Abstract:The inherent toxicity of compounds poses a significant threat to human health and the environment throughout their entire lifecycle, encompassing production, use, storage, and disposal. Compound-induced toxicity evaluation is a critical concern across various fields, including drug discovery and environmental studies. Recent advances in deep learning hold promise for predicting compound-induced toxicities. However, existing works often exhibit poor predictive accuracy and generalizability, especially for rare toxicities with limited data. Most works fail to capture the three-dimensional (3D) spatial arrangement and stereochemical properties of compounds, which are crucial for understanding their toxicological profiles. And the interrelated nature of various toxicities has often been overlooked. Here we propose ToxScan, a novel SE(3)-equivariant multiscale model, as a universal toxicity prediction framework to address these issues. A two-level representation learning protocol, encompassing molecular- and atomic-level information, is introduced to better incorporate the geometry information of 3D conformation. A parallel multiscale modelling and a multi-task learning scheme is applied to learn the universal toxicological characteristics of multiple toxicity categories. Through comprehensive analysis in multiple scenarios such as multi-task toxicity prediction, module effectiveness testing, generalization testing on new small-scale endpoints, distinguishing ability on structurally similar compounds with opposing toxicities, in vitro experimental validation of new predictions, and testing on practical application data, ToxScan demonstrates convincing predictive capability and generalizability, surpassing the state-of-the-art by a remarkable 7.8–37.6% performance boost on four metrics for medium- and small-scale endpoints. Typical case studies demonstrate that ToxScan can identify subtle structural variations by capturing molecular structural features at multiple levels. To facilitate user accessibility, we have integrated the trained ToxScan and developed a user-friendly online prediction platform (https://app.bohrium.dp.tech/toxscan), allowing researchers to obtain multiple toxicity endpoint predictions for new compounds quickly and conveniently. We envision ToxScan as a valuable tool for compound-induced toxicity evaluation, effectively reducing analysis cycles and costs while providing practical utility for researchers in drug discovery, toxicology and environmental studies.

Mining Toxicity Information from Large Amounts of Toxicity Data

Hi-MGT: A hybrid molecule graph transformer for toxicity identification

Using Support Vector Regression Coupled with the Genetic Algorithm for Predicting Acute Toxicity to the Fathead Minnow

MTGL-ADMET: A Novel Multi-task Graph Learning Framework for ADMET Prediction Enhanced by Status-Theory and Maximum Flow.

Multimodal Representation Learning via Graph Isomorphism Network for Toxicity Multitask Learning

Multitask CapsNet: an Imbalanced Data Deep Learning Method for Predicting Toxicants

A Multiscale-Information-Embedded Universal Toxicity Prediction Framework

The Prediction of Molecular Toxicity Based on BiGRU and GraphSAGE

Co‐Model for Chemical Toxicity Prediction Based on Multi‐task Deep Learning

Deep Learning Based Regression and Multi-class Models for Acute Oral Toxicity Prediction with Automatic Chemical Feature Extraction

A Deep Learning Based Multi-Model Approach for Predicting Drug-Like Chemical Compound’s Toxicity

A deep learning based multi-model approach for predicting drug-like chemical compound's toxicity

Drug Toxicity Prediction by Machine Learning Approaches

GraphADT: empowering interpretable predictions of acute dermal toxicity with multi-view graph pooling and structure remapping

Deep Learning Based Regression and Multiclass Models for Acute Oral Toxicity Prediction with Automatic Chemical Feature Extraction.

MolFPG: Multi-level Fingerprint-Based Graph Transformer for Accurate and Robust Drug Toxicity Prediction.

ToxMPNN: A Deep Learning Model for Small Molecule Toxicity Prediction.

ADMET Evaluation in Drug Discovery. 18. Reliable Prediction of Chemical-Induced Urinary Tract Toxicity by Boosting Machine Learning Approaches

Computer‐aided Prediction of Toxicity with Substructure Pattern and Random Forest

Graph Kernel Learning for Predictive Toxicity Models

Integrated Transfer Learning and Multitask Learning Strategies to Construct Graph Neural Network Models for Predicting Bioaccumulation Parameters of Chemicals