CDI-DTI: A Strong Cross-domain Interpretable Drug-Target Interaction Prediction Framework Based on Multi-Strategy Fusion

TL;DR

CDI-DTI addresses cross-domain generalization and cold-start challenges in drug-target interaction prediction by fusing textual, structural, and functional modalities through a four-stage pipeline leveraging Gram Loss alignment, multi-source cross-attention, bidirectional cross-attention, and deep orthogonal fusion. The approach yields state-of-the-art results on BindingDB and DAVIS, including cross-domain transfer and unseen drug/target scenarios, while providing interpretable interaction maps aligned with docking evidence. Ablation and visualization studies confirm that multi-modal fusion and multi-stage fusion are crucial for robust, fine-grained DTI modeling. This work offers a practical, explainable framework with potential utility in drug discovery, resistance analysis, and precision medicine.

Abstract

Accurate prediction of drug-target interactions (DTI) is pivotal for drug discovery, yet existing methods often fail to address challenges like cross-domain generalization, cold-start prediction, and interpretability. In this work, we propose CDI-DTI, a novel cross-domain interpretable framework for DTI prediction, designed to overcome these limitations. By integrating multi-modal features-textual, structural, and functional-through a multi-strategy fusion approach, CDI-DTI ensures robust performance across different domains and in cold-start scenarios. A multi-source cross-attention mechanism is introduced to align and fuse features early, while a bidirectional cross-attention layer captures fine-grained intra-modal drug-target interactions. To enhance model interpretability, we incorporate Gram Loss for feature alignment and a deep orthogonal fusion module to eliminate redundancy. Experimental results on several benchmark datasets demonstrate that CDI-DTI significantly outperforms existing methods, particularly in cross-domain and cold-start tasks, while maintaining high interpretability for practical applications in drug-target interaction prediction.

Figures (6)

• Figure 1: Multi-modal Feature Construction. (a) Textual features are extracted from drug SMILES and protein sequences using ChemBERTa and ESM2, respectively. (b) Structural features are derived from drug molecular graphs constructed via RDKit and protein structure graphs obtained from AlphaFold. (c) Functional features are generated from drug descriptions produced by MolT5 and protein GO annotations predicted by DeepGO, both encoded using BioBERT.
• Figure 2: Overview of the CDI-DTI Framework. (a) Multi-modal feature extraction stage: Textual features are refined using a multi-head self-attention module; structural features are enhanced via a hybrid-pooling graph attention network; functional features are processed similarly to textual features. (b) Early fusion stage: Modal features of drugs and proteins are aligned using Gram Loss, followed by integration through a multi-source cross-attention module. (c) Late fusion stage: A bidirectional cross-attention module captures fine-grained drug-target interactions, and a deep orthogonal fusion module reduces redundancy in fused features. (d) Classification head: The final prediction is made using an MLP classifier based on the fused drug-target representation.
• Figure 3: Deep Orthogonal Fusion (DOF) Module schematic diagram. This module constructs differential representations between modalities using linear transformations and orthogonal projections, followed by a masking mechanism to filter redundant features and enhance discriminative fused representations.
• Figure 4: The CDI-DTI achieves accurate prediction of drug-target interactions. Below are the experimental results on BindingDB and Divas datasets. (a) Sector plots show the performance comparison of CDI-DTI with other five state-of-the-art methods (b) bar plots show the experimental results of ablation for each modality of CDI-DTI (c) 3D line plots show the performance comparison of CDI-DTI with other five state-of-the-art methods in the cold start experiment.
• Figure 5: Visualization of Different Modal Features and Fusion Representations at Each Stage. (a) Text Modality Features. (b) Graph Structural Modality Features. (c) Functional Modality Features. (d) Early Fusion Features. (e) Late Fusion Features. (f) Final Output Features.