split. In some embodiments, the (optionally, graph) convolutional neural network comprises a plurality of convolutional layers configured (optionally, trained) to iteratively aggregate structural information (optionally, from neighboring atoms, motifs or fragments) of the compound to generate the adaptive molecular embedding vector. In some embodiments, the structural information comprises one or more mutagenicity-associated molecular features of the compound.

[0022] In some embodiments of the systems provided herein, the one or more processors are configured to train the (optionally, graph) convolutional neural network using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions. In some embodiments, the training dataset is partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split. In some embodiments, one or more reference compounds in the training dataset comprise one or more mutagenicity-associated molecular features.

[0023] In some embodiments of the systems provided herein, the one or more processors are configured to identify one or more mutagenicity-associated molecular features in the compound. In some embodiments, the one or more processors are configured to provide areport that describes the identified one or more mutagenicity-associated molecular features as present or absent in the compound.

[0024] In some embodiments of the systems provided herein, the one or more processors are configured to provide, prior to (c), the one or more Ames condition vectors.

[0025] In some embodiments of the systems provided herein, the one or more processors are configured to train the task-conditioned machine-learning model (optionally, the multilayer perceptron) using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions. In some embodiments, the one or more processors are configured to train the task-conditioned machine-learning model (optionally, the multilayer perceptron) that comprises: (i) initializing a plurality of parameters of the task-conditioned machine-learning model (optionally, the multilayer perceptron); (ii) processing the training dataset through the task-conditioned machine-learning model (optionally, the multilayer perceptron) to generate predicted mutagenicity probability scores; (iii) computing a loss function based on differences between the predicted mutagenicity probability scores and known mutagenicity classifications; (iv) updating the plurality of parameters of the task-conditioned machine-learning model (optionally, the multilayer perceptron) to minimize the loss function; and (v) repeating steps (ii)-(iv). In some embodiments, the training dataset is partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

[0026] In some embodiments of the systems provided herein, the one or more processors are configured to optimize parameters of the task-conditioned machine-learning model (optionally, the multilayer perceptron) until the task-conditioned machine-learning model (optionally, the multilayer perceptron) achieves an improved or substantially maximal value of sensitivity, specificity, accuracy, balanced accuracy, precision, recall or area under a receiver operating characteristic curve (AUROC) (optionally, on a validation dataset). In some embodiments, the task-conditioned machine-learning model (optionally, the multilayer perceptron) comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

[0027] In some embodiments of the systems provided herein, the (optionally, graph) convolutional neural network and the task-conditioned machine-learning model (optionally, the multilayer perceptron) are configured to be jointly trained. In some embodiments, the (optionally, graph) convolutional neural network and the task-conditioned machine-learning model (optionally, the multilayer perceptron) are configured to be separately trained. In some embodiments, the (optionally, graph) convolutional neural network and the task-conditioned machine-learning model (optionally, the multilayer perceptron) are sequentially or iterativelytrained. In some embodiments, the (optionally, graph) convolutional neural network is configured to be trained prior to or after training the task-conditioned machine-learning model (optionally, the multilayer perceptron). In some embodiments, the (optionally, graph) convolutional neural network and the task-conditioned machine-learning model (optionally, the multilayer perceptron) are configured to be trained using the same dataset. In some embodiments, the (optionally, graph) convolutional neural network and the task-conditioned machine-learning model (optionally, the multilayer perceptron) are configured to be trained using different datasets. In some embodiments, the task-conditioned machine-learning model (optionally, the multilayer perceptron) comprises the convolutional neural network (optionally, the graph convolutional neural network).

[0028] In some embodiments of the systems provided herein, (d) comprises applying an ensemble of neural network models. In some embodiments, each neural network model in the ensemble processes the condition-specific embedding vectors to generate a set of mutagenicity probability scores, and the mutagenicity probability scores are averaged (optionally, with respect to each Ames testing condition) across the ensemble to generate the profile of mutagenicity probability scores for the compound. In some embodiments, the ensemble comprises at least five, six, seven, eight, nine, or ten neural network models.

[0029] In some embodiments of the systems provided herein, the one or more processors are configured, in (a), to receive a molecular structure representation of the compound. In some embodiments, the one or more processors are configured to perform the method on a plurality of compounds or molecular structure representations thereof. In some embodiments, the plurality of compounds comprises at least 100, 200, 300, 400, 500, 600, 700, 800, 900, or 1,000 compounds (optionally, small molecules). In some embodiments, (a) occurs prior to (b). In some embodiments, (a) occurs subsequent to (b).

[0030] In some embodiments of the systems provided herein, the one or more processors are configured to verify a (optionally, mutagenic or non-mutagenic) determination. In some embodiments, the one or more processors are configured to test the compound under at least one of the profiled Ames testing conditions. In some embodiments, the one or more processors are configured to test the compound under each of the profiled Ames testing conditions.

[0031] In some embodiments of the systems provided herein, the one or more processors are configured to determine or stratify the plurality of compounds at a sensitivity of at least 0.70, 0.75, or 0.80. In some embodiments, the one or more processors are configured to determine or stratify the plurality of compounds at an accuracy (optionally, a balanced accuracy) of at least 0.70, 0.75, or 0.80. In some embodiments of the systems provided herein, the one or more processors are configured to determine or stratify the plurality of compounds at a sensitivity of at least 0.70, 0.75, or 0.80 and an accuracy (optionally, a balanced accuracy) of at least 0.70,0.75, or 0.80. In some embodiments, the one or more processors are configured to determine or stratify the plurality of compounds at a specificity of at least 0.70, 0.75, or 0.80.

[0032] In some embodiments of the systems provided herein, the one or more processors are configured to provide a report that that describes the determination or stratification of the compound made in (e). The system can further comprise an output device. In some embodiments, the one or more processors are configured to display on the output device a report that that describes the determination or stratification of the compound made in (e).

[0033] In some embodiments of the methods and systems provided herein, the profile comprises at least two, three, four, five, six, seven, eight, nine, or ten mutagenicity probability scores. In some embodiments, the profile comprises mutagenicity probability scores for at least two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, or sixteen Ames testing conditions. In some embodiments, the profile comprises mutagenicity probability scores for Ames testing conditions corresponding to at least one, two, three, four, five, six, seven, or eight distinct bacterial tester strains, each tested in the presence and / or absence of a metabolic activation. In some embodiments, the profile comprises mutagenicity probability scores for Ames testing conditions corresponding to at least one, two, three, four, five, six, seven, or eight distinct bacterial tester strains, each tested both in the presence and absence of a metabolic activation.

[0034] In some embodiments of the methods and systems provided herein, an Ames testing condition comprises a distinct combination of (1) the identity of a bacterial tester strain and (2) the status (i.e., the presence or absence) of a metabolic activation. In some embodiments, the metabolic activation comprises an S9 fraction. In some embodiments, a bacterial tester strain is a base-pair substitution mutant strain or a frameshift mutant strain. In some embodiments, a bacterial tester strain is a Salmonella typhimurium (S. typhimurium) strain. In some embodiments, a base-pair substitution mutant strain is selected from S. typhimurium TAI 00, TAI 02, TAI 04, and TAI 535. In some embodiments, a frameshift mutant strain is selected from S. typhimurium TA1537, TA1538, TA97, and TA98.

[0035] In some embodiments of the methods and systems provided herein, said determining or stratifying comprises comparing at least one mutagenicity probability score to a threshold. In some embodiments, the compound is determined or stratified as non-mutagenic based on a determination that the compound is negative under all profiled Ames testing conditions. In some embodiments, a mutagenicity probability score is between 0 and 1; and a negative determination is made when the mutagenicity probability score is below (or at or below) 0.3, 0.4, or 0.5. In some embodiments, the compound is determined or stratified as mutagenic based on a determination that the compound is positive (or strongly positive) under at least one profiled Amestesting condition. In some embodiments, the compound is determined or stratified as highly mutagenic based on a determination that the compound is positive (or strongly positive) under at least two profiled Ames testing condition. In some embodiments, a mutagenicity probability score is between 0 and 1; and a positive determination is made when the mutagenicity probability score is above (or at or above) 0.5. In some embodiments, a mutagenicity probability score is between 0 and 1; and a strongly positive determination is made when the mutagenicity probability score is above (or at or above) 0.5. In some embodiments, the compound is determined or stratified as an Ames Class A compound, if the compound is determined to be strongly positive under at least one profiled Ames testing condition. In some embodiments, the compound is determined or stratified as an Ames Class B compound, if the compound is determined to be positive (but not strongly positive) under at least one profiled Ames testing condition. In some embodiments, the compound is determined or stratified as an Ames Class C compound, if the compound is determined to be negative under all profiled Ames testing conditions. In some embodiments, the compound is stratified with a low mutagenicity risk based on a determination that the compound is negative under all profiled Ames testing conditions. In some embodiments, the compound is stratified with a medium mutagenicity risk based on a determination that the compound is positive under only one profiled Ames testing condition. In some embodiments, the compound is stratified with a high mutagenicity risk based on a determination that the compound is positive under at least two profiled Ames testing conditions.

[0036] In some embodiments of the methods and systems provided herein, the compound is a small molecule. A small molecule can have a molecular weight of less than about 1,500 Daltons.

[0037] In some embodiments of the methods and systems provided herein, the molecular structure representation is a (optionally, standardized) (optionally, linear) textual representation. In some embodiments, the molecular structure representation is a standardized textual representation. In some embodiments, the molecular structure representation is a linear textual representation. In some embodiments, the molecular structure representation is a standardized linear textual representation. In some embodiments, the molecular structure representation is a SMILES string. In some embodiments, the adaptive molecular embedding vector is a multi-dimensional vector, optionally, comprising at least 64, 128, 256, 512, or 1,024 dimensions. In some embodiments, the adaptive molecular embedding vector comprises (optionally, local, regional and / or global) structural information (optionally, molecular geometric and / or topological features) of the compound (optionally, at one or more of atomic, motif, crossmotif, or whole-molecule level(s)).

[0038] In some embodiments of the methods and systems provided herein, a (or each)mutagenicity-associated molecular feature is a chemical, geometric, topological, and / or structural feature. In some embodiments, the one or more mutagenicity-associated molecular features comprise one or more chemical, geometric, topological, and / or structural features. In some embodiments, the mutagenicity-associated molecular feature(s) is(are) each independently identified, determined, or suspected to be associated (or correlate) with mutagenicity or nonmutagenicity, optionally, as measured by an Ames test under one or more Ames testing conditions. In some embodiments, the one or more mutagenicity-associated molecular features comprise at least one, two, three, four, five, six, seven, eight, nine or ten features disclosed herein.BRIEF DESCRIPTION OF THE DRAWINGS

[0039] FIG. 1A-FIG. ID depict Ames test and predictive algorithms. FIG. 1A depicts diagram depicting the Ames test, where the rectangles represent eight strains commonly used. Compounds are tested in the presence or absence of a liver fraction (S9), which mimics metabolic activation in the human body. FIG. IB depicts single-task Ames QSAR algorithm (ST) trained to predict a strain-specific feature (in this case, strain TAI 00), with or without metabolic activation (S9). FIG. 1C depicts ungrouped multitask Ames QSAR algorithm (uMT), trained to predict eight strain-specific features, with or without metabolic activation (S9). FIG. ID depicts grouped multitasked Ames QSAR algorithm, trained to predict only substitution strains-specific features with or without metabolic activation (S9).

[0040] FIG. 2 depicts the AmesNet Architecture. AmesNet shared the same graph convolutional neural network as ChemPrint and also contained the addition of Ames specific features.

[0041] FIG. 3 depicts Area Under the Receiver Operating Characteristic Curve (AUROC) for AmesNet trained on in-domain data. AmesNet model 1, which used random splitting with chemical diversity of in-domain data, is in blue (base HPO), while AmesNet model 2, which used t-SNE splitting with chemical diversity of in-domain data, is in orange (tSNE-HPO).

[0042] FIG. 4 depicts Area Under the Receiver Operating Characteristic Curve (AUROC) for AmesNet trained on out-of-domain data. AmesNet model 3, which used random splitting with chemical diversity of out-of-domain data, is in blue (base HPO), while AmesNet model 4, which used t-SNE splitting with chemical diversity of out-of-domain data, is in orange (t-SNE HPO).

[0043] FIG. 5 depicts polyaromatic compounds correctly classified as genotoxic by AmesNet but misclassified by the gMTL model.

[0044] FIG. 6A-FIG.6D depict Ames test and predictive algorithms. FIG.6A depicts diagram depicting the Ames test, where the rectangles represent eight strains commonly used. Compounds are tested in the presence or absence of a liver fraction (+ / -S9), which mimicsmetabolic activation in the human body. FIG. 6B depicts single-task Ames QSAR algorithm (STL) trained to predict a strain-specific feature, with or without metabolic activation (S9). FIG.6C depicts ungrouped multitask Ames QSAR algorithm (uMTL), trained to predict all strainspecific features, with or without metabolic activation (S9). FIG.6D depicts grouped multitasked Ames QSAR algorithm (gMTL), trained to predict subsets such as substitution strains-specific features with or without metabolic activation (S9).

[0045] FIG. 7 depicts sensitivity on the OOD test set estimated by stratified bootstrap resampling (n = 1,000). Points denote macro-averaged sensitivity and bars indicate 90% bootstrap confidence intervals.

[0046] FIG. 8 depicts confusion matrices for AmesNet, STL, uMTL and gMTL models on OOD mutagenicity predictions (n=4,644), showing true negatives, false positives, true positives, and false negatives (clockwise from top-left).

[0047] FIG. 9 depicts enriched substructures in AmesNet true positives vs. unconditioned models false negatives. FIG. 9 depicts top four mutagenic substructures most frequently observed in compounds correctly predicted by AmesNet but misclassified by unconditioned baseline models (gMTL, uMTL, STL). Values (n) indicate the number of unique compounds where AmesNet achieved true positives while comparison models achieved false negatives. AmesNet demonstrated recognition of established mutagenic substructures including aromatic amines, planar aromatic intercalators, nitro-aromatics, and epoxides.

[0048] FIG. 10 depicts sensitivity on the OOD test set estimated by stratified bootstrap resampling (n = 1,000). Points denote macro-averaged sensitivity and bars indicate 90% bootstrap confidence intervals.

[0049] FIG. 11 depicts confusion matrices for AmesNet, ChemProp and GROVER models on OOD mutagenicity predictions (n=4,644), showing true negatives, false positives, true positives, and false negatives (clockwise from top-left).

[0050] FIG. 12 depicts enriched substructures in AmesNet true positives vs. encoderswap models false negatives. FIG. 12 depicts top four mutagenic substructures most frequently observed in compounds correctly predicted by AmesNet but misclassified by encoder-swap baseline models (ChemProp, GROVER). Values (n) indicate the number of unique compounds where AmesNet achieved true positives while comparison models achieved false negatives. AmesNet demonstrated recognition of established mutagenic substructures, including aromatic amines, nitro-aromatics, aromatic ring N-oxides, planar aromatic intercalators, and epoxides across alternative molecular encoding architectures.

[0051] FIG. 13 depicts the AmesNet TCL architecture. AmesNet is a dual branch architecture containing a graph convolutional neural network and an Ames task conditioningchannel conditioned on strain and ±S9.

[0052] FIG. 14 depicts sample-size-weighted sensitivity on the fixed OOD test set. Error bars indicate 95% confidence intervals from a within-task stratified bootstrap (n = 1,000) across the 16 Ames tasks.

[0053] FIG. 15 depicts sample-size-weighted balanced accuracy on the fixed OOD test set. Error bars indicate 95% confidence intervals from a within-task stratified bootstrap (n = 1,000) across the 16 Ames tasks.

[0054] FIG. 16 depicts confusion matrices for AmesNet, gMTL-MLP, uMTL-MLP, STL-MLP, STL-ChemProp, STL-DeepAmes, and STL-GROVER on OOD mutagenicity predictions (n=4,528), showing true negatives, false positives, true positives, and false negatives (clockwise from top-left).

[0055] FIG. 17 depicts enriched substructures in AmesNet true positives vs. unconditioned models’ false negatives. FIG. 17 depicts mutagenic substructures observed in compounds correctly predicted by AmesNet but misclassified by unconditioned baseline models (gMTL-MLP, uMTL-MLP, STL-MLP, STL-ChemProp, STL-DeepAmes, STL-GROVER). Values (n) indicate the number of unique substructure-containing compounds where AmesNet achieved true positives while comparison models achieved false negatives.

[0056] FIG. 18 depicts sample-size-weighted sensitivity on the fixed OOD test set. Error bars indicate 95% confidence intervals from a within-task stratified bootstrap (n = 1,000) across the 16 Ames tasks.

[0057] FIG. 19 depicts sample-size-weighted balanced accuracy on the fixed OOD test set. Error bars indicate 95% confidence intervals from a within-task stratified bootstrap (n = 1,000) across the 16 Ames tasks.

[0058] FIG. 20 depicts confusion matrices for AmesNet, TCL-ChemProp, and TCL-GROVER models on OOD mutagenicity predictions (n=4,528), showing true negatives, false positives, true positives, and false negatives (clockwise from top-left).

[0059] FIG. 21 depicts enriched substructures in AmesNet true positives vs. encoderswap TCL models false negatives. FIG. 21 depicts mutagenic substructures observed in compounds correctly predicted by AmesNet but misclassified by encoder-swap baseline models (optionally, TCL-ChemProp, and TCL-GROVER). Values (n) indicate the number of unique substructure-containing compounds where AmesNet achieved true positives while comparison models achieved false negatives.DETAILED DESCRIPTION

[0060] In the following detailed description, reference is made to the accompanying drawings, which form a part hereof. In the drawings, similar symbols typically identify similarcomponents, unless context dictates otherwise. The illustrative embodiments described in the detailed description, drawings, and claims are not meant to be limiting. Other embodiments may be utilized, and other changes may be made, without departing from the spirit or scope of the subject matter presented herein. It will be readily understood that the aspects of the present disclosure, as generally described herein, and illustrated in the Figures, can be arranged, substituted, combined, separated, and designed in a wide variety of different configurations, all of which are explicitly contemplated herein and made part of the disclosure herein.

[0061] References and citations to other documents, including but not limited to patents, published patent applications, publications, journals, books, papers, web contents, and databases, are made throughout this disclosure. All such documents are hereby incorporated herein by reference in their entirety for all purposes, including with respect to the technology described herein.Definitions

[0062] Unless defined otherwise, technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs. For purposes of the present disclosure, the following terms are defined below.

[0063] As used herein, “molecular structure representation” refers to a digital, symbolic, numerical, or other representation of a compound’s structure. Such representation can be in any form suitable for storage, transmission, processing, interpretation, computational analysis, or other computer-implemented use or manipulation. Molecular structure representations can comprise any machine-readable encodings of a compound or any of its molecular features, including, but not limited to, two-dimensional (2D) or three-dimensional (3D) structural formats, graph-based representations, text-based linear notations, vector or latent representations derived from molecular embeddings, or other encoded descriptors of molecular structure. An example of a molecular structure representation is a Simplified Molecular Input Line Entry System (SMILES) string.

[0064] As used herein, “Ames testing condition” refers to the set of experimental parameters under which an Ames bacterial reverse mutation assay is conducted or modeled. Such parameters can include, for example, the identity of the bacterial tester strain and the status of metabolic activation, such as the presence or absence of a mammalian metabolic activation system (e.g., an S9 liver fraction). Representative tester strains can include any one or combination selected from: Salmonella typhimurium (S. typhimurium) strains TA97, TA98, TA100, TA102, TA104, TA1535, TA1537, and / or TA1538, and / or Escherichia coli (E. coll) strains WP2 uvrA or WP2 uvrA(pKMlOl). Metabolic activation may be provided using an S9 fraction derived from induced mammalian liver preparations (e.g., rat liver S9). A distinct combination of tester strainand metabolic activation status constitutes a particular Ames testing condition. The Ames bacterial reverse mutation assay and representative testing parameters are described, for example, in Ames et al., Proc. Natl. Acad. Sci. USA 70:2281-2285 (1973); Mortelmans & Zeiger, Mutat. Res.455:29-60 (2000); OECD Test Guideline 471 (Bacterial Reverse Mutation Test); and ICH Guideline S2(R1) (Genotoxicity Testing and Data Interpretation for Pharmaceuticals Intended for Human Use), each of which is incorporated herein by reference.

[0065] As used herein, “metabolic activation” refers to a biological preparation used in Ames testing to simulate the metabolic processes of the mammalian liver. It is often derived from a crude liver homogenate (e.g., from rats or humans) and contains microsomal enzymes. In the present application, metabolic activation may be used interchangeably with “S9 fraction.”

[0066] As used herein, “adaptive molecular embedding vector” refers to a multidimensional representation of a compound’s molecular structure. An adaptive molecular embedding vector can be generated by a neural network having parameters obtained and / or updated through iterative development and / or optimization, for example via backpropagation or other gradient-based methods, such that the resulting representation adaptively captures structural information about the compound. This vector can capture structural information about a compound, including atomic-level, regional, and / or global features relevant to its properties or activity, such as mutagenicity.

[0067] As used herein, “Ames condition vector” refers to a numerical representation of one or more Ames testing conditions. It can be generated as a one-hot encoded vector that specifies, for example, the identity of a bacterial tester strain and the status of metabolic activation.

[0068] As used herein, “condition-specific embedding vector” refers to a numerical representation generated by concatenating or otherwise combining an adaptive molecular embedding vector of a compound with one or more Ames condition vectors. This vector encodes both the structure of the compound and the specific assay context.

[0069] As used herein, “sensitivity” refers to the proportion of actual positives that are correctly identified by the model, calculated as True Positives / (True Positives + False Negatives). For example, in a mutagenicity prediction context, sensitivity reflects the proportion of mutagenic compounds that are correctly predicted as mutagenic. High sensitivity reduces the likelihood of false negatives, such as mutagenic compounds being misclassified as non-mutagenic.

[0070] As used herein, “specificity” refers to the proportion of actual negatives that are correctly identified by the model, calculated as True Negatives / (True Negatives + False Positives). For example, in a mutagenicity prediction context, specificity reflects the proportion of non-mutagenic compounds that are correctly predicted as non-mutagenic. High specificity reduces the likelihood of false positives, such as non-mutagenic compounds being misclassifiedas mutagenic.

[0071] As used herein, “accuracy” refers to the proportion of all predictions that are correct, calculated as (True Positives + True Negatives) / (True Positives + True Negatives + False Positives + False Negatives). For example, in a mutagenicity prediction context, accuracy reflects the proportion of compounds correctly classified as mutagenic or non-mutagenic. High accuracy reflects a low overall rate of misclassification.

[0072] As used herein, “balanced accuracy” refers to the average of sensitivity and specificity, calculated as (Sensitivity + Specificity) / 2. High balanced accuracy reflects strong performance across both positive and negative classes, particularly where the dataset contains unequal numbers of mutagenic and non-mutagenic compounds.Methods

[0073] Disclosed herein is a method of assessing compound mutagenicity. The method comprises: (a) providing a compound or a molecular structure representation thereof; (b) providing (optionally, selecting) one or more Ames testing conditions; (c) combining (optionally, concatenating) an adaptive molecular embedding vector corresponding to the compound, with one or more Ames condition vectors corresponding to the one or more Ames testing conditions, to generate one or more condition-specific embedding vectors; (d) processing the one or more condition-specific embedding vectors (optionally, through a task-conditioned machine-learning model, such as a multilayer perceptron (MLP)) to generate a mutagenicity probability score for each of the one or more Ames testing conditions, thereby generating a profile of mutagenicity probability scores for the compound; and (e) determining the compound as mutagenic or non-mutagenic, or stratifying a mutagenicity risk of the compound, based on the profile of mutagenicity probability scores.

[0074] In some embodiments, the compound is a small molecule. In some embodiments, the method is performed on a plurality of compounds or molecular structure representations thereof. In some embodiments, the plurality of compounds is a plurality of small molecules. In some embodiments, the plurality of compounds comprises at least 100, 200, 300, 400, 500, 600, 700, 800, 900, or 1,000 compounds (optionally, small molecules). The compounds can each comprise any mutagenic or non-mutagenic structural features as disclosed herein. The molecular structure representation can be in any format as disclosed herein. In some embodiments, the molecular structure representation is a 2D / 3D representation or a text-based representation. In some embodiments, the molecular structure representation is a textual representation. In some embodiments, the molecular structure representation is a standardized textual representation. In some embodiments, the molecular structure representation is a linear textual representation. In some embodiments, the molecular structure representation is a standardized linear textualrepresentation. In some embodiments, the molecular structure representation is MOL / SDF, PDB, XYZ or Simplified Molecular Input Line Entry System (SMILES) representation. In some embodiments, the molecular structure representation is a SMILES string.

[0075] In some embodiments, the adaptive molecular embedding vector is generated from the molecular structure representation of the compound. In some embodiments, the adaptive molecular embedding vector is a multi-dimensional vector. In some embodiments, the adaptive molecular embedding vector comprises at least 64, 128, 256, 512, or 1,024 dimensions. In some embodiments, the adaptive molecular embedding vector comprises 512 dimensions.

[0076] In some embodiments, the adaptive molecular embedding vector comprises structural information of the compound. In some embodiments, the structural information comprises local, regional and / or global structural information. In some embodiments, the structural information comprises molecular geometric and / or topological features. The structural information of the compound can be atomic, motif, cross-motif, or whole-molecule level(s). In some embodiments, the structural information of the compound is at atomic level. In some embodiments, the structural information comprises structural information at one or more levels. For example, the structural information of the compound is at atomic and cross-motif levels.

[0077] In some embodiments, step (a) occurs prior to step (b). In some embodiments, step (a) occurs subsequent to step (b).

[0078] In some embodiments, providing one or more Ames testing conditions comprises selecting the Ames testing conditions. In some embodiments, the providing one or more Ames testing conditions comprises pre-determining the Ames testing conditions. The Ames testing conditions can comprise the strain of bacteria and the presence or absence of metabolic activation. The bacteria can be of any strains disclosed herein. The method can further comprise generating one or more Ames condition vectors. The one or more Ames condition vectors can be generated based on the selection or pre-determination of the Ames testing conditions. In some embodiments, the method further comprises providing the one or more Ames condition vectors, prior to (c).

[0079] In some embodiments, the method comprises generating the adaptive molecular embedding vector from a molecular structure representation of the compound. The adaptive molecular embedding vector can be generated using a neural network. The neural network can be convolutional neural network. In some embodiments, the method comprises generating, using a convolutional neural network, the adaptive molecular embedding vector from a molecular structure representation of the compound.

[0080] In some embodiments, generating the adaptive molecular embedding vector comprises iteratively aggregating structural information of the compound. The structuralinformation can comprise any structural features or information disclosed herein. The structural information can be from neighboring atoms, motifs or fragments. In some embodiments, (c) comprises iteratively aggregating structural information of the compound to generate the adaptive molecular embedding vector.

[0081] In some embodiments, convolutional neural network comprises a plurality of convolutional layers. In some embodiments, convolutional neural network comprises at least two, three, four or five layers. In some embodiments, convolutional neural network comprises a plurality of convolutional layers configured to iteratively aggregate structural information of the compound to generate the adaptive molecular embedding vector. Examples of such convolutional neural network can include LeNet-5, AlexNet, VGGNet, ResNet, DenseNet, and ConvNeXt.

[0082] In some embodiments, the method further comprises training the convolutional neural network prior to generating the adaptive molecular embedding vector. In some embodiments, the method further comprises training the convolutional neural network using a training dataset. In some embodiments, the training dataset comprises one or more reference compounds. In some embodiments, the one or more reference compounds comprises at least 100, 200, 300, 400, 500, 600, 700, 800, 900, or 1,000 compounds (optionally, small molecules). In some embodiments, the one or more reference compounds in the training dataset comprise one or more mutagenicity-associated molecular features. The mutagenicity-associated molecular features can be a known or suspected mutagenicity-associated molecular features. In some embodiments, the one or more reference compounds in the training dataset comprise one or more molecular features in the compounds to be assessed using the herein disclosed method. In some embodiments, the one or more reference compounds in the training dataset do not comprise any of the molecular features in the compounds to be assessed using the herein disclosed method. In some embodiments, the one or more reference compounds in the training dataset do not comprise any of the mutagenicity-associated molecular features in the compounds to be assessed using the herein disclosed method. In some embodiments, the one or more reference compounds in the training dataset do not comprise any of the mutagenicity-associated molecular features disclosed herein. In some embodiments, the one or more reference compounds in the training dataset comprise one or more mutagenicity-associated molecular features disclosed herein.

[0083] In some embodiments, the training dataset comprises information on mutagenicity of reference compounds under one or more Ames testing conditions. In some embodiments, the training dataset comprises information on mutagenicity of reference compounds in bacterial strains disclosed herein. In some embodiments, the training dataset comprises information on mutagenicity of reference compounds in Salmonella strain. In some embodiments, the training dataset comprises information on mutagenicity of reference compounds in thepresence and / or absence of metabolic activation.

[0084] In some embodiments, the training dataset is dimension reduced. Dimension reduction can be achieved, for example, using t-distributed stochastic neighbor embedding (t-SNE), uniform manifold approximation and projection (UMAP), isometric mapping, locally linear embedding (LLE), potential of heat-diffusion for affinity -based trajectory embedding (PHATE). In some embodiments, the training dataset is partitioned using t-SNE chemical diversity split.

[0085] In some embodiments, the method further comprises training the convolutional neural network using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions. In some embodiments, the convolutional neural network comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

[0086] In some embodiments, processing the one or more condition-specific embedding vectors comprises applying an ensemble of neural network models. In some embodiments, the ensemble comprises at least five, six, seven, eight, nine, or ten neural network models. In some embodiments, each neural network model in the ensemble processes the condition-specific embedding vectors to generate a set of mutagenicity probability scores, which are averaged across the ensemble to generate the profile of mutagenicity probability scores for the compound. In some embodiments, averaging the sets of mutagenicity probability scores is with respect to each Ames testing condition.

[0087] In some embodiments, processing the one or more condition-specific embedding vectors is through a task-conditioned machine-learning model, such as a multilayer perceptron (MLP). In some embodiments, the multilayer perceptron is trained prior to processing the one or more condition-specific embedding vectors. In some embodiments, training the multilayer perceptron comprises: (i) initializing a plurality of parameters of the multilayer perceptron; (ii) processing the training dataset through the multilayer perceptron to generate predicted mutagenicity probability scores; (iii) computing a loss function based on differences between the predicted mutagenicity probability scores and known mutagenicity classifications; (iv) updating the plurality of parameters of the multilayer perceptron to minimize the loss function; and (v) repeating steps (ii)-(iv). In some embodiments, optimizing parameters of the multilayer perceptron until the multilayer perceptron achieves an improved or substantially maximal value of sensitivity, specificity, accuracy, balanced accuracy, precision, recall or area under a receiver operating characteristic curve (AUROC). In some embodiments, optimization of the parameters is verified on a validation dataset.

