Search

The Toxic Substances Control Act (TSCA) requires the US EPA to evaluate the hazard and exposure of new and existing chemicals. New chemical notifications are typically data-poor and EPA has historically relied upon approaches including chemical categories to fill data gaps. As part of a multi-year Research Program, opportunities are being explored to leverage New Approach Methods (NAMs) in hazard and exposure assessments. Data from a battery of in vitro NAMs will be generated to form a case study for an adaptable approach to inform new chemical assessments. Herein, a cheminformatics workflow was developed to identify a set of ∼300 representative candidate chemicals for in vitro screening from the TSCA non-confidential active inventory. The freely available web application ClassyFire was used to categorize all discrete organic structures from the TSCA inventory into one of 68 primary structural categories. Large primary categories were subcategorized into smaller categories using hierarchical agglomerative clustering, ultimately yielding 180 structural terminal categories. The inventory was filtered to substances that lacked previous ToxCast bioactivity screening, were associated with physicochemical property predictions indicating non-volatile solids or liquids, and had a higher chance of procurement. Amenability predictions for liquid chromatography-mass spectrometry were also generated to provide an indication of which chemicals lent themselves to aqueous-based screening and analytical verification in solvated samples. Structures associated with transformation in solvent, potentially explosive or highly reactive, were excluded. Potential candidate substances were selected on the basis of being structurally representative of the terminal category and meeting other screenability conditions. A final set of 318 candidate chemicals were proposed to undergo analytical quality control and screening in a range of broad and targeted biological technologies for human health-relevant end points. Finally, in silico tools were applied to explore predicted hazard profiles of these candidate substances relative to the full inventory.

Regulatory authorities aim to organize substances into groups to facilitate prioritization within hazard and risk assessment processes. Often, such chemical groupings are not explicitly defined by structural rules or physicochemical property information. This is largely due to how these groupings are developed, namely, a manual expert curation process, which in turn makes updating and refining groupings, as new substances are evaluated, a practical challenge. Herein, machine learning methods were leveraged to build models that could preliminarily assign substances to predefined groups. A set of 86 groupings containing 2,184 substances as published on the European Chemicals Agency (ECHA) website were mapped to the U.S. Environmental Protection Agency (EPA) Distributed Toxicity Structure Database (DSSTox) content to extract chemical and structural information. Substances were represented using Morgan fingerprints, and two machine learning approaches were used to classify test substances into 56 groups containing at least 10 substances with a structural representation in the data set: k-nearest neighbor (kNN) and random forest (RF), that led to mean 5-fold cross-validation test accuracies (average F1 scores) of 0.781 and 0.853, respectively. With a 9% improvement, the RF classifier was significantly more accurate than KNN ( p -value = 0.001). The approach offers promise as a means of the initial profiling of new substances into predefined groups to facilitate prioritization efforts and streamline the assessment of new substances when earlier groupings are available. The algorithm to fit and use these models has been made available in the accompanying repository, thereby enabling both use of the produced models and refitting of these models, as new groupings become available by regulatory authorities or industry.

Per- and polyfluoroalkyl substances (PFAS) are a diverse set of commercial chemicals widely detected in humans and the environment. However, only a limited number of PFAS are associated with epidemiological or experimental data for hazard identification. To provide developmental neurotoxicity (DNT) hazard information, the work herein employed DNT new approach methods (NAMs) to generate in vitro screening data for a set of 160 PFAS. The DNT NAMs battery was comprised of the microelectrode array neuronal network formation assay (NFA) and high-content imaging (HCI) assays to evaluate proliferation, apoptosis, and neurite outgrowth. The majority of PFAS (118/160) were inactive or equivocal in the DNT NAMs, leaving 42 active PFAS that decreased measures of neural network connectivity and neurite length. Analytical quality control indicated 43/118 inactive PFAS samples and 10/42 active PFAS samples were degraded; as such, careful interpretation is required as some negatives may have been due to loss of the parent PFAS, and some actives may have resulted from a mixture of parent and/or degradants of PFAS. PFAS containing a perfluorinated carbon (C) chain length ≥8, a high C:fluorine ratio, or a carboxylic acid moiety were more likely to be bioactive in the DNT NAMs. Of the PFAS positives in DNT NAMs, 85% were also active in other EPA ToxCast assays, whereas 79% of PFAS inactives in the DNT NAMs were active in other assays. These data demonstrate that a subset of PFAS perturb neurodevelopmental processes in vitro and suggest focusing future studies of DNT on PFAS with certain structural feature descriptors.