PaperlayerView exact revision on arXiv

Source-linked AI summary

VINCENT: Validated Interaction Network for Cross-drug Explanation of Therapeutics

Fan-Sheng Chuang, Xuchen Li, Yujing Bian, Kaixiong Zhou

arXiv:2608.25841v1cs.LGcs.AI

TL;DR

Existing synergy predictors usually provide a single score without showing which cross-drug molecular regions drive it. VINCENT validates and refines chemically coherent motif-pair explanations, achieving 0.826 mean motif recall versus 0.49–0.66 for baselines and 3.36 TP/TN separation across 71 test pairs.

  • Problem

    Most synergy models provide scalar predictions without identifying chemically coherent cross-drug motif pairs or validating their stability and predictor alignment.

  • Method

    VINCENT is a post-training framework that constructs motif pairs from predictor signals, validates them through repeated local perturbations, and feeds evidence back to refine assignments.

  • Results

    0.826 mean motif recall versus 0.49–0.66 for baselines, while validated interaction scores achieve 3.36 TP/TN separation across 71 test pairs.

  • Takeaways & Limitations

    VINCENT recovers literature-supported molecular regions more accurately than evaluated alternatives while producing interaction scores aligned with predictor behavior.

  • Takeaways & Limitations

    VINCENT explains predictor-level evidence rather than biological causality or clinical efficacy, and its evaluation covers only 25 literature-annotated benchmark entries.

Abstract

from arXiv · show

Drug synergy prediction estimates whether two drugs produce a stronger joint effect than expected from their individual activities. For drug combination discovery, a single synergy score is often not enough: researchers also need to know which molecular regions jointly drive the prediction. We study motif-pair synergy explanation, which identifies pairs of chemically coherent regions, one from each drug, that jointly contribute to predicted synergy. Existing interpretable synergy models expose atom- or substructure-level signals, but their explanations are built into the predictor architecture, and none validates cross-drug region scores under repeated perturbations or feeds that evidence back to refine the explanation. A reliable motif-pair explanation should instead be chemically coherent, perturbation-stable, and aligned with predictor behavior. We introduce VINCENT (Validated Interaction Network for Cross-drug Explanation of Therapeutics), a post-training framework for a fixed interaction-aware synergy predictor. VINCENT extracts atom-pair evidence from attention and gradient signals, groups atoms into chemically coherent motifs, and validates candidate motif pairs through repeated local perturbations. The validated evidence is fed back to refine motif assignments, yielding explanations that satisfy these three criteria. On a 25-pair literature-annotated subset, VINCENT achieves a mean motif recall of 0.826 (95% CI: 0.78-0.87), compared with 0.49-0.66 for baselines. Across all 71 test pairs, its validated interaction scores yield a TP/TN separation of 3.36. These results show that closed-loop perturbation validation recovers literature-supported molecular regions more accurately than existing alternatives while producing cross-drug interaction scores that better reflect predictor behavior.

1 Introduction

VINCENT addresses the need for explanations that identify chemically coherent cross-drug motif pairs driving predicted drug synergy, rather than reporting only scalar synergy scores. Its closed-loop framework validates and refines motif-pair evidence, achieving higher literature-grounded motif recall and stronger predictor alignment.

  • Motivation: Drug-pair synergy prediction supports combination discovery, but a single scalar score does not reveal which molecular regions drive predicted synergy.The application areas include drug repurposing, lead optimization, and antiviral combination screening.
  • Problem formulation: A useful explanation should encode connected cross-drug motif pairs and satisfy chemical coherence, perturbation stability, and predictor alignment.These requirements are formalized as R1–R3, with perturbation stability requiring consistent predictor responses across repeated local perturbations.
  • Method: VINCENT is a post-training framework that extracts atom-level cross-drug evidence, constructs chemically coherent motifs, validates candidate motif pairs, and feeds the evidence back to refine explanations.It operates on a fixed interaction-aware drug synergy predictor whose design is not itself a contribution.
  • Evaluation protocol: 111 pharmacophore-level molecular-region annotations across 25 literature-supported drug pairs form the reference set for evaluating recovery of pharmacologically relevant regions.The annotations are traceable to their supporting sources.
  • Results: 0.826 mean motif recall on 25 pairs exceeds 0.49–0.66 for evaluated explanation baselines, while TP/TN separation reaches 3.36 across 71 test pairs.The recall result measures recovery of literature-supported molecular regions, and the TP/TN result reflects validated interaction scores’ predictor alignment.

