META-Tox: Multiview Ensemble with Topological Aggregation for Robust In Vivo Toxicity Prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accurate prediction of *in vivo* toxicity is crucial but limited by data scarcity and activity cliffs.
• Current methods often fail by relying on semantic representations and ignoring geometric constraints.

Purpose of the Study:

• To develop a novel framework, META-Tox, that integrates semantic and structural information for improved toxicity prediction.
• To address the challenge of distinguishing structurally similar yet toxicologically different compounds.

Main Methods:

• Developed META-Tox, a Multi-view Ensemble framework leveraging Topological Aggregation.
• Employed a tiered meta-learning architecture with adaptive feature selection.
• Fused heterogeneous structural insights from 2D graphs and 3D conformations.

Main Results:

• META-Tox improved performance gains by nearly 3-fold compared to standard LLM fine-tuning (5.4% vs 1.9%).
• Demonstrated ability to navigate activity cliffs by capturing subtle structural nuances through case studies.
• Established a new state-of-the-art benchmark with an AUC of 0.772 on independent external datasets.

Conclusions:

• META-Tox offers a robust solution for minimizing safety attrition in early drug development.
• The framework successfully augments LLM semantics with explicit structural constraints.
• Highlights the importance of integrating diverse data modalities for accurate toxicity prediction.