Machine learning modeling of zebrafish toxicity endpoints after exposure to PROTACs

Zebrafish (Danio rerio) are an ideal system for understanding developmental toxicity as they display similar toxicity outcomes to other vertebrates. Further, many molecules have been tested for developmental toxicity in zebrafish, providing an opportunity for machine learning model development. We curated 1,345 small molecules from ToxCast, flame retardant compounds, per- and polyfluoroalkyl substances (PFAS), and industrial chemicals published by the Superfund Research Program (SRP). Following curation, we trained machine learning models on the zebrafish toxicity endpoints ANY_ = any effect including mortality, ANY_BUT_MORT = any effect excluding mortality, MORT = mortality, i.e. did the embryo die, EDEM = did an edema form, and CRAN = Craniofacial malformation. We demonstrated that these models were better than random when compared with shuffled data. We also fine-tuned the molecular SMILES encoder MolBART to predict on all zebrafish toxicity endpoints and found it generally matched the performance of classical machine learning models for ANY_BUT_MORT, CRAN, and EDEM endpoints. We present new toxicity data for Proteolysis Targeting Chimeras (PROTACs) in zebrafish and machine learning models for these data by fingerprinting different parts of the molecule individually, yielding predictive performance (AUROC 0.6 to 0.7). If we are to reduce animal testing with new approach methodologies like these zebrafish toxicity models they need to be able adapt to new molecular classes like PROTACs.