<p>Cross-dataset transfer learning results.</p>

Drug-drug interactions (DDI) represent a significant clinical challenge in modern healthcare, contributing to over 125,000 deaths annually in the United States alone. Current computational approaches face substantial limitations in capturing long-range molecular dependencies and generalizing to novel drug combinations. Traditional Graph Neural Networks (GNNs) suffer from over-smoothing and locality bias, while sequence-based methods fail to adequately represent three-dimensional molecular structures. To address these limitations, we propose Graph Former-CL, a novel deep learning framework that synergistically combines Graph Transformer architecture with contrastive learning for DDI prediction. Our approach features four key innovations: (1) a hierarchical Graph Transformer with position-aware multi- head self-attention to capture both local and global molecular patterns, (2) a domain-specific contrastive learning module with molecular augmentation strategies, (3) a cross-modal fusion mechanism integrating SMILES sequences with graph representations, and (4) an adaptive pooling strategy for multi-scale molecular representation. Comprehensive evaluation on four benchmark datasets demonstrates superior performance, with Graph Former-CL achieving 98.2% accuracy on DrugBank and 89.4% on TWOSIDES, both representing statistically significant improvements (p < 0.001) over state-of-the-art methods. Notably, the framework achieves 85.6% accuracy for novel drugs in inductive settings, demonstrating robust generalization capabilities essential for real-world clinical applications.药物相互作用（Drug-drug interactions, DDI）是现代医疗领域的重大临床挑战，仅在美国境内每年就造成超12.5万例死亡。现有计算方法在捕捉长程分子依赖关系以及泛化至新型药物组合方面存在显著局限。传统图神经网络（Graph Neural Networks, GNNs）存在过平滑及局部性偏差问题，而基于序列的方法无法充分表征三维分子结构。为解决上述局限，我们提出图Former-CL（Graph Former-CL）这一新型深度学习框架，该框架将图Transformer（Graph Transformer）架构与对比学习方法协同结合，用于药物相互作用预测。本方法具备四项核心创新：（1）引入带位置感知多头自注意力的层级图Transformer，以同时捕获局部与全局分子模式；（2）配备结合分子增强策略的领域专属对比学习模块；（3）设计跨模态融合机制，将简化分子线性输入规范（SMILES）序列与图表征进行融合；（4）提出面向多尺度分子表征的自适应池化策略。在四个基准数据集上的全面评估表明，本框架性能优异：在DrugBank数据集上准确率达98.2%，在TWOSIDES数据集上准确率达89.4%，相较于当前最优方法均实现了统计学意义上的显著提升（p < 0.001）。值得注意的是，该框架在归纳学习设置下针对新型药物的预测准确率可达85.6%，展现出实际临床应用所需的强大泛化能力。