Abstract
Traditional drug discovery processes are both time-consuming and require extensive professional expertise. With the accumulation of drug-target interaction (DTI) data from experimental studies, leveraging modern machine-learning techniques to discern patterns between drugs and target proteins has become increasingly feasible. In this paper, we introduce the Multi-channel Interaction Network (MIN), a novel framework designed to predict DTIs through two primary components: a representation learning module and a multi-channel interaction module. The representation learning module features a C-Score Predictor-assisted screening mechanism, which selects critical residues to enhance prediction accuracy and reduce noise. The multi-channel interaction module incorporates a structure-agnostic channel, a structure-aware channel, and an extended-mixture channel, facilitating the identification of interaction patterns at various levels for optimal complementarity. Additionally, contrastive learning is utilized to harmonize the representations of diverse data types. Our experimental evaluations on public datasets demonstrate that MIN surpasses other strong DTI prediction methods. Furthermore, the case study reveals a high overlap between the residues selected by the C-Score Predictor and those in actual binding pockets, underscoring MIN's explainability capability. These findings affirm that MIN is not only a potent tool for DTI prediction but also offers fresh insights into the prediction of protein binding sites.
Introduction
Predicting whether a drug candidate will bind to a specific protein (DTI) is a cornerstone of modern drug discovery. However, computational models face a massive "needle in a haystack" problem. On average, a protein is significantly larger than a drug molecule (roughly 320 residues vs. 56 atoms). Most of a protein's sequence is "noise" regarding a specific interaction; only a small subset of residues—the binding pocket—actually interacts with the drug.
Existing models often struggle to align disparate data types (1D sequences vs. 3D structures) and fail to filter out the irrelevant parts of the protein, leading to lower accuracy and poor explainability. The Multi-channel Interaction Network (MIN) addresses these challenges by "distilling" protein sequences down to their most important functional residues.
Methodology
MIN utilizes three core strategies to bridge the gap between drugs and proteins:
1. Protein Distillation (C-Score Predictor)
This is the framework's standout feature. The model uses Conservation Scores (C-Scores) derived from Multiple Sequence Alignment (MSA). Evolution tends to preserve residues that are functionally critical. MIN trains a predictor to identify these conserved residues and "masks" (filters out) the unimportant ones. This allows the model to focus exclusively on the potential binding sites, effectively solving the "length disparity" problem.
2. Multi-channel Interaction Module
Instead of a single prediction path, MIN uses three parallel channels to capture interaction patterns at various levels:
- Structure-Agnostic: Compares 1D sequences (SMILES for drugs, amino acids for proteins).
- Structure-Aware: Compares 3D representations (Molecular graphs for drugs, contact maps for proteins).
- Extended-Mixture: Combines both to capture complex, cross-modal patterns.
3. Contrastive Learning
To ensure the model understands that a protein's sequence and its 3D structure represent the same entity, the authors use contrastive loss. This aligns the different representations in a shared latent space, making the final interaction prediction more robust.
Experimental Results
The model was evaluated on two gold-standard benchmarks: DUD-E (102 targets, >1M decoys) and a Human Dataset (1,803 targets, 5,423 interactions).
- State-of-the-Art Performance: MIN outperformed strong baselines (like DrugVQA and HyperPCM) on both datasets. On DUD-E, it achieved an AUC of 0.983.
- High Sensitivity: MIN showed significant gains in ROC Enrichment (RE), identifying true positives much more effectively than other models at very tight false-positive rate budgets.
- Ablation Success: Removing the C-Score Predictor caused the most significant drop in performance (AUC fell from 0.983 to 0.968), proving that distillation is the "secret sauce" of the architecture.
- Biological Validation: In a case study of Mitogen-activated protein kinase kinase 5, the model successfully distilled a 1,608-residue sequence down to 588 critical residues, leading to a high-confidence (0.999) correct prediction.
Conclusion
MIN demonstrates that evolutionary history (conservation) is a highly effective proxy for binding potential. By distilling the protein, the model performs a pre-screening of the binding pocket before heavy computation begins. This design not only achieves state-of-the-art accuracy but also provides a biological rationale for its decisions, offering researchers a clear map of where drugs are likely to bind.
Summary updated