[0088] In some embodiments, the multilayer perceptron is trained using a trainingdataset. In some embodiments, training the multilayer perceptron using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions. In some embodiments, the convolutional neural network and the multilayer perceptron are trained using the same dataset. In some embodiments, the convolutional neural network and the multilayer perceptron are trained using different datasets. In some embodiments, the training dataset is partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split. In some embodiments, the multilayer perceptron comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

[0089] In some embodiments, the convolutional neural network and the multilayer perceptron are jointly trained. In some embodiments, the convolutional neural network and the multilayer perceptron are separately trained. In some embodiments, the convolutional neural network and the multilayer perceptron are sequentially or iteratively trained, optionally wherein the convolutional neural network is trained prior to or after training the multilayer perceptron.

[0090] In some embodiments, the method further comprises identifying one or more mutagenicity-associated molecular features in the compound. In some embodiments, the one or more mutagenicity-associated molecular features identified is not in the training dataset. In some embodiments, the one or more mutagenicity-associated molecular features identified is in the training dataset. In some embodiments, the method further comprises providing a report that describes the identified one or more mutagenicity-associated molecular features as present or absent in the compound.

[0091] In some embodiments, the method further comprises providing a report that describes the determination or stratification of the compound made in step (e). In some embodiments, the method further comprises displaying on a user interface connected to an output device, a report that describes the determination or stratification of the compound made in step (e).

[0092] In some embodiments, the method further comprises verifying a determination, optionally, a mutagenic or non-mutagenic determination. In some embodiments, the verification is by using another model. In some embodiments, the verification is by conducting the Ames test. In some embodiments, the verification comprises testing the compound under at least one of the profiled Ames testing conditions. In some embodiments, the verification comprises testing the compound under each of the profiled Ames testing conditions.Systems

[0093] Disclosed in certain embodiments include a system comprising: one or moreprocessors, and memory storing a convolutional neural network (optionally, a graph convolutional neural network), a task-conditioned machine-learning model (optionally, a multilayer perceptron), and a plurality of lines of instructions; wherein the one or more processors are configured to execute the plurality of lines of instructions to perform the methods disclosed herein.

[0094] Disclosed includes a system comprising: one or more processors, and memory storing a task-conditioned machine-learning model (optionally, a multilayer perceptron) and a plurality of lines of instructions; wherein the one or more processors are configured to execute the plurality of lines of instructions to: (a) receive a compound or a molecular structure representation thereof; (b) receiving one or more Ames testing conditions; (c) combine (optionally, concatenate) an adaptive molecular embedding vector corresponding to the compound, with one or more Ames condition vectors corresponding to the one or more Ames testing conditions, to generate one or more condition-specific embedding vectors; (d) process the one or more condition-specific embedding vectors (optionally, through a task-conditioned machine-learning model, such as a multilayer perceptron (MLP)) to generate a mutagenicity probability score for each of the one or more Ames testing conditions, thereby generating a profile of mutagenicity probability scores for the compound; and (e) determine the compound as mutagenic or non-mutagenic, or stratify a mutagenicity risk of the compound, based on the profile of mutagenicity probability scores.

[0095] In some embodiments, the one or more processors are configured, in (a), to receive a molecular structure representation of the compound. In some embodiments, the compound is a small molecule. In some embodiments, the one or more processors are configured to perform the method on a plurality of compounds or molecular structure representations thereof. In some embodiments, the plurality of compounds is a plurality of small molecules. In some embodiments, the plurality of compounds comprises at least 100, 200, 300, 400, 500, 600, 700, 800, 900, or 1,000 compounds (optionally, small molecules). The compounds can comprise any mutagenic or non-mutagenic structural features as disclosed herein. The molecular structure representation can be in any format as disclosed herein. In some embodiments, the molecular structure representation is a 2D / 3D representation or a text-based representation. In some embodiments, the molecular structure representation is a textual representation. In some embodiments, the molecular structure representation is a standardized textual representation. In some embodiments, the molecular structure representation is a linear textual representation. In some embodiments, the molecular structure representation is a standardized linear textual representation.

[0096] In some embodiments, the adaptive molecular embedding vector is generated from the molecular structure representation of the compound. In some embodiments, the adaptive molecular embedding vector is a multi-dimensional vector. In some embodiments, the adaptivemolecular embedding vector comprises at least 64, 128, 256, 512, or 1,024 dimensions. In some embodiments, the adaptive molecular embedding vector comprises 512 dimensions.

[0097] In some embodiments, the adaptive molecular embedding vector comprises structural information of the compound. In some embodiments, the structural information comprises local, regional and / or global structural information. In some embodiments, the structural information comprises molecular geometric and / or topological features. The structural information of the compound can be atomic, motif, cross-motif, or whole-molecule level(s). In some embodiments, the structural information of the compound is at atomic level. In some embodiments, the structural information comprises structural information at one or more levels. For example, the structural information of the compound is at atomic and cross-motif levels.

[0098] In some embodiments, the system is configured to allow users to select the Ames testing conditions. In some embodiments, the system is configured to store pre-determining the Ames testing conditions. The Ames testing conditions can comprise the strain of bacteria and the presence or absence of metabolic activation. The bacteria can be of any strains disclosed herein. In some embodiments, the one or more processors are configured to provide, prior to step (c). The one or more Ames condition vectors can be generated based on the selection or predetermination of the Ames testing conditions.

[0099] In some embodiments, the memory further stores a convolutional neural network. In some embodiments, the one or more processors are configured to execute the plurality of lines of instructions to generate, using the convolutional neural network, the adaptive molecular embedding vector from a molecular structure representation of the compound.

[0100] In some embodiments, the convolutional neural network is configured to iteratively aggregate structural information of the compound to generate the adaptive molecular embedding vector. In some embodiments, the structural information comprises one or more mutagenicity-associated molecular features of the compound. The structural information can comprise any structural features or information disclosed herein. The structural information can be from neighboring atoms, motifs or fragments.

[0101] In some embodiments, convolutional neural network comprises a plurality of convolutional layers. In some embodiments, convolutional neural network comprises at least two, three, four or five layers. In some embodiments, convolutional neural network comprises a plurality of convolutional layers configured to iteratively aggregate structural information of the compound to generate the adaptive molecular embedding vector. Examples of such convolutional neural network can include LeNet-5, AlexNet, VGGNet, ResNet, DenseNet, and ConvNeXt.

[0102] In some embodiments, the one or more processors are configured to train the convolutional neural network prior to generating the adaptive molecular embedding vector. Insome embodiments, the one or more processors are configured to train the convolutional neural network using a training dataset. In some embodiments, the training dataset comprises one or more reference compounds. In some embodiments, the one or more reference compounds comprises at least 100, 200, 300, 400, 500, 600, 700, 800, 900, or 1,000 compounds (optionally, small molecules). In some embodiments, the one or more reference compounds in the training dataset comprise one or more mutagenicity-associated molecular features. The mutagenicity-associated molecular features can be a known or suspected mutagenicity-associated molecular features. In some embodiments, the one or more reference compounds in the training dataset comprise one or more molecular features in the compounds to be assessed using the herein disclosed method. In some embodiments, the one or more reference compounds in the training dataset do not comprise any of the molecular features in the compounds to be assessed using the herein disclosed method. In some embodiments, the one or more reference compounds in the training dataset do not comprise any of the mutagenicity-associated molecular features in the compounds to be assessed using the herein disclosed method. In some embodiments, the one or more reference compounds in the training dataset do not comprise any of the mutagenicity-associated molecular features disclosed herein. In some embodiments, the one or more reference compounds in the training dataset comprise one or more mutagenicity-associated molecular features disclosed herein.

[0103] In some embodiments, the training dataset comprises information on mutagenicity of reference compounds under one or more Ames testing conditions. In some embodiments, the training dataset comprises information on mutagenicity of reference compounds in bacterial strains disclosed herein. In some embodiments, the training dataset comprises information on mutagenicity of reference compounds in Salmonella strain. In some embodiments, the training dataset comprises information on mutagenicity of reference compounds in the presence and / or absence of metabolic activation.

[0104] In some embodiments, the training dataset is dimension reduced. Dimension reduction can be achieved, for example, using t-distributed stochastic neighbor embedding (t-SNE), uniform manifold approximation and projection (UMAP), isometric mapping, locally linear embedding (LLE), potential of heat-diffusion for affinity -based trajectory embedding (PHATE). In some embodiments, the training dataset is partitioned using t-SNE chemical diversity split.

[0105] In some embodiments, the one or more processors are configured to train the convolutional neural network using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions. In some embodiments, the convolutional neural network comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

[0106] In some embodiments, processing the one or more condition-specific embedding vectors comprises applying an ensemble of neural network models. In some embodiments, the ensemble comprises at least five, six, seven, eight, nine, or ten neural network models. In some embodiments, each neural network model in the ensemble processes the condition-specific embedding vectors to generate a set of mutagenicity probability scores, which are averaged across the ensemble to generate the profile of mutagenicity probability scores for the compound. In some embodiments, averaging the sets of mutagenicity probability scores is with respect to each Ames testing condition.

[0107] In some embodiments, the one or more processors are configured to process the one or more condition-specific embedding vectors using a task-conditioned machine-learning model, such as a multilayer perceptron (MLP)). In some embodiments, the one or more processors are configured to train the multilayer perceptron prior to processing the one or more conditionspecific embedding vectors. In some embodiments, the one or more processors are configured to train the multilayer perceptron that comprises: (i) initializing a plurality of parameters of the multilayer perceptron; (ii) processing the training dataset through the multilayer perceptron to generate predicted mutagenicity probability scores; (iii) computing a loss function based on differences between the predicted mutagenicity probability scores and known mutagenicity classifications; (iv) updating the plurality of parameters of the multilayer perceptron to minimize the loss function; and (v) repeating steps (ii)-(iv). In some embodiments, the one or more processors are configured to optimize parameters of the multilayer perceptron until the multilayer perceptron achieves an improved or substantially maximal value of sensitivity, specificity, accuracy, balanced accuracy, precision, recall or area under a receiver operating characteristic curve (AUROC). In some embodiments, optimization of the parameters is verified on a validation dataset.

[0108] In some embodiments, the multilayer perceptron is configured to be trained using a training dataset. In some embodiments, the multilayer perceptron is configured to be trained using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions. In some embodiments, the convolutional neural network and the multilayer perceptron are configured to be trained using the same dataset. In some embodiments, the convolutional neural network and the multilayer perceptron are configured to be trained using different datasets. In some embodiments, the training dataset is partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split. In some embodiments, the multilayer perceptron comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity

[0109] In some embodiments, the convolutional neural network and the multilayer perceptron are jointly trained. In some embodiments, the convolutional neural network and the multilayer perceptron are separately trained. In some embodiments, the convolutional neural network and the multilayer perceptron are sequentially or iteratively trained, optionally wherein the convolutional neural network is trained prior to or after training the multilayer perceptron.

[0110] In some embodiments, the one or more processors are configured to identify one or more mutagenicity-associated molecular features in the compound. In some embodiments, the one or more mutagenicity-associated molecular features identified is not in the training dataset. In some embodiments, the one or more mutagenicity-associated molecular features identified is in the training dataset. In some embodiments, the one or more processors are configured to provide a report that describes the identified one or more mutagenicity-associated molecular features as present or absent in the compound.[OHl] In some embodiments, the one or more processors are configured to provide a report that that describes the determination or stratification of the compound made in (e). In some embodiments, the system further comprises an output device, and the one or more processors are configured to display on the output device a report that that describes the determination or stratification of the compound made in (e).

[0112] In some embodiments, the one or more processors are configured to verify a determination, optionally, a mutagenic or non-mutagenic determination. In some embodiments, the one or more processors are configured to test the compound under at least one of the profiled Ames testing conditions. In some embodiments, the one or more processors are configured to test the compound under each of the profiled Ames testing conditions.Mutagenicity- Associated Molecular Features

[0113] In some embodiments of the methods and systems provided herein, the structural information comprises one or more mutagenicity-associated molecular features of compound(s). A mutagenicity-associated molecular feature can be a chemical, geometric, topological, or structural feature that is identified, determined, or suspected to be associated or correlate with mutagenicity or non-mutagenicity. In some embodiments, the mutagenicity-associated molecular feature is a feature known to be associated or correlate with mutagenicity. In some embodiments, the mutagenicity-associated molecular feature is feature determined to be associated or correlated with mutagenicity by an Ames test. In some embodiments, the Ames test is carried out under one or more Ames testing conditions disclosed herein. In some embodiments, a molecular feature of the compound to be tested is comprised in one or more reference compounds in the training dataset. In some embodiments, a molecular feature in the compound tobe tested is not comprised of in one or more reference compounds in the training dataset. In some embodiments, the molecular feature associated or correlated with mutagenicity identified by the method or system disclosed herein comprises molecular features identified by a reference model. In some embodiments, the molecular feature associated or correlated with mutagenicity identified by the method or system disclosed herein comprises molecular features not identified by a reference model. The reference model can be a single-task learning (STL), a ungrouped multitask learning (uMTL) or a grouped multitask learning (gMTL). In some embodiments, the reference model is ChemProp, DeepAmes or GROVER.

[0114] In some embodiments of the methods and systems provided herein, a molecular feature associated or correlated with mutagenicity is selected from 1,2-dihydroxy aromatic ring, aliphatic halogen, alpha, beta-unsaturated carbonyl, aromatic amine, aromatic N-heterocycle, aromatic ring N-oxide, aryl halide, azo group, epoxide, nitro-aromatic, nitroso group, planar aromatic intercalator, polyfunctional alkylator.

[0115] In some embodiments of the methods and systems provided herein, the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, six, seven, eight, nine, or ten features each independently selected from: a polyaromatic moiety (optionally, a planar polyaromatic moiety), an aromatic amine moiety, an aromatic N-heterocycle moiety, an a,P-unsaturated carbonyl moiety, an epoxide moiety, a nitro-aromatic moiety, an aromatic ring N-oxide moiety, an azo moiety, an aliphatic halogen moiety, a 1,2-dihydroxy aromatic ring moiety, a polyfunctional alkyl moiety, an aryl halide moiety, and a nitroso moiety.

[0116] In some embodiments of the methods and systems provided herein, the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, six, seven, eight, nine, or ten features each independently selected from: Aziridine, Acrylate / Acrylamide, Alpha-methylene lactone, Epoxide (oxirane), Quinoline, Aromatic imine (Schiff base), Imine (aliphatic C=N), Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Poly functional alkyl: bis-haloalkyl, Tertiary aromatic amine (N,N-dialkyl), Non-aromatic C=C fused / adjacent to aromatic ring, Cyclic dithioacetal / 1,3-dithiane, Bay region (phenanthrene-type angular fusion), Naphthalene, Alpha-diketone (1,2-dicarbonyl), K-region (anthracene-type linear tri-ring fusion), Primary aromatic amine, Aromatic nitro group, Aromatic azo compound, 1 -Aminonaphthalene, Aromatic amine N-oxide, Phenol (general), Diarylamine (Ar-NH-Ar), Coumarin / Benzo-alpha-pyrone, Peroxide, Hydroperoxide, Benzylic methyl / methylene, meta-Diaminoarene, Triazene, Imidazole (general), Acridine scaffold, Quinolone (4-oxoquinoline), Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Aryl halide (single), Diaryl thioether, Benzylic primary alcohol, Pyrrole ring, Indole, Thiazole ring, Alpha, beta-unsaturated ketone (enone), Alpha, beta-unsaturated aldehyde, Aldehyde, Enol ether,Aromatic secondary amine, Nitrofuran, Furan ring, Gem-dihaloalkane, Benzylic secondary alcohol, N-Nitrosamine (dialkyl), Nitroalkene (nitrostyrene-type), Naphthyridine, Thiophene ring, Catechol (1,2-dihydroxybenzene), Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Diazo compound, Anthraquinone, Alpha, beta-unsaturated nitrile, 2-Aminonaphthalene, Phosphoryl halide, Alpha-methylene lactam, para-Aminophenol, and Phenol with para electron-donating group.

[0117] In some embodiments of the methods and systems provided herein, the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, six, seven, eight, nine, or ten features each independently selected from: Naphthalene, Benzylic secondary alcohol, Catechol (1,2-dihydroxybenzene), Phenol (general), Anthraquinone, Benzylic primary alcohol, Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Non-aromatic C=C fused / adjacent to aromatic ring, Epoxide (oxirane), Acridine scaffold, Quinoline, Diazo compound, Imine (aliphatic C=N), Alpha, beta-unsaturated nitrile, Enol ether, Tertiary aromatic amine (N,N-dialkyl), Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Polyfunctional alkyl: bis-haloalkyl, Acrylate / Acrylamide, Alpha-methylene lactone, Aromatic nitro group, Aromatic amine N-oxide, Imidazole (general), Alpha, beta-unsaturated ketone (enone), Aromatic imine (Schiff base), Primary aromatic amine, Aromatic azo compound, 2-Aminonaphthalene, Phosphoryl halide, Cyclic dithioacetal / 1,3-dithiane, Alpha-diketone (1,2-dicarbonyl), Thiazole ring, Bay region (phenanthrene-type angular fusion), K-region (anthracene-type linear tri-ring fusion), Gem-dihaloalkane, Aziridine, Pyrrole ring, Indole, Alpha-methylene lactam, 1 -Aminonaphthalene, Coumarin / Benzo-alpha-pyrone, N-Nitrosamine (dialkyl), Methylenedioxybenzene (safrole-type), Naphthyridine, Hydrazine, Aromatic hydrazine (phenylhydrazine-type), Quinolone (4-oxoquinoline), Aromatic secondary amine, Benzylic methyl / methylene, Nitrofuran, Furan ring, Hydrazone, Thioether on aromatic ring, Aryl halide (single), meta-Diaminoarene, Thiophene ring, and Triazene.

[0118] In some embodiments of the methods and systems provided herein, the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, six, seven, eight, nine, or ten features each independently selected from: Aziridine, Aromatic nitro group, Alpha, beta-unsaturated ketone (enone), Aromatic amine N-oxide, Non-aromatic C=C fused / adjacent to aromatic ring, Imidazole (general), Aromatic imine (Schiff base), Imine (aliphatic C=N), Cyclic dithioacetal / 1,3-dithiane, Nitrofuran, Furan ring, Hydrazone, Epoxide (oxirane), Bay region (phenanthrene-type angular fusion), Naphthalene, Alpha-diketone (1,2-dicarbonyl), K-region (anthracene-type linear tri-ring fusion), Primary aromatic amine, Aromatic azo compound, 1 -Aminonaphthalene, Phenol (general), Alpha, beta-unsaturated aldehyde, Aldehyde, Acetal / Hemiacetal (potential aldehyde release), Peroxide, Hydroperoxide, meta-Diaminoarene, Thioether on aromatic ring, Triazene, Acridine scaffold, Quinoline, Quinolone (4-oxoquinoline), Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Pyrrole ring, Indole, Thiazole ring, Enol ether, Methylenedi oxybenzene (safrole-type), Aromatic secondary amine, Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Polyfunctional alkyl: bis-haloalkyl, Acrylate / Acrylamide, Alpha-methylene lactone, Tertiary aromatic amine (N,N-dialkyl), N-Nitrosamine (dialkyl), Nitroalkene (nitrostyrene-type), Naphthyridine, ortho-Aminophenol, Phenazine scaffold, Coumarin / Benzo-alpha-pyrone, Anthraquinone, Thiophene ring, Catechol (1,2-dihydroxybenzene), Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Diazo compound, Benzylic primary alcohol, Alpha, beta-unsaturated nitrile, 2-Aminonaphthalene, Phosphoryl halide, Alpha-methylene lactam, Hydrazine, Aromatic hydrazine (phenylhydrazine-type), para-Aminophenol, Phenol with para electron-donating group, Hydroquinone (1,4-dihydroxybenzene), and para-Diaminoarene (non-biphenyl).

[0119] In some embodiments of the methods and systems provided herein, the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, six, seven, eight, nine, or ten features each independently selected from: Aziridine, Acrylate / Acrylamide, Alpha-methylene lactone, Aromatic nitro group, Alpha, beta-unsaturated ketone (enone), Aromatic amine N-oxide, Non-aromatic C=C fused / adjacent to aromatic ring, Aromatic imine (Schiff base), Imine (aliphatic C=N), Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Poly functional alkyl: bis-haloalkyl, Tertiary aromatic amine (N,N-dialkyl), Cyclic dithioacetal / 1,3-dithiane, Alpha-diketone (1,2-dicarbonyl), Naphthalene, K-region (anthracene-type linear tri-ring fusion), Primary aromatic amine, Aromatic azo compound, 1-Aminonaphthalene, Phenol (general), Alpha, beta-unsaturated aldehyde, Aldehyde, Acetal / Hemiacetal (potential aldehyde release), Diarylamine (Ar-NH-Ar), Peroxide, Hydroperoxide, Benzylic methyl / methylene, meta-Diaminoarene, Thioether on aromatic ring, Acridine scaffold, Quinoline, Quinolone (4-oxoquinoline), Hydroquinone (1,4-dihydroxybenzene), Phenol with para electron-donating group, Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Epoxide (oxirane), Aryl halide (single), Diaryl thioether, Benzylic primary alcohol, Pyrrole ring, Indole, Thiazole ring, Imidazole (general), Enol ether, N-Acyloxy group, Methylenedioxybenzene (safrole-type), Aromatic secondary amine, Nitrofuran, Furan ring, Hydrazone, Anthraquinone, N-Nitrosamine (dialkyl), Naphthyridine, ortho-Aminophenol, Phenazine scaffold, Coumarin / Benzo-alpha-pyrone, Alkyl sulfonate ester, Propane sultone, Tosylate / Mesylate / Tritiate ester, Thiophene ring, Triazene, Catechol (1,2-dihydroxybenzene), Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Diazo compound, Alpha, beta-unsaturated nitrile, 2-Aminonaphthalene, Phosphoryl halide, Benzylic secondary alcohol, Bay region (phenanthrene-type angular fusion), Alpha-methylene lactam, and para-Aminophenol.

[0120] In some embodiments of the methods and systems provided herein, the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, six, seven, eight, nine, or ten features each independently selected from: Aziridine, Acrylate / Acrylamide, Alpha-methylene lactone, Aromatic nitro group, Alpha, beta-unsaturated ketone (enone), Aromatic amine N-oxide, Non-aromatic C=C fused / adjacent to aromatic ring, Imidazole (general), Aromatic imine (Schiff base), Imine (aliphatic C=N), Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Poly functional alkyl: bis-haloalkyl, Tertiary aromatic amine (N,N-dialkyl), Cyclic dithioacetal / 1,3-dithiane, Epoxide (oxirane), Bay region (phenanthrene-type angular fusion), Naphthalene, Alpha-diketone (1,2-dicarbonyl), K-region (anthracene-type linear tri-ring fusion), Primary aromatic amine, Aromatic azo compound, 1-Aminonaphthalene, Phenol (general), Alpha, beta-unsaturated aldehyde, Aldehyde, Acetal / Hemiacetal (potential aldehyde release), Diarylamine (Ar-NH-Ar), Coumarin / Benzo-alpha-pyrone, Benzylic methyl / methylene, Peroxide, Hydroperoxide, meta-Diaminoarene, Thioether on aromatic ring, Triazene, Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Aryl halide (single), Diaryl thioether, Benzylic primary alcohol, Pyrrole ring, Indole, Thiazole ring, Enol ether, Benzylic secondary alcohol, N-Acyloxy group, Methylenedi oxybenzene (safrole-type), Aromatic secondary amine, Acridine scaffold, Quinoline, Nitrofuran, Furan ring, Hydrazone, N-Nitrosamine (dialkyl), Nitroalkene (nitrostyrene-type), ortho-Aminophenol, Phenazine scaffold, Anthraquinone, Alkyl sulfonate ester, Propane sultone, Tosylate / Mesylate / Tritiate ester, 2-Aminonaphthalene, Thiophene ring, Catechol (1,2-dihydroxybenzene), Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Diazo compound, Alpha, beta-unsaturated nitrile, Phosphoryl halide, Alpha-methylene lactam, para-Aminophenol, Phenol with para electrondonating group, Hydroquinone (1,4-dihydroxybenzene), and para-Diaminoarene (non-biphenyl).

[0121] In some embodiments of the methods and systems provided herein, the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, six, seven, eight, nine, or ten features each independently selected from: Aziridine, Aromatic nitro group, Alpha, beta-unsaturated ketone (enone), Aromatic amine N-oxide, Non-aromatic C=C fused / adjacent to aromatic ring, Aromatic imine (Schiff base), Imine (aliphatic C=N), Primary aromatic amine, Aromatic azo compound, 2-Aminonaphthalene, Naphthalene, Phenol (general), Cyclic dithioacetal / 1,3-dithiane, Epoxide (oxirane), Bay region (phenanthrene-type angular fusion), Alpha-diketone (1,2-dicarbonyl), K-region (anthracene-type linear tri-ring fusion), Coumarin / Benzo-alpha-pyrone, Thiazole ring, Peroxide, Hydroperoxide, Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Polyfunctional alkyl: bis-haloalkyl, Tertiary aromatic amine (N,N-dialkyl), Benzylic methyl / methylene, 1 -Aminonaphthalene, meta-Diaminoarene, Thioether on aromatic ring, Triazene, Imidazole (general), Aromatic N-oxide,Pyridinium / Quaternary aromatic nitrogen, Pyrrole ring, Indole, Quinoline, Quinolone (4-oxoquinoline), Alpha, beta-unsaturated aldehyde, Aldehyde, Enol ether, Methylenedi oxybenzene (safrole-type), Aromatic secondary amine, Acridine scaffold, Benzylic secondary alcohol, Acrylate / Acrylamide, Alpha-methylene lactone, N-Nitrosamine (dialkyl), Nitroalkene (nitrostyrene-type), Naphthyridine, ortho-Aminophenol, Phenazine scaffold, Anthraquinone, Alkyl sulfonate ester, Propane sultone, Tosylate / Mesylate / Tritiate ester, Thiophene ring, Catechol (1,2-dihydroxybenzene), Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Benzylic primary alcohol, Alpha, beta-unsaturated nitrile, Phosphoryl halide, Alpha-methylene lactam, para- Aminophenol, Phenol with para electron-donating group, Hydroquinone (1,4-dihydroxybenzene), and para-Diaminoarene (non-biphenyl).

[0122] In some embodiments of the methods and systems provided herein, the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, six, seven, eight, nine, or ten features each independently selected from: Epoxide (oxirane), Acrylate / Acrylamide, Alpha-methylene lactone, Non-aromatic C=C fused / adjacent to aromatic ring, Bay region (phenanthrene-type angular fusion), Naphthalene, Enol ether, K-region (anthracene-type linear tri-ring fusion), Cyclic dithioacetal / 1,3-dithiane, Alpha-methylene lactam, Phenol (general), N-Acyloxy group, Aromatic azo compound, Aryl halide (single), Tertiary aromatic amine (N,N-dialkyl), Thiazole ring, Peroxide, Hydroperoxide, Alpha, beta-unsaturated aldehyde, Aldehyde, Acetal / Hemiacetal (potential aldehyde release), Alpha, beta-unsaturated nitrile, Alpha, beta-unsaturated ketone (enone), Aromatic secondary amine, Aromatic nitro group, Nitrofuran, Furan ring, Aromatic amine N-oxide, Hydrazone, Alpha-diketone (1,2-dicarbonyl), Aromatic N-oxide, Diaryl thioether, Benzylic primary alcohol, Pyridinium / Quaternary aromatic nitrogen, Primary alkyl halide, Half-mustard (2-chloroethylamine), Acridine scaffold, Quinoline, Benzylic methyl / methylene, Thiophene ring, Hydroquinone (1,4-dihydroxybenzene), Phenol with para electron-donating group, Anthraquinone, Quinolone (4-oxoquinoline), Naphthyridine, Imidazole (general), Pyrrole ring, Indole, Coumarin / Benzo-alpha-pyrone, Methylenedioxybenzene (safrole-type), N-Nitrosamine (dialkyl), Nitrogen mustard, Polyfunctional alkyl: bis-haloalkyl, Benzylic secondary alcohol, Triazene, Diarylamine (Ar-NH-Ar), Primary aromatic amine, 2-Aminonaphthalene, Aziridine, Hydrazine, Aromatic hydrazine (phenylhydrazine-type), 1 -Aminonaphthalene, para- Aminophenol, ortho-Aminophenol, Phenazine scaffold, para-Diaminoarene (non-biphenyl), meta-Diaminoarene, Thioether on aromatic ring, Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Aromatic imine (Schiff base), Imine (aliphatic C=N), Gem-dihaloalkane, Phosphoryl halide, Alkyl sulfonate ester, Propane sultone, Tosylate / Mesylate / Tritiate ester, Nitroalkene (nitrostyrene-type), Catechol (1,2-dihydroxybenzene), and Diazo compound.

[0123] In some embodiments of the methods and systems provided herein, the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, six, seven, eight, nine, or ten features each independently selected from: Cyclic dithioacetal / 1,3-dithiane, Epoxide (oxirane), Acrylate / Acrylamide, Alpha-methylene lactam, Alpha, betaunsaturated nitrile, Enol ether, Tertiary aromatic amine (N,N-dialkyl), Non-aromatic C=C fused / adjacentto aromatic ring, Alpha-diketone (1,2-dicarbonyl), Thiophene ring, N-Nitrosamine (dialkyl), Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Polyfunctional alkyl: bis-haloalkyl, Alpha-methylene lactone, Triazene, Primary aromatic amine, Aromatic azo compound, 2-Aminonaphthalene, Naphthalene, Phenol (general), Aromatic imine (Schiff base), Imine (aliphatic C=N), Phosphoryl halide, Bay region (phenanthrene-type angular fusion), Catechol (1,2-dihydroxybenzene), Imidazole (general), Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, and Thiazole ring.

[0124] In some embodiments of the methods and systems provided herein, the one or more mutagenicity-associated molecular features comprise at least one, two, three, four, five, six, seven, eight, nine or ten features disclosed in Tables 1A-1H, or any portion thereof. In some, the one or more mutagenicity-associated molecular features comprise at least one, two, three, four, five, six, seven, eight, nine or ten features disclosed in any one of Table 1A. In some, the one or more mutagenicity-associated molecular features comprise at least one, two, three, four, five, six, seven, eight, nine or ten features disclosed in any one of Table IB. In some, the one or more mutagenicity-associated molecular features comprise at least one, two, three, four, five, six, seven, eight, nine or ten features disclosed in any one of Table 1C. In some, the one or more mutagenicity-associated molecular features comprise at least one, two, three, four, five, six, seven, eight, nine or ten features disclosed in any one of Table ID. In some, the one or more mutagenicity-associated molecular features comprise at least one, two, three, four, five, six, seven, eight, nine or ten features disclosed in any one of Table IE. In some, the one or more mutagenicity-associated molecular features comprise at least one, two, three, four, five, six, seven, eight, nine or ten features disclosed in any one of Table IF. In some, the one or more mutagenicity-associated molecular features comprise at least one, two, three, four, five, six, seven, eight, nine or ten features disclosed in any one of Table 1G. In some, the one or more mutagenicity-associated molecular features comprise at least one, two, three, four, five, six, seven, eight, nine or ten features disclosed in any one of Table 1H.

