Mechanistic Task Groupings Enhance Multitask Deep Learning of Strain-Specific Ames Mutagenicity.

The Ames test is a gold standard mutagenicity assay that utilizes various Salmonella typhimurium strains with and without S9 fraction to provide insights into the mechanisms by which a chemical can mutate DNA. Multitask deep learning is an ideal framework for developing QSAR models with multiple end points, such as the Ames test, as the joint training of multiple predictive tasks may synergistically improve the prediction accuracy of each task. This work investigated how toxicology domain knowledge can be used to handcraft task groupings that better guide the training of multitask neural networks compared to a naïve ungrouped multitask neural network developed on a complete set of tasks. Sixteen S. typhimurium ± S9 strain tasks were used to generate groupings based on mutagenic and metabolic mechanisms that were reflected in correlation data analyses. Both grouped and ungrouped multitask neural networks predicted the 16 strain tasks with a higher balanced accuracy compared with single task controls, with grouped multitask neural networks consistently featuring incremental increases in predictivity over the ungrouped approach. We conclude that the main variable driving these performance improvements is the general multitask effect with mechanistic task groupings acting as an enhancement step to further concentrate synergistic training signals united by a common biological mechanism. This approach enables incorporation of toxicology domain knowledge into multitask QSAR model development allowing for more transparent and accurate Ames mutagenicity prediction.

Multiple Instance Learning Improves Ames Mutagenicity Prediction for Problematic Molecular Species.

The prediction of Ames mutagenicity continues to be a concern in both regulatory and pharmacological toxicology. Traditional quantitative structure-activity relationship (QSAR) models of mutagenicity make predictions based on molecular descriptors calculated on a chemical data set used in their training. However, it is known that molecules such as aromatic amines can be non-mutagenic themselves but metabolically activated by S9 rodent liver enzyme in Ames tests forming molecules such as iminoquinones or amine substituents that better stabilize mutagenic nitrenium ions in known pathways of mutagenicity. Modern in silico modeling methods can implicitly model these metabolites through consideration of the structural elements relevant to their formation but do not include explicit modeling of these metabolites' potential activity. These metabolites do not have a known individual mutagenicity label and, in their current state, cannot be fitted into a traditional QSAR model. Multiple instance learning (MIL) however can be applied to a group of metabolites and their parent under a single mutagenicity label. Here we trained MIL models on Ames data, first with an aromatic amines data set (n = 457), a class known to require metabolic activation, and subsequently on a larger data set (n = 6505) incorporating multiple molecular species. MIL was shown to be able to predict Ames mutagenicity with performance in line with previously established models (balanced accuracy = 0.778), suggesting its potential utility in Ames prediction applications. Furthermore, the MIL model predicted well on identified hard-to-predict molecule groups relative to the models in which these molecule groups were identified. These results are presumably due to the increased consideration of the metabolic contribution to the mutagenic outcome. Further exploration of MIL as a supplement to existing models could aid in the prediction of chemicals where implicit modeling of metabolites cannot fully grasp their characteristics. This paper demonstrates the potential of an MIL approach to modeling Ames tests with S9 and is particularly relevant to metabolically activated xenobiotic mutagens.

A comparison of molecular representations for lipophilicity quantitative structure-property relationships with results from the SAMPL6 logP Prediction Challenge.

Effective representation of a molecule is required to develop useful quantitative structure-property relationships (QSPR) for accurate prediction of chemical properties. The octanol-water partition coefficient logP, a measure of lipophilicity, is an important property for pharmacological and toxicological endpoints used in the pharmaceutical and regulatory spheres. We compare physicochemical descriptors, structural keys, and circular fingerprints in their ability to effectively represent a chemical space and characterise molecular features to correlate with lipophilicity. Exploratory landscape continuity analyses revealed that whole-molecule physicochemical descriptors could map together compounds that were similar in both molecular features and logP, indicating higher potential for use in logP QSPRs compared to the substructural approach of structural keys and circular fingerprints. Indeed, logP QSPR models parameterised by physicochemical descriptors consistently performed with the lowest error. Our best performing model was a stochastic gradient descent-optimised multilinear regression with 1438 descriptors, returning an internal benchmark RMSE of 1.03 log units. This corroborates the well-established notion that lipophilicity is an additive, whole-molecule property. We externally tested the model by participating in the 2019 SAMPL6 logP Prediction Challenge and blindly predicting for 11 protein kinase inhibitor fragment-like molecules. Our model returned an RMSE of 0.49 log units, placing eighth overall and third in the empirical methods category (submission ID 'hdpuj'). Permutation feature importance analyses revealed that physicochemical descriptors could characterise predictive molecular features highly relevant to the kinase inhibitor fragment-like molecules.

