Graph attention-guided multi-modal learning for high-throughput screening of environmental chemicals associated with breast cancer risk.

The rapid expansion of chemical production and use has increased human exposure to environmental chemicals that can impact biological pathways relevant to mammary carcinogenesis. This trend creates an urgent need for scalable screening tools that can identify chemicals of potential concern before extensive animal testing or widespread exposure occurs. Conventional predictive models built solely on molecular structures do not directly capture complementary bioassay and receptor-interaction information, which can limit toxicological interpretation. Here, we introduce BCPredictor, a multi-modal deep learning framework that integrates a graph attention network (GAT) and a fully connected neural network (FCNN) for prioritizing chemicals according to rodent mammary-tumor-related hazard evidence. The model integrates molecular graph representations, structural descriptors, breast cancer-relevant bioassay variables, and estrogen receptor/progesterone receptor docking scores within a unified architecture. Trained on a curated set of 897 compounds, BCPredictor achieves an area under the curve (AUC) of 0.928 and a recall of 0.892, and was further evaluated using an independently assembled external benchmark. The incorporation of biologically and mechanistically anchored features improves AUC by up to 7.8%. BCPredictor is accompanied by a cloud-based research interface (http://www.ai4environ.cn/bcpredictor) for single-compound and batch prioritization. This study provides a mechanism informed computational toxicology framework for breast cancer hazard prioritization, supports next generation new approach methodologies (NAMs), and helps prioritize chemicals that warrant further experimental or regulatory evaluation.