DSA-DeepFM: a dual-stage attention-enhanced DeepFM model for predicting anticancer synergistic drug combinations

Drug combinations are crucial in combating drug resistance, reducing toxicity, and improving therapeutic outcomes in disease management. Because a large number of drugs are available, the potential combinations increase exponentially, making it impractical to rely solely on biological experiments to identify synergistic combinations. Consequently, machine learning methods are increasingly being used to find synergistic drug combinations. Most existing methods focus on predictive performance through auxiliary data or complex models, but neglecting underlying biological mechanisms limits their accuracy in predicting synergistic drug combinations.

We present DSA-DeepFM, a deep learning model that integrates a dual-stage attention (DSA) mechanism with Factorization Machines (FMs) to predict synergistic two-drug combinations by addressing complex biological feature interactions. The model incorporates categorical and auxiliary numerical inputs to capture both field-aware and embedding-aware patterns. These patterns are then processed by a deep FM module, which captures low- and high-order feature interactions before making the final predictions. Validation testing demonstrates that DSA-DeepFM significantly outperforms traditional machine learning and state-of-the-art deep learning models. Furthermore, t-SNE visualizations confirm the discriminative power of the model at various stages. Additionally, we use our model to identify eight novel synergistic drug combinations, underscoring its practical utility and potential for future applications.

Source code is available at https://github.com/gracygyx/DSA-DeepFM.