2 Related Work

Prior work spans diverse drug-synergy predictors, general graph-neural-network explanation methods, and interpretable models that identify molecular substructures or cross-drug regions. These approaches motivate VINCENT’s post-hoc focus on explaining cross-drug motif pairs.

  • Drug synergy prediction: Drug-synergy predictors include descriptor- and fingerprint-based neural networks, graph architectures, and dual-view, mechanism-informed, similarity-network, knowledge-graph, and multimodal formulations.DrugComb and DrugCombDB have accelerated model development, with a recent survey providing broader coverage.
  • GNN explainability: Graph-neural-network explanation methods include Integrated Gradients, mask learning, subgraph search, counterfactual explanations, and cooperative explanation strategies.Attention weights alone are not necessarily faithful explanations.
  • Substructure-level synergy explanation: Interpretable synergy models explain predictions through adaptive molecular substructures, predefined substructures, connected graphlets, multiscale structures, or pharmacophoric regions.These approaches provide structural and biological interpretations at several molecular levels.

3 Preliminaries

The preliminaries define synergy as excess combination activity over the Bliss-independence baseline and frame motif-pair explanation as localized cross-drug evidence. Reliable explanations require chemical coherence, perturbation stability, and alignment with predictor behavior.

  • Drug synergy prediction: Synergy score s_AB measures predicted combination activity beyond the Bliss-independence baseline, with s_AB > 0.5 defining a synergistic pair.Positive s_AB indicates excess above independent action.
  • Motif-pair synergy explanation: A motif-pair explanation localizes cross-drug evidence through validated interaction scores r_kl between chemically coherent motifs from drugs A and B.The interaction matrix R represents motif pairs rather than isolated atom scores.
  • Explanation requirements: Each motif-pair score must remain supported across repeated perturbations, rather than relying on a noisy or unrepeatable single forward pass.This criterion is termed motif-pair perturbation stability.
  • Explanation requirements: Aggregate interaction strength should be positively associated with predicted synergy s_AB across drug pairs, ensuring the explanation tracks the predictor’s decision behavior.This requirement is predictor-explainer alignment.

4 Method

VINCENT is a post-training framework that explains a fixed interaction-aware synergy predictor through chemically coherent, perturbation-stable cross-drug motif pairs. It combines predictor-derived atom-pair evidence, multi-view motif assignment, repeated perturbation validation, and feedback refinement.

  • Phases 1–2: VINCENT extracts atom-pair evidence by intersecting predictor cross-drug associations with positive Integrated Gradients contributions to the synergy score.The resulting evidence map retains atom pairs supported by both signals and remains fixed during explanation.
  • Phases 1–2: Three affinity views guide motif discovery through chemical locality, partner-conditioned interaction patterns, and perturbation-validated feedback.Structural and pattern views are fixed, whereas the feedback view is initialized at zero and updated after validation.
  • Phase 3: Soft motif assignments combine multi-view smoothness with entropy and minimum-mass regularization, then enforce connectivity and ring completion to produce coherent candidate regions.Each atom is assigned to its highest-weight motif before graph-based postprocessing prevents disconnected regions and split rings.
  • Phase 4: Candidate motif pairs are screened using aggregated fixed evidence, and only the highest-ranked pairs undergo repeated local perturbation trials.Each trial masks different local subsets and fractions of atoms in both motifs to generate nearby perturbed realizations.
  • Phase 4: Validated interaction scores summarize perturbation effects by magnitude, variability, activation, and directional consistency, then project validated profiles back to atoms for the next assignment iteration.This feedback encourages atoms with repeated similar cross-drug effects to form common motifs without modifying the original evidence map.

5 Experiments

VINCENT is evaluated for literature-grounded motif recovery, alignment with a fixed predictor’s synergy behavior, component contributions, sensitivity, and qualitative interaction evidence. It achieves strong motif coverage and predictor alignment, with iterative feedback providing the largest ablation gain.

