A new computational explainer recovered substantially more literature-backed molecular regions behind drug-pair synergy predictions than the comparison methods tested in a drug-combination benchmark. VINCENT recorded a mean recall of 0.826, with a 95% confidence interval of 0.782 to 0.868; baseline recall ranged from 0.49 to 0.66. In plain terms, the method recovered more of the reference regions used to judge whether its explanations matched published molecular clues.
The strongest coverage result came from a narrow slice of the data. Of the 71 test pairs, 25 were retained for literature-grounded evaluation, producing 111 independently constructed molecular-region annotations, or an average of 4.4 annotations per pair. Those 25 pairs were a subset of the test set, not an additional independent group.
What VINCENT is trying to show
VINCENT is designed to answer a specific question: which chemically coherent regions in two drugs jointly explain a model's predicted synergy? It starts with evidence at the level of atom pairs, organizes those signals into molecular motifs, tests motif pairs through repeated local perturbations, and feeds the validated evidence back into motif assignment. The perturbations are small, repeated changes to the molecular inputs used to check whether the model responds consistently to the proposed interaction.
The method is applied after training to a fixed predictor. The predictor parameters were frozen after training. A one-time, mask-aware calibration of components used at the perturbation interface was performed, after which those components were frozen as well. That setup evaluated the explanation procedure against a stable predictor rather than one continually altered during the analysis.
A focused test of the explanations
After canonicalization and deduplication, the benchmark contained 88 training pairs, 19 validation pairs and 71 test pairs. VINCENT was run on all 71 test pairs. The coverage analysis used recall, the share of reference regions recovered, alongside precision, Jaccard overlap and a hit rate for pairs reaching recall of at least 0.7. Confidence intervals came from pair-level bootstrap resampling with 2,000 resamples.
VINCENT's mean precision was 0.790, and its Jaccard overlap was 0.689. The Jaccard score measures the overlap between the regions selected by the method and those in the reference annotations. Its hit rate was 76.6% at recall of at least 0.7.
Closer to the model's own signal
VINCENT also showed a closer match to the behavior of the predictor itself. Across all 71 test pairs, total validated interaction strength had a Pearson correlation of 0.423 with predicted synergy. The mean top interaction score was 4.82 for predictor true positives and 1.43 for true negatives, producing a TP/TN separation of 3.36. The controlled baselines had correlation/separation values of 0.043/1.49 and -0.238/0.48. No uncertainty interval was reported for these alignment measures.
That distinction matters because an explanation can accurately reflect a model without showing that the model is biologically correct. On the test set, the fixed predictor's ROC-AUC, a measure of how well it separates positive and negative pairs, was 0.85, compared with 0.82 for original ComboNet, 0.80 for DeepDDS, 0.68 for DeepSynergy and 0.62 for random forest. The reported scores measure discrimination and agreement with the predictor; they do not establish that the highlighted regions are biologically causal or clinically effective.
Testing the moving parts
When feedback refinement was removed, the no-feedback version recorded recall of 0.724 instead of 0.826, TP/TN separation of 1.95 instead of 3.36, and synergy correlation of 0.352 instead of 0.423. With feedback disabled, the multi-view variant reached recall of 0.724 under the same perturbation validation, compared with 0.586 for one single-signal clustering-and-perturbation comparator and 0.646 for another.
Sensitivity checks found recall above 0.80 for motif cluster sizes from 5 to 8 atoms. Across perturbation-trial settings of 4, 8, 16 and 32, recall varied by only 0.014, while TP/TN separation ranged from 2.38 to 3.44. Within those tested settings, the outputs changed little.
A useful warning in one false negative
One diagnostic case shows why explanation and classification should not be treated as interchangeable. For the Nitazoxanide-Remdesivir pair, the predicted synergy score was 0.487, below the 0.5 threshold, making it a false negative. Yet the pair's top validated interaction score was 5.9, above the mean of 4.82 for predictor true positives and well above 1.43 for true negatives.
That example is only a case study and does not establish general performance. More broadly, the coverage result rests on 25 literature-annotated test pairs, while the alignment scores describe evidence produced by a fixed predictor. The method needs a predictor with atom-level representations and cross-drug signals, and a predictor that has learned spurious correlations could pass those correlations into the explanation. VINCENT therefore offers a computational audit of predictor behavior, not experimental validation of the molecular interactions it highlights.
Paper data and sources
Original title: VINCENT: Validated Interaction Network for Cross-drug Explanation of Therapeutics
Authors: Fan-Sheng Chuang, Xuchen Li, Yujing Bian, Kaixiong Zhou
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text