DART Predictor: A Multi-Label Attention Model for High-Throughput Screening of Chemicals with Developmental and Reproductive Toxicity (DART)

Chemicals
with developmental and reproductive toxicity (DART) pose
significant risks to human health, particularly exposure during critical
windows of embryonic and fetal development. Therefore, rapid and accurate
identification of DART chemicals is urgently needed. Existing predictive
models are predominantly limited to binary classification and lack
explicit integration of exposure information, hindering the precise
risk extrapolation across realistic exposure scenarios. Herein, we
present DART Predictor, a multilabel deep learning model trained using
a label-aware attention mechanism to predict six DART outcomes (Growth
Disorders, Malformation, Fetal Viability Loss, Maternal Systemic Toxicity,
Maternal Pathology, and Fertility Impairment). Trained on 25,175 chemically
diverse records integrating molecular descriptors and bioassay activity
features calibrated with exposure parameters, DART Predictor achieves
state-of-the-art performance (average AUC: 0.964, average recall:
0.923) and strong interpretability and generalizability (AUC: 0.889,
recall: 0.959) on two external validation data sets. The exposure
parameters enhance model performance by up to 8.6% gain of AUC across
multiple DART outcomes, indicating the vital role of realistic exposure
information for model improvement. DART Predictor is further deployed
into a cloud platform (http://www.ai4environ.cn/dartpredictor) to provide high-throughput screening service. Our study provides
a novel framework for exposure-informed DART risk assessment, advancing
the development of DART-related new approach methodologies.