[0125] Table 1A. Example compounds correctly characterized by AmesNet but not by ChemPropExample compounds missed from ChemProp Example compounds missed from ChemProp clccc([C@H]2N[C@@H]2c2ccccc2)ccl Cl=Cc2ccc3c4c(ccclc24)-clcccccl-3C=C / C 1=C / [C@H] (C) [C@] (C)(OC(C)=O)C(=O)O O=Clc2ccccc2-c2cc3ccccc3cc21 CC2=CC[N+]3(C)CC[C@@H](OC1=O)C23[O-] [O]S(=O)(=O)clccc(Oc2ccc3c4c(cccc24)Cc2ccccc2 clcnc2ccc3c(c2cl)C10ClCC3 -3)cclO=C 1 CC(=Nc2ccccc2)C(=O)c2ccccc21 clccc2cc(C3 C03 )ccc2c 1 C[C@]12CC[C@@H](OC(=O) / C=C / c3cccc(N(CCC O=[N+]([O-])clccc2c(cl)CCclcc3c(ccl-2)CCCC3 l)CCCl)c3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[ clccc([C@H]2N[C@H]2c2ccccc2)ccl C@]2(C)NC(=O)CC[C@@H] 12 clccc2c(cl)-clc(c3ccccc3c3cccccl3)ClNC21 C1CSCSC1 CCN(CC)CCnlnc2c3c(c(CO)ccc31)Sclcc(Cl)cccl-2 Cl=CC2OC2c2cc3ccc4cccc5ccc(c21)c3c45 Nnlnnc2c([nH]c3ccccc32)cl=O CC1OCC(=O)C(=O)C1O CCnlcc(C(=0)0)c(=0)c2cc(F)c(- Cl=Cc2c3ccccc3cc3cccclc23 c3ccc(O)cc3)c(F)c21Nc 1 c(N=Nc2ccc( [N+] (=0) [0- 0=C 1 C(=O)c2cccc3 cccc 1 c23 ])cc2)c(S([O])(=O)=O)cc2cc(S([O])(=O)=O)c(N=Nc c 1 ccc2 [nH] c(-c3 cscn3 )nc2c 1 3ccccc3)c(O)cl2 COc 1CC2C(C3OC(=O)C4C(C 13)CCC4=O) [C@H] 1 [C@ clccc2c(cl)cclc3c2ccc2cccc(c23)C2NC12 @H](O2)O[C@@H]2O[C@@H]21 Cclccc(Nc2ccc3c4c(cccc24)C(=O)c2ccccc2-3)ccl Cclnc2c(c3[nH]c4ccccc4cl3)C(=O)C(NlCCN(C)C C(CCC1CO1)CC1CO1 C1)=CC2=OO=ClCCC2(O)c3c(ccc(O)c31)- CC1=CC2=C(C=O)C(=O)C(C)(O)C(O)C2=CO1 clccc(0)c3clC2C10ClC3=0 Cnlc(-c2ccccc2)c( / N=N / c2scc[n+]2C)c2ccccc21 COclccc2cc([N+](=O)[O-])c3ccc(=O)oc3c2cl Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- CC(C)CCCC(C) [C@H] 1 CC [C@H] 2 [C@@H] 3 [C@ c4ccc(N=Nc5c(O)ccc6cc(S([O])(=O)=O)cc(S([O])(= H](OO)C=C4C[C@@H](O)CC[C@]4(C)[C@H]3C O)=O)c56)c(C)c4)cc3 C)cc2)cc 1C[C@]12C CCN(CCCl)clccc(CCCNc2c3ccccc3nc3ccccc23)ccl C[C@]12CC[C@H](OC(=O)Cc3ccc(N(CCCl)CCCl COclcc(OC)c2c(cl)O[C@@H]lOC=C[C@H]21 )cc3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[C@]2 O=[N+]([O-])clccc( / C=N / C2SNc3ccccc32)ol (C)NC(=O)CC[C@@H] 12 CCN(CC)CCnlc2ccc(CO)c3c2c([n+] 1 [0-])- Nclccc(N=Nc2ccc(N=Nc3c(S([O])(=O)=O)cc4cc(S( clccc(Cl)cclS3 [O])(=O)=O)c(N=Nc5ccc(N=Nc6ccc(N)cc6N)cc5)c( C1C 1 C=Cc2cc3c4c(ccc5cccc(c54)C=3)c2=C 1 O)c4c3N)cc2)c(N)cl C[C@]12CC[C@H](OC(=O) / C=C / c3cccc(N(CCCl) Cclccc(N=NNCc2ccc([N+](=O)[O-])cc2)ccl CCC1)C3)CC1=CC[C@@H]1[C@@H]2CC[C@]2( O=C(Cnl ccnc 1 [N+] (=0) [0-] )NCc 1 ccccc 1 C)NC(=O)CC[C@@H] 12 O=clc2ccccc2[nH]c2ccc3c(cl2)OCCO3 C / C=C 1 / CC(C)C(C)(OC(C)=O)C(=O)OCC2=C3 C( 0=[N+]([0- CC[N+] 3 (C)C=C2)0C 1=0 ])clccc(C2OCC3(CO)COC(c4ccc([N+](=O)[O- Cl=Cc2c3ccc4cccc5ccc(c6cccclc26)c3c45 ])cc4)N23)ccl CN(C)CCNC(=O)clcccn2cc(-c3ccccc3)ncl2 COclcc(-c2nc(-c3ccccc3)c(-c3ccccc3)[nH]2)ccclO 0=C 1 CCc2c 1 cc(C(F)(F)F)c 1 c3 c(ccc21 ) [C@H] (0) [ [0-] [n+] Iccc2cnncc2cl C@@H](0)C=C3Example compounds missed from ChemProp Example compounds missed from ChemProp 0=c ln(CC2CO2)c(=O)n(CC2CO2)c(=O)nl CC 1 CO CCl=CC(=C(C#N)C#N)C=C( / C=C / c2ccc(N(C)C)cc 1 2)01COclcc2c(cclOC)CN(N=O)CC2 Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- 0=[N+]([0- c4ccc(N=Nc5c(N)c(S([O])(=O)=O)cc6cc(S([O])(=O ])C1=CC2C=C3C(=CC2C=C1)C=C[C@H](O)[C@ )=O)ccc56)cc4)cc3)cc2)cclH]30 O=C1OCCN1P(=O)(C1)N1CCOC1=O CCnlcc(C(=O)O)c(=O)c2ccc(C)nc21 C[n+] Iccc(-c2ccccn2)ccl COclcc2c(c3oc(=O)c4c(cl3)CCOC4=O)ClC=COC 0=S 1(=O)C2CCCC3CCC4C5CCCCC5CC Ic4c23102 clccc2c(cl)clccccclclccccc21 Cclccc(C)c2c3c(cccl2)C(0)C(0)C10C31 Cclnc2c(c3[nH]c4ccccc4cl3)C(=O)C=C(NlCCl)C O[C@H] 1[C@H]2O[C@H]2C=C[C@@H] 10 2=0 CC(=O)OCCN(CCOC(C)=O)clccc(N=Nc2nc3ccc( C=CC(=O)N(CC 1 CO 1 )CC 1 CO 1Cl)cc3 s2)c(NC(=O)c2ccccc2)c 1 CCnlcc(C(=O)O)c(=O)c2cnc(N3CCNCC3)nc21 CCOC(=O)C 1 (CCCCCCOCc2cccs2)CO 1 COC(=O)clcc2c([nH]c3ccccc32)c(- clcc2ccc3ccc(C4CO4)c4ccc(cl)c2c34 c2ccc3c(n2)C(=O)C=CC3=O)nlCc 1 ccc(N=NNCc2ccccc2)cc 1 C 1=CC2CCCCC2C2CCCCC21c 1 ccc2c(c 1 )CCc 1 cc3 ccc4c(c3 cc 1 -2)CCCC4 Nclc(Oc2ccccc2)cc(O)c2clC(=O)clccccclC2=O 0=c lc(OC2OC(CO)C(O)C(O)C2O)c(- O=clc2ccc3c4c(ccc(c24)c2nc4ccccc4nl2)c(=O)nlc c2ccc(O)c(O)c2)oc2cc(O)cc(O)c 12 2ccccc2nc310=C( [0-] )c 1 c [n+] (=O)c2ccccc2nl 0 clcc2c3c(cccc3cl)ClNC21[N-]=[N+]=C1C=NC(=O)NC1=O Cclccc2nc(-c3ccc(N)cc3)sc2cl O=Clc2ccccc2C(=O)c2clcc(OClOC(COC3OCC(O Cnl c(N)nc2c3 ccc(=O) [nH] c3 ccc21)C(O)C3O)C(O)C(O)ClO)c(CO)c2OTable IB. Example compounds correctly characterized by AmesNet but not by DeepAmes Example compounds missed from DeepAmes Example compounds missed from DeepAmes Cc 1 ccc2cc3 c(c4c2c 1 CC4)C(0)C(0)c 1 ccccc 1 -3 2)010=c lc(OC2OC(CO)C(O)C(O)C2O)c(- C[C@]12CC[C@H](OC(=O) / C=C / c3cccc(N(CCCl) c2ccc(O)c(O)c2)oc2cc(O)cc(O)c 12 CCCl)c3)CCl=CC[C@@H]l[C@@H]2CC[C@]2( O=Clc2ccccc2C(=O)c2clcc(O[C@H]lO[C@H](C C)NC(=O)CC[C@@H] 12O [C@@H] 3 OC [C@@H] (O) [C@H] (O) [C@H] 3 O) [ O=C(Cnl ccnc 1 [N+] (=0) [0-] )NCc 1 ccccc 1 C@@H](O)[C@H](O)[C@H] 10)c(C0)c20 O=C(C=Cc 1 ccc( [N+] (=0) [O-])cc 1 )c 1 ccccc 1[0-] [n+] Iccc2cnncc2cl 0=C 1 CC(=Nc2ccccc2)C(=O)c2ccccc21 Cl=Cc2ccc3c4c(ccclc24)-clcccccl-3 Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- O=Clc2ccccc2-c2cc3ccccc3cc21 c4ccc(N=Nc5c(N)c(S([O])(=O)=O)cc6cc(S([O])(=O O=clc2c(c3c(ccc4cc5c(ccc6ccccc65)[nH]c43)clO) )=O)ccc56)cc4)cc3)cc2)ccl02 O=C1OCCN1P(=O)(C1)N1CCOC1=O[N-]=[N+]=C1C=NC(=O)NC1=O C1CSCSC1 CCl=CC(=C(C#N)C#N)C=C( / C=C / c2ccc(N(C)C)cc C[C@]12CC[C@@H](OC(=O) / C=C / c3cccc(N(CCCExample compounds missed from DeepAmes Example compounds missed from DeepAmes l)CCCl)c3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[ c4ccc(N=Nc5c(O)ccc6cc(S([O])(=O)=O)cc(S([O])(= C@]2(C)NC(=O)CC[C@@H] 12 O)=O)c56)c(C)c4)cc3 C)cc2)cc 1O=C 1 C(=O)c2cccc3 cccc 1 c23 COclcc2c(cclO)CCNC21CS[C@@H]2c3c(OC(C)= c 1 ccc2 [nH] c(-c3 cscn3 )nc2c 1 O)c(C)c4c(c3 [C@H] (COC1=O)N 1 C(0)C3 Cc5cc(C) Cl=COc2c(c3cc4ccccc4cc3c3ccccc23)C=Cl c(OC)c(O)c5C([C@H]21)N3C)OCO4 clccc2c(cl)cclc3ccccc3c3cccc4ccc2clc43 c 1 cnc2c(c 1 )ccc 1 c2ccc2cccnc21 CC1OCC(=O)C(=O)C1O clccc2c(cl)ccclcc3c(ccc4ccncc43)ccl2 Cl=Cc2c3ccccc3cc3cccclc23 O=ClCCC2(O)c3c(ccc(O)c31)- 0ClC=Cc2c(ccc3ccccc23)C10 clccc(0)c3clC2C10ClC3=0O=S 1 (=O)c2cccc3ccc4c5ccccc5cc Ic4c23 0=C 1 CCCc2cc([N+] (=0) [O-])c3ccccc3c21 O=c ln(CC2CO2)c(=O)n(CC2CO2)c(=O)nl CC 1 CO O=clc2ccc3c4c(ccc(c24)c2nc4ccccc4nl2)c(=O)nlc 1 2ccccc2nc31 O=ClCCc2clcc(C(F)(F)F)clc3c(ccc21)[C@H](O)[ CCnlcc(C(=O)O)c(=O)c2ccc(C)nc21 C@@H](O)C=C3 NNclnc2ccc3[nH]c4ccccc4c(=O)c3c2sl Cclnc2c(c3[nH]c4ccccc4cl3)C(=O)C=C(NlCCl)C COclcc2c(c3oc(=O)c4c(cl3)CCOC4=O)ClC=COC 2=0 102C=CC(=O)N(CC 1 CO 1 )CC 1 CO 1 CCN(CCCl)clccc(CCCNc2c3ccccc3nc3ccccc23)ccl Nc 1 c(N=Nc2ccc( [N+] (=0) [0- C[C@]12CC[C@H](OC(=O)Cc3ccc(N(CCCl)CCCl ])cc2)c(S([O])(=O)=O)cc2cc(S([O])(=O)=O)c(N=Nc )cc3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[C@]2 3ccccc3)c(O)cl2 (C)NC(=O)CC[C@@H] 12 clcnc2c(cl)ccclc3cccnc3cccl2 CClCS(=O)(=O)CCNl / N=C / clccc([N+](=O)[O- COc 1CC2C(C3OC(=O)C4C(C 13)CCC4=O) [C@H] 1 [C@ ])ol@H](O2)O[C@@H]2O[C@@H]21 Nc Iccccc 1 SCCSc Iccccc IN COclcc2c(cclOC)CN(N=O)CC2 CC(=O)OCCN(CCOC(C)=O)clccc(N=Nc2nc3ccc( CCnlcc(C(=O)O)c(=O)c2cnc(N3 CCNCC3)nc21 Cl)cc3 s2)c(NC(=O)c2ccccc2)c 1c 1 ccc2c(c 1 )c 1 ccccc 1 c 1 ccccc21 Nclccc(N=Nc2cccc(N=Nc3ccc(N)cc3N)c2)c(N)cl Cclnc2c(c3[nH]c4ccccc4cl3)C(=O)C(NlCCN(C)C COclcc(OC)c2c(cl)O[C@@H]lOC=C[C@H]21 C1)=CC2=O CCOC(=O)C 1 (CCCCCCOCc2cccs2)CO 1 CC(C)(COCC 1 CO 1)COCC 1 CO 1 O=[N+]([O-])clccc( / C=N / C2SNc3ccccc32)ol C !=CC2CCCCC2c2ccccc21 Cc 1 ccc(N=NNCc2ccccc2)cc 1c 1 ccc2c(c 1 )c 1 ccccc 1c 1 cc3 cnccc3 cc21 0=[N+]([0- C(CCC1CO1)CC1CO1 ])c lccc(C2OCC3 (C0)C0C(c4ccc([N+] (=0) [0- O=clc(=O)c2cc3ccccc3c3ccc4cccclc4c23 ])cc4)N23)ccl Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(-Table 1C. Example compounds correctly characterized by AmesNet but not by grouped multitask learning (gMTL)Example compounds missed from gMTL Example compounds missed from gMTL clccc([C@H]2N[C@@H]2c2ccccc2)ccl O=C(C=Cclccc([N+](=O)[O-])ccl)clccccclExample compounds missed from gMTL Example compounds missed from gMTL Cc 1 c [nH] c2ccc3 c(nc(=NO)n3 C)c2nl -3)cclO=C 1 CC(=Nc2ccccc2)C(=O)c2ccccc21 clccc2cc(C3 C03 )ccc2c 1C1CSCSC1 Cl=CC2OC2c2clcclccc3cccc4ccc2clc34 O=C 1NCCCN1 / N=C / C lccc([N+] (=0) [O-])o 1 clccc([C@H]2N[C@H]2c2ccccc2)ccl Cl=CC2OC2c2cc3ccc4cccc5ccc(c21)c3c45 Nnlnnc2c([nH]c3ccccc32)cl=O CC10CC(=0)C(=0)C10 C=c 1 nccc2c 1 =Nc 1 cc(O)ccc 1 -2 O=clc(=O)c2cc3ccccc3c3ccc4cccclc4c23 0=C 1 C(=O)c2cccc3 cccc 1 c23Nc 1 c(N=Nc2ccc( [N+] (=0) [0- c 1 ccc2 [nH] c(-c3 cscn3 )nc2c 1 ])cc2)c(S([O])(=O)=O)cc2cc(S([O])(=O)=O)c(N=Nc Cnl c(N)nc2nc3 ccccc3 cc213ccccc3)c(O)cl2 CC1=CC2=C(C=O)C(=O)C(C)(O)C(O)C2=CO1 clccc2c(cl)cclc3c2ccc2cccc(c23)C2NC12 Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- c 1 ccc2c(c 1 )c 1 ccccc 1c 1 cc3 cnccc3 cc21 c4ccc(N=Nc5c(O)ccc6cc(S([O])(=O)=O)cc(S([O])(= CC 1 (C)CC2C(O)(C=C(C=O)C34CC23 C(=0)0C40 O)=O)c56)c(C)c4)cc3 C)cc2)cc 1)C1 COclcc2c(cclO)CCNC21CS[C@@H]2c3c(OC(C)= C(CCC1CO1)CC1CO1 O)c(C)c4c(c3 [C@H] (COC1=O)N 1 C(0)C3 Cc5cc(C) O=ClCCC2(O)c3c(ccc(O)c31)- c(OC)c(O)c5C([C@H]21)N3C)OCO4 clccc(0)c3clC2C10ClC3=0 CClCNc2cc(O)c3ccc4ccccc4c3c21 CC(C)CCCC(C) [C@H] 1 CC [C@H] 2 [C@@H] 3 [C@ 0=S (=O)(Cc Iccccc 1 )N 1 C2c3 ccccc3 -c3 ccccc3 C21 H](OO)C=C4C[C@@H](O)CC[C@]4(C)[C@H]3C COclcc(OC)c2c(cl)O[C@@H]lOC=C[C@H]21 C[C@]12C O=[N+]([O-])clccc( / C=N / C2SNc3ccccc32)ol Nclccc(N=Nc2ccc(N=Nc3c(S([O])(=O)=O)cc4cc(S( clccc2c(cl)ccclc3ccncc3ccc21 [O])(=O)=O)c(N=Nc5ccc(N=Nc6ccc(N)cc6N)cc5)c( clcc2ccc3cccc4ncc(cl)c2c34 O)c4c3N)cc2)c(N)cl C[C@]12CC[C@H](OC(=O) / C=C / c3cccc(N(CCCl) NclccccclSCCSclccccclN CCCl)c3)CCl=CC[C@@H]l[C@@H]2CC[C@]2( Cclccc(N=NNCc2ccc([N+](=O)[O-])cc2)ccl C)NC(=O)CC[C@@H] 12 clcc2cc3c4c(ccc5cc6c(clCCC6)c2c54)CCC3 Cl=Cc2c3ccc4cccc5ccc(c6cccclc26)c3c45 c Icc2c3c(c l)cc 1CCC4CCCC5CC(C3C1C45)CC2 CN(C)CCNC(=O)clcccn2cc(-c3ccccc3)ncl2 O=C(Cnl ccnc 1 [N+] (=0) [0-] )NCc 1 ccccc 1 0=c ln(CC2CO2)c(=O)n(CC2CO2)c(=O)nl CC 1 CO O=clc2ccccc2[nH]c2ccc3c(cl2)OCCO3 10=[N+]([0- COclcc2c(cclOC)CN(N=O)CC2 ])clccc(C2OCC3(CO)COC(c4ccc([N+](=O)[O- 0=[N+]([0- ])cc4)N23)ccl ])C1=CC2C=C3C(=CC2C=C1)C=C[C@H](O)[C@ c 1 cc2c3 c(c 1 )=C 1 OC 1 C=3 Cc 1 cc3 ccccc3 cc 1 -2 H]30[0-] [n+] Iccc2cnncc2cl clcc2ccc3cc4c(c5ccc(cl)c2c35)CCC10C41 Cl=Cc2ccc3c4c(ccclc24)-clcccccl-3 C 1=C2CCCCC2=C2CC3CCCCC3 C3=C4OC4=C 1 C32 O=Clc2ccccc2-c2cc3ccccc3cc21 0=C 1 CCCc2cc([N+] (=0) [O-])c3ccccc3c21 Cl=CC2OC2c2ncccc21 CCnlcc(C(=O)O)c(=O)c2ccc(C)nc21[0] S(=O)(=O)c 1CCC(OC2CCC3C4C(CCCC24)CC2CCCCC2 Nclcc2nc3ccccc3nc2cclOExample compounds missed from gMTL Example compounds missed from gMTL COclcc2c(c3oc(=O)c4c(cl3)CCOC4=O)ClC=COC c4ccc(N=Nc5c(N)c(S([O])(=O)=O)cc6cc(S([O])(=O 102 )=O)ccc56)cc4)cc3)cc2)ccl Cclcc(O)c2c(cl)C(=O)clcc(O)cc(OC3OC(CO)C(O) O=C1OCCN1P(=O)(C1)N1CCOC1=O C(O)C3O)clC2=O C[n+] Iccc(-c2ccccn2)cclCCOC(=O)C 1 (CCCCCCOCc2cccs2)CO 1 0=S 1(=O)C2CCCC3CCC4C5CCCCC5CC Ic4c23Cc 1 ccc(N=NNCc2ccccc2)cc 1 clccc2c(cl)clccccclclccccc210=c lc(OC2OC(CO)C(O)C(O)C2O)c(- C=CC(=O)N(CC 1 CO 1 )CC 1 CO 1 c2ccc(O)c(O)c2)oc2cc(O)cc(O)c 12 COC(=O)clcc2c([nH]c3ccccc32)c(- 0=C( [0-] )c 1 c [n+] (=O)c2ccccc2nl 0 c2ccc3c(n2)C(=O)C=CC3=O)nlc Iccc2c3c(ccc2c l)Cc lc(ccc2ccccc 12)C3 Cnl c(N)nc2nc3 cnccc3 cc21[N-]=[N+]=C1C=NC(=O)NC1=O NNclnc2ccc3[nH]c4ccccc4c(=O)c3c2sl O=Clc2ccccc2C(=O)c2clcc(OClOC(COC3OCC(O Nclc(Oc2ccccc2)cc(O)c2clC(=O)clccccclC2=O )C(O)C3O)C(O)C(O)ClO)c(CO)c2O O=clc2ccc3c4c(ccc(c24)c2nc4ccccc4nl2)c(=O)nlc CCl=CC(=C(C#N)C#N)C=C( / C=C / c2ccc(N(C)C)cc 2ccccc2nc312)01 Cclccc2nc(-c3ccc(N)cc3)sc2clc 1 ccc2c(c 1 )Cc 1 c(c3 ccccc3 c3 ccccc 13 )C2 0=NN 1 CCCN(N=O)CC 1 clcc2c3oc3cc3ccc4c5oc5cclc4c32 Nclccc(N)c2c(O)c3ccccc3c(O)cl2 Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- Cnl c(N)nc2c3 ccc(=O) [nH] c3 ccc21Table ID. Example compounds correctly characterized by AmesNet but not by GROVER Example compounds missed from GROVER Example compounds missed from GROVER clccc([C@H]2N[C@@H]2c2ccccc2)ccl Cclccc(Nc2ccc3c4c(cccc24)C(=O)c2ccccc2-3)ccl C=C / C 1=C / [C@H] (C) [C@] (C)(OC(C)=O)C(=O)O CC(C)CCCC(C) [C@H] 1 CC [C@H] 2 [C@@H] 3 [C@ CC2=CC[N+]3(C)CC[C@@H](OC1=O)C23[O-] H](OO)C=C4C[C@@H](O)CC[C@]4(C)[C@H]3C O=C(C=Cc lccc( [N+] (=0) [0 -] )cc 1 )c 1 ccccc 1 C[C@]12C0=C 1 CC(=Nc2ccccc2)C(=O)c2ccccc21 C[C@]12CC[C@H](OC(=O)Cc3ccc(N(CCCl)CCCl C[C@]12CC[C@@H](OC(=O) / C=C / c3cccc(N(CCC )cc3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[C@]2 l)CCCl)c3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[ (C)NC(=O)CC[C@@H] 12 C@]2(C)NC(=0)CC[C@@H] 12 Nclccc(N=Nc2ccc(N=Nc3c(S([O])(=O)=O)cc4cc(S( C1CSCSC1 [O])(=O)=O)c(N=Nc5ccc(N=Nc6ccc(N)cc6N)cc5)c( CC10CC(=0)C(=0)C10 O)c4c3N)cc2)c(N)cl Cl=Cc2c3ccccc3cc3cccclc23 Nc Iccccc 1 SCCSc Iccccc INNc 1 c(N=Nc2ccc( [N+] (=0) [0- O=c 1 c2ccccc2 [nH] c2ccc3 c(c 12)OCCO3 ])cc2)c(S([O])(=O)=O)cc2cc(S([O])(=O)=O)c(N=Nc O=[N+]([O- 3ccccc3)c(O)cl2 ])c lccc(C2OCC3 (CO)COC(c4ccc([N+] (=0) [O- O=Clc2ccccc2-c2clccclccccc21 ])cc4)N23)cclCC 1 (C)CC2C(O)(C=C(C=O)C34CC23 C(=0)0C40 CC1(C)C[C@H]2[C@](O)(C=C(C=O)[C@]3(C=O) )C1 C[C@]23C)C1Example compounds missed from GROVER Example compounds missed from GROVER CC[C@@]1(O)C[C@H](OC2CC(N(C)C)C(OC3CC CCN(CC)CCnlc2ccc(CO)c3c2c([n+] 1 [0-])- (O)C(OC4CC(O)C(O)C(C)O4)C(C)O3)C(C)O2)c2c clccc(Cl)cclS3 (CC3C(O)C4C(C(O)C3C2O)C(=O)C=CC4=O)C1C(=O) O=clc2c(c3c(ccc4cc5c(ccc6ccccc65)[nH]c43)clO) OC 02[O-] [n+] Iccc2cnncc2cl C[C@]12CC[C@H](OC(=O) / C=C / c3cccc(N(CCCl) O=Clc2ccccc2-c2cc3ccccc3cc21 CCCl)c3)CCl=CC[C@@H]l[C@@H]2CC[C@]2( Cl=CC2OC2c2ncccc21 C)NC(=O)CC[C@@H] 12[0] S(=O)(=O)c 1CCC(OC2CCC3C4C(CCCC24)CC2CCCCC2 CN(C)CCNC(=O)clcccn2cc(-c3ccccc3)ncl2 -3)ccl 0=c ln(CC2CO2)c(=O)n(CC2CO2)c(=O)nl CC 1 CO O=[N+]([O-])clccc2c(cl)CCclcc3c(ccl-2)CCCC3 1c 1CCC([C@H]2N[C@H]2C2CCCCC2)CC 1 COclcc2c(cclOC)CN(N=O)CC2 CCN(CC)CCnlnc2c3c(c(CO)ccc31)Sclcc(Cl)cccl-2 0=C 1 CCCc2cc([N+] (=0) [O-])c3ccccc3c21 Nnlnnc2c([nH]c3ccccc32)cl=O CCnlcc(C(=O)O)c(=O)c2ccc(C)nc21C=c 1 nccc2c 1 =Nc 1 cc(O)ccc 1 -2 Nclcc2nc3ccccc3nc2cclO0=C 1 C(=O)c2cccc3 cccc 1 c23 COclcc2c(c3oc(=O)c4c(cl3)CCOC4=O)ClC=COC c 1 ccc2 [nH] c(-c3 cscn3 )nc2c 1 102 CCC / C=C / C(=O)O[C@@H]1C(C)(C)C[C@@H]2C O=S1(=O)CCCO1 34CC3(C(C=O)=C[C@]12O)[C@@H](O)OC4=O CC(=O)OCCN(CCOC(C)=O)clccc(N=Nc2nc3ccc( 0=C 1 C=Cc2cccc3 cccc 1 c23 Cl)cc3 s2)c(NC(=O)c2ccccc2)c 1 Cclnc2c(c3[nH]c4ccccc4cl3)C(=O)C(NlCCN(C)C O=ClOCCNl / N=C / clccc([N+](=O)[O-])ol C1)=CC2=O CCOC(=O)C 1 (CCCCCCOCc2cccs2)CO 1CC 1=CC2=C(C=O)C(=O)C(C)(O)C(O)C2=CO 1 Cc 1 ccc(N=NNCc2ccccc2)cc 1Cnlc(-c2ccccc2)c( / N=N / c2scc[n+]2C)c2ccccc21 CC1 CS(=O)(=O)CCN lN=Cc lccc([N+] (=0) [O-])o 1 CC(=0)N(0C(=0)clcccccl)clccc2c(cl)Cclcccccl- O=clc(OC2OC(CO)C(O)C(O)C2O)c(- 2 c2ccc(O)c(O)c2)oc2cc(O)cc(O)cl2 Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- 0=C( [0-] )c 1 c [n+] (=O)c2ccccc2nl 0 c4ccc(N=Nc5c(O)ccc6cc(S([O])(=O)=O)cc(S([O])(= [N-]=[N+]=C1C=NC(=O)NC1=O O)=O)c56)c(C)c4)cc3 C)cc2)cc 1 O=Clc2ccccc2C(=O)c2clcc(OClOC(COC3OCC(O COclcc2c(cclO)CCNC21CS[C@@H]2c3c(OC(C)= )C(O)C3O)C(O)C(O)ClO)c(CO)c2O O)c(C)c4c(c3[C@H](COCl=O)NlC(O)C3Cc5cc(C) CCl=CC(=C(C#N)C#N)C=C( / C=C / c2ccc(N(C)C)cc c(OC)c(O)c5C([C@H]21)N3C)OCO4 2)01 CCN(CCCl)clccc(CCCNc2c3ccccc3nc3ccccc23)ccl CC[C@@] 1 (O)C[C@H](OC2CC(N(C)C)C(OC3 CC CC 1 CS(=O)(=O)CCN l / N=C / c lccc([N+] (=0) [0- (O)C(OC4CCC(=O)C(C)O4)C(C)O3)C(C)O2)c2c(c ])ol C3C(C2O)C(=O)C2C(O)CCCC2C3=O)C1C(=O)OC O=[N+]([O-])clccc( / C=N / C2SNc3ccccc32)ol Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- O=Clc2ccccc2C(=O)c2clcc(O[C@H]lO[C@H](C c4ccc(N=Nc5c(N)c(S([O])(=O)=O)cc6cc(S([O])(=O 0 [C@@H] 3 OC [C@@H] (0) [C@H] (0) [C@H] 30) [ )=O)ccc56)cc4)cc3)cc2)ccl C@@H](O)[C@H](O)[C@H] 10)c(C0)c20 O=C1OCCN1P(=O)(C1)N1CCOC1=OExample compounds missed from GROVER Example compounds missed from GROVER C[n+] Iccc(-c2ccccn2)ccl C 1=CC2CCCCC2C2CCCCC21 0ClC=Cc2c(ccc3ccccc23)C10 Nclc(Oc2ccccc2)cc(O)c2clC(=O)clccccclC2=O O=S 1 (=O)c2cccc3ccc4c5ccccc5cc Ic4c23 O=clc2ccc3c4c(ccc(c24)c2nc4ccccc4nl2)c(=O)nlc C=CC(=O)N(CC 1 CO 1 )CC 1 CO 1 2ccccc2nc31O=C(OCC1 CCC2OC2C 1 )C 1 CCC2OC2C 1 O=ClC=Cc2ccc3c(ccc4ccccc43)c2Cl=O CCnlcc(C(=O)O)c(=O)c2cnc(N3 CCNCC3)nc21 Cclccc2nc(-c3ccc(N)cc3)sc2cl NclccccclSCSclccccclN CC1 (C)C[C@H]2C=C(C=O) [C@] 34C[C@] 3 (C(=O COC(=O)clcc2c([nH]c3ccccc32)c(- )OC4O)[C@H]2C1 c2ccc3c(n2)C(=O)C=CC3=O)nl Cnlc(N)nc2c3ccc(=O)[nH]c3ccc21Table IE. Example compounds correctly characterized by AmesNet but not by single-task learning (STL)Example compounds missed from STL Example compounds missed from STL clccc([C@H]2N[C@@H]2c2ccccc2)ccl c 1 ccc2c(CCC3 C03 )cccc2c 1C=C / C 1=C / [C@H] (C) [C@] (C)(OC(C)=O)C(=O)O CC(C)CCCC(C) [C@H] 1 CC [C@H] 2 [C@@H] 3 [C@ CC2=CC[N+]3(C)CC[C@@H](OC1=O)C23[O-] H](OO)C=C4C[C@@H](O)CC[C@]4(C)[C@H]3C O=C(C=Cc lccc( [N+] (=0) [0 -] )cc 1 )c 1 ccccc 1 C[C@]12CCc 1 c [nH] c2ccc3 c(nc(=N0)n3 C)c2nl C[C@]12CC[C@H](OC(=O)Cc3ccc(N(CCCl)CCCl 0=C 1 CC(=Nc2ccccc2)C(=O)c2ccccc21 )cc3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[C@]2 