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
TL;DR
Existing DTI methods often do not explicitly retain intermediate biochemical patterns while matching drug and protein representations. ProbeMatchDTI uses IterProbe and BindingProbe to model those patterns and achieves the strongest AUC-ROC and AUC-PR across four public benchmarks.
Problem
Existing DTI methods often match final representations without explicitly retaining intermediate contextual states, limiting joint modeling of local biochemical patterns and broader binding context.
Source: §1 · p. 2, §2.1 · p. 3Method
ProbeMatchDTI uses IterProbe to retain and positionwise select contextual states before matching, while BindingProbe models bidirectional atom–residue correspondences and local-to-global complementarity.
Source: §3.1 · p. 3, §1 · p. 3Results
Takeaways & Limitations
The puerarin case shows how ProbeMatchDTI predictions can extend into evidence-guided, executable validation.
Source: §5 · p. 8The study relies on public datasets and resources that may inherit biases and incompleteness, and real-world drug-discovery applications remain future work.
Source: §6 · p. 8
Abstract
from arXiv · showhide
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 multiscale drug–protein complementarity by actively preserving and matching binding-relevant biochemical patterns. It combines entity-level refinement probes with pair-level correspondence and investigates predictive performance and downstream evidence-guided drug discovery.
Introduction
- DTI prediction supports virtual screening and drug discovery [30, 49], but accurate prediction requires assessing cross-scale binding-pattern complementarity beyond representing each entity [12, 20, 24].Drug functional groups and local motifs form contacts with protein residues within the broader molecular and sequence context.Source: §1 · p. 2
- Hierarchical encoders passively aggregate biochemical evidence [9, 12, 20, 36], potentially obscuring weak functional-group, motif, or residue-context signals before cross-entity matching [18, 21, 32].Direct fusion can also blend atom–residue correspondence with broader molecular context before modeling the two scales separately [18, 37, 44, 47].Source: §1 · p. 2
- IterProbe retains initial, intermediate, and current contextual states at each drug-token or protein-residue position, then uses learned probes to select binding-relevant patterns before cross-entity matching.The input-specific selection yields sample- and position-dependent weights, keeping functional-group, local-motif, and residue-context patterns accessible across refinement depths.Source: §1 · p. 2
- BindingProbe models counterpart-conditioned bidirectional atom–residue correspondence within covalent and sequence-local environments, then probes pair evidence for local-to-global biochemical complementarity.Its macroscopic probe relates local biochemical correspondence to the broader molecular scaffold and protein environment.Source: §1 · pp. 2–3
- The study evaluates four public DTI benchmarks, ablations, feature-level pattern analyses, mutation sensitivity, and an integrated downstream case study connecting predictions to evidence-guided refinement and validation planning.The case study uses DTI-Agent to integrate domain knowledge and predictions, and DPL to translate refined hypotheses into structured validation procedures; molecular-dynamics analyses illustrate the workflow’s application.Source: §1 · pp. 2–3
2 Related Work
DTI representation learning has progressed from sequence, SMILES, and graph models to Transformer-based, attention-based, and multimodal approaches. Related work identifies passive aggregation and limited explicit multi-scale matching as gaps, while AI-assisted workflows connect predictions to downstream evidence analysis.
2.1 Biological Representation Learning and Multi-Scale DTI Modeling
- Prior DTI models progressed from sequence, SMILES, and graph representations to Transformer-based, attention-based, and multimodal methods, motivating probe-based selection of fine-grained biochemical representations.Examples include MolTrans [13] and CPIformer.Source: §2.1 · p. 3
- Existing methods can let dominant scaffolds or global contexts overwhelm weaker binding-relevant signals, and often match final representations without explicitly probing intermediate states or separating interaction scales.The underrepresented signals include functional groups, local chemical motifs, and residue-level environments; microscopic motif–residue correspondence and macroscopic complementarity are often modeled in a unified interaction space.Source: §2.1 · p. 3
2.2 AI-Assisted Drug Discovery Workflows
- AI-assisted biomedical agents and scientific assistants organize heterogeneous evidence, prioritize candidates, and support computational experiment planning beyond core DTI representation and matching.These workflows connect predicted interactions with interpretation and validation processes [1–3, 14, 16, 34].Source: §2.2 · p. 3
3 Method
ProbeMatchDTI uses IterProbe to enhance entity-specific biochemical patterns and BindingProbe to model drug–protein evidence at local and whole-pair scales, then predicts interactions and supports downstream candidate refinement and validation planning.
