New AI Model Predicts Molecular Toxicity Using 3D Geometry

Researchers have developed a new artificial intelligence framework named MolMMP to improve the prediction of molecular toxicity in early-stage drug development. Published in BMC Bioinformatics, the study introduces a method that integrates structure-aware cross-modal alignment with geometry-biased attention. The primary goal of this research is to address significant limitations in current toxicity prediction models, which often rely on single data modalities and fail to capture the full complexity of molecular structures. By combining sequence semantics with three-dimensional geometric relationships, the new model aims to provide a more robust tool for evaluating the safety of lead compounds before they enter costly clinical trials.

Drug toxicity prediction is a critical step in modern pharmaceutical development, serving as a vital mechanism for avoiding late-stage clinical failures and reducing both financial costs and risks. However, existing computational methods face notable challenges. Many current models depend primarily on a single modality, making it difficult to simultaneously capture the semantics of SMILES sequences and account for molecular topological structure. Furthermore, the absence of three-dimensional molecular conformation and interatomic geometric relationships prevents these models from capturing the critical influence of spatial structure on toxicity outcomes.

To overcome these obstacles, the researchers developed a multimodal approach that employs a bidirectional cross-attention mechanism. This mechanism is designed to deeply align SMILES sequence semantics with graph topological structure, thereby combining the complementary strengths of global semantic information and local structural features. In parallel, the framework includes a specialized three-dimensional module that encodes interatomic distances. This module utilizes a geometric transformation that leverages radial basis functions, effectively compensating for the inadequate representation of molecular conformation found in conventional approaches. By integrating these distinct data types, the model can better account for the spatial arrangement of atoms within a molecule.

A key innovation in the MolMMP framework is a hierarchical gating fusion strategy. This design dynamically balances fingerprint features, topological features, and geometric information, providing a flexible strategy for integrating heterogeneous molecular representations. The gating mechanism allows the model to weigh the importance of different data sources adaptively, which contributes to improved predictive performance. This approach addresses the challenges associated with multi-source feature fusion and imbalanced class distributions, areas where current models frequently exhibit insufficient predictive robustness.

The study evaluated the performance of MolMMP on multiple toxicity benchmarks under both random and scaffold-based data splits. The results indicate that the model achieves competitive predictive performance across these diverse testing conditions. Notably, the framework showed particularly notable gains on tasks that benefit from the joint modeling of molecular topology and three-dimensional geometry.

The proposed multimodal framework demonstrates robust and versatile performance across diverse toxicity endpoints.

Subject of Research: Development of a multimodal AI framework for molecular toxicity prediction using 3D geometry and cross-modal alignment.

Article Title: MolMMP: structure-aware cross-modal alignment with geometry-biased attention for molecular toxicity prediction

Article References: Hou, Z., Zhao, W., Zhai, H., Shen, J., & Luo, J. (2026). MolMMP: structure-aware cross-modal alignment with geometry-biased attention for molecular toxicity prediction. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06745-0

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06745-0

Keywords: Molecular Toxicity, Artificial Intelligence, Drug Discovery, 3D Geometry, Cross-Modal Alignment, BMC Bioinformatics, Machine Learning, Chemical Safety, MolMMP, structure-aware, cross-modal, alignment