  • Motif coverage: 0.826 mean recall (95% CI: 0.782–0.868), 0.790 precision, 0.689 Jaccard, and 76.6% hit rate at recall ≥0.7 establish VINCENT’s strongest literature motif coverage.Evaluation uses 25 literature-annotated test pairs and 111 reference motifs.
  • Predictor alignment: 0.423 Pearson correlation and 3.36 TP/TN separation show that validated interaction scores align with predictor synergy behavior across 71 test pairs.Mean top interaction scores are 4.82 for predictor true positives and 1.43 for true negatives.
  • Ablation: Removing iterative feedback reduces recall from 0.826 to 0.724, TP/TN separation from 3.36 to 1.95, and synergy correlation from 0.423 to 0.352.The ablation retains multi-view assignment, isolating the contribution of validation feedback to motif recovery and predictor alignment.
  • Ablation: 0.724 recall from multi-view assignment exceeds 0.586 for IG Clust.+Pert. and 0.646 for ATT Clust.+Pert. under the same perturbation-based validation.The comparison isolates grouping and attributes the improvement to combining partner-conditioned interaction patterns with molecular locality.
  • Hyperparameter sensitivity: 0.002 recall change and 0.08 TP/TN-separation change when perturbation trials increase from 16 to 32 indicate stable aggregate metrics beyond the default setting.Performance rises rapidly through the first three outer-loop iterations and then stabilizes.
  • Case study: 5.9 top interaction score for Nitazoxanide + Remdesivir, a predictor false negative, links the nitrothiazole region to Remdesivir’s adenine ring above the synergistic-group mean.The example shows cross-drug evidence in internal representations despite an incorrect final binary decision.

6 Limitations and Ethical Considerations

VINCENT explains the fixed predictor’s learned evidence, not biological causality or clinical efficacy, and its conclusions are limited by available annotations, compounds, assays, and model interfaces. Because the study uses public data without human intervention, outputs should support mechanistic inspection and experimental prioritization alongside domain expertise and evidence, not treatment decisions.

  • Predictor dependence and scope: VINCENT inherits the behavior and assumptions of the fixed predictor, so region-pair scores should not be interpreted as biological causality or clinical efficacy.Its scores characterize predictor-level evidence.
  • Predictor dependence and scope: 25 benchmark entries had published evidence supporting region-level annotation, limiting conclusions to the evaluated compounds and assay settings.The literature-grounded evaluation remains conditioned on those benchmark conditions.
  • Predictor dependence and scope: Spurious correlations learned by the predictor can propagate into VINCENT explanations instead of true pharmacological signals.Predictive uncertainty estimates could help identify such cases.
  • Predictor dependence and scope: VINCENT requires an interaction-aware predictor exposing atom-level representations and cross-drug signals; predictors without these interfaces would require architectural adaptation.This limits direct applicability to models lacking those interfaces.
  • Ethical considerations: The study uses publicly available molecular and drug-combination data without identifiable personal data or human-subject intervention.VINCENT outputs should be interpreted with domain expertise and experimental evidence, not used directly for treatment selection or clinical decision making.

8 Conclusion

VINCENT answers how to identify cross-drug motif pairs that jointly contribute to a fixed drug synergy predictor’s decision through a learned, perturbation-validated explanation–validation loop.

  • Conclusion: VINCENT targets cross-drug motif pairs that jointly contribute to a fixed drug synergy predictor’s synergy decision.The framework addresses identification that is learned and perturbation-validated rather than predefined or based on a single forward pass.
  • Conclusion: The framework constructs chemically coherent motifs within a closed explanation–validation loop.
  • Conclusion: VINCENT validates retained motif pairs through repeated local perturbations.

A Literature-Grounded Evaluation Protocol: Detailed Construction … A.6 Reference-Set Statistics

VINCENT constructs a literature-grounded molecular-region reference independently of explanation outputs, then maps and matches predicted regions under shared constraints. The reference supports localization evaluation across 25 annotated entries containing 111 regions.