3.1 Entity-Specific Representation Enhancement with IterProbe
- IterProbe retains contextual states across refinement depths and uses position-specific learnable probes to select biochemical patterns before cross-entity matching.Its independent drug and protein paths keep functional-group, local-motif, and residue-context patterns accessible as context expands.Source: §3.1 · pp. 3–4
3.2 Cross-Scale Complementarity Modeling with BindingProbe
- BindingProbe repeatedly updates reciprocal atom–residue correspondences under drug covalent-neighborhood and protein sequence-context constraints, allowing functional-group and residue patterns to refine one another.The bidirectional correspondences preserve the entities’ asymmetric environments rather than collapsing them into one undirected score.Source: §3.2 · p. 5
- BindingProbe keeps semantic and structure-derived microscopic evidence distinct, then uses pair-specific macroscopic probing to weight their whole-pair summaries for interaction prediction.The semantic view retains contextual biochemical patterns, while the structure-derived view captures atom–residue correspondence under graph and sequence-neighborhood organization.Source: §3.2 · p. 5
3.3 Prediction Head
- The prediction head maps the fused representation to interaction logits, while auxiliary classifiers keep semantic and structure-derived summaries individually predictive.A fixed ensemble combines fused and auxiliary logits to produce the interaction probability.Source: §3.3 · p. 5
3.4 DTI-Agent Integration
- At inference, ProbeMatchDTI converts interaction logits into probabilities and ranked drug–target candidates for evidence integration, candidate refinement, and validation planning.Source: §3.4 · p. 5
4 Experiments
Experiments report leading results across four public DTI benchmarks, complementary effects from IterProbe and BindingProbe, and feature-level evidence of residue localization. The paper also evaluates evidence-guided candidate refinement, molecular-dynamics validation, and mutation sensitivity on a private ABPP dataset.
4.2 Comparison with State-of-the-Art (SOTA) Models on Public DTI Benchmarks
- 0.991 AUC-ROC and 0.992 AUC-PR on BindingDB, and 0.906 AUC-ROC and 0.906 AUC-PR on DrugBank, were ProbeMatchDTI’s best benchmark results.ProbeMatchDTI led all four public benchmarks; on BindingDB the strongest baseline, MMDG-DTI, scored 0.971 and 0.972 on AUC-ROC and AUC-PR, while on DrugBank the best baseline scored 0.901 and 0.902. On C. elegans it scored 0.997 on both metrics versus MMDG-DTI’s 0.996; on Human it scored 0.992 AUC-ROC and 0.994 AUC-PR.Source: Table 1 · p. 6, §4.2 · p. 6
4.1.1 Datasets.
- Evaluation used Human, C. elegans, BindingDB, and DrugBank public benchmarks, plus a private ABPP dataset of experimentally identified drug–target interactions for mutation-sensitivity analysis.DTI prediction was evaluated with AUC-ROC and AUC-PR.Source: §4.1.1 · pp. 5–6, §4.1.2 · p. 6
4.6 Performance and Mutation Sensitivity on the ABPP Dataset
- On the ABPP dataset, ProbeMatchDTI’s best AUC-ROC/AUC-PR were 0.591655/0.563363, versus MMDG-DTI’s 0.580163/0.561936; its mutation responses also varied by target and compound.Scores declined near-monotonically with accumulating XPO2 and RUVBL1 mutations, whereas MMDG-DTI was nearly insensitive to RUVBL1 mutations; for Compound 26–EGFR, ProbeMatchDTI stayed below wild type under both mutation settings while MMDG-DTI increased.Source: Table 3 · p. 6, §4.6 · p. 8
4.3 Ablation Study
- Combining both probes achieved the best AUC-ROC and AUC-PR on every dataset, including 0.991/0.992 on BindingDB and 0.906/0.906 on DrugBank.IterProbe improved C. elegans AUC-ROC from 0.992 to 0.995 and Human Precision from 0.958 to 0.970; BindingProbe improved BindingDB AUC-ROC from 0.971 to 0.988 and DrugBank Precision from 0.794 to 0.822. BindingProbe alone lowered DrugBank AUC-ROC/AUC-PR from 0.897/0.897 to 0.872/0.872, whereas after IterProbe they rose to 0.906/0.906.Source: Table 2 · p. 6, §4.3 · p. 7
4.4 Feature-Level Biochemical Pattern Analysis
- The puerarin–NOS3 attribution linked a BindingProbe hotspot at Glu361 to IterProbe-strengthened drug and protein patterns, while the predicted P367 site lay 6, 11, and 29 residues from three MD contacts.Those deviations from Glu361, Trp356, and Asn338 corresponded to 0.59%, 1.07%, and 2.83% of the NOS3 sequence; the analysis aligned the MD contact, BindingProbe hotspot, and IterProbe-enhanced patterns.Source: Figure 3 · p. 7, §4.4 · p. 8
