Skip to content
Conference Open access

Beyond Pairwise Risk: A Corrected Aggregation Model for Higher-Order Drug Interaction Prediction

Sep 2026 · 2026 10th International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (I-SMAC) · pp. 944-953 · 0 citations · 16 references

Abstract

Adverse events from multi-drug regimens are difficult to predict computationally because most models score only drug pairs, requiring pairwise results for combinations of three or more drugs to be aggregated by hand -- an approach that cannot capture risk emerging only from the full combination. We present HO-DDI-SE, which decomposes higher-order interaction risk into a pairwise-aggregation baseline plus a learned, side-effect-conditioned correction term, trained with an auxiliary loss preserving baseline interpretability during joint fine-tuning. The model is trained and evaluated on HODDI (820,892 FAERS-derived records, 3,003 drugs, 3,968 adverse events) using a temporal split. The correction term improves AUROC over the pairwise baseline by 0.0067 to 0.0158 depending on combination size, significant after Benjamini-Hochberg correction and stable across random seeds and held-out quarters. A controlled comparison against a hypergraph-based correction shows no reliable advantage for the more complex alternative. Testing generalization on 100,000 independent FAERS reports spanning combination sizes up to 100 drugs, the correction term's advantage holds within HODDI's training-size range and degrades progressively beyond it, a precise, mechanistically grounded limitation rather than an unexplained failure. A game-theoretic Shapley attribution, grounded in empirical drug-outcome rates, recovers documented clinical regimens, including the co-formulated trifluridine-tipiracil combination. An independent validation against 15,511 real interactions from TWOSIDES confirms the pairwise component assigns significantly higher risk to known dangerous combinations than to random drug pairs (p < 0.000001). These results support explicit, interpretable decomposition as a viable, auditable path for higher-order interaction prediction, while precisely characterizing where that advantage currently holds.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.