Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rethinking Drug-Drug Interaction Modeling as Generalizable Relation Learning

Dong Xu, Jiantao Wu, Qihua Pan, Sisi Yuan, Zexuan Zhu, Junkai Ji

TL;DR

The paper tackles the generalization gap in drug–drug interaction prediction under real-world polypharmacy by moving from molecule-centric representations to a generalizable relation-learning framework. GenRel-DDI keeps pretrained molecular encoders as stable anchors while training a lightweight, partner-conditioned relation trunk via an anchor–adapter setup, enabling transfer to unseen drugs and novel drug pairs. Across seven public benchmarks with diverse label spaces, GenRel-DDI achieves state-of-the-art or best-reported performance, with pronounced gains in strict entity-disjoint settings and cross-domain transfer, and analysis shows that role assignment and freezing choices matter more than encoder scale under overlap restrictions. This work demonstrates the practical value of relation learning for robust DDI prediction and provides a scalable pathway for generalizable safety assessment in drug development.

Abstract

Drug-drug interaction (DDI) prediction is central to drug discovery and clinical development, particularly in the context of increasingly prevalent polypharmacy. Although existing computational methods achieve strong performance on standard benchmarks, they often fail to generalize to realistic deployment scenarios, where most candidate drug pairs involve previously unseen drugs and validated interactions are scarce. We demonstrate that proximity in the embedding spaces of prevailing molecule-centric DDI models does not reliably correspond to interaction labels, and that simply scaling up model capacity therefore fails to improve generalization. To address these limitations, we propose GenRel-DDI, a generalizable relation learning framework that reformulates DDI prediction as a relation-centric learning problem, in which interaction representations are learned independently of drug identities. This relation-level abstraction enables the capture of transferable interaction patterns that generalize to unseen drugs and novel drug pairs. Extensive experiments across multiple benchmark demonstrate that GenRel-DDI consistently and significantly outperforms state-of-the-art methods, with particularly large gains on strict entity-disjoint evaluations, highlighting the effectiveness and practical utility of relation learning for robust DDI prediction. The code is available at https://github.com/SZU-ADDG/GenRel-DDI.

Rethinking Drug-Drug Interaction Modeling as Generalizable Relation Learning

TL;DR

The paper tackles the generalization gap in drug–drug interaction prediction under real-world polypharmacy by moving from molecule-centric representations to a generalizable relation-learning framework. GenRel-DDI keeps pretrained molecular encoders as stable anchors while training a lightweight, partner-conditioned relation trunk via an anchor–adapter setup, enabling transfer to unseen drugs and novel drug pairs. Across seven public benchmarks with diverse label spaces, GenRel-DDI achieves state-of-the-art or best-reported performance, with pronounced gains in strict entity-disjoint settings and cross-domain transfer, and analysis shows that role assignment and freezing choices matter more than encoder scale under overlap restrictions. This work demonstrates the practical value of relation learning for robust DDI prediction and provides a scalable pathway for generalizable safety assessment in drug development.

Abstract

Drug-drug interaction (DDI) prediction is central to drug discovery and clinical development, particularly in the context of increasingly prevalent polypharmacy. Although existing computational methods achieve strong performance on standard benchmarks, they often fail to generalize to realistic deployment scenarios, where most candidate drug pairs involve previously unseen drugs and validated interactions are scarce. We demonstrate that proximity in the embedding spaces of prevailing molecule-centric DDI models does not reliably correspond to interaction labels, and that simply scaling up model capacity therefore fails to improve generalization. To address these limitations, we propose GenRel-DDI, a generalizable relation learning framework that reformulates DDI prediction as a relation-centric learning problem, in which interaction representations are learned independently of drug identities. This relation-level abstraction enables the capture of transferable interaction patterns that generalize to unseen drugs and novel drug pairs. Extensive experiments across multiple benchmark demonstrate that GenRel-DDI consistently and significantly outperforms state-of-the-art methods, with particularly large gains on strict entity-disjoint evaluations, highlighting the effectiveness and practical utility of relation learning for robust DDI prediction. The code is available at https://github.com/SZU-ADDG/GenRel-DDI.
Paper Structure (30 sections, 19 equations, 5 figures, 3 tables)

This paper contains 30 sections, 19 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Molecule-centric
  4. Evaluation and generalization
  5. Problem Setup
  6. Task Definition.
  7. Evaluation Splits by Train--Test Drug Overlap.
  8. Method: GenRel-DDI
  9. Pretrained Reference Encoders
  10. Prediction drift under selective freezing.
  11. Assumption 1 (Lipschitz in encoder token sequences).
  12. Anchoring reduces prediction drift.
  13. Within-Drug Conditioning
  14. Role-Separated Relation Trunk
  15. Prediction Head and Training
  16. ...and 15 more sections

Figures (5)

  • Figure 1: Proximity does not determine DDI labels. Closest pairs are not non-interacting (cyan), while farthest pairs still exhibit interactions (red) in both ChemBERTa-3 embedding and Morgan fingerprint spaces.
  • Figure 2: Illustration of evaluation splits by train--test drug overlap. (a) Unseen-pair (S1): both drugs are seen in training ($a,b\in\mathcal{V}_{\mathrm{tr}}$) but the test pair is held out. (b) Unseen-drug (S2): exactly one drug is seen ($a\in\mathcal{V}_{\mathrm{tr}}$ XOR $b\in\mathcal{V}_{\mathrm{tr}}$). (c) Entity-disjoint (S3): neither drug is seen ($a,b\notin\mathcal{V}_{\mathrm{tr}}$). Real drug examples are shown for illustration.
  • Figure 3: Overview of GenRel-DDI. Panels (a–c) illustrate common molecule-centric DDI pipelines that encode drugs independently and fuse pair features. Panel (d) shows GenRel-DDI, which keeps a pretrained encoder as a fixed anchor and trains a lightweight relation trunk with cross-attention to model partner-conditioned relational factors, reusing the same trunk across tasks by swapping only the head.
  • Figure 4: Scaling and tuning under overlap restriction on MeTDDI: AUPR on S1–S3 for three pretrained encoders under frozen and full fine-tuning settings. Published baselines are shown for reference; $(\ast)$ indicates a frozen encoder. Error bars denote mean ± std over five runs.
  • Figure 5: Role assignment and encoder pairing results of GenRel-DDI on MeTDDI under overlap-restricted splits (S2/S3). Each column is an ordered encoder pair $X{+}Y$, where $X$ is assigned to the anchor stream $(r)$ and $Y$ to the adapter stream $(t)$. $(*)$ indicates a frozen encoder stream. Top: per-metric ranks (1 = best) across all configurations on S2 and S3 for ACC/AUROC/AUPR. Bottom: absolute differences (points) relative to the MeTDDI baseline (the column labeled MeTDDI). Columns are sorted by S3 AUPR in ascending order.