C[C@]12CC[C@@H](OC(=O) / C=C / c3cccc(N(CCC (C)NC(=O)CC[C@@H] 12 l)CCCl)c3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[ Nclccc(N=Nc2ccc(N=Nc3c(S([O])(=O)=O)cc4cc(S( C@]2(C)NC(=0)CC[C@@H] 12 [O])(=O)=O)c(N=Nc5ccc(N=Nc6ccc(N)cc6N)cc5)c( C1CSCSC1 O)c4c3N)cc2)c(N)cl Cl=CC2OC2c2cc3ccc4cccc5ccc(c21)c3c45 Nc Iccccc 1 SCCSc Iccccc IN CC10CC(=0)C(=0)C10 Cclccc(N=NNCc2ccc([N+](=O)[O-])cc2)ccl Cl=Cc2c3ccccc3cc3cccclc23 O=C(Cnl ccnc 1 [N+] (=0) [0-] )NCc 1 ccccc 1 Nc 1 c(N=Nc2ccc( [N+] (=0) [0- clcc2c3c(cl)cclccc4cccc5cc(c3clc45)CC2 ])cc2)c(S([O])(=O)=O)cc2cc(S([O])(=O)=O)c(N=Nc 0=[N+]([0- 3ccccc3)c(O)cl2 ])c lccc(C2OCC3 (C0)C0C(c4ccc([N+] (=0) [0- clccc2c(cl)cclc3c2ccc2cccc(c23)C2NC12 ])cc4)N23)cclc 1 ccc2c(c 1 )c 1 ccccc 1c 1 cc3 cnccc3 cc21 clcc2c3c(cl)=C10ClC=3Cclcc3ccccc3ccl-2 CC 1 (C)CC2C(O)(C=C(C=O)C34CC23 C(=0)0C40 [0-] [n+] Iccc2cnncc2cl)C1 Cl=Cc2ccc3c4c(ccclc24)-clcccccl-3 Cclccc(Nc2ccc3c4c(cccc24)C(=O)c2ccccc2-3)ccl O=Clc2ccccc2-c2cc3ccccc3cc21 C(CCC1CO1)CC1CO1 Cl=CC2OC2c2ncccc21 O=ClCCC2(O)c3c(ccc(O)c31)- [O]S(=O)(=O)clccc(Oc2ccc3c4c(cccc24)Cc2ccccc2 clccc(0)c3clC2C10ClC3=0 -3)cclCOclccc2cc([N+](=O)[O-])c3ccc(=O)oc3c2cl clccc2cc(C3 C03 )ccc2c 1Example compounds missed from STL Example compounds missed from STL Cl=CC2OC2c2clcclccc3cccc4ccc2clc34 0=c ln(CC2CO2)c(=O)n(CC2CO2)c(=O)nl CC 1 CO c 1CCC([C@H]2N[C@H]2C2CCCCC2)CC 1 1clccc2c(cl)-clc(c3ccccc3c3cccccl3)ClNC21 COclcc2c(cclOC)CN(N=O)CC2 CCN(CC)CCnlnc2c3c(c(CO)ccc31)Sclcc(Cl)cccl-2 0=[N+]([0- Nnlnnc2c([nH]c3ccccc32)cl=O ])C1=CC2C=C3C(=CC2C=C1)C=C[C@H](O)[C@ C=c 1 nccc2c 1 =Nc 1 cc(O)ccc 1 -2 H]30O=C 1 C(=O)c2cccc3 cccc 1 c23 C 1=C2CCCCC2=C2CC3CCCCC3 C3=C4OC4=C 1 C32 c 1 ccc2 [nH] c(-c3 cscn3 )nc2c 1 0=C 1 CCCc2cc([N+] (=0) [O-])c3ccccc3c21 COc 1CC2C(C3OC(=O)C4C(C 13)CCC4=O) [C@H] 1 [C@ Nclcc2nc3ccccc3nc2cclO @H](O2)O[C@@H]2O[C@@H]21 COclcc2c(c3oc(=O)c4c(cl3)CCOC4=O)ClC=COC Cclnc2c(c3[nH]c4ccccc4cl3)C(=O)C(NlCCN(C)C 102C1)=CC2=O Cclcc(O)c2c(cl)C(=O)clcc(O)cc(OC3OC(CO)C(O) CC 1=CC2=C(C=O)C(=O)C(C)(O)C(O)C2=CO 1 C(O)C3O)clC2=OO=[N+]([O- O=S1(=O)CCCO1 ])clccc2c3c4c(cccc4c4ccccc4cl3)[C@H](O)[C@H] Nclccc2cc(S([O])(=O)=O)cc(O)c2clN=Nclccc(- 20 c2ccc(N=Nc3cc(S([O])(=O)=O)c4ccccc4c3N)cc2)cc CC(=0)N(0C(=0)clcccccl)clccc2c(cl)Cclcccccl- 12 CC(=O)OCCN(CCOC(C)=O)clccc(N=Nc2nc3ccc( Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- Cl)cc3 s2)c(NC(=O)c2ccccc2)c 1 c4ccc(N=Nc5c(O)ccc6cc(S([O])(=O)=O)cc(S([O])(= CCOC(=O)C 1 (CCCCCCOCc2cccs2)CO 1 O)=O)c56)c(C)c4)cc3 C)cc2)cc 1 Cc 1 ccc(N=NNCc2ccccc2)cc 1 COclcc2c(cclO)CCNC21CS[C@@H]2c3c(OC(C)= CC1 CS(=O)(=O)CCN lN=Cc lccc([N+] (=0) [O-])o 1 O)c(C)c4c(c3[C@H](COCl=O)NlC(O)C3Cc5cc(C) clccc2c(cl)CCclcc3ccc4c(c3ccl-2)CCCC4 c(OC)c(O)c5C([C@H]21)N3C)OCO4 O=clc(OC2OC(CO)C(O)C(O)C2O)c(- CCN(CCCl)clccc(CCCNc2c3ccccc3nc3ccccc23)ccl c2ccc(O)c(O)c2)oc2cc(O)cc(O)cl2CC 1 CNc2cc(O)c3ccc4ccccc4c3c21 0=C( [0-] )c 1 c [n+] (=O)c2ccccc2nl 0CC 1 CS(=O)(=O)CCN l / N=C / c lccc([N+] (=0) [0- clccc2c3c(ccc2c l)Cc lc(ccc2ccccc 12)C3])ol [N-]=[N+]=C1C=NC(=O)NC1=O COclcc(OC)c2c(cl)O[C@@H]lOC=C[C@H]21 O=Clc2ccccc2C(=O)c2clcc(OClOC(COC3OCC(O O=[N+]([O-])clccc( / C=N / C2SNc3ccccc32)ol )C(O)C3O)C(O)C(O)ClO)c(CO)c2OCc 1 ccc2cc3 c(c4c2c 1 CC4)C(0)C(0)c 1 ccccc 1 -3 CCl=CC(=C(C#N)C#N)C=C( / C=C / c2ccc(N(C)C)cc c 1 ccc2c(c 1 )ccc 1 c3 ccncc3 ccc21 2)01 C[C@]12CC[C@H](OC(=O) / C=C / c3cccc(N(CCCl) clcc2c3oc3cc3ccc4c5oc5cclc4c32 CCCl)c3)CCl=CC[C@@H] 1 [C@@H]2CC[C@]2( Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- C)NC(=O)CC[C@@H] 12 c4ccc(N=Nc5c(N)c(S([O])(=O)=O)cc6cc(S([O])(=O C / C=C 1 / CC(C)C(C)(OC(C)=O)C(=O)OCC2=C3 C( )=O)ccc56)cc4)cc3)cc2)ccl CC[N+]3(C)C=C2)OC1=O O=C1OCCN1P(=O)(C1)N1CCOC1=O Cnlccc2ccc3ccc4ccc(O)cc4c3c21 C[n+] Iccc(-c2ccccn2)cclExample compounds missed from STL Example compounds missed from STL 0=S 1 (=O)c2cccc3ccc4c5ccccc5cc Ic4c23 c 1 cnc2c(c 1 )ccc 1 c2ccc2cccnc21c 1 ccc2c(c 1 )c 1 ccccc 1 c 1 ccccc21 Nclc(Oc2ccccc2)cc(O)c2clC(=O)clccccclC2=O C=CC(=O)N(CC 1 CO 1 )CC 1 CO 1 O=clc2ccc3c4c(ccc(c24)c2nc4ccccc4nl2)c(=O)nlc CC 1 CC2(OC2C)C(=O)OC2CCN(C)C / C=C( / COC(= 2ccccc2nc31O)C1(C)O)C2=O Cclccc2nc(-c3ccc(N)cc3)sc2cl COC(=O)clcc2c([nH]c3ccccc32)c(- Nclccc(N)c2c(O)c3ccccc3c(O)cl2 c2ccc3c(n2)C(=O)C=CC3=O)nl Cnl c(N)nc2c3 ccc(=O) [nH] c3 ccc21Table IF. Example compounds correctly characterized by AmesNet but not by grouped ungrouped multitask learning (uMTL)Example compounds missed from uMTL Example compounds missed from uMTL clccc([C@H]2N[C@@H]2c2ccccc2)ccl O=C(Cnl ccnc 1 [N+] (=0) [0-] )NCc 1 ccccc 1 O=C(C=Cclccc([N+](=O)[O-])ccl)clcccccl O=ClC=Cc2ccc3c4c(ccclc24)C=CC3=O 0=C 1 CC(=Nc2ccccc2)C(=O)c2ccccc21 [0-] [n+] Iccc2cnncc2cl Nclccc2c(S([O])(=O)=O)ccc(O)c2clN=Nclccc(- Cl=Cc2ccc3c4c(ccclc24)-clcccccl-3 c2ccc(N=Nc3c(N)ccc4c(S([O])(=O)=O)ccc(O)c34)c O=Clc2ccccc2-c2cc3ccccc3cc21c2)ccl Cl=CC2OC2c2ncccc21C1CSCSC1 [O]S(=O)(=O)clccc(Oc2ccc3c4c(cccc24)Cc2ccccc2 Cl=CC2OC2c2cc3ccc4cccc5ccc(c21)c3c45 -3)cclCC10CC(=0)C(=0)C10 clccc2cc(C3 C03 )ccc2c 1 Cl=Cc2c3ccccc3cc3cccclc23 Cl=CC2OC2c2clcclccc3cccc4ccc2clc34 O=Clc2ccccc2-c2clccclccccc21 C(CC1CO1)C1CO1 clccc2c(cl)cclc3c2ccc2cccc(c23)C2NC12 clccc([C@H]2N[C@H]2c2ccccc2)ccl c 1 ccc2c(c 1 )c 1 ccccc 1c 1 cc3 cnccc3 cc21 Nnlnnc2c([nH]c3ccccc32)cl=O C(CCC1CO1)CC1CO1 CCnlcc(C(=0)0)c(=0)c2cc(F)c(- COclccc2cc([N+](=O)[O-])c3ccc(=O)oc3c2cl c3ccc(O)cc3)c(F)c21 O=ClCCCNlclncc([N+](=O)[O-])sl C=c 1 nccc2c 1 =Nc 1 cc(O)ccc 1 -2 CC(C)CCCC(C) [C@H] 1 CC [C@H] 2 [C@@H] 3 [C@ c 1 ccc2 [nH] c(-c3 cscn3 )nc2c 1 H](OO)C=C4C[C@@H](O)CC[C@]4(C)[C@H]3C CC1=CC2=C(C=O)C(=O)C(C)(O)C(O)C2=CO1 C[C@]12C Cnlc(-c2ccccc2)c( / N=N / c2scc[n+]2C)c2ccccc21 C[C@]12CC[C@H](OC(=O)Cc3ccc(N(CCCl)CCCl Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- )cc3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[C@]2 c4ccc(N=Nc5c(O)ccc6cc(S([O])(=O)=O)cc(S([O])(= (C)NC(=O)CC[C@@H] 12 O)=O)c56)c(C)c4)cc3 C)cc2)cc 1 Nclccc(N=Nc2ccc(N=Nc3c(S([O])(=O)=O)cc4cc(S( COclcc2c(cclO)CCNC21CS[C@@H]2c3c(OC(C)= [O])(=O)=O)c(N=Nc5ccc(N=Nc6ccc(N)cc6N)cc5)c( O)c(C)c4c(c3 [C@H] (COC1=O)N 1 C(0)C3 Cc5cc(C) O)c4c3N)cc2)c(N)cl c(OC)c(O)c5C([C@H]21)N3C)OCO4 NclccccclSCCSclccccclN CCN(CCCl)clccc(CCCNc2c3ccccc3nc3ccccc23)cclCclccc(N=NNCc2ccc([N+](=O)[O-])cc2)cclExample compounds missed from uMTL Example compounds missed from uMTL CC 1 CNc2cc(O)c3ccc4ccccc4c3c21 clccc2c(cl)CCclcc3ccc4c(c3ccl-2)CCCC4 COclcc(OC)c2c(cl)O[C@@H]lOC=C[C@H]21 O=C( [O-] )c 1 c [n+] (=O)c2ccccc2nl 0Cc 1 ccc2cc3 c(c4c2c 1 CC4)C(O)C(O)c 1 ccccc 1 -3 clccc2c3c(ccc2c l)Cc lc(ccc2ccccc 12)C3 clcc2ccc3cccc4ncc(cl)c2c34 O=Clc2ccccc2C(=O)c2clcc(OClOC(COC3OCC(O C[C@]12CC[C@H](OC(=O) / C=C / c3cccc(N(CCCl) )C(O)C3O)C(O)C(O)ClO)c(CO)c2O CCCl)c3)CCl=CC[C@@H] 1 [C@@H]2CC[C@]2( CCl=CC(=C(C#N)C#N)C=C( / C=C / c2ccc(N(C)C)cc C)NC(=O)CC[C@@H] 12 2)01CN(C)CCNC(=O)clcccn2cc(-c3ccccc3)ncl2 clccc2c(cl)Cclc(c3ccccc3c3cccccl3)C2 O=c ln(CC2CO2)c(=O)n(CC2CO2)c(=O)nl CC 1 CO clcc2c3oc3cc3ccc4c5oc5cclc4c321 Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- COclcc2c(cclOC)CN(N=O)CC2 c4ccc(N=Nc5c(N)c(S([O])(=O)=O)cc6cc(S([O])(=O O=[N+]([O- )=O)ccc56)cc4)cc3)cc2)ccl ])C1=CC2C=C3C(=CC2C=C1)C=C[C@H](O)[C@ O=C1OCCN1P(=O)(C1)N1CCOC1=OH]3O C[n+] Iccc(-c2ccccn2)ccl Cl=c2ccccc2=C2Cc3ccccc3C3=C4OC4=ClC32 0=S 1(=O)C2CCCC3CCC4C5CCCCC5CC Ic4c23 CCnlcc(C(=O)O)c(=O)c2ccc(C)nc21 C=CC(=O)N(CC 1 CO 1 )CC 1 CO 1 Nclcc2nc3ccccc3nc2cclO O=C(OCC 1 CCC2OC2C 1 )C 1 CCC2OC2C1 Cclcc(O)c2c(cl)C(=O)clcc(O)cc(OC3OC(CO)C(O) COC(=O)clcc2c([nH]c3ccccc32)c(- C(O)C3O)clC2=O c2ccc3c(n2)C(=O)C=CC3=O)nl O=S1(=O)CCCO1 NclccccclSCSclccccclN Nclccc2cc(S([O])(=O)=O)cc(O)c2clN=Nclccc(- c 1 cnc2c(c 1 )ccc 1 c2ccc2cccnc21 c2ccc(N=Nc3cc(S([O])(=O)=O)c4ccccc4c3N)cc2)cc Nclc(Oc2ccccc2)cc(O)c2clC(=O)clccccclC2=O 1 O=clc2ccc3c4c(ccc(c24)c2nc4ccccc4nl2)c(=O)nlc CCOC(=O)C 1 (CCCCCCOCc2cccs2)CO 1 2ccccc2nc31Cc 1 ccc(N=NNCc2ccccc2)cc 1 Cclccc2nc(-c3ccc(N)cc3)sc2clO=c lc(OC2OC(CO)C(O)C(O)C2O)c(- Nclccc(N)c2c(O)c3ccccc3c(O)cl2 c2ccc(O)c(O)c2)oc2cc(O)cc(O)c 12 Cnl c(N)nc2c3 ccc(=O) [nH] c3 ccc21Table 1G. Example compounds correctly characterized by AmesNet but missed by at least one of the six models referenced in Tables 1A-1FExample compounds missed from at least one Example compounds missed from at least one other model other model C(CC1CO1)C1CO1 Cl=CC2OC2c2clcclccc3cccc4ccc2clc34 C(CCC1CO1)CC1CO1 Cl=CC2OC2c2cc3ccc4cccc5ccc(c21)c3c45 C / C=C 1 / CC(C)C(C)(OC(C)=O)C(=O)OCC2=C3 C( Cl=CC2OC2c2ncccc21 CC[N+]3(C)C=C2)OC1=O Cl=COc2c(c3cc4ccccc4cc3c3ccccc23)C=Cl C 1=CC2CCCCC2C2CCCCC21 Cl=Cc2c3ccc4cccc5ccc(c6cccclc26)c3c45Example compounds missed from at least one Example compounds missed from at least one other model other model Cl=Cc2c3ccccc3cc3cccclc23 CCN(CCCl)clccc(CCCNc2c3ccccc3nc3ccccc23)ccl Cl=Cc2ccc3c4c(ccclc24)-clcccccl-3 CCOC(=O)C 1 (CCCCCCOCc2cccs2)CO 1 Cl=c2ccccc2=C2Cc3ccccc3C3=C4OC4=ClC32 CC[C@@] 1 (O)C[C@H](OC2CC(N(C)C)C(OC3 CC C1CSCSC1 (O)C(OC4CC(O)C(O)C(C)O4)C(C)O3)C(C)O2)c2c C=C / C 1=C / [C@H] (C) [C@] (C)(OC(C)=O)C(=O)O (CC3C(O)C4C(C(O)C3C2O)C(=O)C=CC4=O)C1C(=O) CC2=CC[N+]3(C)CC[C@@H](OC1=O)C23[O-] OCC=CC(=O)N(CC 1 CO 1 )CC 1 CO 1 CC[C@@] 1 (O)C[C@H](OC2CC(N(C)C)C(OC3 CC C=c 1 nccc2c 1 =Nc 1 cc(O)ccc 1 -2 (O)C(OC4CCC(=O)C(C)O4)C(C)O3)C(C)O2)c2c(c CC(=O)N(OC(=O)clcccccl)clccc2c(cl)Cclcccccl- C3C(C2O)C(=O)C2C(O)CCCC2C3=O)C1C(=O)OC 2 CCnlcc(C(=0)0)c(=0)c2cc(F)c(- CC(=O)OCCN(CCOC(C)=O)clccc(N=Nc2nc3ccc( c3ccc(O)cc3)c(F)c21Cl)cc3 s2)c(NC(=O)c2ccccc2)c 1 CCnlcc(C(=O)O)c(=O)c2ccc(C)nc21 CC(C)(COCC 1 CO 1)COCC 1 CO 1 CCnlcc(C(=O)O)c(=O)c2cnc(N3CCNCC3)nc21 CC(C)CCCC(C) [C@H] 1 CC [C@H] 2 [C@@H] 3 [C@ CN(C)CCNC(=O)clcccn2cc(-c3ccccc3)ncl2 H](OO)C=C4C[C@@H](O)CC[C@]4(C)[C@H]3C COC(=O)clcc2c([nH]c3ccccc32)c(- C[C@]12C c2ccc3c(n2)C(=O)C=CC3=O)nlCC 1 (C)CC2C(O)(C=C(C=O)C34CC23 C(=O)OC4O COclcc(-c2nc(-c3ccccc3)c(-c3ccccc3)[nH]2)ccclO )C1 COclcc(OC)c2c(cl)O[C@@H]lOC=C[C@H]21 CC 1 (C)C[C@H]2C=C(C=O) [C@] 34C[C@] 3 (C(=O COc 1CC2C(C3OC(=O)C4C(C 13)CCC4=O) [C@H] 1 [C@ )OC4O)[C@H]2C1 @H](O2)O[C@@H]2O[C@@H]21CC 1 (C)C[C@H]2 [C@] (O)(C=C(C=O) [C@] 3 (C=O) COclcc2c(c3oc(=O)c4c(cl3)CCOC4=O)ClC=COC C[C@]23C)C1 102 CCl=CC(=C(C#N)C#N)C=C( / C=C / c2ccc(N(C)C)cc COclcc2c(cclO)CCNC21CS[C@@H]2c3c(OC(C)= 2)01 O)c(C)c4c(c3 [C@H] (COC1=O)N 1 C(0)C3 Cc5cc(C) CC 1=CC2=C(C=O)C(=O)C(C)(O)C(O)C2=CO 1 c(OC)c(O)c5C([C@H]21)N3C)OCO4CC 1 CC2(OC2C)C(=O)OC2CCN(C)C / C=C( / COC(= COclcc2c(cclOC)CN(N=O)CC2 O)C1(C)O)C2=O COclccc2cc([N+](=O)[O-])c3ccc(=O)oc3c2cl CC 1 CNc2cc(O)c3ccc4ccccc4c3c21 C[C@]12CC[C@@H](OC(=O) / C=C / c3cccc(N(CCC CC 1 CS(=O)(=O)CCN l / N=C / c lccc([N+] (=0) [0- l)CCCl)c3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[ ])ol C@]2(C)NC(=0)CC[C@@H] 12 CClCS(=O)(=O)CCNlN=Cclccc([N+](=O)[O-])ol C[C@]12CC[C@H](OC(=O) / C=C / c3cccc(N(CCCl) CC10CC(=0)C(=0)C10 CCC1)C3)CC1=CC[C@@H]1[C@@H]2CC[C@]2( CCC / C=C / C(=O)O[C@@H]1C(C)(C)C[C@@H]2C C)NC(=O)CC[C@@H] 12 34CC3(C(C=O)=C[C@]12O)[C@@H](O)OC4=O C[C@]12CC[C@H](OC(=O)Cc3ccc(N(CCCl)CCCl CCN(CC)CCnlc2ccc(CO)c3c2c([n+] 1 [0-])- )cc3)C[C@@H] 1CC[C@@H] 1 [C@@H]2CC[C@]2 clccc(Cl)cclS3 (C)NC(=O)CC[C@@H] 12 CCN(CC)CCnlnc2c3c(c(CO)ccc31)Sclcc(Cl)cccl-2 C[n+] Iccc(-c2ccccn2)cclExample compounds missed from at least one Example compounds missed from at least one other model other model Cc 1 c [nH] c2ccc3 c(nc(=NO)n3 C)c2nl c2ccc(N=Nc3c(N)ccc4c(S([O])(=O)=O)ccc(O)c34)c Cclcc(O)c2c(cl)C(=O)clcc(O)cc(OC3OC(CO)C(O) c2)cclC(O)C3O)clC2=O Nclccc2cc(S([O])(=O)=O)cc(O)c2clN=Nclccc(- Cclccc(C)c2c3c(cccl2)C(0)C(0)C10C31 c2ccc(N=Nc3cc(S([O])(=O)=O)c4ccccc4c3N)cc2)cc Cclccc(N=NNCc2ccc([N+](=O)[O-])cc2)ccl 1Cc 1 ccc(N=NNCc2ccccc2)cc 1 Nc Iccccc 1 SCCSc Iccccc IN Cclccc(Nc2ccc3c4c(cccc24)C(=O)c2ccccc2-3)ccl NclccccclSCSclccccclN Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- Nnlnnc2c([nH]c3ccccc32)cl=O c4ccc(N=Nc5c(N)c(S([O])(=O)=O)cc6cc(S([O])(=O O=C(C=Cclccc([N+](=O)[O-])ccl)clcccccl )=O)ccc56)cc4)cc3)cc2)ccl O=C(Cnlccnc 1 [N+] (=0) [0-] )NCc 1 ccccc 1 Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- O=C(OCC 1 CCC2OC2C 1)C 1 CCC2OC2C1 c4ccc(N=Nc5c(O)ccc6cc(S([O])(=O)=O)cc(S([O])(= 0=C( [0-] )c 1 c [n+] (=O)c2ccccc2nl 0 O)=O)c56)c(C)c4)cc3 C)cc2)cc 1 0=C 1 C(=O)c2cccc3 cccc 1 c23Cc 1 ccc2cc3 c(c4c2c 1 CC4)C(O)C(O)c 1 ccccc 1 -3 O=ClC=Cc2ccc3c(ccc4ccccc43)c2Cl=O Cclccc2nc(-c3ccc(N)cc3)sc2cl O=ClC=Cc2ccc3c4c(ccclc24)C=CC3=O Cclnc2c(c3[nH]c4ccccc4cl3)C(=O)C(NlCCN(C)C 0=C 1 C=Cc2cccc3 cccc 1 c23C1)=CC2=O 0=C 1 CC(=Nc2ccccc2)C(=O)c2ccccc21 Cclnc2c(c3[nH]c4ccccc4cl3)C(=O)C=C(NlCCl)C O=ClCCC2(O)c3c(ccc(O)c31)- 2=0 clccc(0)c3clC2C10ClC3=0C1C 1 C=Cc2cc3c4c(ccc5cccc(c54)C=3)c2=C 1 O=ClCCCNlclncc([N+](=O)[O-])slCnlc(-c2ccccc2)c( / N=N / c2scc[n+]2C)c2ccccc21 0=C 1 CCCc2cc([N+] (=0) [O-])c3ccccc3c21 Cnlc(N)nc2c3ccc(=O)[nH]c3ccc21 0=C 1 CCc2c 1 cc(C(F)(F)F)c 1 c3 c(ccc21 ) [C@H] (0) [ Cnlc(N)nc2nc3ccccc3cc21 C@@H](0)C=C3 Cnlc(N)nc2nc3cnccc3cc21 0=C 1NCCCN l / N=C / c 1 ccc( [N+] (=0) [0-] )o 1 Cnlccc2ccc3ccc4ccc(O)cc4c3c21 O=ClOCCNl / N=C / clccc([N+](=O)[O-])ol NNclnc2ccc3 [nH]c4ccccc4c(=O)c3c2s 1 O=C1OCCN1P(=O)(C1)N1CCOC1=ONc 1 c(N=Nc2ccc( [N+] (=0) [0- O=Clc2ccccc2-c2clccclccccc21 ])cc2)c(S([O])(=O)=O)cc2cc(S([O])(=O)=O)c(N=Nc O=Clc2ccccc2-c2cc3ccccc3cc21 3ccccc3)c(O)cl2 O=Clc2ccccc2C(=O)c2clcc(OClOC(COC3OCC(O Nc 1 c(Oc2ccccc2)cc(O)c2c 1 C(=O)c 1 ccccc 1 C2=0 )C(O)C3O)C(O)C(O)ClO)c(CO)c2O Nclcc2nc3ccccc3nc2cclO O=Clc2ccccc2C(=O)c2clcc(O[C@H]lO[C@H](C Nclccc(N)c2c(O)c3ccccc3c(O)cl2 O[C@@H]3OC[C@@H](O)[C@H](O)[C@H]3O)[ Nclccc(N=Nc2ccc(N=Nc3c(S([O])(=O)=O)cc4cc(S( C@@H](O)[C@H](O)[C@H] 10)c(C0)c20 [O])(=O)=O)c(N=Nc5ccc(N=Nc6ccc(N)cc6N)cc5)c( O=NN 1 CCCN(N=O)CC 1 O)c4c3N)cc2)c(N)cl 0=S (=O)(Cc Iccccc 1 )N 1 C2c3 ccccc3 -c3 ccccc3 C21 Nclccc(N=Nc2cccc(N=Nc3ccc(N)cc3N)c2)c(N)cl O=S1(=O)CCCO1 Nclccc2c(S([O])(=O)=O)ccc(O)c2clN=Nclccc(- 0=S 1(=O)C2CCCC3CCC4C5CCCCC5CC Ic4c23Example compounds missed from at least one Example compounds missed from at least one other model other model O=[N+]([O- clcc2c3c(cl)=C10ClC=3Cclcc3ccccc3ccl-2 ])C1=CC2C=C3C(=CC2C=C1)C=C[C@H](O)[C@ clcc2c3c(cl)cclccc4cccc5cc(c3clc45)CC2 H]3O clcc2c3c(cccc3cl)ClNC21O=[N+]([O-])clccc( / C=N / C2SNc3ccccc32)ol clcc2c3oc3cc3ccc4c5oc5cclc4c32 O=[N+]([O- clcc2cc3c4c(ccc5cc6c(clCCC6)c2c54)CCC3 ])clccc(C2OCC3(CO)COC(c4ccc([N+](=O)[O- clcc2ccc3cc4c(c5ccc(cl)c2c35)CCC10C41 ])cc4)N23)ccl clcc2ccc3ccc(C4CO4)c4ccc(cl)c2c34 O=[N+]([O-])clccc2c(cl)CCclcc3c(ccl-2)CCCC3 clcc2ccc3cccc4ncc(cl)c2c34O=[N+]([O- clccc([C@H]2N[C@@H]2c2ccccc2)ccl ])clccc2c3c4c(cccc4c4ccccc4cl3)[C@H](O)[C@H] clccc([C@H]2N[C@H]2c2ccccc2)ccl20 c 1 ccc2 [nH] c(-c3 cscn3 )nc2c 1 O=clc(=O)c2cc3ccccc3c3ccc4cccclc4c23 c 1 ccc2c(CCC3 C03 )cccc2c 10=c lc(OC2OC(CO)C(O)C(O)C2O)c(- clccc2c(cl)-clc(c3ccccc3c3cccccl3)ClNC21 c2ccc(O)c(O)c2)oc2cc(O)cc(O)c 12 clccc2c(cl)CCclcc3ccc4c(c3ccl-2)CCCC4 O=clc2c(c3c(ccc4cc5c(ccc6ccccc65)[nH]c43)clO) clccc2c(cl)Cclc(c3ccccc3c3cccccl3)C202 clccc2c(cl)clccccclclcc3cnccc3cc21 O=clc2ccc3c4c(ccc(c24)c2nc4ccccc4nl2)c(=O)nlc clccc2c(cl)clccccclclccccc212ccccc2nc31 clccc2c(cl)cclc3c2ccc2cccc(c23)C2NC12 O=clc2ccccc2[nH]c2ccc3c(cl2)OCCO3 clccc2c(cl)cclc3ccccc3c3cccc4ccc2clc43 0=c ln(CC2CO2)c(=O)n(CC2CO2)c(=O)nl CC 1 CO clccc2c(cl)ccclc3ccncc3ccc211 clccc2c(cl)ccclcc3c(ccc4ccncc43)ccl2 0ClC=Cc2c(ccc3ccccc23)C10 clccc2c3c(ccc2c l)Cc lc(ccc2ccccc 12)C3 O[C@H] 1[C@H]2O[C@H]2C=C[C@@H] 10 clccc2cc(C3 C03 )ccc2c 1[N-]=[N+]=C1C=NC(=O)NC1=O c 1 cnc2c(c 1 )ccc 1 c2ccc2cccnc21[0-] [n+] Iccc2cnncc2cl c 1 cnc2c(c 1 )ccc 1 c3 cccnc3 ccc 12[0] S(=O)(=O)c 1CCC(OC2CCC3C4C(CCCC24)CC2CCCCC2 clcnc2ccc3c(c2cl)C10ClCC3-3)cclTable 1H. Example compounds correctly characterized by AmesNet but missed by all six models referenced in Tables 1A-1FExample compounds missed by all other models Example compounds missed by all other models C1CSCSC1 COclcc2c(cclOC)CN(N=O)CC2 C=CC(=O)N(CC 1 CO 1 )CC 1 CO 1 C[C@]12CC[C@H](OC(=O) / C=C / c3cccc(N(CCCl) CCl=CC(=C(C#N)C#N)C=C( / C=C / c2ccc(N(C)C)cc CCCl)c3)CCl=CC[C@@H]l[C@@H]2CC[C@]2( 2)01 C)NC(=O)CC[C@@H] 12 CC10CC(=0)C(=0)C10 Cc 1 ccc(N=NNCc2ccccc2)cc 1CCOC(=O)C 1 (CCCCCCOCc2cccs2)CO 1 Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(-Example compounds missed by all other models Example compounds missed by all other models c4ccc(N=Nc5c(N)c(S([O])(=O)=O)cc6cc(S([O])(=O O=clc(OC2OC(CO)C(O)C(O)C2O)c(- )=O)ccc56)cc4)cc3)cc2)ccl c2ccc(O)c(O)c2)oc2cc(O)cc(O)c!2 Cclccc(S(=O)(=O)Oc2ccc( / N=N / c3ccc(- O=clc2ccc3c4c(ccc(c24)c2nc4ccccc4nl2)c(=O)nlc c4ccc(N=Nc5c(O)ccc6cc(S([O])(=O)=O)cc(S([O])(= 2ccccc2nc31O)=O)c56)c(C)c4)cc3 C)cc2)cc 1 O=c ln(CC2CO2)c(=O)n(CC2CO2)c(=O)nl CC 1 CO O=C 1 CC(=Nc2ccccc2)C(=O)c2ccccc21 1O=C1OCCN1P(=O)(C1)N1CCOC1=O [O-] [n+] Iccc2cnncc2clO=Clc2ccccc2-c2cc3ccccc3cc21 c 1 ccc2 [nH] c(-c3 cscn3 )nc2c 1O=S 1 (=O)c2cccc3ccc4c5ccccc5cc Ic4c23Ames Testing Conditions

