PaperlayerView exact revision on arXiv

Source-linked AI summary

ProbeMatchDTI: Probe-Driven Multi-Scale Biochemical Pattern Matching for Drug-Target Interaction Prediction

Quan Hao, Mengyue Fan, Zifan Dong, Youru Li, Jianduo Zhao, Lechuan Xu, Hao Zhang, Fei Xia, Jigang Wang, Chong Qiu, Liguo Zhang

arXiv:2609.02549v1cs.LGcs.AI

TL;DR

DTI prediction must capture complementary drug–protein patterns across biochemical scales, but passive aggregation can suppress weak binding-relevant signals. ProbeMatchDTI uses IterProbe and BindingProbe to retain and actively probe multi-scale patterns and correspondences, achieving the strongest reported benchmark performance while supporting downstream candidate refinement and validation planning.

  • Problem

    Passive feature transformation and aggregation can overwhelm weak binding-relevant patterns, limiting explicit modeling of multi-scale drug–target correspondences.

  • Method

    ProbeMatchDTI combines IterProbe, which retains and selects contextual states across refinement depths, with BindingProbe, which models atom–residue correspondence and local-to-global complementarity.

  • Results

    ProbeMatchDTI achieves the strongest AUC-ROC and AUC-PR across four public benchmarks, with both probes showing complementary effects in ablation studies.

  • Takeaways & Limitations

    The framework’s predictions support evidence-guided candidate refinement, validation planning, and structured downstream drug-discovery workflows.

  • Takeaways & Limitations

    The study uses public datasets and resources that may inherit biases and incompleteness, while real-world drug-discovery applications remain future work.

Abstract

from arXiv · show

Drug-target interaction (DTI) prediction is an important task in AI-driven drug discovery. Although recent biochemical representation learning methods have improved DTI prediction, their passive feature aggregation tends to favor dominant molecular patterns while suppressing weak yet binding-relevant signals, such as functional groups and residue-context patterns, limiting the modeling of multi-scale biochemical correspondences. To address this issue, we propose ProbeMatchDTI, a pattern-probe-driven framework comprising IterProbe and BindingProbe. IterProbe explicitly retains contextual states across refinement depths and uses learnable probes to select them at each position before cross-entity matching, thereby preserving weak biochemical patterns and strengthening associations among functional groups, local motifs, and molecular scaffolds. BindingProbe then characterizes cross-entity drug-protein complementarity at local biochemical-unit and whole-pair levels, jointly modeling fine-grained interactions and multi-scale correspondences while preserving weaker binding-relevant associations. Extensive experiments demonstrate the superiority of ProbeMatchDTI, achieving 2.0% and 0.5% higher AUC-ROC on BindingDB and DrugBank, respectively. Feature-level pattern analyses further characterize its probe-driven behavior in cross-scale biochemical pattern matching. We further connect ProbeMatchDTI predictions with an evidence-guided downstream drug-discovery workflow, demonstrating their utility for candidate refinement and validation planning. Our code is available at https://github.com/developer-hq/ProbeMatchDTI

1 Introduction

ProbeMatchDTI targets a gap in DTI modeling: passive aggregation can suppress weak binding-relevant patterns, so the framework probes biochemical representations before and during cross-entity matching. Its experiments and analyses evaluate predictive performance, component complementarity, and downstream use for candidate refinement and validation planning.

  • DTI prediction supports virtual screening, candidate prioritization, and drug discovery, but accurate prediction requires complementary drug–protein patterns across biochemical scales.
  • Local binding involves drug functional groups and motifs matching residue patterns within broader chemical and sequence contexts.The puerarin–NOS3 example includes hydrogen bonds with ASN338, TRP356, and GLU361 and van der Waals stabilization.
  • Passive hierarchical aggregation may let dominant scaffold or sequence-environment patterns overshadow weaker functional-group and residue-context signals.
  • IterProbe retains initial, intermediate, and current contextual states and learns position- and sample-dependent weights before cross-entity matching.It is designed to keep functional-group, local-motif, residue-context, and related binding patterns accessible across refinement depths.
  • Experiments across four public benchmarks report strongest ROC and PR performance, with ablations, feature-level analyses, mutation perturbations, and a downstream case study examining the model’s behavior and utility.
  • BindingProbe models bidirectional atom–residue correspondence in local environments and probes pair evidence for local-to-global biochemical complementarity.The framework uses active probing at both entity and pair levels rather than relying only on passive final aggregation.

2 Related Work

Related work has expanded DTI representations from handcrafted descriptors to sequence, molecular-string, graph, and interaction-oriented models. However, existing approaches generally transform and match final representations, while AI evidence systems address downstream analysis separately from biochemical pattern matching.

