Accurate prediction of in vivo toxicity end points is a pivotal challenge in drug discovery, primarily constrained by data scarcity and sharp activity cliffs. Existing knowledge transfer approaches often fail because they rely heavily on implicit semantic representations while neglecting the explicit geometric constraints required to distinguish structurally similar but toxically distinct compounds. To bridge this gap, we present META-Tox, a Multi-view Ensemble framework leveraging Topological Aggregation to explicitly augment LLM-derived semantics with comprehensive structural constraints (spanning 2D graphs and 3D conformations). Specifically, META-Tox employs a tiered meta-learning architecture with adaptive feature selection to rigorously filter redundant noise and fuse heterogeneous structural insights. Notably, our topological aggregation strategy amplifies the performance gains of standard LLM fine-tuning by nearly 3-fold (5.4% vs 1.9% gain over the baseline). Furthermore, qualitative analysis via case studies demonstrates the framework's potential to navigate activity cliffs by capturing subtle structural nuances. Ultimately, extensive evaluations confirm that META-Tox establishes a new state-of-the-art benchmark on independent external data sets with an AUC of 0.772, offering a robust solution for minimizing safety attrition in early drug development.
Liu et al. (Thu,) studied this question.