[0126] In some embodiments of the methods and systems provided herein, an Ames testing condition comprises (1) the identity of a bacterial tester strain or (2) the status of a metabolic activation. In some embodiments, the Ames testing condition comprises (1) the identity of a bacterial tester strain and (2) the status of a metabolic activation. In some embodiments, the method or system disclosed herein comprises providing one or more Ames testing conditions. In some embodiments, the method or system disclosed herein comprises providing two or more Ames testing conditions. For example, each Ames testing condition can be distinct from another. In some embodiments, the status of the metabolic activation comprises the presence or absence of a metabolic activation. In some embodiments, the metabolic activation comprises an S9 fraction. The S9 fraction can be a crude liver homogenate from animal. In some embodiments, the S9 fraction is from rat, hamster or human.

[0127] In some embodiments of the methods and systems provided herein, a bacterial tester strain is a strain of Salmonella or Escherichia. In some embodiments, the bacterial tester strain is a Salmonella strain. In some embodiments, the bacterial tester strain is a Salmonella typhimurium (S. typhimurium) strain. In some embodiments, the one or more Ames testing conditions comprise more than one bacterial tester strains. In some embodiments, the one or more Ames testing conditions comprise both Salmonella and Escherichia strains. In some embodiments, the one or more Ames testing conditions comprise more than one Salmonella strains. In some embodiments, a bacterial tester strain is a base-pair substitution mutant strain or a frameshift mutant strain. In some embodiments, a base-pair substitution mutant strain is selected from S. typhimurium TAI 00, TAI 02, TAI 04, and TAI 535. In some embodiments, a frameshift mutant strain is selected from S. typhimurium TA1537, TA1538, TA97, and TA98. In some embodiments, the one or more Ames testing conditions comprise at least one base-pair substitution mutant strain and / or at least one frameshift mutant strain. For example, the one or more Amestesting conditions comprise S. typhimurium TA100 and TA1537. In some embodiments, the one or more Ames testing conditions comprise two or more base-pair substitution mutant strains. For example, the one or more Ames testing conditions comprise S. typhimurium TAI 00 and TAI 535. In some embodiments, the one or more Ames testing conditions comprise two or more frameshift mutant strains. For example, the one or more Ames testing conditions comprise S. typhimurium TAI 537 and TA98.Mutagenicity Assessment

[0128] In some embodiments of the methods and systems provided herein, the profile comprises at least two, three, four, five, six, seven, eight, nine, or ten mutagenicity probability scores. In some embodiments, the mutagenicity probability score is between 0 and 1. For example, the mutagenicity probability score can be 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9 or 1. In some embodiments, the profile comprises mutagenicity probability scores for at least two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, or sixteen Ames testing conditions. In some embodiments, the profile comprises mutagenicity probability scores for Ames testing conditions corresponding to at least one, two, three, four, five, six, seven, or eight distinct bacterial tester strains, each tested in the presence and / or absence of a metabolic activation. For example, the profile comprises mutagenicity probability scores for Ames testing conditions corresponding to Salmonella strains TA100, TA102, TA104, TA1535, TA1537, TA1538, TA97, and TA98, each tested both in the presence or absence of a metabolic activation. In some embodiments, the profile comprises mutagenicity probability scores for Ames testing conditions corresponding to at least one, two, three, four, five, six, seven, or eight distinct bacterial tester strains, each tested both in the presence and absence of a metabolic activation. In some embodiments, the profile comprises mutagenicity probability scores for Ames testing conditions corresponding to at least one, two, three, four, five, six, seven, or eight Salmonella strains, each tested both in the presence and absence of a metabolic activation. For example, the profile comprises mutagenicity probability scores for Ames testing conditions corresponding to Salmonella strains TA100, TA102, TA104, TA1535, TA1537, TA1538, TA97, and TA98, each tested both in the presence and absence of a metabolic activation.

[0129] In some embodiments, the method or system disclosed herein comprises determining or stratifying mutagenicity. In some embodiments, determining or stratifying mutagenicity comprises comparing at least one mutagenicity probability score to a threshold. In some embodiments, the threshold is 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8 or 0.9. In some embodiments, the threshold is 0.3, 0.4 or 0.5.

[0130] In some embodiments of the methods and systems provided herein, a negative determination is made when the mutagenicity probability score is below the threshold. In someembodiments, a mutagenicity probability score is between 0 and 1; and a negative determination is made when the mutagenicity probability score is below (or at or below) 0.3, 0.4, or 0.5. For example, the mutagenicity probability score of a compound is 0.2 in S. typhimurium TAI 537 in the presence of a metabolic activation and the threshold is 0.3. Thus, a negative determination is made for the compound in S. typhimurium TAI 537 in the presence of a metabolic activation.

[0131] In some embodiments of the methods and systems provided herein, a positive determination is made when the mutagenicity probability score is above the threshold. In some embodiments, a mutagenicity probability score is between 0 and 1; and a positive determination is made when the mutagenicity probability score is above (or at or above) 0.3, 0.4, or 0.5. In some embodiments, a mutagenicity probability score is between 0 and 1; and a positive determination is made when the mutagenicity probability score is above (or at or above) 0.5. For example, the mutagenicity probability score of a compound is 0.8 in S. typhimurium TA1537 in the presence of a metabolic activation and the threshold is 0.3. Thus, a positive determination is made for the compound in S. typhimurium TAI 537 in the presence of a metabolic activation.

[0132] In some embodiments of the methods and systems provided herein, a strongly positive determination is made when the mutagenicity probability score is above the threshold for strongly positive determination. In some embodiments, the threshold for strongly positive determination is 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8 or 0.9. In some embodiments, the threshold for strongly positive determination is 0.5, 0.6, 0.7, 0.8 or 0.9. In some embodiments, a mutagenicity probability score is between 0 and 1; and a strongly positive determination is made when the mutagenicity probability score is above (or at or above) 0.5.

[0133] In some embodiments of the methods and systems provided herein, the determination or stratification of the compound is based on the determination under at least one profiled Ames testing condition. In some embodiments, the determination or stratification of the compound is based on the determination under one profiled Ames testing condition. In some embodiments, the compound is determined or stratified as mutagenic based on a determination that the compound is positive (or strongly positive) under at least one profiled Ames testing condition. In some embodiments, the compound is determined or stratified as non-mutagenic based on a determination that the compound is negative under at least one profiled Ames testing condition.

[0134] In some embodiments of the methods and systems provided herein, the determination or stratification of the compound is based on the determination under more than one (optionally, at least two, three, four, five, six, seven, eight, nine or ten) profiled Ames testing conditions. In some embodiments, the compound is determined or stratified as mutagenic based on a determination that the compound is positive (or strongly positive) under at least two profiledAmes testing conditions. In some embodiments, the compound is determined or stratified as non-mutagenic based on a determination that the compound is negative under at least two profiled Ames testing conditions.

[0135] In some embodiments of the methods and systems provided herein, the compound is determined or stratified as highly mutagenic based on a determination that the compound is strongly positive under at least one profiled Ames testing condition. In some embodiments, the compound is determined or stratified as highly mutagenic based on a determination that the compound is positive (or strongly positive) under at least two profiled Ames testing conditions. In some embodiments, the compound is determined or stratified as highly mutagenic based on a determination that the compound is strongly positive under at least two profiled Ames testing conditions. In some embodiments, the compound is determined or stratified as highly mutagenic based on a determination that the compound is strongly positive under all profiled Ames testing conditions. For example, the compound is determined or stratified as highly mutagenic based on a determination that the compound is strongly positive under five profiled Ames testing conditions.

[0136] In some embodiments of the methods and systems provided herein, the determination or stratification of the compound is based on the determination under all profiled Ames testing conditions. In some embodiments, the compound is determined or stratified as mutagenic based on a determination that the compound is positive (or strongly positive) under all profiled Ames testing conditions. In some embodiments, the compound is determined or stratified as non-mutagenic based on a determination that the compound is negative under all profiled Ames testing conditions.

[0137] In some embodiments of the methods and systems provided herein, the compound is further classified into an Ames class. Compounds can be classified into one of the three Ames classes: Class A, Class B and Class C. Class A compounds are strongly positive or mutagenic. The Class A compound generally induces more than 1,000 revertant colonies per mg of at least one Ames test strain in the presence or absence of S9. Class B compounds are positive. The Class B compound induces at least a 2-fold increase in revertant colonies (but less than for Class A compounds) compared to the negative control in at least one Ames strain in the presence or absence of S9. Class C compounds are negative, with no revertants increased (<2-fold) (or neither Class A nor B).

[0138] In some embodiments of the methods and systems provided herein, the compound is determined or stratified as an Ames Class A compound. In some embodiments, the compound is determined or stratified as an Ames Class A compound, if the compound is determined to be strongly positive under at least one profiled Ames testing condition. In someembodiments, the compound is determined or stratified as an Ames Class A compound, if the compound is determined to be strongly positive under two or more profiled Ames testing conditions. In some embodiments, the compound is determined or stratified as an Ames Class A compound, if the compound is determined to be strongly positive under all profiled Ames testing conditions.

[0139] In some embodiments of the methods and systems provided herein, the compound is determined or stratified as an Ames Class B compound. In some embodiments, the compound is determined or stratified as an Ames Class B compound, if the compound is determined to be positive (but not strongly positive) under at least one profiled Ames testing condition. In some embodiments, the compound is determined or stratified as an Ames Class B compound, if the compound is determined to be positive (but not strongly positive) under two or more profiled Ames testing conditions. In some embodiments, the compound is determined or stratified as an Ames Class B compound, if the compound is determined to be positive (but not strongly positive) under all profiled Ames testing conditions.

[0140] In some embodiments of the methods and systems provided herein, the compound is determined or stratified as an Ames Class C compound. In some embodiments, the compound is determined or stratified as an Ames Class C compound, if the compound is determined to be negative under at least one profiled Ames testing condition. In some embodiments, the compound is determined or stratified as an Ames Class C compound, if the compound is determined to be negative under two or more profiled Ames testing conditions. In some embodiments, the compound is determined or stratified as an Ames Class C compound, if the compound is determined to be negative under all profiled Ames testing conditions.

[0141] In some embodiments of the methods and systems provided herein, the compound is further classified into one of the mutagenicity risk levels. In some embodiments, the compound is determined or stratified with a low mutagenicity risk. In some embodiments, the compound is stratified with a low mutagenicity risk based on a determination that the compound is negative under at least one profiled Ames testing condition. In some embodiments, the compound is stratified with a low mutagenicity risk based on a determination that the compound is negative under two or more profiled Ames testing conditions. In some embodiments, the compound is stratified with a low mutagenicity risk based on a determination that the compound is negative under all profiled Ames testing conditions.

[0142] In some embodiments of the methods and systems provided herein, the compound is determined or stratified with a medium mutagenicity risk. In some embodiments, the compound is stratified with a medium mutagenicity risk based on a determination that the compound is positive under only one profiled Ames testing condition. In some embodiments, thecompound is stratified with a medium mutagenicity risk based on a determination that the compound is positive under two, three, or four profiled Ames testing conditions. In some embodiments, the compound is stratified with a medium mutagenicity risk based on a determination that the compound is positive under no more than half of the profiled Ames testing conditions, if at least two Ames testing conditions are profiled.

[0143] In some embodiments of the methods and systems provided herein, the compound is determined or stratified with a high mutagenicity risk. In some embodiments, the compound is stratified with a high mutagenicity risk based on a determination that the compound is positive under at least two profiled Ames testing conditions. In some embodiments, the compound is stratified with a high mutagenicity risk based on a determination that the compound is positive under all profiled Ames testing conditions. In some embodiments, the compound is stratified with a high mutagenicity risk based on a determination that the compound is positive under at least half of the profiled Ames testing conditions, if at least two Ames testing conditions are profiled.

[0144] In some embodiments of the methods and systems provided herein, the method or system disclosed herein determines or stratifies the plurality of compounds at a sensitivity of at least 0.70, 0.75, or 0.80. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at a sensitivity of at least 0.70. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at a sensitivity of at least 0.75. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at a sensitivity of at least 0.80. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at a specificity of at least 0.70, 0.75, or 0.80. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at a specificity of at least 0.70. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at a specificity of at least 0.75. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at a specificity of at least 0.80. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at an accuracy (optionally, a balanced accuracy) of at least 0.70, 0.75, or 0.80. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at an accuracy (optionally, a balanced accuracy) of at least 0.70. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at an accuracy (optionally, a balanced accuracy) of at least 0.75. In some embodiments, the method or system disclosed herein determines or stratifies the plurality of compounds at an accuracy (optionally, a balanced accuracy) of at least 0.80.EXAMPLES

[0145] Some aspects of the embodiments discussed above are disclosed in further detail in the following examples, which are not in any way intended to limit the scope of the present disclosure.

[0146] Various modifications of the invention and many further embodiments thereof, in addition to those shown and described herein, will become apparent to those skilled in the art from the full contents of this document, including references to the scientific and patent literature cited herein. The subject matter herein contains important information, exemplification and guidance that can be adapted to the practice of this invention in its various embodiments and equivalents thereof.

[0147] Example 1: AmesNet: A Deep Learning Model Enhancing Generalization in Ames Mutagenicity PredictionMethods

[0148] Data availability: The best available publicly accessible dataset was used for this study.

[0149] Dataset source: For model training, the reference dataset used in this study included those described in this example, used to compare AmesNet against, and for which the authors provided a link in their GitHub repository. Compiled from four individual data sources (e.g., ISSSTY, OASIS, EFSAP, and MHLW), downloaded from the OECD QSAR Toolbox, the dataset used in this study for model training, included Ames annotations from eight Salmonella typhimurium strains, each indicative of a specific mutation type: TAI 00 (base-pair substitution type), TAI 02 (base-pair substitution type), TAI 04 and TAI 535 (base-pair substitution type); TA1537 (frameshift type); and TA1538, TA97, and TA98 (frameshift type). Tested in the presence or absence of S9 fraction, which was an extract that mimics compound activation in the liver.

[0150] Dataset clean up approach: The small molecules in the dataset were converted into SMILES format and standardized using a custom methodology. High-quality SMILES strings were obtained having canonical tautomerism and consistent stereochemistry throughout the dataset used in this study. After SMILES standardization, the data was merged with the prepartitioned in-domain and out-of-domain datasets using the compound identifier. The merged dataset was further cleaned by organizing strain and S9 condition information into separate columns. Both the in-domain and out-of-domain datasets, each containing 48,735 rows, were prepared for model training, validation, and testing according to the predefined data splits.

[0151] Dataset partition approach: For the data partition, t-SNE chemical diversity split was used. The splitting proportions were 45% training, 45% validation, and 10% test. Thepartition proportions used in the comparator publication were 70% training, 20% validation, and 10% test.

[0152] Hyperparameter optimization approach: Grid search was used for hyperparameter optimization, using the highest area under the receiver operating characteristic curve (AUROC), value as the target parameter.

[0153] Model Retraining: Having obtained optimized hyperparameters, the four AmesNet models were retrained with the recombination of the train and validation data. The probabilities used were averaged over 10 models, and thereafter bootstrapped.

[0154] Reference Model Comparison: AmesNet was benchmarked against three reference models: single-task (ST), ungrouped multitask (uMT), and grouped multitask (gMT). The “overall” model was excluded as it benefited from post hoc selection, making it an ensemblelike construct rather than a single, directly comparable model. Although the “overall” model’s results are not shown, AmesNet still achieved higher balanced accuracy and sensitivity.

[0155] In this Example, the experimental materials and methods described above were used in the studies and analyses described below.

[0156] For a drug candidate to make it past the preclinical trial stage, mutagenicity must be ruled out. In the preclinical pipeline, drug mutagenicity is assessed using the Salmonella / microsome mutagenicity assay, colloquially called the Ames test. In nearly all instances, the FDA necessitates the Ames test for IND submission and drug approval. Difficult to scale up, the Ames test has given way to artificial intelligence (Al)-driven computational alternatives, which predict strain-specific Ames mutations in quantitative structure-activity relationship (QSAR) experiments. A multitask Ames QSAR algorithm, which predicts eight strain-specific mutations, matched state-of-the-art performance as measured by balanced accuracy. But, like most other neural network applications used to predict Ames genotoxicity, these multitask algorithms struggle with sensitivity and fail to generalize efficiently to out-of-domain data. In this study, ChemPrint, a molecular geometric convolutional neural network, was developed, which generalized well to out-of-domain data. Validated across multiple protein targets in the drug discovery programs disclosed herein, ChemPrint achieved a 46% (19 / 41) zeroshot hit rate in AXL and BRD4 oncology program, a 100% (12 / 12) one-shot hit rate in RdRp Thumb- 1 antiviral program, and enabled off-target binding reductions ranging from 800-fold to over 15,000-fold. In this study, AmesNet, an adaptation of ChemPrint, was provided, which predicted strain-specific Ames mutations. AmesNet learned from the concatenation of ChemPrint’ s molecular embeddings with Ames mutagenicity features. When benchmarked against leading Ames QSAR algorithms, AmesNet achieved state-of-the-art balanced accuracy and sensitivity across both in-domain and out-of-domain data. AmesNet attained a sensitivity of0.75, marking a 30% improvement over the best-performing ML method on the test set. AmesNet was embedded in GALILEO, an Al-based drug discovery platform.Introduction

[0157] Many potential drugs with promising targeted activity do not make it into clinical studies because of unfavorable in vivo interactions. Measurements such as absorption, distribution, metabolism, excretion, toxicity (ADMET), and chemical mutagenicity are a bottleneck in the drug discovery pipeline. Mutagenicity is typically assessed experimentally only after a significant investment of time and money in optimizing on-target activity. Mutagenicity is screened for in the preclinical stage using the Salmonella / microsome mutagenicity test, colloquially known as the Ames test. The Ames test is required for IND submission and drug approval by the FDA in nearly all cases. The study must follow GLP guidelines and be conducted using five different bacterial strains, each tested with and without S9 metabolic activation, for a total of 10 tests. This mutagenicity assay comprises a battery of histidine auxotroph Salmonella strains, which have a reversible mutation in the his gene that regulates histidine biosynthesis. Compounds are tested in the presence or absence of a liver fraction (S9), which mimics metabolic activation in the human body. In each Salmonella strain, the mutation is either a single substitution or a frameshift with a unique sensitivity to mutagens. If mutagenic, a drug candidate will revert specific mutation types. Having recovered histidine synthesis capabilities, the revertant Salmonella strains grow in histidine-free media (FIG. 1A).

[0158] The Ames test is expensive and difficult to scale, costing more than $10,000 per compound under GLP conditions. This makes it impractical for routine screening of large compound libraries in drug discovery campaigns. An alternative to the Ames test is quantitative structure-activity relationship (QSAR) algorithms, which can leverage deep learning networks to predict strain-specific Ames mutagenicity. Most Ames QSAR algorithms are limited in scope as they restrict their predictions to one strain-specific mutation type (FIG. IB). Published recently, a multitask Ames QSAR algorithm outperforms its single-task predecessors in balanced accuracy and sensitivity, by predicting the eight strain-specific mutation types commonly assessed in the Ames test (FIG. 1C and FIG. ID). Despite these advances, current Ames QSAR algorithms demonstrate low sensitivity and generalize poorly to out-of-domain data. Ames QSAR algorithms with low sensitivity are particularly concerning, as they produce more false negatives. A false negative in this context is the most serious type of error, as it allows a genotoxic compound to advance further in the drug discovery pipeline undetected. This results in wasted resources and costly follow-up studies. More critically, if such a compound also escapes detection in later Ames testing, it may advance to clinical trials, where it could lead to patient harm. Additionally, Ames QSAR algorithms show reduced performance when predicting genotoxicity for novel compoundclasses that were absent from the training set, a challenge known as out-of-domain generalization. This drop in performance on novel chemical structures greatly reduces the model’s practical utility and trustworthiness, as it fails to provide reliable predictions for the majority of untested compounds encountered in real-world drug discovery. Without the ability to generalize beyond its training data, the model offers little value in screening efforts, where identifying previously unrecognized genotoxic liabilities is critical. These deficiencies were evident in the first and second Ames / QSAR International Challenge Projects, where widely used models such as Sarah Nexus and Derek Nexus, as well as newer models like the FDA’s DeepAmes, demonstrated limited success.

[0159] Suboptimal deep learning algorithms in drug discovery lead to costly and misguided decisions, producing inaccurate predictions that fail to detect critical molecular properties, as seen in the current state of Ames QSAR algorithms. To address this problem in drug discovery programs, GALILEO, a computational drug development platform with Al-based generative chemistry capabilities, was developed. Embedded in the GALILEO ecosystem, ChemPrint is a molecular geometric convolutional neural network that uniquely generalizes out-of-domain data effectively. ChemPrint has been extensively validated across multiple protein targets within drug discovery programs. In out-of-domain settings, it achieved a 46% (19 / 41) zeroshot hit rate in AXL and BRD4 oncology program and a 100% (12 / 12) one-shot hit rate in RdRp Thumb- 1 antiviral program. Additionally, ChemPrint enabled the elimination of off-target binding, achieving selectivity improvements ranging from 800-fold to over 15,000-fold. ChemPrint’ s out-of-domain generalization was achieved through adaptive molecular embeddings and a hyperparameter optimization strategy on t-distributed Stochastic Neighbor Embedding (t-SNE) data splits. ChemPrint is a moldable algorithm and it can perform different tasks when trained accordingly, without losing predictive power.

[0160] AmesNet, an innovative version of ChemPrint, which predicts strain-specific Ames mutations, is provided in this study. AmesNet learned to predict compound mutagenicity from the concatenation of ChemPrint’ s adaptive molecular embeddings to Ames-specific features (FIG. 2). To assess AmesNet’ s performance, it was compared to a set of single-task and multitask Ames QSAR algorithms illustrated in FIG. 1B-FIG. ID. The results showed that when predicting features from in-domain or out-of-domain data, AmesNet outperformed the comparator Ames QSAR algorithms in balanced accuracy and sensitivity. Like ChemPrint, AmesNet can also generalize well.Results and Discussion

[0161] The goal of this study was to benchmark AmesNet against the three Ames QSAR algorithms illustrated in FIG. 1B-FIG. ID: single task (ST), ungrouped multitask (uMT),and grouped multitask (gMT). All data used in this study were sourced directly from the reference publication. The test datasets, both in-domain and out-of-domain, remained unchanged across all experiments to ensure consistency in evaluation. AmesNet models were trained and optimized using two different approaches. In the first approach, the same training and validation datasets were used as those reported in the reference publication. In the second approach, the training and validation datasets were partitioned using the t-SNE-based methodology. Four metrics were used in these comparisons: balanced accuracy, sensitivity, specificity, and Area Under the Receiver Operating Characteristic Curve (AUROC).

[0162] When using the in-domain datasets from the previous publication, AmesNet outperformed the three Ames QSAR models used as comparators (ST, uMT, and gMT). AmesNet performed even better when using the t-SNE approach to partition in-domain data (FIG. 3 and Table 2). These results mean that in the case of in-domain data, tuning on the reference paper’s data split by AmesNet was enough to outperform current QSAR algorithms in balanced accuracy and sensitivity. The t-SNE data split for hyperparameter optimization offered AmesNet an extra advantage over the Ames QSAR comparators by increasing balanced accuracy and sensitivity rate. Like ChemPrint, AmesNet was designed for maintaining predictive power on out-of-domain data. When compared against the three Ames QSAR comparators on out-of-domain data, AmesNet was superior, with and without using the t-SNE hyperparameter optimization approach (FIG. 4 and Table 3) AmesNet attained a balanced accuracy of 0.82 and a sensitivity of 0.75, representing improvements of approximately 10% and 30%, respectively, over the best-performing ML method on the test set. The ability to generalize on unseen data with high balanced accuracy and sensitivity made AmesNet a unique algorithm. Sensitivity is the most critical metric in Ames mutagenicity prediction, as false negatives, or missed detections of genotoxic compounds, pose the greatest risk. A model with low sensitivity fails to identify hazardous compounds early, allowing them to advance through the drug discovery pipeline undetected. This leads to wasted resources, costly follow-up studies, and an increased risk of genotoxic compounds progressing into clinical trials, where late-stage failures are significantly more consequential. Again, the t-SNE data split for hyperparameter optimization provided AmesNet with an extra generalization advantage, which was more pronounced when presented with out-of-domain data. These results put AmesNet ahead of the comparator QSARs, even those that competed in this year’s Ames / QSAR International Challenge Projects.Table 2: Averaged In-Domain Test Set Prediction Values by AmesNet versus Multitask QSAR algorithmsDL Architecture Balanced Accuracy Sensitivity Specificity ROC AUCSTL (reference) 0.610 0.273 0.948 0.745(0.536-0.689) (0.128-0.423) (0.911-0.979) (0.635-0.842) 0.754 0.571 0.937 0.894uMTL (reference) (0.662-0.842) (0.391-0.738) (0.895-0.972) (0.831-0.944) 0.771 0.603 0.940 0.87gMTL (reference) (0.676-0.857) (0.425-0.767) (0.898-0.975) (0.800-0.934) 0.789 0.620 0.959 0.880 AmesNet - base HPO (0.772-0.806) (0.586-0.653) (0.953-0.965) (0.864-0.894) 0.796 0.668 0.924 0.878 AmesNet - t-SNE HPO (0.780-0.812) (0.635-0.700) (0.916-0.932) (0.863-0.893) Values are presented with 95% confidence intervals in parentheses.The best learning architecture for each strain task is in bold, whether STL, uMTL, gMTL, orAmesNet.Table 3: Averaged Out-of-Domain Test Set Prediction Values by AmesNet versus Multitask OSAR algorithmsDL Architecture Balanced Accuracy Sensitivity Specificity ROC AUC 0.700 0.448 0.951 0.799STL (reference) (0.592-0.813) (0.238-0.675) (0.913-0.979) (0.664-0.915) 0.716 0.542 0.890 0.823uMTL (reference) (0.602-0.828) (0.317-0.749) (0.834-0.943) (0.713-0.913) 0.745 0.575 0.915 0.832gMTL (reference) (0.628-0.858) (0.347-0.788) (0.863-0.957) (0.721-0.923) 0.788 0.669 0.907 0.861 AmesNet - base HPO (0.771-0.808) (0.636-0.708) (0.898-0.916) (0.843-0.879) AmesNet - t-SNE 0.818 0.749 0.886 0.882HPO (0.800-0.835) (0.718-0.783) (0.877-0.896) (0.865-0.900) Values are presented with 95% confidence intervals in parentheses.The best learning architecture for each strain task is in bold, whether STL, uMTL, gMTL,or AmesNet.

