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

Drug toxicity prediction is a core step in evaluating the safety of lead compounds in modern drug development, playing a vital role in avoiding late-stage clinical failures and reducing both cost and risk. However, existing methods still face notable limitations. Many models rely primarily on a single modality, making it difficult to capture SMILES sequence semantics while simultaneously accounting for molecular topological structure. This narrow focus restricts the completeness of feature representation. Moreover, the absence of three-dimensional molecular conformation and interatomic geometric relationships prevents models from capturing the critical influence of spatial structure on toxicity. When confronted with multi-source feature fusion and imbalanced class distributions, current models frequently exhibit insufficient predictive robustness. To address these challenges, we developed a toxicity prediction method that integrates structure-aware cross-modal alignment with geometry-biased attention. The method employs a bidirectional cross-attention mechanism 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, a 3D module encodes interatomic distances through a geometric transformation module that leverages radial basis functions, effectively compensating for the inadequate representation of molecular conformation found in conventional approaches. We designed a hierarchical gating fusion strategy that dynamically balances fingerprint features, topological features, and geometric information. This design provides a flexible strategy for integrating heterogeneous molecular representations and contributes to improved predictive performance. MolMMP achieves competitive predictive performance on multiple toxicity benchmarks under both random and scaffold-based data splits, with particularly notable gains on tasks that benefit from joint modeling of molecular topology and 3D geometry. The proposed multimodal framework shows robust and versatile performance across diverse toxicity endpoints, offering a useful tool for early-stage drug toxicity assessment.