MIN: Multi-Channel Interaction Network for Drug-Target Interaction With Protein Distillation

Shuqi Li, Shufang Xie, Hongda Sun, Yuhan Chen, Tao Qin, Tianjun Ke, Rui Yan

TL;DR

The Multi-channel Interaction Network (MIN) improves drug-target interaction (DTI) prediction by "distilling" massive protein sequences down to their most important functional residues. By combining evolutionary conservation scores with a multi-channel architecture that aligns sequence and structural data, MIN achieves state-of-the-art accuracy while providing a clear "map" of where drugs are likely to bind.

What Problem Does This Paper Address?

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. Furthermore, residues that are far apart in a linear sequence often end up side-by-side once the protein folds into its 3D shape. Existing models often struggle to align these 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.

Methods

The authors introduce MIN, which utilizes three core strategies to bridge the gap between drugs and proteins:

  1. Protein Distillation (C-Score Predictor): This is the paper'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.
  2. Multi-channel Interaction Module: Instead of a single prediction path, MIN uses three parallel channels:
  3. Structure-Agnostic: Compares 1D sequences (SMILES for drugs, amino acids for proteins).
  4. Structure-Aware: Compares 3D representations (Molecular graphs for drugs, contact maps for proteins).
  5. Extended-Mixture: Combines both to capture complex, cross-modal patterns.
  6. 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.

Datasets

The model was evaluated on two gold-standard benchmarks:

  • DUD-E: A diverse dataset containing 102 targets and over a million "decoys" (molecules that look like they should bind but don't), testing the model's ability to avoid false positives.
  • Human Dataset: A high-credibility dataset featuring 1,803 target proteins and 5,423 verified interactions.

Key Insights and Ideas

  • Evolution as a Filter: The most powerful insight is that evolutionary history (conservation) is a highly effective proxy for binding potential. By "distilling" the protein, the model effectively performs a pre-screening of the binding pocket before the heavy computation begins.
  • The Length Disparity Solution: By masking irrelevant residues, MIN solves the "length disparity" problem. It prevents the drug's signal from being drowned out by the protein's massive, non-functional sequence.
  • Explainability by Design: Unlike "black box" models, MIN’s C-Score Predictor provides a biological rationale for its decisions. If the model predicts an interaction, researchers can look at which residues it "kept" to see if they align with known binding pockets.

Key Results

  • 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 at Low False Positive Rates: In drug discovery, you want to be very sure about your hits. MIN showed significant gains in ROC Enrichment (RE), meaning it identifies true positives much more effectively than other models when the "budget" for false positives is very tight (e.g., at a 0.5% FPR).
  • 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 indeed the "secret sauce" of the architecture.
  • Biological Validation: In a case study of the protein 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.

Limitations and Future Work

While MIN is powerful, the authors note that conservation is a strong indicator but not the only factor in binding. Some binding sites may involve less-conserved, highly specific residues.

Future directions include: * Integrating more granular 3D geometric information beyond simple contact maps. * Testing the model on even larger, multi-protein complexes where the "noise" problem is even more severe. * Exploring how the C-Score Predictor might be used as a standalone tool for de novo binding site discovery.

Summary updated

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.
BibTeX
@article{li2025min,
  author = {Shuqi Li and Shufang Xie and Hongda Sun and Yuhan Chen and Tao Qin and Tianjun Ke and Rui Yan},
  title = {{MIN: Multi-Channel Interaction Network for Drug-Target Interaction With Protein Distillation}},
  journal = {IEEE Trans. Comput. Biol. Bioinform.},
  year = {2025},
  doi = {10.1109/tcbbio.2025.3549588},
}

Loading related content...

Figure