  • A Literature-Grounded Evaluation Protocol: Detailed Construction: The protocol evaluates whether explanations localize pharmacologically relevant regions in both drugs of SARS-CoV-2 combinations.The reference is constructed independently of explanation outputs and fixed before method comparison.
  • A Literature-Grounded Evaluation Protocol: Detailed Construction: From 71 test entries, literature screening retained 25 entries with sufficient molecular specificity, yielding 111 drug-specific molecular-region annotations.The annotations were fixed before comparing explanation methods.
  • A.1 Level 1: Literature Evidence and Provenance: The evidence hierarchy comprises T1 direct combination-response evidence, T2 mechanistic characterization with combination rationale, and T3m mechanism-supported pharmacological context.The final set contains 10 T1 entries, 3 T2 entries, and 12 T3m entries.
  • A.2 Level 2: Pair-Level Mechanistic Context: Pair-level literature summaries establish observed combination behavior and pharmacological context before drug-specific molecular regions are localized.Sources include published combination studies, antiviral screens, mechanistic studies, and structure–activity analyses.
  • A.3 Level 3: Drug-Specific Literature-Supported Regions: Each drug-specific region is a chemically coherent functional group, scaffold component, or pharmacophore-level locus linked to published pharmacology.The annotation is region-based because determinants may span aromatic systems, basic side chains, modified nucleosides, or prodrug moieties.
  • A.4 Level 4: Atom-Index Reference Mapping: Literature-supported regions are converted into zero-based RDKit atom-index reference sets while preserving functional-group or scaffold boundaries.Indices need not be consecutive when a region spans multiple molecular-graph branches.
  • A.5 Region-Aware Matching Protocol: Matching restricts candidates to the same benchmark entry and drug, allowing either one predicted cluster or a union of at most two clusters.All methods use the same admissibility rule and select the best matching region.
  • A.6 Reference-Set Statistics: 111 molecular regions across 25 annotated benchmark entries correspond to 4.44 regions per entry on average.Coverage evaluation reports region recall, precision, Jaccard overlap, and the fraction reaching recall ≥0.7.

B Case Study … E.5 Stable Interaction Score

VINCENT’s case study links chemically coherent cross-drug segmentation to validated interaction scores and literature-supported regions, while its appendices specify the perturbation, affinity, assignment, feedback, and stability operators. The Amodiaquine–Nitazoxanide example shows strong validated region-pair interactions and high literature-region recall.

  • B Case Study: 0.690 is the predicted synergy score for the true-positive Amodiaquine + Nitazoxanide case study.VINCENT uses this pair to connect segmentation, cross-drug interaction scoring, and literature-supported region localization.
  • B Case Study: 3.8 is the strongest validated interaction score, linking Amodiaquine’s phenol-containing C2 with Nitazoxanide’s salicylamide-bridge C2.Other prominent scores are C2×C1 = 2.6, C1×C1 = 2.5, and C3×C1 = 1.9.
  • B Case Study: 0.941 is VINCENT’s mean recall across four literature-supported reference regions in this benchmark entry.The predicted regions recover the chloroquinoline core and diethylamino sidechain on Amodiaquine and overlap Nitazoxanide’s annotated nitrothiazole and salicylamide regions.
  • C External Baseline Adaptation: External baselines explain the same fixed predictor by separately converting each drug’s atom-level importance into candidate regions using connected components or spectral clustering.The best globally swept configuration is reported for each baseline; PGExplainer, SubgraphX, CF-GNNExplainer, and random substructure use distinct training, ranking, sparsity, or sampling procedures.
  • D Method Implementation Details; D.6 Feedback Operator; D.7 Complete Hyperparameter Table: Validated interaction profiles feed back into atom affinities through graph locality and safe cosine similarity, followed by exponential-moving-average smoothing.The framework’s operators and hyperparameters are defined for drug A and applied identically to drug B, with complete hyperparameters listed separately.
  • D.1 Structural Affinity; D.2 Interaction-Pattern Affinity: VINCENT’s structural affinity uses Gaussian decay over shortest-path molecular-graph distance, with d_s controlling neighborhood radius and σ_s controlling decay rate.The interaction-pattern affinity combines local connectivity, cross-drug evidence-profile similarity, and an activity gate whose threshold is the 60th percentile of {a_i}.
  • D.3 Normalized Laplacian; D.4 Assignment Regularization; D.5 Ring Completion: VINCENT converts affinity matrices into self-looped symmetric normalized Laplacians and regularizes assignments to prevent premature one-hot collapse and pathologically small motifs.After hard assignment, graph connectivity is enforced and partially assigned rings are completed to keep chemically indivisible ring systems intact.
  • E Perturbation-Based Validation Details; E.1 Local Subset Sampling; E.2 Feature Substitution; E.3 Local Reconditioning; E.4 Perturbation States and Interaction Effect; E.5 Stable Interaction Score: T=16 trials vary local atom subsets and masking fractions, preserving molecular topology while locally reconditioning nearby representations.Each trial evaluates four predictor states—both regions retained, either region alone, or both masked—and computes a second-order finite-difference interaction effect; the stable score combines four across-trial statistics.

