An exploration of the use of hybrid fingerprints in Generalized Read-Across and their impact on predictive performance for selected in vivo toxicity outcomes

Abstract

Read-across is a cost-efficient means of generating information for hazard assessment. Approaches such as Generalized Read-Across (GenRA) facilitate objective and reproducible read-across for untested substances. GenRA is a web application, and its prediction engine is also available as a python package (genra-py). Recent updates permit source analogues to be identified using ‘hybrid’ fingerprints, i.e. analogues identified based on more than one type of similarity measure. Herein, the performance of hybrid fingerprints relative to Morgan chemical fingerprints was evaluated for a selection of acute and chronic in vivo toxicity outcomes. Grid search and cross-validation on a dataset of 5,830 chemicals with rodent acute oral toxicity (LD₅₀) values were used to tune the hybrid weight hyperparameter for up to four chemical fingerprints (Morgan, Torsion, ToxPrint and Analog Identification Methodology (AIM)). The optimal hybrid fingerprint derived (52.12% Morgan, 23.40% ToxPrint, 12.44% AIM, 12.04% Torsion) outperformed Morgan fingerprints across all 10 folds of a cross-validation procedure (mean test set coefficient of determination (R²) 0.517 (Morgan) vs. 0.557 (hybrid)). The hybrid fingerprint was then used to make toxicity predictions for 2 other datasets, a set of 3,266 chemicals with oral chronic human equivalent benchmark dose values (mean test set R² 0.445 vs. 0.417 for Morgan) and a set of 9,443 chemicals with acute mammalian oral hazard classifications (mean balanced accuracy (BA) 0.577 vs 0.553 for Morgan). Overall, performance improved when using the hybrid fingerprint tuned for the acute toxicity dataset. Using the custom hybrid option in GenRA results in improved read-across predictions relative to current defaults.