[0163] Analysis of compound structures in the out-of-domain test data revealed a distinct advantage of AmesNet over the gMTL model: its ability to detect genotoxic polyaromatic compounds. Polyaromatic compounds are a well-known source of false negatives in Ames QSAR mutagenicity prediction and have proven difficult for models to classify accurately. AmesNet correctly identified all polyaromatic compounds in FIG.5 as genotoxic. In comparison, the gMTL model misclassified all compounds in FIG. 5 as non-genotoxic, failing to detect their mutagenic potential. AmesNet’ s performance on the polyaromatic compound class demonstrated its improved sensitivity. These results validated AmesNet’ s ability to address limitations in existing models and improve predictive accuracy for high-risk chemical classes.

[0164] AmesNet was a variation of ChemPrint. It was shown that ChemPrint, which leveraged t-SNE splitting approach for data training and hyperparameter optimization, exhibited remarkable superiority over existing algorithms. Its ability to generate highly selective ligands with potent on-target activity while minimizing off-target interactions highlighted its versatility in drug discovery. ChemPrint has demonstrated strong predictive performance across multipleprograms, achieving a 46% (19 / 41) zero-shot hit rate in the AXL and BRD4 oncology program and a 100% (12 / 12) one-shot hit rate in the RdRp Thumb-1 antiviral program. Additionally, it significantly improved selectivity, reducing off-target binding by factors ranging from 800-fold to over 15,000-fold. These results showed that ChemPrint can be trained to perform different functions (e.g., identify ligands, and thereafter remove those with poly-affinity features) within the same drug discovery pipeline. Leveraging ChemPrint’ s unique t-SNE splitting approach was instrumental in this case. Without the t-SNE splitting approach, ChemPrint would have failed to generalize beyond known chemical entities.Conclusions

[0165] As shown in this Example, AmesNet, an adaptation of ChemPrint for the prediction of compound mutagenicity, outperformed existing multitask Ames QSAR algorithms (FIG. 2, Table 2, Table 3). Sensitivity is the most important metric in mutagenicity prediction, as low sensitivity increases the risk of false negatives, allowing genotoxic compounds to advance undetected through the drug discovery pipeline. AmesNet differentiated itself from existing Ames QSAR models by achieving superior sensitivity while maintaining high balanced accuracy, minimizing the likelihood of toxic compounds evading detection.

[0166] Example 2: AmesNet: A Task-Conditioned Deep Learning Model with Enhanced Sensitivity and Generalization in Ames Mutagenicity PredictionMethodsData

[0167] Dataset Source: The best available publicly accessible dataset was used for this study. The dataset was compiled from four individual data sources: ISSSTY, OASIS, EFSAP, and MHLW. The dataset included Ames test annotations from eight Salmonella strains, each corresponding to a specific mutation type. The base-pair substitution strains were TAI 00, TAI 02, TA104, and TA1535. The frameshift mutation strains were TA1537, TA1538, TA97, and TA98. Assays were conducted in the presence or absence of S9 fraction.

[0168] Dataset Cleaning: The small molecules in the dataset were converted from SDF to SMILES format and standardized. The OOD dataset was prepared for model training, validation, and testing according to the predefined data splits in the comparator publication. All (compound, strain, S9) triplets with conflicting Ames endpoints were removed from the dataset. Duplicate triplets sharing identical endpoints were collapsed to a single instance. Data leakage across dataset splits was identified and eliminated, including 22 SMILES that appeared in more than one split. These filtering steps reduced the training set from 43,897 to 43,257 data points and the test set from 4,838 to 4,644 data points.

[0169] Dataset Partition: The 10% test set from the comparator publication was heldconstant. The training and validation portions were reconstituted then re-split using our t-SNE chemical diversity approach.

[0170] Dataset Featurization: Each compound was represented as a geometric molecular graph. Nodes corresponded to atoms, and edges corresponded to chemical bonds. This representation captured the compound’s structural information. Atomic features were encoded as one-hot vectors and provided as node attributes. Nine one-hot encoded context features represented the Ames assay conditions and were provided through the state conditioning channel. The experimental conditions consisted of the Salmonella strain and the presence or absence of S9 metabolic activation. Each data point corresponded to a (compound, strain, S9) triplet with a binary ground-truth Ames label (+ / -).AmesNet Architecture

[0171] Task-Conditioned Dual Branch: AmesNet consisted of two input branches. The first branch encoded molecular structure using ChemPrint, an atomic-level graph encoder that produced a 512-dimensional molecular embedding. The second branch encoded nine one-hot Ames assay context features specifying the Salmonella strain and the presence or absence of S9 metabolic activation. The molecular embedding and context feature vectors were concatenated and passed to a multilayer perceptron for prediction. Both branches were required for Task-Conditioned Modeling. The molecular branch captured structural information, while the context branch enabled conditional mutagenicity prediction. All AmesNet parameters were learned end-to-end from random initialization, and no ChemPrint weights were imported or frozen.Model Development Operations

[0172] Hyperparameter Optimization: Grid search was used to optimize hyperparameters on the training and validation datasets.

[0173] Model Final Training: AmesNet was retrained on the reconstituted training and validation datasets using hyperparameters selected by grid search.

[0174] Compute Resources: All model training and inference were performed on an NVIDIA Al 0040GB GPU.Reference Models

[0175] Unconditioned Models: AmesNet was benchmarked against three reference unconditioned Ames QSAR models: single-task learning (STL), ungrouped multitask learning (uMTL), and grouped multitask learning (gMTL).

[0176] Task-Conditioned Encoder-Swap Models: Task-conditioned encoder- swap models were trained and evaluated on the OOD test set. AmesNet’ s encoder was replaced with two alternatives: ChemProp and GROVERiarge. ChemProp was trained without frozen or imported weights. GROVERiarge was instantiated with frozen 5000-bit graph embeddings. The Ames-specific feature block and multilayer perceptron classifier were held constant. These replacements isolated the molecular representation and attributed performance effects to the encoders.OOD Benchmark Design

[0177] All models were evaluated using an OOD benchmark to simulate chemical space unseen during training. Test data contained Bemis-Murcko scaffolds was absent from the training and validation sets. A two-tiered benchmarking strategy was used to assess AmesNet. Benchmark I compared task-conditioned and unconditioned paradigms to determine the impact of task-conditioning via three unconditioned Ames QSAR models: STL, uMTL and gMTL. Benchmark II evaluated molecular encoders within the task-conditioned framework to decouple the impact of ChemPrint. Encoder swaps using ChemProp and GROVER represented the two leading molecular representation paradigms of message-passing networks and self-supervised graph transformers. Simultaneous assessment of sensitivity and generalization across both benchmarks utilized three complementary approaches: stratified bootstrap resampling with macro-averaged confidence intervals, confusion matrix analysis, and structural enrichment analysis. Stratified bootstrap resampling enabled compare-to-best evaluation of model sensitivity under OOD generalization using 90% confidence intervals. This bootstrapping established the statistically meaningful sensitivity differences between AmesNet and comparator models. Confusion matrix analysis summarized predictions relative to ground truth labels and contextualized sensitivity by assessing classification trade-offs. Structural enrichment analysis identified chemical classes and functional groups within AmesNet’ s true-positive predictions. These three methods provided the evidence for comparative analysis of model performance across both benchmarking tiers.

[0178] In this Example, the experimental materials and methods described above were used in the studies and analyses described below.

[0179] Regulatory agencies expect comprehensive genotoxicity assessments for all novel small-molecule therapeutics prior to human trials. The Ames test is one such assay used to detect a molecule’s mutagenic potential. Developers often delay these studies until a candidate is nearing regulatory submission because the Ames test is expensive and secondary to bioactivity. This timing creates a bottleneck where later-stage failures can jeopardize >$10 million in capital and multiple years of developmental progress per candidate. While computational methods offer a potential alternative, current models suffer from a dramatic sensitivity drop-off when they evaluate molecules outside their training domain. Sensitivity is the most important metric in Ames prediction because false negatives allow mutagenic compounds to advance undetected and trigger the most costly late-stage failures. AmesNet, a novel Task-Conditioned modeling paradigm that achieved materially improved sensitivity and robust generalization in novel chemical spaces, isprovided in this study. AmesNet utilized ChemPrint for its molecular encoder branch and contains a dedicated channel to condition Ames assay context such as metabolic activation and strain type. In comparative benchmarks, AmesNet reached a macro-averaged out-of-domain sensitivity of 0.71 (90% confidence interval: 0.66-0.76), which exceeded the sensitivity of standard single-task and multitask learning approaches by 51% and 20%, respectively. Structural analysis demonstrated that AmesNet accurately identified difficult-to-detect mutagenic compounds containing substructures. This framework provided a high-confidence filtering mechanism that enabled drug developers to turn a costly late-stage safety bottleneck into a proactive decisionmaking edge.Introduction

[0180] The pervasive sensitivity gap across current Ames QSAR models necessitates a new modeling framework for maintaining robust performance in novel chemical space. All existing Al models for Ames prediction fall under the archetype of unconditioned modeling. In this archetype, predictions are generated without the use of an input-based conditioning channel for explicit assay states. These states include bacterial strain identity or metabolic activation. AmesNet, as a novel task-conditioned paradigm, was driven by explicit input-based conditioning that diverged from established methods in Ames QSAR modeling. AmesNet predicted strainspecific Ames mutagenicity from the concatenation of an assay state conditioning channel and a molecular encoder branch. AmesNet utilized an adaptation of ChemPrint for the molecular encoder (FIG. 2). ChemPrint is an experimentally validated geometric convolutional neural network within the Model Medicines’ GALILEO ecosystem. The model has documented success across oncology and antiviral programs. These successes included on-target hit rates up to 100%, selectivity improvements exceeding 15,000-fold, and the discovery of MDL-001. AmesNet was designed to leverage ChemPrint’ s OOD generalization through its molecular encoder branch. This framework closed the existing performance gap of Ames QSAR models and enabled high-confidence mutagenicity screening in novel chemical space.

[0181] A two-tiered benchmarking strategy was executed to assess the efficacy of AmesNet’ s task-tonditioned modeling for Ames mutagenicity prediction. First, the architectural impact of task-conditioning was isolated by comparing AmesNet against three recently reported and widely utilized unconditioned modeling standards: Single-Task Learning (STL), Ungrouped Multitask Learning (uMTL), and Grouped Multitask Learning (gMTL) (FIG. 6B- FIG.6D). This comparison served to quantify the specific performance gains attributed to task-conditioning. Second, the representational impact of ChemPrint was decoupled by performing molecular encoder-swaps within the AmesNet framework against state-of-the-art (SOTA) alternatives.These included ChemProp and GROVER. These models spanned the two leading paradigms: message-passing networks and self-supervised graph transformers. By holding the task-conditioned design constant while alternating encoders, the representational advantage of ChemPrint was isolated. Performance was assessed for each benchmark strategy through the concurrent evaluation of sensitivity and generalization on a withheld OOD test set. Confusion matrices were additionally utilized to verify the model’s overall predictive robustness. Across all validation benchmarks, AmesNet demonstrated substantial improvements in sensitivity and OOD generalization, outperforming both the unconditioned models in the first benchmark and the encoder-swaps in the second benchmark. These results suggested that AmesNet’ s task-conditioned model framework offered enhanced predictive performance in novel chemical space over alternatives engineered up to this point.

[0182] The performance of AmesNet substantially addressed the OOD sensitivity gap that has limited the utility of Ames QSAR models. AmesNet provided a more reliable pathway for higher-confidence safety screening at the earliest stages of discovery. This framework offered a notably improved alternative to current in silico approaches and can serve to mitigate the $10 million three-year risk of late-stage failure due to undetected genotoxicity. This task-conditioned paradigm supported higher-confidence screening of massive chemical libraries and can transform a late-stage safety liability into a primary strategic advantage.Benchmark I: Task-Conditioned vs. Unconditioned Modelins Paradigms

[0183] Benchmark I: Sensitivity on OOD Test Set: Benchmark I evaluated the impact of Task-Conditioning by comparing AmesNet against established unconditioned Ames QSAR paradigms (STL, uMTL, and gMTL) on sensitivity and generalization in an OOD setting. Stratified bootstrap resampling with macro-averaged 90% confidence intervals estimated sensitivity on a fixed OOD test set (FIG. 7). AmesNet achieved a macro-averaged sensitivity of 0.71 with a 90% CI of 0.66-0.76. Unconditioned models showed consistently lower sensitivity. STL achieved 0.47 (0.41-0.52), uMTL achieved 0.56 (0.51-0.61), and gMTL achieved 0.59 (0.54-0.64). AmesNet reported sensitivity improvements of +0.24 (51%) over STL, +0.15 (27%) over uMTL, and +0.12 (20%) over gMTL. These results demonstrated that AmesNet’s task-conditioned framework materially improved mutagenic compound identification in novel chemical space compared to unconditioned odeling paradigms.

[0184] Benchmark L Confusion Matrix on OOD Test Set: Confusion matrix analysis in FIG.8 indicated that AmesNet achieved substantial sensitivity gains without sacrificing overall predictive robustness. AmesNet reported respective TP / FN deltas of ±131 over STL, ±94 over uMTL, and ±78 over gMTL. The model maintained a robust true negative (TN) count of 3,571 throughout these tests. These results demonstrated that the task-conditioned paradigmsignificantly increased mutagen identification in novel chemical space without a detrimental trade-off in classification performance.

[0185] Benchmark I: Structural Enrichment of True-Positive Predictions : Structural enrichment analysis demonstrated that AmesNet’s task-conditioned framework resolved established mutagenic substructures that escaped detection under the traditional unconditioned modeling paradigm. AmesNet’s true-positive predictions were enriched for the four dominant mutagenic alerts reported in FIG.9: aromatic amines, planar aromatic intercalators, nitro-aromatic groups, and epoxides. These four substructures are well-established sources of false negatives in Ames QSAR modeling. All benchmarked unconditioned models exhibited systematic difficulty in identifying compounds containing these alerts under OOD evaluation. However, the relative burden of false negatives differed across modeling paradigms. Aromatic amines and epoxides accounted for the largest proportion of missed mutagenic compounds in gMTL and uMTL. These were followed by planar aromatic intercalators and nitro-aromatic groups. STL exhibited its most pronounced failure mode among planar aromatic intercalators. Aromatic amines and epoxides comprised the next most frequently missed substructures in the STL model. These patterns indicated that AmesNet’s task-conditioned representation enabled consistent identification of these systematically underrepresented mutagenic chemical classes in unconditioned models. This result established the chemical basis for the sensitivity gains observed under OOD generalization.Benchmark II: Molecular Encoders within Task-Conditioning

[0186] Benchmark IL Sensitivity on OOD Test Set: Benchmark II isolated the molecular encoder’s contribution by fixing the Ames-specific feature block and MLP while substituting ChemProp and GROVER for ChemPrint. AmesNet achieved the highest sensitivity on the OOD test set in FIG. 10. The model reached a macro-averaged sensitivity of 0.71 with a 90% CI of 0.66-0.76. The ChemProp and GROVER encoder-swap models both yielded a lower sensitivity of 0.67. The 90% CI for ChemProp was 0.62-0.71 and the 90% CI for GROVER was 0.62-0.72. These results demonstrated that the ChemPrint encoder provided a performance advantage over leading message-passing and graph transformer architectures within the Task-Conditioned paradigm.

[0187] Benchmark II: Confusion Matrix on OOD Test Set: Encoder swap confusion matrices in FIG. 11 show that AmesNet improved every performance category by increasing true predictions and decreasing errors. AmesNet yielded respective TP / FN deltas of ±17 over ChemProp and ±10 over GROVER. AmesNet achieved FP / TN deltas of ±44 over ChemProp and ±133 over GROVER. These results demonstrated that the ChemPrint molecular encoder within the AmesNet task-conditioned framework enhanced predictive accuracy over state-of-the-art message-passing and graph transformer architectures.

[0188] Benchmark II: Structural Enrichment of True-Positive Predictions : Structural enrichment analysis in FIG. 12 demonstrated that AmesNet’s task-conditioned framework identified mutagenic substructures that remained undetected by alternative molecular encoding architectures. AmesNet’s true-positive predictions were enriched for five established mutagenic chemical classes that passed undetected as false negatives for ChemProp and GROVER: aromatic amines, nitro-aromatic groups, aromatic ring N-oxides, planar aromatic intercalators, and epoxides. These chemical groups are historically difficult to correctly classify as mutagenic. Both benchmarked encoder-swap models exhibited conserved identification failures for these five toxic substructures. Aromatic amines and nitro-aromatic groups accounted for the highest frequency of missed mutagenic compounds in both ChemProp and GROVER. ChemProp exhibited additional failure modes specifically among aromatic ring N-oxides and planar aromatic intercalators. GROVER showed a similar struggle with aromatic ring N-oxides and epoxides. These patterns indicated that the ChemPrint molecular encoder within the AmesNet framework enabled superior recognition of these mutagenic chemical classes compared to state-of-the-art message-passing and graph transformer architectures. This result established the chemical basis for the sensitivity gains achieved through AmesNet and the Task-Conditioned paradigm.Discussion

[0189] Addressing the Sensitivity Gap: The First and Second Ames / QSAR International Challenge Projects established a consistent failure mode for Ames QSAR models on OOD test data. Sensitivity fails to hold under chemical novelty assessment as demonstrated by the participant average sensitivity of 0.46. ChemProp and DeepAmes yielded sensitivities of 0.32 and 0.47 in that competition setting. Those competition results show that existing models do not deliver the sensitivity required for reliable safety screening in chemically novel space. This failure mode can be overcome directly using an explicit OOD benchmark. AmesNet’s performance showed that sensitivity drop-off was not inevitable under chemical novelty. Instead, this pattern reflected the limitations of the unconditioned modeling paradigm that defined all prior approaches. AmesNet achieved a macro-averaged OOD sensitivity of 0.71 (90% CI: 0.66-0.76) in Benchmark I (FIG. 7). This exceeded all unconditioned model comparators (STL: 0.47, uMTL: 0.56, gMTL: 0.59). Confusion-matrix shifts confirmed robust classification performance (FIG. 8). Structural enrichment analysis highlighted classes of mutagenic chemicals that AmesNet uniquely flagged, including aromatic amines, planar aromatic intercalators, nitro-aromatics, and epoxides. These specific groups resulted in false negatives in unconditioned models under OOD evaluation. This results indicated that sensitivity gains resulted from improved recognition of historically challenging toxic substructures. Benchmark II decoupled the impact of task-conditioning from the molecular encoder. Swapping encoders while holding the task-conditioned framework fixedshowed that ChemPrint provided an advantage over leading message-passing and graphtransformer paradigms such as ChemProp and GROVER. This configuration improved OOD sensitivity and error profiles (FIG. 10-FIG. 11). Encoder-swap enrichment patterns remained dominated by the same difficult chemical classes and ChemPrint reduces that false-negative burden (FIG. 12). Task-Conditioning was required to preserve sensitivity under novelty. ChemPrint contributed to an additional performance edge within that paradigm.

[0190] Operational Utility: AmesNet provided the operational utility required for high-confidence mutagenicity screening in early drug discovery. Sensitivity is the most critical metric in Ames prediction and false negatives pose the greatest risk to drug development. A model with low sensitivity fails to identify hazardous compounds and allows them to advance through the pipeline undetected. This failure could waste millions in resources and years in development time. AmesNet enabled developers to identify these genotoxic liabilities before significant investment occurs. The task-conditioned framework can transform a potential late-stage safety liability into a primary strategic advantage. It provided a high-confidence pathway for screening compound libraries at massive throughput.

[0191] Beyond Ames: The success of AmesNet established task-conditioning as a superior design paradigm for Ames modeling. AmesNet leveraged the OOD generalization capabilities of ChemPrint. This encoder achieved on-target hit rates up to 100% and selectivity improvements exceeding 15,000-fold in oncology and antiviral programs. These results demonstrated that ChemPrint performed diverse functions within a single pipeline.Conclusions

[0192] Toxicity assessment remains a prerequisite for advancing novel smallmolecule therapeutics to human trials. The Ames test is a core genotoxicity assay for identifying mutagenic risk. Developers often wait to complete Ames studies until a candidate approaches regulatory submission because these experiments are costly and frequently deprioritized relative to bioactivity optimization. This timing concentrates risk late in development, where an Ames failure can jeopardize >$10 million in capital and multiple years of progress per candidate. In response, the FDA and international regulatory agencies have issued a clear call to action to operationalize credible in silico toxicology approaches that can reduce these late-stage bottlenecks. In silico Ames prediction has not reliably mitigated this bottleneck because existing Ames QSAR models do not sustain high sensitivity under chemical novelty. Low sensitivity is a critical failure because false negatives allow mutagenic compounds to advance undetected and recreate the same late-stage bottleneck these in silico approaches are intended to prevent. AmesNet addressed this sensitivity failure by introducing a task-conditioned modeling paradigm that conditioned predictions on assay state variables. These states included strain identity andmetabolic activation (±S9). All prior Ames Al models follow the unconditioned modeling paradigm and generate predictions without explicit conditioning on assay state. The AmesNet design used an assay state conditioning channel and ChemPrint as its molecular encoder. ChemPrint has demonstrated experimental generalization to chemically novel space across various antiviral and oncology programs. This study evaluated an OOD benchmark to test sensitivity preservation under chemical novelty. AmesNet achieved a macro-averaged sensitivity of 0.71 (90% CI: 0.66-0.76) and improved sensitivity by up to 51% relative to all evaluated unconditioned baselines. Bootstrap and confusion-matrix analyses were used to demonstrate sensitivity gains driven by reduced false negatives without a significant tradeoff in overall classification performance. Structural enrichment analysis further linked these improvements to enhanced recognition of historically challenging mutagenic chemotypes, including aromatic amines, planar aromatic intercalators, nitro-aromatics, and epoxides. A complementary benchmark was used to separate the contributions of molecular encoders within the taskconditioning framework. Results showed that ChemPrint provided an additional sensitivity contribution over state-of-the-art message-passing (ChemProp) and graph-transformer (GROVER) alternatives. These results support task-conditioned modeling as a practical advance for reliable in silico mutagenicity screening and positioned AmesNet as a high-confidence early screening tool to reduce late-stage safety attrition risk in modern drug discovery pipelines.

[0193] Example 3: AmesNet: A Task-Conditioned Deep Learning Model with Enhanced Sensitivity and Generalization in Ames Mutagenicity PredictionMethodsData

[0194] Dataset Source: The best available publicly accessible dataset was used for this study. The dataset was compiled from four individual data sources: ISSSTY, OASIS, EFSAP and MHLW. The dataset included Ames test annotations from eight Salmonella strains, each corresponding to a specific mutation type. The base-pair substitution strains were TAI 00, TAI 02, TA104 and TA1535. The frameshift mutation strains were TA1537, TA1538, TA97 and TA98. Assays were conducted in the presence or absence of S9 fraction.

[0195] Dataset Cleaning: The molecules in the dataset were converted from SDF to SMILES format and standardized with respect to charge, stereochemistry, and tautomerism. The dataset was prepared for model training, validation, and testing according to the predefined OOD data splits in the comparator publication. All (compound, strain, S9) triplets with conflicting Ames endpoints were removed from the dataset. Duplicate triplets sharing identical endpoints were collapsed to a single instance. Data leakage across dataset splits was identified and eliminated. Compounds were filtered by molecular weight 100-1000 Da to retain small molecules only. Thesefiltering steps reduced the training / validation set from 43,897 to 40,129 data points and the test set from 4,838 to 4,528 data points.

[0196] Dataset Featurization: Each compound was represented as a geometric molecular graph. Nodes corresponded to atoms, and edges corresponded to chemical bonds. This representation captured the compound’s structural information. Atomic features were encoded as one-hot vectors and provided as node attributes. Nine one-hot encoded context features represented the Ames assay conditions and were provided through the state conditioning channel. The experimental conditions consisted of the Salmonella strain and the presence or absence of S9 metabolic activation. Each data point corresponded to a (compound, strain, S9) triplet with a binary ground-truth Ames label (+ / -).AmesNet Architecture

[0197] Task-Conditioned Dual Branch: AmesNet consisted of two input branches. The first branch encodes molecular structure using an atomic-level graph encoder that produced a 512-dimensional molecular embedding. The second branch encoded nine one-hot Ames assay context features specifying the Salmonella strain and the presence or absence of S9 metabolic activation. The molecular embedding and context feature vectors were concatenated and passed to a multilayer perceptron for prediction. The molecular branch captured structural information, while the context branch enabled conditional mutagenicity prediction. All AmesNet parameters were learned end-to-end from random initialization, and no weights were imported or frozen.Model Development Operations

[0198] Hyperparameter Optimization: Grid search was used to optimize hyperparameters during model training and validation.Reference Models

[0199] Unconditioned Models: AmesNet was benchmarked against three unconditioned Ames QSAR frameworks: single-task learning (STL), ungrouped multitask learning (uMTL), and grouped multitask learning (gMTL). The multilayer perceptron (MLP) architectures for STL, uMTL, and gMTL followed the reference implementation. In addition, three architectures (DeepAmes, ChemProp and GROVER) were evaluated as STL baselines by training 16 task-specific models per architecture (bacterial strain x ±S9). Each model was trained end-to-end.

[0200] Task-Conditioned Encoder-Swap Models: AmesNet was also benchmarked by injecting the task conditioning channel into ChemProp and GROVERlarge. Each encoder-swap model was trained end-to-end.Metrics and Bootstrapping

[0201] Multi-task binary classification performance was evaluated across 16 Amestasks. Each task was defined as a unique combination of bacterial strain and ±S9. Sensitivity and balanced accuracy were computed per task and were the principal metrics reported. Specificity and Matthews correlation coefficient (MCC) were also computed per task and were analyzed. Each metric was aggregated using a sample-size weighted average, with weights proportional to the number of observations per task. Uncertainty was estimated with a within-task stratified bootstrap (n=l,000). Positives and negatives were resampled with replacement while also retaining the same task size and class counts. Metrics were recomputed per task and aggregated using the same weights to yield one bootstrap replicate of the estimator. The 95% confidence interval was defined by the 2.5th and 97.5th percentiles of the bootstrap distribution.

[0202] In this Example, the experimental materials and methods described above were used in the studies and analyses described below.

[0203] Regulatory agencies require comprehensive genotoxicity assessments for all novel small-molecule therapeutics prior to human trials. Developers often delay these studies until a candidate is nearing regulatory submission because they are expensive and secondary to bioactivity. This timing creates a bottleneck where late-stage failures can jeopardize >$10 million in capital and multiple years of developmental progress per candidate. The Ames assay is used to detect a molecule’s mutagenic potential. Regulators now explicitly support the use of in silico Ames mutagenicity models through enabling legislation, dedicated FDA Al toxicology programs, internationally harmonized guidelines, and benchmark challenges. However, current Ames models suffer from a dramatic sensitivity drop-off when they evaluate molecules outside their training domain. Sensitivity is the most important metric in Ames prediction because false negatives allow mutagenic compounds to advance undetected and trigger the most costly latestage failures. Attempts to fix this sensitivity drop-off often reduce overall model performance, which can be represented by balanced accuracy. For example, DeepAmes reports high levels of sensitivity only by sacrificing its balanced accuracy. AmesNet, a novel Task-Conditioned modeling paradigm that achieved both class-leading sensitivity and balanced accuracy in novel chemical spaces, is provided in this Example. AmesNet utilized a dual branch architecture containing a molecular encoder and a dedicated channel to condition Ames assay context such as metabolic activation and bacterial strain type. In comparative benchmarks, AmesNet reached a sensitivity of 0.73 (95% confidence interval: 0.68-0.77) and a simultaneous balanced accuracy of 0.81 (95% confidence interval: 0.79-0.83) on the out-of-domain test data. This represented an improvement in sensitivity of up to 46% over existing approaches without a trade-off in balanced accuracy. Structural analysis demonstrated that AmesNet recovered difficult-to-detect mutagenic compounds missed by existing models. This framework provided a high-confidence filtering mechanism that enabled drug developers to turn a costly late-stage safety bottleneck into aproactive decision-making edge.Introduction

[0204] The Ames test uses specialized bacterial strains in growth-restrictive conditions to detect if a drug candidate might cause genetic mutations. Mutagenic potential is evidenced by bacterial growth after exposure to either the parent compound or its metabolic byproducts (FIG.6A). Effective screening therefore depends on high sensitivity to avoid misclassifying a dangerous molecule as safe. Meeting this requirement remains a challenge for Ames QSAR models. Low sensitivity is a critical failure because it leads to more genotoxic compounds advancing undetected. Such failures are amplified when Al models encounter novel chemical structures distinct from their training data. This challenge in novel chemical space is known as out-of-domain (OOD) generalization. Achieving operational Ames model utility simultaneously requires high sensitivity and OOD generalization so that no genotoxic compound escapes detection when screening novel chemical libraries. Benchmark data from the Second Ames / QSAR International Challenge Projects revealed a significant collapse in sensitivity for current Ames models when encountering novel chemotypes. The project yielded a participant average sensitivity of only 0.46 on the OOD test set. High-profile models like the FDA’s DeepAmes achieved a sensitivity of 0.47 and MIT’s ChemProp yielded 0.32. Attempts to increase sensitivity have forced trade-offs in overall model performance, which can be measured by balanced accuracy. DeepAmes reported a sensitivity of 0.87 with a corresponding balanced accuracy of 0.52, providing little predictive signal. These results demonstrate that current methods cannot maintain the sensitivity required for the reliable safety screening of novel drug candidates. Such unreliability creates a pathway for unidentified genotoxic molecules to advance into later stage drug development and potentially waste years and millions of dollars in resources.