  • DTI representation learning progressed from handcrafted descriptors to deep sequence-, molecular-string-, graph-, and multimodal representations.
  • Passive feature transformation and aggregation can allow dominant scaffold or global sequence patterns to overwhelm weaker functional groups, local motifs, and residue environments.
  • Interaction-oriented methods establish drug–target correspondences, but generally match final representations instead of retaining and probing intermediate contextual states during refinement.
  • Evidence-oriented AI systems organize heterogeneous biomedical knowledge, assist candidate prioritization, and facilitate computational experiment planning.
  • These downstream evidence systems complement DTI models while remaining separate from biochemical representation enhancement and drug–target pattern matching.

3 Method

ProbeMatchDTI combines IterProbe and BindingProbe to preserve entity-specific biochemical patterns and match drug–protein evidence across microscopic and macroscopic scales. IterProbe selects contextual states before matching, while BindingProbe integrates atom–residue correspondence with broader molecular and protein context.

  • Entity-Specific Representation Enhancement with IterProbe: IterProbe retains contextual states across refinement depths and positionwise selects among them before cross-entity matching.This keeps functional-group, local-motif, and residue-context patterns accessible instead of exposing only terminal representations.
  • Entity-Specific Representation Enhancement with IterProbe: Each refinement step expands the biochemical environment around the same drug-token or protein-residue position, from localized patterns toward broader scaffold or sequence context.Drug positions incorporate surrounding SMILES patterns, while protein positions incorporate sequence-derived biochemical environments.
  • Entity-Specific Representation Enhancement with IterProbe: IterProbe assigns sample- and position-specific relevance weights to contextual candidates without fixing refinement depths to biochemical categories.The same depth can contribute differently for functional groups, scaffold-associated tokens, isolated residue patterns, or residues in broader sequence environments.
  • Cross-Scale Complementarity Modeling with BindingProbe: BindingProbe first establishes counterpart-conditioned atom–residue correspondences while preserving drug covalent neighborhoods and protein sequence-local environments.Its reciprocal updates refine functional-group and residue contexts without collapsing their asymmetric environments into one undirected score.
  • Cross-Scale Complementarity Modeling with BindingProbe: BindingProbe then probes semantic and structure-derived pair summaries, weighting the two evidence sources for each prediction.The resulting pair representation combines local atom–residue correspondence with broader drug-scaffold and sequence-derived protein context.

4 Experiments

Experiments evaluate ProbeMatchDTI on public benchmarks, ablations, feature-level analyses, mutation sensitivity, and a puerarin downstream case study. The method achieves strong benchmark performance, while analyses connect its probe behavior to biochemical patterns and candidate refinement.

  • Public benchmark comparison: ProbeMatchDTI achieves the highest AUC-ROC and AUC-PR on all four public benchmarks.On BindingDB, it reaches 0.991 AUC-ROC and 0.992 AUC-PR; on DrugBank, it reaches 0.906 for both metrics.
  • Public benchmark comparison: 2.0 percentage points: BindingDB AUC-ROC rises from MMDG-DTI’s 0.971 to 0.991, while DrugBank rises from 0.901 to 0.906.The corresponding AUC-PR values are 0.972 versus 0.992 on BindingDB and 0.902 versus 0.906 on DrugBank.
  • Ablation study: IterProbe improves C. elegans AUC-ROC from 0.992 to 0.995, while BindingProbe improves BindingDB AUC-ROC from 0.971 to 0.988.BindingProbe alone reduces DrugBank AUC-ROC/AUC-PR from 0.897/0.897 to 0.872/0.872, but the combined model reaches 0.906/0.906.
  • Ablation study: Combining IterProbe and BindingProbe produces the best AUC-ROC and AUC-PR on all datasets, including 0.991/0.992 on BindingDB and 0.906/0.906 on DrugBank.C. elegans Precision also increases from 0.979 to 0.982 when both probes are combined.
  • Feature-level analysis: Feature attribution links a BindingProbe hotspot to NOS3 Glu361 and shows IterProbe strengthening the corresponding puerarin functional-group and protein-context patterns.The analysis traces learned features to SMILES tokens and residues using an MD-identified contact as an external anchor.
  • Feature-level analysis: 6, 11, and 29 residues: predicted position P367 deviates from MD contacts Glu361, Trp356, and Asn338 by those amounts.These correspond to 0.59%, 1.07%, and 2.83% of the full-length NOS3 sequence.
  • Downstream case study: DTI-Agent filters 500 candidates to 291, retains 192 of 239 known targets, raises precision from 0.4780 to 0.6598, and reduces screening burden by 41.8%.It prioritizes NOS3 and PPARG as puerarin targets and connects prediction scores with evidence integration and downstream validation workflows.
  • ABPP evaluation: On the ABPP dataset, ProbeMatchDTI reaches 0.591655 AUC-ROC and 0.563363 AUC-PR, exceeding MMDG-DTI’s 0.580163 and 0.561936.Its mutation responses show residue-context sensitivity across XPO2, RUVBL1, and Compound 26–EGFR.

