Artificial Intelligence Driven Deep Learning Framework Integrating Molecular Descriptors and Multiview Images for Bioactivity Prediction

Accurately predicting chemical bioactivity across large and structurally diverse chemical spaces remains a central challenge in computational toxicology and molecular design. This work presents a unified framework that systematically benchmarks three complementary modeling strategies for predicting PercentToxCast, an aggregate bioactivity index derived from the U.S. EPA ToxCast program. We integrate twenty seven physicochemical and topological descriptors with standardized three dimensional molecular renderings generated from RDKit conformers and evaluate descriptor based learning, multiview image based deep learning, and multimodal fusion of numerical and image derived features. Across 7,211 compounds, gradient boosted descriptor models deliver the strongest generalization, with CatBoost achieving a hold out R 2 of 0.398 and providing interpretable feature attributions that highlight hydrophobicity, polarity, and electronic structure as dominant contributors to broad assay engagement. A six view convolutional neural network captures geometric and steric information but attains lower external performance (R 2 of 0.217), indicating that image derived structure alone does not fully encode the physicochemical determinants of bioactivity. Multimodal fusion produces extremely strong cross validated performance but yields only modest hold out gains (R 2 of 0.241 to 0.259), revealing overfitting challenges associated with the high dimensional fused representation. This study establishes one of the first rigorously controlled evaluations of descriptor, image, and fused feature spaces for broad bioactivity modeling and demonstrates that descriptors remain the most informative single modality while image based geometry provides complementary but underexploited structural cues. By unifying these approaches within a transparent and reproducible benchmark, this work defines clear performance baselines and identifies key methodological limitations that must be addressed for next generation multimodal architectures. These findings provide a foundation for chemically intelligent models that integrate three dimensional structure with physicochemical properties to advance predictive toxicology, risk assessment, and data driven molecular design.