LogP prediction performance with the SMD solvation model and the M06 density functional family for SAMPL6 blind prediction challenge molecules.

This work presents a quantum mechanical model for predicting octanol-water partition coefficients of small protein-kinase inhibitor fragments as part of the SAMPL6 LogP Prediction Challenge. The model calculates solvation free energy differences using the M06-2X functional with SMD implicit solvation and the def2-SVP basis set. This model was identified as dqxk4 in the SAMPL6 Challenge and was the third highest performing model in the physical methods category with 0.49 log Root Mean Squared Error (RMSE) for predicting the 11 compounds in SAMPL6 blind prediction set. We also collaboratively investigated the use of empirical models to address model deficiencies for halogenated compounds at minimal additional computational cost. A mixed model consisting of the dqxk4 physical and hdpuj empirical models found improved performance at 0.34 log RMSE on the SAMPL6 dataset. This collaborative mixed model approach shows how empirical models can be leveraged to expediently improve performance in chemical spaces that are difficult for ab initio methods to simulate.

Combining machine learning models of in vitro and in vivo bioassays improves rat carcinogenicity prediction.

In vitro genotoxicity bioassays are cost-efficient methods of assessing potential carcinogens. However, many genotoxicity bioassays are inappropriate for detecting chemicals eliciting non-genotoxic mechanisms, such as tumour promotion, this necessitates the use of in vivo rodent carcinogenicity (IVRC) assays. In silico IVRC modelling could potentially address the low throughput and high cost of this assay. We aimed to develop and combine computational QSAR models of novel bioassays for the prediction of IVRC results and compare with existing software. QSAR models were generated from existing Ames (n = 6512), Syrian Hamster Embryonic (SHE, n = 410), ISSCAN rodent carcinogenicity (ISC, n = 834) and GreenScreen GADD45a-GFP (n = 1415) chemical datasets. These models mapped the molecular descriptors of each compound to their respective assay result using machine learning algorithms (adaboost, k-Nearest Neighbours, C.45 Decision Tree, Multilayer Perceptron, Random Forest). The best performing models were combined with k-Nearest Neighbours to create a cascade model for IVRC prediction. High QSAR model performance was observed from ten time 10-fold cross-validation with above 80% accuracy and 0.85 AUC for each assay dataset. The cascade model predicted rat carcinogenicity with 69.3% accuracy and 0.700 AUC. This study demonstrates the novelty of a combined approach for IVRC prediction, with higher performance than existing software.

Bridging informatics and medicinal inorganic chemistry: Toward a database of metallodrugs and metallodrug candidates.

Metallodrug discovery has evolved in recent years, yielding several compounds in the clinic for therapeutic and medical imaging diagnostic applications. As reviewed here, several research groups in well-established medicinal inorganic chemistry groups are consistently generating high-quality SAR data representing an ideal starting point in the use of computational methods to advance the development of new drugs. Although there are representative chemical structures of metallodrugs in public databases annotated with biological activity, there is currently no public compound database dedicated to metallodrugs. Here, we also discuss the significance, viability, applications and challenges of developing a public compound database of metallodrugs - with consistent representation of metallodrug structure being a crucial obstacle. A curated metallo-compound database would substantially benefit metallodrug discovery and development.

An Open Drug Discovery Competition: Experimental Validation of Predictive Models in a Series of Novel Antimalarials.

The Open Source Malaria (OSM) consortium is developing compounds that kill the human malaria parasite, Plasmodium falciparum, by targeting PfATP4, an essential ion pump on the parasite surface. The structure of PfATP4 has not been determined. Here, we describe a public competition created to develop a predictive model for the identification of PfATP4 inhibitors, thereby reducing project costs associated with the synthesis of inactive compounds. Competition participants could see all entries as they were submitted. In the final round, featuring private sector entrants specializing in machine learning methods, the best-performing models were used to predict novel inhibitors, of which several were synthesized and evaluated against the parasite. Half possessed biological activity, with one featuring a motif that the human chemists familiar with this series would have dismissed as "ill-advised". Since all data and participant interactions remain in the public domain, this research project "lives" and may be improved by others.

QSAR ligand dataset for modelling mutagenicity, genotoxicity, and rodent carcinogenicity.

Five datasets were constructed from ligand and bioassay result data from the literature. These datasets include bioassay results from the Ames mutagenicity assay, Greenscreen GADD-45a-GFP assay, Syrian Hamster Embryo (SHE) assay, and 2 year rat carcinogenicity assay results. These datasets provide information about chemical mutagenicity, genotoxicity and carcinogenicity.