5 Conclusion

The paper concludes that ProbeMatchDTI combines weak binding-pattern preservation with microscopic correspondence and macroscopic complementarity matching. It reports strongest public-benchmark metrics, biochemical sensitivity analyses, and a puerarin workflow extending predictions toward validation.

  • ProbeMatchDTI combines IterProbe for weak binding-pattern preservation with BindingProbe for microscopic correspondence and macroscopic complementarity.
  • The method achieves the strongest AUC-ROC and AUC-PR across four public benchmarks, while ABPP and molecular-dynamics-anchored analyses support its biochemical sensitivity.
  • The puerarin case connects ProbeMatchDTI predictions with DTI-Agent and DPL for evidence-guided, executable validation.

6 Limitations and Ethical Considerations

The study notes that its use of public datasets and resources may inherit their biases and incompleteness, while real-world drug-discovery applications remain future work.

  • Public datasets and resources may transmit biases and incompleteness into the study.
  • Real-world drug-discovery applications remain future work despite the case study demonstrating the system’s potential.
  • The study involved no identifiable personal data, human participants, or animal experiments.

7 Generative AI Usage

Generative AI assisted with language polishing, while the authors developed and verified the research design, analysis, interpretation, and final decisions.

  • Generative AI assisted with language polishing.
  • The authors developed and verified the research design, analysis, interpretation, and final decisions.

A.1 Public Benchmark Statistics and Preprocessing

ProbeMatchDTI is evaluated on four public DTI benchmarks using aligned drug–target samples, with preprocessing that preserves multimodal tuples but does not enforce drug- or protein-disjoint splits. Its pipeline combines IterProbe-based entity encoding with semantic and microscopic BindingProbe evidence before producing a DTI score.

  • Public benchmarks: The evaluation uses Human, C. elegans, BindingDB, and DrugBank public DTI benchmarks.Each sample contains a drug, target protein, and binary interaction label.
  • Preprocessing: DrugBank, Human, and C. elegans samples use an 83.3%/16.7% train–test split after shuffling with seed 1234.BindingDB uses its provided training/test partition.
  • Model pipeline: IterProbe encodes drug and protein features through entity-specific refinement, while BindingProbe forms semantic and microscopic pair-level evidence.The high-level inference pipeline returns a DTI prediction score from these branches and their fused evidence.
  • Model pipeline: The inference procedure combines semantic, microscopic, and fused logits for the final prediction.The microscopic branch models atom–residue complementarity, while the semantic branch produces a pair-level semantic feature.

A.3 Training Settings and Training Dynamics

ProbeMatchDTI is trained with AdamW and evaluated using ranking and classification metrics, with branch-level analyses examining semantic and microscopic convergence. Training metrics stabilize early on three benchmarks, whereas DrugBank improves more gradually.

  • Training settings: ProbeMatchDTI reports AUROC, AUPR, Precision, and Recall, using mean and standard deviation across random seeds for multi-seed experiments.Training uses batch size 16, weight decay 0.07, and gradient clipping at 10.
  • Training dynamics: Semantic and microscopic branches show complementary convergence behaviors, while the final output remains competitive or superior across most epochs.The comparison covers Human, C. elegans, BindingDB, and DrugBank.
  • Training dynamics: Normalized training loss decreases steadily, while AUROC and AUPR stabilize after early training on Human, C. elegans, and BindingDB.DrugBank improves more gradually, indicating a more challenging interaction distribution.
  • Supplementary evaluation: ProbeMatchDTI achieves the highest Precision on BindingDB and C. elegans while remaining close to the best method on DrugBank.Precision adds a threshold-dependent view of positive-prediction reliability alongside AUROC and AUPR.

B.1 Dataset Overview

The private ABPP dataset contains experimentally validated drug–target relationships from laboratory target-fishing studies of bioactive compounds. Its records document targets, disease coverage, literature provenance, and mutation metadata, providing a high-confidence reference beyond large databases dominated by indirect or inferred associations.

  • Dataset provenance: The private ABPP dataset contains experimentally validated drug–target relationships collected through laboratory target-fishing studies.The compounds include Eupatorium lindleyanum sesquiterpene lactones and related bioactive molecules.
  • Dataset contents: ABPP records document experimentally verified targets, disease coverage, literature provenance, and mutation metadata.The complete dataset is retained as a private research resource.
  • Dataset provenance: The dataset was curated using activity-based or photoaffinity probes, click chemistry, and quantitative proteomics to identify direct protein targets.These records provide biological evidence beyond large-scale databases dominated by indirect or inferred associations.
Loading 2609.02549v1…