F Hyperparameter Sensitivity Analysis

The sensitivity analysis varies each hyperparameter independently while holding others at defaults and evaluates the same six metrics used in the main text. Results show graceful degradation, with feedback-related parameters producing the strongest effects.

  • Analysis design: Each hyperparameter is varied individually with all others held at their defaults, using the same six evaluation metrics as the main text.The analysis covers five hyperparameter sweeps and early outer-loop convergence.
  • Feedback sensitivity: Disabling feedback (𝜆fb = 0) reduces recall by approximately 0.10 and TP/TN separation by 1.4.These are the strongest reported sensitivity effects and align with the Section 5.5 ablation findings.
  • Parameter effects: Feedback-related parameters, including 𝜆fb and outer-loop depth, have stronger effects than segmentation granularity.Cluster sizes from 5 to 8 atoms are described as having a moderate effect on recall.

G Evaluation Robustness Analysis · H Predictor Implementation and Adequacy

VINCENT’s evaluation robustness analyses show strong performance for directly supported reference pairs and stability against pair removal. The available passages report that no small subset dominates aggregate recall, while predictor implementation and adequacy are not detailed in the supplied text.

  • G Evaluation Robustness Analysis: The appendix reports evidence-tier stratification and pair-selection stability analyses summarized in Section 5.2.Table 12 covers evidence-tier motif coverage, while Table 13 covers leave-k-out mean-recall stability.
  • G Evaluation Robustness Analysis: 0.842 recall and 85.8% hit rate were achieved by Tier-1 pairs supported by direct experimental synergy evidence.These pairs aligned most closely with the predictor’s learned motif boundaries.
  • G Evaluation Robustness Analysis: No individual pair shifted the mean recall by more than 1.5 percentage points.This supports the conclusion that aggregate results were not driven by a small number of favorable pairs.
  • G Evaluation Robustness Analysis: At k=5, removing 20% of the 25-pair set left the 95% range of mean recall at 0.810–0.852.Leave-k-out analysis randomly removed k pairs and recomputed mean recall over 2,000 repetitions.
  • G Evaluation Robustness Analysis: Pair_003, Nitazoxanide + Remdesivir, was the most influential single pair with recall 0.553.Removing it raised the mean recall by +0.012.
  • G Evaluation Robustness Analysis: Pair_061, Camostat + Remdesivir, was the most favorable pair with recall 0.975.Removing it lowered the mean recall by −0.006.

H.1 Architecture … H.4 Predictive Adequacy

The fixed synergy predictor integrates 2D, 3D, and cross-drug interaction modeling, while VINCENT calibrates perturbation interfaces after training and evaluates explanations against a meaningful predictive target. Its test ROC-AUC is 0.85, although predictor accuracy is not itself a contribution.

  • H.1 Architecture: The architecture fuses D-MPNN graph representations with EGNN conformer representations before cross-drug interaction modeling.The 2D branch produces atom-level representations from bond-level messages, while the 3D branch provides geometry-aware atom representations.
  • H.1 Architecture: Bidirectional atom-level cross-attention conditions each drug on its partner and exposes the association matrix used by VINCENT.Separate heads predict P_A, P_B, and P_AB; synergy is s_AB = P_AB − P_bliss, with synergistic classification when s_AB > 0.5.
  • H.2 Training: The predictor jointly optimizes drug–target interaction, single-agent antiviral activity, and drug-combination synergy objectives.This multi-task setting shares molecular representations to address the limited 88 SARS-CoV-2 combination training pairs after deduplication.
  • H.3 Mask-Aware Calibration: After training, predictor parameters are frozen while calibration learns only the neutral mask embedding e_mask and local reconditioning operator.Random connected molecular regions are replaced by e_mask during calibration, reducing distribution shift from motif masking; both components are then frozen for explanations.
  • H.4 Predictive Adequacy: 0.85 test ROC-AUC demonstrates that the fixed predictor provides a sufficiently informative target for post-training explanation.The adequacy comparison uses the SARS-CoV-2 combination benchmark.
  • H.4 Predictive Adequacy: Predictor accuracy is not a contribution of VINCENT; the comparison only verifies that the explanation target is meaningful.The predictor is treated as a fixed target for the post-training explanation framework.
Loading 2608.25841v1…