[0205] The pervasive sensitivity gap across current Ames QSAR models necessitates a new modeling framework for maintaining robust performance in novel chemical space. All existing Al models for Ames prediction fall under the archetype of unconditioned modeling. In this archetype, predictions are generated without the use of an input conditioning channel for explicit assay states like bacterial strain or metabolic activation.

[0206] AmesNet is provided herein as a novel Task-Conditioned Learning (TCL) paradigm driven by explicit input-based conditioning that diverged from established methods in Ames QSAR modeling. AmesNet predicted strain-specific Ames mutagenicity from the concatenation of an assay state conditioning channel and a molecular encoder branch. AmesNet utilized an adaptation of ChemPrint for the molecular encoder (FIG. 13). ChemPrint is an experimentally validated geometric convolutional neural network within the Model Medicines’ GALILEO platform. The model has documented success across oncology and antiviral programs.These successes included on-target hit rates up to 100%, selectivity improvements exceeding 15,000-fold, and the discovery of a proprietary therapeutic candidate. AmesNet was designed to leverage ChemPrint’s OOD generalization and power this first instantiation of the TCL paradigm for Ames prediction. This framework can close the existing performance gap of Ames QSAR models and enable high-confidence mutagenicity screening in novel chemical space.

[0207] A two-tiered benchmarking strategy was executed to assess AmesNet’ s overall performance and to isolate the contribution of TCL for Ames prediction. First, we AmesNet’ s end-to-end performance was measured relative to established learning frameworks: Single-Task Learning (STL), Ungrouped Multitask Learning (uMTL), and Grouped Multitask Learning (gMTL) (FIG. 6B-FIG.6D). In this benchmark, STL instantiations of ChemProp, GROVER, and DeepAmes were included. This comparison quantified AmesNet’ s overall performance advantage relative to these comparators. Second, the contribution of the TCL framework was isolated by augmenting established encoders with the conditioning channel: message-passing networks and graph transformers. In this benchmark, TCL instantiations of ChemProp and GROVER were included. This comparison quantified the performance gains attributable to TCL. Performance was assessed for each benchmark through the concurrent evaluation of generalization and classification metrics on a withheld OOD test set. Sensitivity is reported as the primary metric. Additionally, balanced accuracy (BA) and confusion matrices were utilized to verify the model’s overall predictive robustness. Across all benchmarks, AmesNet demonstrated substantial improvements in sensitivity, BA, and OOD generalization, outperforming both the unconditioned models in the first benchmark and the encoder-swaps in the second benchmark. These results demonstrated that AmesNet and its TCL framework offer enhanced predictive performance in novel chemical space over alternatives.

[0208] The performance of AmesNet substantially addressed the OOD sensitivity gap that has limited the utility of Ames QSAR models. AmesNet provided a more reliable pathway for higher-confidence safety screening at the earliest stages of discovery. This TCL paradigm supported higher-confidence screening of massive chemical libraries and can transform a latestage safety liability into a primary strategic advantage.Benchmark I: AmesNet vs. Unconditioned Modeling ParadigmsBenchmark I: Sensitivity on OOD Test Set

[0209] Benchmark I evaluated the performance of AmesNet against established unconditioned Ames QSAR frameworks (STL, uMTL and gMTL) on sensitivity in an OOD setting. OOD sensitivity and 95% confidence intervals were estimated (FIG. 14). AmesNet achieved a sensitivity of 0.73 (95% CL 0.68-0.77), the highest reported among all models evaluated in Benchmark I. Unconditioned models showed consistently lower sensitivity. Withinthe STL framework, the STL-MLP achieved 0.50 (95% CI: 0.45-0.55), STL-ChemProp achieved 0.54 (95% CI: 0.49-0.59), STL-GROVER achieved 0.55 (95% CI: 0.50-0.60), and STL-DeepAmes achieved 0.67 (95% CI: 0.62-0.71). Within the uMTL framework, the uMTL-MLP achieved 0.57 (95% CI: 0.52-0.62). Within the gMTL framework, the gMTL-MLP achieved 0.59 (95% CI: 0.55-0.63). AmesNet achieved sensitivity improvements of +0.23 (46%) over STL-MLP, +0.19 (35%) over STL-ChemProp, +0.18 (33%) over STL-GROVER, +0.06 (9%) over STL-DeepAmes, +0.16 (28%) over uMTL-MLP, and +0.14 (24%) over gMTL-MLP. AmesNet exhibited non-overlapping 95% confidence intervals across sensitivity measurements relative to STL-GROVER, STL-ChemProp, STL-MLP, uMTL-MLP, and gMTL-MLP baselines.Benchmark I: Balanced Accuracy on OOP Test Set

[0210] OOD balanced accuracy was calculated with 95% confidence intervals (FIG.15). AmesNet achieved a balanced accuracy of 0.81 (95% CI: 0.79-0.83). AmesNet demonstrated a substantial improvement in balanced accuracy over all models, as evidenced by non-overlapping 95% bootstrap confidence intervals. Within the STL framework, the STL-MLP achieved 0.72 (95% CL 0.70-0.75), STL-ChemProp achieved 0.74 (95% CI: 0.72-0.77), STL-GROVER achieved 0.75 (95% CL 0.73-0.78), and STL-DeepAmes achieved 0.75 (95% CI: 0.72-0.77). Within the uMTL framework, the uMTL-MLP achieved 0.73 (95% CI: 0.70-0.75). Within the gMTL framework, the gMTL-MLP achieved 0.75 (95% CI: 0.73-0.77). AmesNet achieved balanced accuracy improvements of +0.09 (13%) over STL-MLP, +0.07 (9%) over STL-ChemProp, +0.06 (8%) over STL-GROVER, +0.06 (8%) over STL-DeepAmes, +0.08 (11%) over uMTL-MLP, and +0.06 (8%) over gMTL-MLP.Benchmark I: Confusion Matrix on OOD Test Set

[0211] FIG. 16 shows OOD test set confusion matrices. Compared with each baseline, AmesNet increased true positives and decreased false negatives by 123 compared to STL-MLP, 96 compared to STL-ChemProp, 91 compared to STL-GROVER, 27 compared to STL-DeepAmes, 85 compared to uMTL-MLP, and 76 compared to gMTL-MLP.Benchmark I: Structural Enrichment of True-Positive Predictions

[0212] FIG. 17 summarizes mutagenic substructures present in compounds that AmesNet classified as true positives but that unconditioned models misclassified as false negatives. The largest counts were observed for planar aromatic intercalators (n = 27-46), aromatic amines (n = 20-33), and aromatic N-heterocycles (n = 17-25). Smaller but recurrent categories included a,P-unsaturated carbonyls (n = 6-15), epoxides (n = 5-14), nitro-aromatics (n = 5-11), aromatic ring N-oxides (n = 5-11), and azo groups (n = 4-6). Several alerts were rare across models such as nitroso groups (n = 1-2). These patterns showed that a substantial fraction of OOD false negatives in unconditioned models are concentrated in a small set of well-known mutagenic alertclasses, and that AmesNet correctly identified many of these cases.Benchmark II: Encoder-Swaps within Task-ConditioningBenchmark II: Sensitivity on OOP Test Set

[0213] Benchmark II isolated the contribution of the TCL framework by augmenting two established encoders with the conditioning channel to create TCL-ChemProp and TCL-GROVER. AmesNet achieved the highest sensitivity on the OOD test set (FIG. 18), reaching a sensitivity of 0.73 (95% CI: 0.68-0.77). TCL-ChemProp improved sensitivity relative to STL-ChemProp, increasing from 0.54 (95% CI: 0.49-0.59) to 0.65 (95% CI: 0.60-0.70), a +0.11 absolute gain (20% relative). By contrast, TCL-GROVER showed no material sensitivity change relative to STL-GROVER (0.54, 95% CI: 0.50-0.58 vs. 0.55, 95% CI: 0.50-0.60). Together, these results indicated that within the TCL paradigm, AmesNet’ s molecular representation provided a sensitivity advantage over leading message-passing and graph transformer alternatives. The TCL framework also confered a material sensitivity lift for TCL-ChemProp relative to STL-ChemProp.Benchmark II: Balanced Accuracy on OOD Test Set

[0214] Balanced accuracy on the OOD test set was calculated and 95% confidence intervals were reported (FIG. 19). AmesNet achieved a balanced accuracy of 0.81 (95% CI: 0.79-0.83). STL-ChemProp achieved 0.74 (95% CL 0.72-0.77) and TCL-ChemProp achieved 0.79 (95% CI: 0.76-0.81). TCL-GROVER showed no material change in balanced accuracy compared to STL-GROVER with 0.75 (95% CI: 0.73-0.77) and 0.75 (95% CI: 0.73-0.78), respectively. Together, these results showed that AmesNet achieved the strongest balanced accuracy and that the TCL framework confered a material balanced accuracy lift for TCL-ChemProp relative to STL-ChemProp.Benchmark II: Confusion Matrix on OOD Test Set

[0215] TCL model confusion matrices on the OOD test set are reported in FIG. 20.AmesNet increased true positives and reduced false negatives relative to TCL-ChemProp and TCL-GROVER, demonstrating improved positive-class identification in an OOD setting. AmesNet increased true positives and decreased false negatives by 44, relative to TCL-ChemProp and by 102, relative to TCL-GROVER. TCL-ChemProp increased true positives and decreased false negatives by 52, relative to STL-ChemProp, demonstrating improvements associated with the TCL framework. Together, these results indicated that within the TCL paradigm, AmesNet achieved stronger recovery of mutagenic compounds than state-of-the-art message-passing and graph transformer alternatives, and that moving from STL to TCL provided a measurable improvement.Benchmark II: Structural Enrichment of True-Positive Predictions

[0216] Structural enrichment analysis (FIG. 21) highlighted mutagenic substructuresthat AmesNet recovered as true positives but that TCL-ChemProp and TCL-GROVER miss as false negatives. The most frequent missed classes were planar aromatic intercalators (n=26 for both models) and aromatic amines (n=24 for TCL-ChemProp; n=26 for TCL-GROVER), followed by aromatic N-heterocycles (n=16-17). Additional recurrent alerts included a,P-unsaturated carbonyls (n=6-9), epoxides (n=l-5), nitro-aromatics and aromatic ring N-oxides (n=3-4 each), azo groups (n=5-8), and 1,2-dihydroxy aromatic rings (n=4-7). Relative to STL-ChemProp, TCL-ChemProp showed broad recovery for planar aromatic intercalators (38 to 26), aromatic amines (31 to 24), aromatic N-heterocycles (25 to 16), epoxides (12 to 1), a,P-unsaturated carbonyls (9 to 6), nitro-aromatics (7 to 3), and aromatic ring N-oxides (7 to 3). Recovery from STL-GROVERto TCL-GROVER was mixed: TCL-GROVER reduced misses for planar aromatic intercalators (31 to 26), aromatic amines (30 to 26), aromatic N-heterocycles (24 to 17), a,P-unsaturated carbonyls (15 to 9), nitro-aromatics (9 to 4), and aromatic ring N-oxides (9 to 4). TCL-GROVER showed no change for epoxides (5 to 5) and increased for azo groups (6 to 8), 1,2-dihydroxy aromatic rings (4 to 7), and nitroso groups (1 to 2). Together, these patterns indicated that within the TCL paradigm, AmesNet’ s molecular representation more reliably recognized mutagenic alert classes than TCL-ChemProp and TCL-GROVER. These results also showed that the TCL framework can reduce conserved false-negative failures in established encoders.Addressing the Performance GapOOP Sensitivity is a Failure Mode in Ames OSAR

[0217] The First and Second Ames / QSAR International Challenge Projects established a consistent failure mode for Ames QSAR models under OOD evaluation. Sensitivity fails to hold under chemical novelty assessment as demonstrated by the participant average sensitivity of 0.46. ChemProp and DeepAmes yielded sensitivities of 0.32 and 0.47 in that competition setting. Those competition results show that existing models do not deliver the sensitivity required for reliable safety screening in chemically novel space.Benchmark I: AmesNet Performance vs. Unconditioned Modeling

[0218] This failure mode was targeted directly using an OOD benchmark. AmesNet proved that sensitivity drop-off was not inevitable under chemical novelty. Instead, this pattern reflected the limitations of the unconditioned modeling paradigm that defined all prior approaches. AmesNet achieved a sensitivity of 0.73 (95% CL 0.68-0.77), exceeding all unconditioned model comparators (FIG. 14). AmesNet exhibited non-overlapping 95% confidence intervals against all models except DeepAmes, which still had a lower point-estimate sensitivity. Importantly, the sensitivity improvement of AmesNet was not achieved through a tradeoff in overall discrimination. Balanced accuracy and confusion-matrix analysis were used to assess tradeoffs.AmesNet achieved a balanced accuracy of 0.81 (95% CI: 0.79-0.83) and has non-overlapping 95% confidence intervals against all models (FIG. 15). AmesNet’s performance advantage over DeepAmes became clear with balanced accuracy assessed. AmesNet achieved substantial sensitivity gains without sacrificing overall predictive robustness.Benchmark I: Interpreting the Chemistry of False-Negative Recovery

[0219] Structural enrichment analysis demonstrated that AmesNet recovered established mutagenic substructures that escaped detection under the traditional unconditioned modeling paradigm. AmesNet’s true-positive predictions were enriched for the mutagenic alerts reported in FIG. 17. These substructures are well-established sources of false negatives in Ames QSAR modeling. All benchmarked unconditioned models exhibited systematic difficulty identifying compounds containing these alerts under OOD evaluation. However, the relative burden of false negatives differed across modeling paradigms. Planar aromatic intercalators and many aromatic N-heterocycles often cause mutagenicity through DNA intercalation and are associated with a frameshift mutation signature. These chemotypes are detected in the Ames frameshift-sensitive bacterial strains (e.g., TA1537, TA1538, TA97 and TA98) rather than in base-substitution strains. AmesNet’s recovery of these motifs indicated that it captured straindependent positive signals that unconditioned models missed. Similar context dependence arises for metabolic activation. Aromatic amines require metabolic activation (+S9) to express mutagenicity. AmesNet’s recovery of aromatic amines demonstrated that it captured Ames positive risk that was dependent on the S9 activation assay context. AmesNet learned different decision rules across assay states rather than averaging them into a single unconditional boundary, providing a plausible mechanistic explanation for the OOD sensitivity gains.Benchmark II: Encoder-Swaps within Task-Conditioning

[0220] The contribution of the TCL framework to the overall success of AmesNet was isolated by injecting the same assay-conditioning channel (strain and ±S9) into two well-established molecular encoders to form TCL-ChemProp and TCL-GROVER. Similarly to AmesNet’s molecular encoder, ChemProp can be classified as a message-passing network encoder, whereas GROVER is classified as a graph transformer architecture. Performance of TCL-ChemProp and TCL-GROVER variants was compared to their STL counterparts to evaluate the improvements attributed to the TCL framework (FIG. 18-FIG.20). TCL-ChemProp exhibited a clear sensitivity increase over STL-ChemProp, increasing from 0.54 (95% CL 0.49-0.59) to 0.65 (95% CI: 0.60-0.70) with non-overlapping 95% confidence intervals. This sensitivity increase was accomplished without an apparent trade-off in overall performance as supported by an increase in balanced accuracy from 0.74 (95% CI: 0.72-0.77) to 0.79 (95% CL 0.76-0.81) for STL-ChemProp and TCL-ChemProp, respectively. However, TCL-GROVER showed no material change inperformance relative to STL-GROVER. This suggested that the concatenated conditioning channel may couple more effectively with message-passing network embeddings than to transformer-derived embeddings. Another plausible explanation can be that GROVER’s pretrained features were less responsive to explicit metadata concatenation. Even with the performance lift from the TCL framework, no model performed as well as AmesNet in both sensitivity and balanced accuracy. This indicated that the success of AmesNet came both from its TCL framework and its molecular encoder. Together, these results supported TCL as a practical mechanism for improving OOD mutagen recovery without an apparent trade-off in overall classification robustness.Benchmark II: Structural Enrichment Under Task-Conditioning

[0221] Structural enrichment analysis in FIG. 21 showed that TCL produced a clear recovery for TCL-ChemProp compared to STL-ChemProp. TCL-GROVER showed a weaker mixed effect compared to its STL equivalent. AmesNet remained the best model overall. Relative to STL-ChemProp, TCL-ChemProp reduced conserved false negatives across several major alert classes, including planar aromatic intercalators (38 to 26), aromatic N-heterocycles (25 to 16), and aromatic amines (31 to 24). The shift from STL-GROVER to TCL-GROVER was less consistently favorable, although TCL-GROVER recovered the same planar aromatic intercalators (31 to 26), aromatic amines (30 to 26), aromatic N-heterocycles (24 to 17). This showed that the TCL framework was responsible for the recovery of these mutagenic toxicophores that were dependent on the bacterial strain and S9 activation assay context. Even after these TCL gains, AmesNet still recovered additional mutagenic compounds that both encoder-swap TCL models missed. These false negatives only recovered by AmesNet were concentrated in high-impact chemotypes such as planar aromatic intercalators (n=26 in both TCL baselines), aromatic amines (n=24-26), and aromatic N-heterocycles (n=16-17). This supported the conclusion that AmesNet’ s advantage was attributed to both the TCL framework and a molecular representation that more reliably encoded mutagenicity-relevant structure compared to state-of-the-art message-passing and graph transformer architectures. This result established the chemical basis for the sensitivity gains achieved through AmesNet and the TCL paradigm.Operational Utility

[0222] AmesNet provided the operational utility required for high-confidence mutagenicity screening in early drug discovery. Sensitivity is the most critical metric in Ames prediction and false negatives pose the greatest risk to drug development. A model with low sensitivity fails to identify hazardous compounds and allows them to advance through the pipeline undetected. This failure could waste millions in resources and years in development time. AmesNet enabled developers to identify these genotoxic liabilities before significant investmentoccurs. The TCL framework transformed a potential late-stage safety liability into a primary strategic advantage. It provided a high-confidence pathway for screening compound libraries at high-throughput.

[0223] Reported sensitivity values can be difficult to interpret without corresponding trade-off metrics. For example, DeepAmes reports a headline sensitivity of 0.87, but the corresponding trade-off performance metrics in DeepAmes’ additional materials show markedly reduced discrimination (specificity = 0.18, MCC = 0.04, balanced accuracy = 0.52). A balanced accuracy of 0.52 means the model is effectively guessing and provides no reliable predictive signal. In practice, DeepAmes would flag the majority of compounds as mutagenic, limiting utility for screening triage. Additionally, DeepAmes’ sensitivity of 0.87 was achieved by sweeping classweights that maximizes sensitivity on the test set (weight = 16). This reflects post-hoc operatingpoint selection and is not directly comparable to a standard held-out test evaluation.Conclusions

[0224] Toxicity assessment remains a prerequisite for advancing novel smallmolecule therapeutics to human trials. The Ames test is a core genotoxicity assay for identifying mutagenic risk. Developers often wait to complete Ames studies until a candidate approaches regulatory submission because these experiments are costly and frequently deprioritized relative to bioactivity optimization. This timing concentrates risk late in development, where an Ames failure can jeopardize >$10 million in capital and multiple years of progress per candidate.

[0225] In response, the FDA and international regulatory agencies have issued a clear call to action to operationalize credible in silico toxicology approaches that can reduce these latestage bottlenecks. In silico Ames prediction has not reliably mitigated this bottleneck because existing Ames QSAR models do not sustain high sensitivity under chemical novelty. Optimizing for higher sensitivity has historically forced a trade-off in overall model performance, which can be measured by balanced accuracy. Low sensitivity is a critical failure because false negatives allow mutagenic compounds to advance undetected and recreate the same late-stage bottleneck these in silico approaches are intended to prevent.

[0226] AmesNet addressed this sensitivity failure by introducing a TCL modeling paradigm that conditioned predictions on assay state variables. These states included strain identity and metabolic activation (±S9). All prior Ames Al models follow the unconditioned modeling paradigm and generate predictions without explicit conditioning on assay state. AmesNet had an assay-state conditioning channel and a ChemPrint-derived molecular encoder. ChemPrint showed experimental generalization to chemically novel space in variout antiviral and oncology programs. In this Example, sensitivity and balanced accuracy were evaluated on a held-out OOD benchmark designed to measure model performance under chemical novelty.

[0227] AmesNet achieved a sensitivity of 0.73 (95% CI: 0.68-0.77) with a balanced accuracy of 0.81 (95% CI: 0.79-0.83), improving sensitivity by up to 46% relative to evaluated unconditioned models. AmesNet outperformed the FDA’s DeepAmes in both categories, as the latter achieved a sensitivity of 0.67 (95% CI: 0.62-0.71) and a balanced accuracy of 0.75 (95% CI: 0.72-0.77). By simultaneously improving sensitivity and balanced accuracy, AmesNet successfully overcame the historical trade-off between sensitivity and overall model performance. Structural enrichment analysis further attributed these improvements to enhanced recognition of historically challenging chemotypes whose mutagenicity was dependent on specific assay conditions, such as strain and metabolic activation.

[0228] The performance improvements provided by the TCL framework were evaluated by injecting the assay-conditioning channel into state-of-the-art message-passing (ChemProp) and graph-transformer (GROVER) architectures. TCL simultaneously improved sensitivity and balanced accuracy when injected into message-passing architectures. Results showed that AmesNet still outperformed these TCL adjusted state-of-the-art models, indicating that AmesNet’ s encoder provided an additional contribution. These results supported TCL modeling as a practical advance for reliable in silico mutagenicity screening and positioned AmesNet as a high-confidence early screening tool to reduce late-stage safety attrition risk in modern drug discovery pipelines.

[0229] In at least some of the previously described embodiments, one or more elements used in an embodiment can interchangeably be used in another embodiment unless such a replacement is not technically feasible. It will be appreciated by those skilled in the art that various other omissions, additions and modifications may be made to the methods and structures described above without departing from the scope of the claimed subject matter. All such modifications and changes are intended to fall within the scope of the subject matter, as defined by the appended claims.

[0230] With respect to the use of substantially any plural and / or singular terms herein, those having skill in the art can translate from the plural to the singular and / or from the singular to the plural as is appropriate to the context and / or application. The various singular / plural permutations may be expressly set forth herein for sake of clarity. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” include plural references unless the context clearly dictates otherwise. Any reference to “or” herein is intended to encompass “and / or” unless otherwise stated.

[0231] It will be understood by those within the art that, in general, terms used herein, and especially in the appended claims (e.g., bodies of the appended claims) are generally intended as “open” terms (e.g., the term “including” should be interpreted as “including but not limited to,”the term “having” should be interpreted as “having at least,” the term “includes” should be interpreted as “includes but is not limited to,” etc.). It will be further understood by those within the art that if a specific number of an introduced claim recitation is intended, such an intent will be explicitly recited in the claim, and in the absence of such recitation no such intent is present. For example, as an aid to understanding, the following appended claims may contain usage of the introductory phrases “at least one” and “one or more” to introduce claim recitations. However, the use of such phrases should not be construed to imply that the introduction of a claim recitation by the indefinite articles “a” or “an” limits any particular claim containing such introduced claim recitation to embodiments containing only one such recitation, even when the same claim includes the introductory phrases “one or more” or “at least one” and indefinite articles such as “a” or “an” (e.g., “a” and / or “an” should be interpreted to mean “at least one” or “one or more”); the same holds true for the use of definite articles used to introduce claim recitations. In addition, even if a specific number of an introduced claim recitation is explicitly recited, those skilled in the art will recognize that such recitation should be interpreted to mean at least the recited number (e.g., the bare recitation of “two recitations,” without other modifiers, means at least two recitations, or two or more recitations). Furthermore, in those instances where a convention analogous to “at least one of A, B, and C, etc.” is used, in general such a construction is intended in the sense one having skill in the art would understand the convention (e.g., “ a system having at least one of A, B, and C” would include but not be limited to systems that have A alone, B alone, C alone, A and B together, A and C together, B and C together, and / or A, B, and C together, etc.). In those instances where a convention analogous to “at least one of A, B, or C, etc.” is used, in general such a construction is intended in the sense one having skill in the art would understand the convention (e.g., “ a system having at least one of A, B, or C” would include but not be limited to systems that have A alone, B alone, C alone, A and B together, A and C together, B and C together, and / or A, B, and C together, etc.). It will be further understood by those within the art that virtually any disjunctive word and / or phrase presenting two or more alternative terms, whether in the description, claims, or drawings, should be understood to contemplate the possibilities of including one of the terms, either of the terms, or both terms.

[0232] In addition, where features or aspects of the disclosure are described in terms of Markush groups, those skilled in the art will recognize that the disclosure is also thereby described in terms of any individual member or subgroup of members of the Markush group.

[0233] As will be understood by one skilled in the art, for any and all purposes, such as in terms of providing a written description, all ranges disclosed herein also encompass any and all possible sub-ranges and combinations of sub-ranges thereof. Any listed range can be easily recognized as sufficiently describing and enabling the same range being broken down into at least-n-equal halves, thirds, quarters, fifths, tenths, etc. As a non-limiting example, each range discussed herein can be readily broken down into a lower third, middle third and upper third, etc. As will also be understood by one skilled in the art all language such as “up to,” “at least,” “greater than,” “less than,” and the like include the number recited and refer to ranges which can be subsequently broken down into sub-ranges as discussed above. Finally, as will be understood by one skilled in the art, a range includes each individual member. Thus, for example, a group having 1-3 articles refers to groups having 1, 2, or 3 articles. Similarly, a group having 1-5 articles refers to groups having 1, 2, 3, 4, or 5 articles, and so forth.

[0234] While various aspects and embodiments have been disclosed herein, other aspects and embodiments will be apparent to those skilled in the art. The various aspects and embodiments disclosed herein are for purposes of illustration and are not intended to be limiting, with the true scope and spirit being indicated by the following claims.

Claims

WHAT IS CLAIMED IS:

1. A method of assessing compound mutagenicity, comprising:(a) providing a compound or a molecular structure representation thereof; (b) providing (optionally, selecting) one or more Ames testing conditions; (c) combining (optionally, concatenating) an adaptive molecular embedding vector corresponding to the compound, with one or more Ames condition vectors corresponding to the one or more Ames testing conditions, to generate one or more condition-specific embedding vectors;(d) processing the one or more condition-specific embedding vectors through a task-conditioned machine-learning model (optionally, through a multilayer perceptron (MLP)) to generate a mutagenicity probability score for each of the one or more Ames testing conditions, thereby generating a profile of mutagenicity probability scores for the compound; and(e) determining the compound as mutagenic or non-mutagenic, or stratifying a mutagenicity risk of the compound, based on the profile of mutagenicity probability scores.

2. The method of claim 1 , further comprising generating, using a convolutional neural network (optionally, a graph convolutional neural network), the adaptive molecular embedding vector from a molecular structure representation of the compound.

3. The method of claim 1 or 2, wherein (c) comprises iteratively aggregating structural information (optionally, from neighboring atoms, motifs or fragments) of the compound to generate the adaptive molecular embedding vector.

4. The method of any one of claims 1-3, wherein the convolutional neural network comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

5. The method of any one of claims 1-4, wherein the convolutional neural network comprises a plurality of convolutional layers configured (optionally, trained) to iteratively aggregate structural information (optionally, from neighboring atoms, motifs or fragments) of the compound to generate the adaptive molecular embedding vector.

6. The method of any one of claims 2-5, wherein the structural information comprises one or more mutagenicity-associated molecular features of the compound.

7. The method of any one of claims 1-6, comprising training the convolutional neural network using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions.

8. The method of claim 7, wherein the training dataset is partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

9. The method of claim 7 or 8, wherein one or more reference compounds in the training dataset comprise one or more mutagenicity-associated molecular features.

10. The method of any one of claims 6-9, wherein a mutagenicity-associated molecular feature is a chemical, geometric, topological, or structural feature that is identified, determined, or suspected to be associated (or correlate) with mutagenicity or non-mutagenicity (optionally, as measured by an Ames test under one or more Ames testing conditions).

11. The method of any one of claims 6-10, wherein the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, six, seven, eight, nine or ten features each independently selected from: a polyaromatic moiety (optionally, a planar polyaromatic moiety), an aromatic amine moiety, an aromatic N-heterocycle moiety, an a,P-unsaturated carbonyl moiety, an epoxide moiety, a nitro-aromatic moiety, an aromatic ring N-oxide moiety, an azo moiety, an aliphatic halogen moiety, a 1,2-dihydroxy aromatic ring moiety, a polyfunctional alkyl moiety, an aryl halide moiety, and a nitroso moiety.

12. The method of any one of claims 1-11, comprising identifying one or more mutagenicity-associated molecular features in the compound.

13. The method of claim 12, comprising providing a report that describes the identified one or more mutagenicity-associated molecular features as present or absent in the compound.

14. The method of any one of claims 1-13, comprising, prior to (c), providing the one or more Ames condition vectors.

15. The method of any one of claims 1-14, comprising training the task-conditioned machine-learning model using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions.

16. The method of claim 15, wherein said training the task-conditioned machinelearning model comprises:(i) initializing a plurality of parameters of the task-conditioned machinelearning model;(ii) processing the training dataset through the task-conditioned machinelearning model to generate predicted mutagenicity probability scores;(iii) computing a loss function based on differences between the predicted mutagenicity probability scores and known mutagenicity classifications;(iv) updating the plurality of parameters of the task-conditioned machinelearning model to minimize the loss function; and(v) repeating steps (ii)-(iv).

17. The method of claim 15 or 16, wherein the training dataset is partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

18. The method of any one of claims 1-17, comprising optimizing parameters of the task-conditioned machine-learning model until the task-conditioned machine-learning model achieves an improved or substantially maximal value of sensitivity, specificity, accuracy, balanced accuracy, precision, recall or area under a receiver operating characteristic curve (AUROC) (optionally, on a validation dataset).

19. The method of any one of claims 1-18, wherein the task-conditioned machinelearning model comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

20. The method of any one of claims 1-19, wherein the convolutional neural network and the task-conditioned machine-learning model are jointly trained.

