RESUMEN

Animal toxicity testing is time and resource intensive, making it difficult to keep pace with the number of substances requiring assessment. Machine learning (ML) models that use chemical structure information and high-throughput experimental data can be helpful in predicting potential toxicity . However, much of the toxicity data used to train ML models is biased with an unequal balance of positives and negatives primarily since substances selected for in vivo testing are expected to elicit some toxicity effect. To investigate the impact this bias had on predictive performance, various sampling approaches were used to balance in vivo toxicity data as part of a supervised ML workflow to predict hepatotoxicity outcomes from chemical structure and/or targeted transcriptomic data. From the chronic, subchronic, developmental, multigenerational reproductive, and subacute repeat-dose testing toxicity outcomes with a minimum of 50 positive and 50 negative substances, 18 different study-toxicity outcome combinations were evaluated in up to 7 ML models. These included Artificial Neural Networks, Random Forests, Bernouilli Naïve Bayes, Gradient Boosting, and Support Vector classification algorithms which were compared with a local approach, Generalised Read-Across (GenRA), a similarity-weighted k-Nearest Neighbour (k-NN) method. The mean CV F1 performance for unbalanced data across all classifiers and descriptors for chronic liver effects was 0.735 (0.0395 SD). Mean CV F1 performance dropped to 0.639 (0.073 SD) with over-sampling approaches though the poorer performance of KNN approaches in some cases contributed to the observed decrease (mean CV F1 performance excluding KNN was 0.697 (0.072 SD)). With under-sampling approaches, the mean CV F1 was 0.523 (0.083 SD). For developmental liver effects, the mean CV F1 performance was much lower with 0.089 (0.111 SD) for unbalanced approaches and 0.149 (0.084 SD) for under-sampling. Over-sampling approaches led to an increase in mean CV F1 performance (0.234, (0.107 SD)) for developmental liver toxicity. Model performance was found to be dependent on dataset, model type, balancing approach and feature selection. Accordingly tailoring ML workflows for predicting toxicity should consider class imbalance and rely on simpler classifiers first.

RESUMEN

Per- and polyfluoroalkyl substances (PFAS) is a large compound class (n > 12,000) that is extensively present in food, drinking water, and aquatic environments. Reduced serum triglycerides and hepatosteatosis appear to be the common phenotypes for different PFAS chemicals. However, the hepatosteatosis potential of most PFAS chemicals remains largely unknown. This study aims to investigate PFAS-induced hepatosteatosis using in vitro high-throughput phenotype profiling (HTPP) and high-throughput transcriptomic (HTTr) data. We quantified the in vitro hepatosteatosis effects and mitochondrial damage using high-content imaging, curated the transcriptomic data from the Gene Expression Omnibus (GEO) database, and then calculated the point of departure (POD) values for HTPP phenotypes or HTTr transcripts, using the Bayesian benchmark dose modeling approach. Our results indicated that PFAS compounds with fully saturated C-F bonds, sulfur- and nitrogen-containing functional groups, and a fluorinated carbon chain length greater than 8 have the potential to produce biological effects consistent with hepatosteatosis. PFAS primarily induced hepatosteatosis via disturbance in lipid transport and storage. The potency rankings of PFAS compounds are highly concordant among in vitro HTPP, HTTr, and in vivo hepatosteatosis phenotypes (ρ = 0.60-0.73). In conclusion, integrating the information from in vitro HTPP and HTTr analyses can accurately project in vivo hepatosteatosis effects induced by PFAS compounds.

Asunto(s)

Fluorocarburos , Perfilación de la Expresión Génica , Teorema de Bayes , Transcriptoma , Fenotipo , Fluorocarburos/toxicidad

RESUMEN

Open access new approach methods (NAM) in the US EPA ToxCast program and NTP Integrated Chemical Environment (ICE) were used to investigate activities of four neurotoxic pesticides: endosulfan, fipronil, propyzamide and carbaryl. Concordance of in vivo regulatory points of departure (POD) adjusted for interspecies extrapolation (AdjPOD) to modelled human Administered Equivalent Dose (AEDHuman) was assessed using 3-compartment or Adult/Fetal PBTK in vitro to in vivo extrapolation. Model inputs were from Tier 1 (High throughput transcriptomics: HTTr, high throughput phenotypic profiling: HTPP) and Tier 2 (single target: ToxCast) assays. HTTr identified gene expression signatures associated with potential neurotoxicity for endosulfan, propyzamide and carbaryl in non-neuronal MCF-7 and HepaRG cells. The HTPP assay in U-2 OS cells detected potent effects on DNA endpoints for endosulfan and carbaryl, and mitochondria with fipronil (propyzamide was inactive). The most potent ToxCast assays were concordant with specific components of each chemical mode of action (MOA). Predictive adult IVIVE models produced fold differences (FD) < 10 between the AEDHuman and the measured in vivo AdjPOD. The 3-compartment model was concordant (i.e., smallest FD) for endosulfan, fipronil and carbaryl, and PBTK was concordant for propyzamide. The most potent AEDHuman predictions for each chemical showed HTTr, HTPP and ToxCast were mainly concordant with in vivo AdjPODs but assays were less concordant with MOAs. This was likely due to the cell types used for testing and/or lack of metabolic capabilities and pathways available in vivo. The Fetal PBTK model had larger FDs than adult models and was less predictive overall.

RESUMEN

Read-across is a data gap filling technique utilized to predict the toxicity of a target chemical using data from similar analogues. Recent efforts such as Generalized Read-Across (GenRA) facilitate automated read-across predictions for untested chemicals. GenRA makes predictions of toxicity outcomes based on "neighboring" chemicals characterized by chemical and bioactivity fingerprints. Here we investigated the impact of biological similarities on neighborhood formation and read-across performance in predicting hazard (based on repeat-dose testing outcomes from US EPA ToxRefDB v2.0). We used targeted transcriptomic data on 93 genes for 1060 chemicals in HepaRG™ cells that measure nuclear receptor activation, xenobiotic metabolism, cellular stress, cell cycle progression, and apoptosis. Transcriptomic similarity between chemicals was calculated using binary hit-calls from concentration-response data for each gene. We evaluated GenRA performance in predicting ToxRefDB v2.0 hazard outcomes using the area under the Receiver Operating Characteristic (ROC) curve (AUC) for the baseline approach (chemical fingerprints) versus transcriptomic fingerprints and a combination of both (hybrid). For all endpoints, there were significant but only modest improvements in ROC AUC scores of 0.01 (2.1%) and 0.04 (7.3%) with transcriptomic and hybrid descriptors, respectively. However, for liver-specific toxicity endpoints, ROC AUC scores improved by 10% and 17% for transcriptomic and hybrid descriptors, respectively. Our findings suggest that using hybrid descriptors formed by combining chemical and targeted transcriptomic information can improve in vivo toxicity predictions in the right context.