4.5.1 Evidence-Guided Candidate Refinement.
- 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.Source: §4.5.1 · p. 8
4.5.2 Molecular Dynamics Validation.
- In 100-ns simulations, puerarin maintained 2–4 hydrogen bonds with NOS3 versus 0–6 with PPARG, supporting more stable ligand binding in NOS3 despite lower PPARG complex and backbone RMSDs.PPARG complex and backbone RMSDs were approximately 0.30 and 0.20 nm, while puerarin fluctuated more in PPARG (0.35–0.50 nm) than in NOS3 (approximately 0.25 nm after 60 ns).Source: §4.5.2 · p. 8
4.5.3 End-to-End Puerarin Case Study.
- DTI-Agent connected ProbeMatchDTI rankings to traceable evidence-supported hypotheses, computational validation, and executable virtual-screening, molecular-dynamics, and robotic-assay workflows.The puerarin case study reported distinct but stable binding characteristics with NOS3 and PPARG and described the workflow as modular and adaptable to other compounds, targets, and experimental settings.Source: §4.5.3 · p. 8
5 Conclusion
ProbeMatchDTI combines probe-driven preservation of weak binding patterns with multi-scale biochemical correspondence and complementarity. Its predictions are linked to evidence-guided validation, with broader experimental data and prospective assays identified as future directions.
5 Conclusion
- ProbeMatchDTI achieves the strongest AUC-ROC and AUC-PR across four public benchmarks.ABPP and molecular-dynamics-anchored analyses also support its biochemical sensitivity.Source: §5 · p. 8
- IterProbe preserves weak binding patterns, while BindingProbe models microscopic correspondence and macroscopic complementarity.Source: §5 · p. 8
- ABPP and molecular-dynamics-anchored analyses support ProbeMatchDTI’s biochemical sensitivity.Source: §5 · p. 8
- The puerarin case shows how DTI-Agent and DPL extend predictions into evidence-guided, executable validation.Source: §5 · p. 8
- Future work will incorporate broader experimentally validated interactions and prospective assays to advance scalable prediction–evidence–execution pipelines for drug discovery.Source: §5 · p. 8
6 Limitations and Ethical Considerations · A Appendix
ProbeMatchDTI’s evaluation is limited by potential biases and incompleteness in public resources, and its real-world drug-discovery utility remains future work; no human participants, animal experiments, or identifiable personal data were involved.
6 Limitations and Ethical Considerations
- Public datasets and resources may carry biases and be incomplete.These limitations may affect the study’s evidence base.Source: §6 · p. 8
- Real-world drug-discovery applications remain future work despite the case study demonstrating the system’s potential.The case study does not establish real-world application.Source: §6 · p. 8
- The study involved no identifiable personal data, human participants, or animal experiments.Source: §6 · p. 8
A.1 Public Benchmark Statistics and Preprocessing
ProbeMatchDTI is evaluated on four public DTI benchmarks, with samples defined by a drug, target protein, and binary interaction label.