21. The method of any one of claims 1-19, wherein the convolutional neural network and the task-conditioned machine-learning model are separately trained.

22. The method of any one of claims 1-19, wherein the convolutional neural network and the task-conditioned machine-learning model are sequentially or iteratively trained, optionally wherein the convolutional neural network is trained prior to or after training the task-conditioned machine-learning model.

23. The method of any one of claims 1-22, wherein the convolutional neural network and the task-conditioned machine-learning model are trained using the same dataset.

24. The method of any one of claims 1-22, wherein the convolutional neural network and the task-conditioned machine-learning model are trained using different datasets.

25. The method of any one of claims 1-24, wherein (d) comprises applying an ensemble of neural network models, wherein each neural network model in the ensemble processes the condition-specific embedding vectors to generate a set of mutagenicity probability scores, and wherein the mutagenicity probability scores are averaged (optionally, with respect to each Ames testing condition) across the ensemble to generate the profile of mutagenicity probability scores for the compound.

26. The method of claim 25, wherein the ensemble comprises at least five, six, seven, eight, nine, or ten neural network models.

27. The method of any one of claims 1-26, comprising verifying a determination, optionally, a mutagenic or non-mutagenic determination.

28. The method of any one of claims 1-27, comprising testing the compound under at least one of the profiled Ames testing conditions.

29. The method of any one of claims 1-28, comprising testing the compound under each of the profiled Ames testing conditions.

30. The method of any one of claims 1-29, wherein (a) comprises providing a molecular structure representation of the compound.

31. The method of any one of claims 1-30, wherein the method is performed on a plurality of compounds or molecular structure representations thereof.

32. The method of claim 31, wherein the plurality of compounds comprises at least 100, 200, 300, 400, 500, 600, 700, 800, 900, or 1,000 compounds (optionally, small molecules).

33. The method of claim 31 or 32, wherein the method determines or stratifies the plurality of compounds at a sensitivity of at least 0.70, 0.75, or 0.80.

34. The method of any one of claims 31-33, wherein the method determines or stratifies the plurality of compounds at an accuracy (optionally, a balanced accuracy) of at least 0.70, 0.75, or 0.80.

35. The method of any one of claims 31-34, wherein the method determines or stratifies the plurality of compounds at a specificity of at least 0.70, 0.75, or 0.80.

36. The method of any one of claims 1-35, comprising providing a report that describes the determination or stratification of the compound made in (e).

37. The method of any one of claims 1-36, comprising, displaying on a user interface connected to an output device, a report that describes the determination or stratification of the compound made in (e).

38. A system comprising: one or more processors, and memory storing a convolutional neural network (optionally, a graph convolutional neural network), a task-conditioned machinelearning model (optionally, a multilayer perceptron), and a plurality of lines of instructions; wherein the one or more processors are configured to execute the plurality of lines of instructions to perform a method of any one of claims 1-37.

39. A system comprising: one or more processors, and memory storing a task-conditioned machine-learning model (optionally, a multilayer perceptron) and a plurality of lines of instructions; wherein the one or more processors are configured to execute the plurality of lines of instructions to:(a) receive a compound or a molecular structure representation thereof;(b) receiving one or more Ames testing conditions;(c) combine (optionally, concatenate) an adaptive molecular embedding vector corresponding to the compound, with one or more Ames condition vectors corresponding to the one or more Ames testing conditions, to generate one or more condition-specific embedding vectors;(d) process the one or more condition-specific embedding vectors through the task-conditioned machine-learning model (optionally, through the multilayer perceptron (MLP))to generate a mutagenicity probability score for each of the one or more Ames testing conditions, thereby generating a profile of mutagenicity probability scores for the compound; and(e) determine the compound as mutagenic or non-mutagenic, or stratify a mutagenicity risk of the compound, based on the profile of mutagenicity probability scores.

40. The system of claim 39, wherein the memory further stores a convolutional neural network (optionally, a graph convolutional neural network), and wherein the one or more processors are configured to execute the plurality of lines of instructions to generate, using the convolutional neural network, the adaptive molecular embedding vector from a molecular structure representation of the compound.

41. The system of claim 38 or 40, wherein the convolutional neural network is configured to iteratively aggregate structural information (optionally, from neighboring atoms, motifs or fragments) of the compound to generate the adaptive molecular embedding vector.

42. The system of any one of claims 38-40, wherein the convolutional neural network comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

43. The system of any one of claims 38-42, wherein the convolutional neural network comprises a plurality of convolutional layers configured (optionally, trained) to iteratively aggregate structural information (optionally, from neighboring atoms, motifs or fragments) of the compound to generate the adaptive molecular embedding vector.

44. The system of claim 40 or 43, wherein the structural information comprises one or more mutagenicity-associated molecular features of the compound.

45. The system of any one of claims 38-44, wherein the one or more processors are configured to train the convolutional neural network using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions.

46. The system of claim 45, wherein the training dataset is partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

47. The system of claim 45 or 46, wherein one or more reference compounds in the training dataset comprise one or more mutagenicity-associated molecular features.

48. The system of any one of claims 44-47, wherein a mutagenicity-associated molecular feature is a chemical, geometric, topological, or structural feature that is identified, determined, or suspected to be associated (or correlate) with mutagenicity or non-mutagenicity (optionally, as measured by an Ames test under one or more Ames testing conditions).

49. The system of any one of claims 44-48, wherein the one or more mutagenicity-associated molecular features comprise: at least one, two, three, four, five, or six features eachindependently selected from: a polyaromatic moiety (optionally, a planar polyaromatic moiety), an aromatic amine moiety, an aromatic N-heterocycle moiety, an a,P-unsaturated carbonyl moiety, an epoxide moiety, a nitro-aromatic moiety, an aromatic ring N-oxide moiety, an azo moiety, an aliphatic halogen moiety, a 1,2-dihydroxy aromatic ring moiety, a polyfunctional alkyl moiety, an aryl halide moiety, and a nitroso moiety; or50. The system of any one of claims 44-49, wherein the one or more processors are configured to identify one or more mutagenicity-associated molecular features in the compound.

51. The system of claim 50, wherein the one or more processors are configured to provide a report that describes the identified one or more mutagenicity-associated molecular features as present or absent in the compound.

52. The system of any one of claims 38-51, wherein the one or more processors are configured to provide, prior to (c), the one or more Ames condition vectors.

53. The system of any one of claims 38-52, wherein the one or more processors are configured to train the task-conditioned machine-learning model using a training dataset comprising molecular structure representations of reference compounds and, for each reference compound, one or more known mutagenicity classifications under one or more Ames testing conditions.

54. The system of any one of claims 38-53, wherein the one or more processors are configured to train the task-conditioned machine-learning model that comprises:(i) initializing a plurality of parameters of the task-conditioned machinelearning model;(ii) processing the training dataset through the task-conditioned machinelearning model to generate predicted mutagenicity probability scores;(iii) computing a loss function based on differences between the predicted mutagenicity probability scores and known mutagenicity classifications;(iv) updating the plurality of parameters of the task-conditioned machinelearning model to minimize the loss function; and(v) repeating steps (ii)-(iv).

55. The system of claim 53 or 54, wherein the training dataset is partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

56. The system of any one of claims 38-55, wherein the one or more processors are configured to optimize parameters of the task-conditioned machine-learning model until the task-conditioned machine-learning model achieves an improved or substantially maximal value of sensitivity, specificity, accuracy, balanced accuracy, precision, recall or area under a receiver operating characteristic curve (AUROC) (optionally, on a validation dataset).

57. The system of any one of claims 38-56, wherein the task-conditioned machinelearning model comprises a set of parameters obtained from training on a dataset partitioned using t-distributed stochastic neighbor embedding (t-SNE) chemical diversity split.

58. The system of any one of claims 38-57, wherein the convolutional neural network and the task-conditioned machine-learning model are configured to be jointly trained.

59. The system of any one of claims 38-57, wherein the convolutional neural network and the task-conditioned machine-learning model are configured to be separately trained.

60. The system of any one of claims 38-57, wherein the convolutional neural network and the task-conditioned machine-learning model are sequentially or iteratively trained, optionally wherein the convolutional neural network is configured to be trained prior to or after training the task-conditioned machine-learning model.

61. The system of any one of claims 38-60, wherein the convolutional neural network and the task-conditioned machine-learning model are configured to be trained using the same dataset.

62. The system of any one of claims 38-60, wherein the convolutional neural network and the task-conditioned machine-learning model are configured to be trained using different datasets.

63. The system of any one of claims 38-62, wherein (d) comprises applying an ensemble of neural network models, wherein each neural network model in the ensemble processes the condition-specific embedding vectors to generate a set of mutagenicity probability scores, and wherein the mutagenicity probability scores are averaged (optionally, with respect to each Ames testing condition) across the ensemble to generate the profile of mutagenicity probability scores for the compound.

64. The system of claim 63, wherein the ensemble comprises at least five, six, seven, eight, nine, or ten neural network models65. The system of any one of claims 38-64, wherein the one or more processors are configured, in (a), to receive a molecular structure representation of the compound.

66. The system of any one of claims 38-65, wherein the one or more processors are configured to verify a determination, optionally, a mutagenic or non-mutagenic determination.

67. The system of any one of claims 38-66, wherein the one or more processors are configured to test the compound under at least one of the profiled Ames testing conditions.

68. The system of any one of claims 38-67, wherein the one or more processors are configured to test the compound under each of the profiled Ames testing conditions.

69. The system of any one of claims 38-68, wherein the one or more processors are configured to perform the method on a plurality of compounds.

70. The system of claim 69, wherein the plurality of compounds comprises at least 100, 200, 300, 400, 500, 600, 700, 800, 900, or 1,000 compounds (optionally, small molecules).

71. The system of any one of claims 38-70, wherein the one or more processors are configured to determine or stratify the plurality of compounds at a sensitivity of at least 0.70, 0.75, or 0.80.

72. The system of any one of claims 38-71, wherein the one or more processors are configured to determine or stratify the plurality of compounds at an accuracy (optionally, a balanced accuracy) of at least 0.70, 0.75, or 0.80.

73. The system of any one of claims 38-72, wherein the one or more processors are configured to determine or stratify the plurality of compounds at a specificity of at least 0.70, 0.75, or 0.80.

74. The system of any one of claims 38-73, wherein the one or more processors are configured to provide a report that that describes the determination or stratification of the compound made in (e).

75. The system of any one of claims 38-74, further comprising an output device, wherein the one or more processors are configured to display on the output device a report that that describes the determination or stratification of the compound made in (e).

76. The method or system of any one of claims 1-75, wherein the compound is a small molecule.

77. The method or system of any one of claims 1-76, wherein the molecular structure representation is a textual representation, optionally, a standardized and / or linear textual representation, further optionally, a standardized linear textual representation; wherein, optionally, the textual representation is a SMILES string.

78. The method or system of any one of claims 1-77, wherein the adaptive molecular embedding vector is a multi-dimensional vector, optionally, comprising at least 64, 128, 256, 512, or 1,024 dimensions.

79. The method or system of any one of claims 1-78, wherein the adaptive molecular embedding vector comprises (optionally, local, regional and / or global) structural information (optionally, molecular geometric and / or topological features) of the compound (optionally, at one or more of: atomic, motif, cross-motif, or whole-molecule level(s)).

80. The method or system of any one of claims 1-79, wherein the profile comprises at least two, three, four, five, six, seven, eight, nine, or ten mutagenicity probability scores.

81. The method or system of any one of claims 1-80, wherein the profile comprises mutagenicity probability scores for at least two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, or sixteen Ames testing conditions.

82. The method or system of any one of claims 1-81, wherein the profile comprises mutagenicity probability scores for Ames testing conditions corresponding to at least one, two, three, four, five, six, seven, or eight distinct bacterial tester strains, each tested in the presence and / or absence of a metabolic activation.

83. The method or system of any one of claims 1-82, wherein the profile comprises mutagenicity probability scores for Ames testing conditions corresponding to at least one, two, three, four, five, six, seven, or eight distinct bacterial tester strains, each tested both in the presence and absence of a metabolic activation.

84. The method or system of any one of claims 1-83, wherein an Ames testing condition comprises a distinct combination of (1) the identity of a bacterial tester strain and (2) the status (i.e., the presence or absence) of a metabolic activation.

85. The method or system of any one of claims 82-84, wherein the metabolic activation comprises an S9 fraction.

86. The method or system of any one of claims 82-85, wherein a bacterial tester strain is a base-pair substitution mutant strain or a frameshift mutant strain.

87. The method or system of any one of claims 82-86, wherein a bacterial tester strain is a Salmonella typhimurium (S. typhimurium) strain.

88. The method or system of claim 86 or 87, wherein a base-pair substitution mutant strain is selected from S. typhimurium TAI 00, TAI 02, TAI 04, and TAI 535.

89. The method or system of any one of claims 86-88, wherein a frameshift mutant strain is selected from S. typhimurium TA1537, TA1538, TA97, and TA98.

90. The method or system of any one of claims 1-89, wherein said determining or stratifying comprises comparing at least one mutagenicity probability score to a threshold.

91. The method or system of any one of claims 1-90, wherein the compound is determined or stratified as non-mutagenic based on a determination that the compound is negative under all profiled Ames testing conditions.

92. The method or system of claim 91, wherein a mutagenicity probability score is between 0 and 1; and wherein a negative determination is made when the mutagenicity probability score is below (or at or below) 0.3, 0.4, or 0.5.

93. The method or system of any one of claims 1-90, wherein the compound is determined or stratified as mutagenic based on a determination that the compound is positive (or strongly positive) under at least one profiled Ames testing condition.

94. The method or system of any one of claims 1-90 and 93, wherein the compound is determined or stratified as highly mutagenic based on a determination that the compound is positive (or strongly positive) under at least two profiled Ames testing condition.

95. The method or system of claim 93 or 94, wherein a mutagenicity probability score is between 0 and 1; and wherein a positive determination is made when the mutagenicity probability score is above (or at or above) 0.5.

96. The method or system of claim 93 or 94, wherein a mutagenicity probability score is between 0 and 1; and wherein a strongly positive determination is made when the mutagenicity probability score is above (or at or above) 0.5.

97. The method or system of any one of claims 1-96, wherein the compound is determined or stratified as an Ames Class A compound, if the compound is determined to be strongly positive under at least one profiled Ames testing condition.

98. The method or system of any one of claims 1-96, wherein the compound is determined or stratified as an Ames Class B compound, if the compound is determined to be positive (but not strongly positive) under at least one profiled Ames testing condition.

99. The method or system of any one of claims 1-96, wherein the compound is determined or stratified as an Ames Class C compound, if the compound is determined to be negative under all profiled Ames testing conditions.

100. The method or system of any one of claims 1-99, wherein the compound is stratified with a low mutagenicity risk based on a determination that the compound is negative under all profiled Ames testing conditions.

101. The method or system of any one of claims 1-99, wherein the compound is stratified with a medium mutagenicity risk based on a determination that the compound is positive under only one profiled Ames testing condition.

102. The method or system of any one of claims 1-99, wherein the compound is stratified with a high mutagenicity risk based on a determination that the compound is positive under at least two profiled Ames testing conditions.

103. The method or system of any one of claims 6-38 and 44-102, wherein the one or more mutagenicity-associated molecular features comprise:(i) at least one, two, three, four, five, or six features each independently selected from: Aziridine, Acrylate / Acrylamide, Alpha-methylene lactone, Epoxide (oxirane), Quinoline, Aromatic imine (Schiff base), Imine (aliphatic C=N), Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Poly functional alkyl: bis-haloalkyl, Tertiary aromatic amine (N,N-dialkyl), Non-aromatic C=C fused / adjacent to aromatic ring, Cyclic dithioacetal / 1,3-dithiane, Bay region (phenanthrene-type angular fusion), Naphthalene, Alpha-diketone (1,2-dicarbonyl), K-region (anthracene-type linear tri-ring fusion), Primary aromatic amine, Aromatic nitro group, Aromatic azo compound, 1 -Aminonaphthalene, Aromatic amine N-oxide, Phenol (general), Diarylamine (Ar-NH-Ar), Coumarin / Benzo-alpha-pyrone, Peroxide, Hydroperoxide,Benzylic methyl / methylene, meta-Diaminoarene, Triazene, Imidazole (general), Acridine scaffold, Quinolone (4-oxoquinoline), Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Aryl halide (single), Diaryl thioether, Benzylic primary alcohol, Pyrrole ring, Indole, Thiazole ring, Alpha, beta-unsaturated ketone (enone), Alpha, beta-unsaturated aldehyde, Aldehyde, Enol ether, Aromatic secondary amine, Nitrofuran, Furan ring, Gem-dihaloalkane, Benzylic secondary alcohol, N-Nitrosamine (dialkyl), Nitroalkene (nitrostyrene-type), Naphthyridine, Thiophene ring, Catechol (1,2-dihydroxybenzene), Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Diazo compound, Anthraquinone, Alpha, beta-unsaturated nitrile, 2-Aminonaphthalene, Phosphoryl halide, Alpha-methylene lactam, para-Aminophenol, and Phenol with para electron-donating group; or(ii) at least one, two, three, four, five, or six features each independently selected from: Naphthalene, Benzylic secondary alcohol, Catechol (1,2-dihydroxybenzene), Phenol (general), Anthraquinone, Benzylic primary alcohol, Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Non-aromatic C=C fused / adjacent to aromatic ring, Epoxide (oxirane), Acridine scaffold, Quinoline, Diazo compound, Imine (aliphatic C=N), Alpha, beta-unsaturated nitrile, Enol ether, Tertiary aromatic amine (N,N-dialkyl), Primary alkyl halide, Nitrogen mustard, Halfmustard (2-chloroethylamine), Polyfunctional alkyl: bis-haloalkyl, Acrylate / Acrylamide, Alphamethylene lactone, Aromatic nitro group, Aromatic amine N-oxide, Imidazole (general), Alpha, beta-unsaturated ketone (enone), Aromatic imine (Schiff base), Primary aromatic amine, Aromatic azo compound, 2- Aminonaphthalene, Phosphoryl halide, Cyclic dithioacetal / 1,3-dithiane, Alpha-diketone (1,2-dicarbonyl), Thiazole ring, Bay region (phenanthrene-type angular fusion), K-region (anthracene-type linear tri-ring fusion), Gem-dihaloalkane, Aziridine, Pyrrole ring, Indole, Alpha-methylene lactam, 1 -Aminonaphthalene, Coumarin / Benzo-alpha-pyrone, N-Nitrosamine (dialkyl), Methylenedi oxybenzene (safrole-type), Naphthyridine, Hydrazine, Aromatic hydrazine (phenylhydrazine-type), Quinolone (4-oxoquinoline), Aromatic secondary amine, Benzylic methyl / methylene, Nitrofuran, Furan ring, Hydrazone, Thioether on aromatic ring, Aryl halide (single), meta-Diaminoarene, Thiophene ring, and Triazene; or(iii) at least one, two, three, four, five, or six features each independently selected from: Aziridine, Aromatic nitro group, Alpha, beta-unsaturated ketone (enone), Aromatic amine N-oxide, Non-aromatic C=C fused / adjacent to aromatic ring, Imidazole (general), Aromatic imine (Schiff base), Imine (aliphatic C=N), Cyclic dithioacetal / 1,3-dithiane, Nitrofuran, Furan ring, Hydrazone, Epoxide (oxirane), Bay region (phenanthrene-type angular fusion), Naphthalene, Alpha-diketone (1,2-dicarbonyl), K-region (anthracene-type linear tri-ring fusion), Primary aromatic amine, Aromatic azo compound, 1 -Aminonaphthalene, Phenol (general), Alpha, betaunsaturated aldehyde, Aldehyde, Acetal / Hemiacetal (potential aldehyde release), Peroxide,Hydroperoxide, meta-Diaminoarene, Thioether on aromatic ring, Triazene, Acridine scaffold, Quinoline, Quinolone (4-oxoquinoline), Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Pyrrole ring, Indole, Thiazole ring, Enol ether, Methylenedioxybenzene (safrole-type), Aromatic secondary amine, Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Polyfunctional alkyl: bis-haloalkyl, Acrylate / Acrylamide, Alpha-methylene lactone, Tertiary aromatic amine (N,N-dialkyl), N-Nitrosamine (dialkyl), Nitroalkene (nitrostyrene-type), Naphthyridine, ortho-Aminophenol, Phenazine scaffold, Coumarin / Benzo-alpha-pyrone, Anthraquinone, Thiophene ring, Catechol (1,2-dihydroxybenzene), Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Diazo compound, Benzylic primary alcohol, Alpha, beta-unsaturated nitrile, 2-Aminonaphthalene, Phosphoryl halide, Alpha-methylene lactam, Hydrazine, Aromatic hydrazine (phenylhydrazine-type), para-Aminophenol, Phenol with para electron-donating group, Hydroquinone (1,4-dihydroxybenzene), and para-Diaminoarene (non-biphenyl); or(iv) at least one, two, three, four, five, or six features each independently selected from: Aziridine, Acrylate / Acrylamide, Alpha-methylene lactone, Aromatic nitro group, Alpha, betaunsaturated ketone (enone), Aromatic amine N-oxide, Non-aromatic C=C fused / adjacent to aromatic ring, Aromatic imine (Schiff base), Imine (aliphatic C=N), Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Polyfunctional alkyl: bis-haloalkyl, Tertiary aromatic amine (N,N-dialkyl), Cyclic dithioacetal / 1,3-dithiane, Alpha-diketone (1,2-dicarbonyl), Naphthalene, K-region (anthracene-type linear tri-ring fusion), Primary aromatic amine, Aromatic azo compound, 1 -Aminonaphthalene, Phenol (general), Alpha, beta-unsaturated aldehyde, Aldehyde, Acetal / Hemiacetal (potential aldehyde release), Diarylamine (Ar-NH-Ar), Peroxide, Hydroperoxide, Benzylic methyl / methylene, meta-Diaminoarene, Thioether on aromatic ring, Acridine scaffold, Quinoline, Quinolone (4-oxoquinoline), Hydroquinone (1,4-dihydroxybenzene), Phenol with para electron-donating group, Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Epoxide (oxirane), Aryl halide (single), Diaryl thioether, Benzylic primary alcohol, Pyrrole ring, Indole, Thiazole ring, Imidazole (general), Enol ether, N-Acyloxy group, Methylenedioxybenzene (safrole-type), Aromatic secondary amine, Nitrofuran, Furan ring, Hydrazone, Anthraquinone, N-Nitrosamine (dialkyl), Naphthyridine, ortho-Aminophenol, Phenazine scaffold, Coumarin / Benzo-alpha-pyrone, Alkyl sulfonate ester, Propane sultone, Tosylate / Mesylate / Tritiate ester, Thiophene ring, Triazene, Catechol (1,2-dihydroxybenzene), Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Diazo compound, Alpha, betaunsaturated nitrile, 2-Aminonaphthalene, Phosphoryl halide, Benzylic secondary alcohol, Bay region (phenanthrene-type angular fusion), Alpha-methylene lactam, and para-Aminophenol; or (v) at least one, two, three, four, five, or six features each independently selected from:Aziridine, Acrylate / Acrylamide, Alpha-methylene lactone, Aromatic nitro group, Alpha, betaunsaturated ketone (enone), Aromatic amine N-oxide, Non-aromatic C=C fused / adjacent to aromatic ring, Imidazole (general), Aromatic imine (Schiff base), Imine (aliphatic C=N), Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Polyfunctional alkyl: bis-haloalkyl, Tertiary aromatic amine (N,N-dialkyl), Cyclic dithioacetal / 1,3-dithiane, Epoxide (oxirane), Bay region (phenanthrene-type angular fusion), Naphthalene, Alpha-diketone (1,2-dicarbonyl), K-region (anthracene-type linear tri-ring fusion), Primary aromatic amine, Aromatic azo compound, 1 -Aminonaphthalene, Phenol (general), Alpha, beta-unsaturated aldehyde, Aldehyde, Acetal / Hemiacetal (potential aldehyde release), Diarylamine (Ar-NH-Ar), Coumarin / Benzo-alpha-pyrone, Benzylic methyl / methylene, Peroxide, Hydroperoxide, meta-Diaminoarene, Thioether on aromatic ring, Triazene, Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Aryl halide (single), Diaryl thioether, Benzylic primary alcohol, Pyrrole ring, Indole, Thiazole ring, Enol ether, Benzylic secondary alcohol, N-Acyloxy group, Methylenedioxybenzene (safrole-type), Aromatic secondary amine, Acridine scaffold, Quinoline, Nitrofuran, Furan ring, Hydrazone, N-Nitrosamine (dialkyl), Nitroalkene (nitrostyrene-type), ortho-Aminophenol, Phenazine scaffold, Anthraquinone, Alkyl sulfonate ester, Propane sultone, Tosylate / Mesylate / Tritiate ester, 2-Aminonaphthalene, Thiophene ring, Catechol (1,2-dihydroxybenzene), Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Diazo compound, Alpha, beta-unsaturated nitrile, Phosphoryl halide, Alpha-methylene lactam, para-Aminophenol, Phenol with para electron-donating group, Hydroquinone (1,4-dihydroxybenzene), and para-Diaminoarene (non-biphenyl); or(vi) at least one, two, three, four, five, or six features each independently selected from: Aziridine, Aromatic nitro group, Alpha, beta-unsaturated ketone (enone), Aromatic amine N-oxide, Non-aromatic C=C fused / adjacent to aromatic ring, Aromatic imine (Schiff base), Imine (aliphatic C=N), Primary aromatic amine, Aromatic azo compound, 2-Aminonaphthalene, Naphthalene, Phenol (general), Cyclic dithioacetal / 1,3-dithiane, Epoxide (oxirane), Bay region (phenanthrene-type angular fusion), Alpha-diketone (1,2-dicarbonyl), K-region (anthracene-type linear tri-ring fusion), Coumarin / Benzo-alpha-pyrone, Thiazole ring, Peroxide, Hydroperoxide, Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Polyfunctional alkyl: bis-haloalkyl, Tertiary aromatic amine (N,N-dialkyl), Benzylic methyl / methylene, 1-Aminonaphthalene, meta-Diaminoarene, Thioether on aromatic ring, Triazene, Imidazole (general), Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, Pyrrole ring, Indole, Quinoline, Quinolone (4-oxoquinoline), Alpha, beta-unsaturated aldehyde, Aldehyde, Enol ether, Methylenedioxybenzene (safrole-type), Aromatic secondary amine, Acridine scaffold, Benzylic secondary alcohol, Acrylate / Acrylamide, Alpha-methylene lactone, N-Nitrosamine (dialkyl),Nitroalkene (nitrostyrene-type), Naphthyridine, ortho-Aminophenol, Phenazine scaffold, Anthraquinone, Alkyl sulfonate ester, Propane sultone, Tosylate / Mesylate / Tritiate ester, Thiophene ring, Catechol (1,2-dihydroxybenzene), Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Benzylic primary alcohol, Alpha, beta-unsaturated nitrile, Phosphoryl halide, Alpha-methylene lactam, para-Aminophenol, Phenol with para electron-donating group, Hydroquinone (1,4-dihydroxybenzene), and para-Diaminoarene (non-biphenyl); or(vii) at least one, two, three, four, five, or six features each independently selected from: Epoxide (oxirane), Acrylate / Acrylamide, Alpha-methylene lactone, Non-aromatic C=C fused / adjacent to aromatic ring, Bay region (phenanthrene-type angular fusion), Naphthalene, Enol ether, K-region (anthracene-type linear tri-ring fusion), Cyclic dithioacetal / 1,3-dithiane, Alpha-methylene lactam, Phenol (general), N-Acyloxy group, Aromatic azo compound, Aryl halide (single), Tertiary aromatic amine (N,N-dialkyl), Thiazole ring, Peroxide, Hydroperoxide, Alpha, beta-unsaturated aldehyde, Aldehyde, Acetal / Hemiacetal (potential aldehyde release), Alpha, beta-unsaturated nitrile, Alpha, beta-unsaturated ketone (enone), Aromatic secondary amine, Aromatic nitro group, Nitrofuran, Furan ring, Aromatic amine N-oxide, Hydrazone, Alpha-diketone (1,2-dicarbonyl), Aromatic N-oxide, Diaryl thioether, Benzylic primary alcohol, Pyridinium / Quaternary aromatic nitrogen, Primary alkyl halide, Half-mustard (2-chloroethylamine), Acridine scaffold, Quinoline, Benzylic methyl / methylene, Thiophene ring, Hydroquinone (1,4-dihydroxybenzene), Phenol with para electron-donating group, Anthraquinone, Quinolone (4-oxoquinoline), Naphthyridine, Imidazole (general), Pyrrole ring, Indole, Coumarin / Benzo-alpha-pyrone, Methylenedioxybenzene (safrole-type), N-Nitrosamine (dialkyl), Nitrogen mustard, Poly functional alkyl: bis-haloalkyl, Benzylic secondary alcohol, Triazene, Diarylamine (Ar-NH-Ar), Primary aromatic amine, 2-Aminonaphthalene, Aziridine, Hydrazine, Aromatic hydrazine (phenylhydrazine-type), 1 -Aminonaphthalene, para-Aminophenol, ortho-Aminophenol, Phenazine scaffold, para-Diaminoarene (non-biphenyl), meta-Diaminoarene, Thioether on aromatic ring, Aromatic N-oxide (=0 form), N-Hydroxy heteroaromatic, Aromatic imine (Schiff base), Imine (aliphatic C=N), Gem-dihaloalkane, Phosphoryl halide, Alkyl sulfonate ester, Propane sultone, Tosylate / Mesylate / Tritiate ester, Nitroalkene (nitrostyrene-type), Catechol (1,2-dihydroxybenzene), and Diazo compound; or (viii) at least one, two, three, four, five, or six features each independently selected from: Cyclic dithioacetal / 1,3-dithiane, Epoxide (oxirane), Acrylate / Acrylamide, Alpha-methylene lactam, Alpha, beta-unsaturated nitrile, Enol ether, Tertiary aromatic amine (N,N-dialkyl), Non-aromatic C=C fused / adjacent to aromatic ring, Alpha-diketone (1,2-dicarbonyl), Thiophene ring, N-Nitrosamine (dialkyl), Primary alkyl halide, Nitrogen mustard, Half-mustard (2-chloroethylamine), Polyfunctional alkyl: bis-haloalkyl, Alpha-methylene lactone, Triazene,Primary aromatic amine, Aromatic azo compound, 2-Aminonaphthalene, Naphthalene, Phenol (general), Aromatic imine (Schiff base), Imine (aliphatic C=N), Phosphoryl halide, Bay region (phenanthrene-type angular fusion), Catechol (1,2-dihydroxybenzene), Imidazole (general), Aromatic N-oxide, Pyridinium / Quaternary aromatic nitrogen, and Thiazole ring.

Images