Adaptive Drug-Drug Interaction Prediction via Gauge-Aware Graph Representation and Distribution Alignment

ICLR 2026 Conference Submission9874 Authors

Keywords: Graph Representation, Domain Adaptation

Abstract: We re-study drug-drug interaction (DDI) prediction under the conditions of data scarcity and distribution shift. In this paper, we propose a practical framework that links a compact gauge-aware graph encoder to light-weight distribution alignment objectives, which we refer to as GraphPharmNet. On the modeling side, GraphPharmNet is based on the message passing mechanism with the per-edge orthogonal transports that align the neighbor features before aggregation, offering a robust and stable mechanism that is especially implementation-friendly given local coordinate choices. (Contrary to the claim of the strict O(d)-equivariance with general nonlinearities being a theoretical ideal, it is important to note that the notion of gauge-aware is a device of stability, as opposed to being a theoretical identity.) Orthogonal transports are on-the-fly generated by a shared edge MLP, and not edge learnable parameters. Note that the memory estimates we report are an \emph{upper bound} when you have to cache all the coefficients at each edge, in practice we materialize them per mini-batch and we do not perform caching. On the learning side, we develop so-called training-only domain alignment based on MMD and optionally entropic-OT between partitions induced inside the training data. In order to prevent target leakage at all, we use a so-called leakage-safe training protocol: for predicting a batch of training edges, the target edges are (temporarily) removed from the message passing adjacency and feature propagation pipeline (''drop-edge-by-target''). Thus, relation labels of the target edges never enter the encoder which is used to generate their embeddings. We, furthermore, guarantee feature parity across the train/validation/test phases by utilizing the same training-only adjacency across all phases, and maintaining the above target-edge exclusion at training time. Empirically, on a DrugBank-based DDI graph (merged with Hetionet; $33{,}765$ nodes and $1{,}690{,}693$ edges), GraphPharmNet shows very strong performance under $6{:}1{:}3$ split outperforming the competitive GNN baselines. Furthermore, we record leakage-safe evaluation, deterministic single-label mapping, handling directed edges, and training-only alignment so that one can reproduce our setup.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 9874