A.1 Public Benchmark Statistics and Preprocessing
- The evaluation covers Human, C. elegans, BindingDB, and DrugBank, with each sample pairing a drug and target protein with a binary interaction label.Table 4 summarizes the benchmark dataset statistics.Source: §A.1 · p. 11, Table 4 · p. 11
- DrugBank, Human, and C. elegans samples were shuffled with seed 1234 and split 83.3%/16.7%, preserving modality-label alignment without drug- or protein-disjoint constraints.BindingDB used its provided training/test partition.Source: §A.1 · p. 11
A.2 Implementation Details
Implementation uses IterProbe to enhance drug and protein representations, then combines pair-level semantic and atom–residue features through BindingProbe to produce DTI predictions.
A.2 Implementation Details
- Drug and protein language-model features are projected to 256 dimensions, then processed by an initial 1D convolution and three probe-guided refinement steps that retain candidate states while preserving sequence length and hidden dimensionality.The input projections are from 768 dimensions for drugs and 1024 for proteins.Source: §A.2 · p. 11
- Enhanced drug and protein representations form a pair-level semantic sequence, where target-aware and drug-aware evidence retrieval precedes a prediction encoder yielding a 128-dimensional semantic DTI feature.Source: §A.2 · p. 11
- BindingProbe’s microscopic stage models atom–residue complementarity using atom features, molecular adjacency, and protein residues, with bidirectional matching and local-context updates producing a 128-dimensional microscopic DTI feature.Source: §A.2 · p. 11
- BindingProbe’s macroscopic probe retrieves complementary cross-scale evidence from semantic and microscopic features; the fused representation uses the microscopic feature as its residual backbone, and prediction combines fused, semantic, and microscopic logits.Source: §A.2 · p. 11
A.3 Training Settings and Training Dynamics
This section reports training dynamics, branch-level convergence behavior, supplementary Precision comparisons, and implementation settings for ProbeMatchDTI.
A.3 Training Settings and Training Dynamics
- ProbeMatchDTI is implemented in PyTorch and trained with AdamW using batch size 16, weight decay 0.07, and gradient-norm clipping at 10.Multi-seed results are reported as means and standard deviations; reported metrics are AUROC, AUPR, Precision, and Recall.Source: §A.3 · p. 11
- Across four benchmarks, semantic and microscopic evidence show complementary convergence behaviors, while the final output remains competitive or superior across most epochs.Source: §A.3 · pp. 11–12, Figure 5 · p. 12
- On Human, C. elegans, and BindingDB, AUROC and AUPR stabilize after early training, while DrugBank improves more gradually.Normalized training loss decreases steadily across the four benchmarks.Source: §A.3 · p. 12, Figure 4 · p. 12
- ProbeMatchDTI achieves the highest Precision on BindingDB and C. elegans and remains close to the best method on DrugBank.Precision supplements the main comparison and provides a threshold-dependent view of positive-prediction reliability.Source: §A.3 · p. 12, Table 6 · p. 12
B Private ABPP Target-Fishing Dataset · B.1 Dataset Overview
The private ABPP dataset compiles experimentally validated drug–target relationships from laboratory target-fishing studies. Its records provide biological context and a high-confidence reference for computational DTI prediction.
B.1 Dataset Overview
- The private dataset contains experimentally validated drug–target relationships collected through laboratory target-fishing studies of Eupatorium lindleyanum sesquiterpene lactones and related bioactive compounds.Source: §B.1 · p. 12
- Table 7 presents selected records documenting experimentally verified targets, disease coverage, literature provenance, and mutation metadata.The table’s “Mutation” field indicates whether mutation information is available for each record.Source: §B.1 · p. 12, Table 7 · p. 13
- The dataset was curated from recent high-quality studies using activity-based or photoaffinity probes, click chemistry, and quantitative proteomics to identify direct protein targets.Source: §B.1 · p. 12
- The target-fishing studies identified targets of bioactive constituents from traditional Chinese medicine and other small molecules.Source: §B.1 · p. 12
- Experimentally resolved direct-binding annotations are scarce and costly to obtain, making these records valuable biological evidence beyond large-scale databases dominated by indirect or inferred associations.Source: §B.1 · pp. 12–13
- The collection serves as a high-confidence reference for training and evaluating computational DTI prediction models.Its annotations provide evidence beyond large-scale databases dominated by indirect or inferred associations.Source: §B.1 · p. 13