Aligning Drugs, Not Interactions: Structured Adversarial Adaptation for New-Drug Interaction Typing

Predicting how two drugs interact means more than saying whether they interact. A useful model has to name the kind of interaction, for example "raises serum concentration," "reduces efficacy," or one of hundreds of specific side effects. Current models do this well when the test drugs resemble the training drugs. They fail badly on newly approved drugs, which tend to sit in unfamiliar regions of chemical space. Domain adaptation, a family of methods that uses unlabeled examples from the new setting to make a model's internal features look the same across old and new settings, is the obvious remedy. The usual version, however, aligns the wrong object. It aligns the representation of the drug pair. The mix of interaction types among new drugs differs from the mix among old ones, so forcing pair features to match pushes rare interaction types into the clusters of common ones.

This note proposes moving the adversary from the pair to the individual drug, and aligning drugs along a graph of chemical neighborhoods rather than all at once. The model has three parts. Each drug's molecule is encoded separately. The adversary sees only these single-drug embeddings, including those of new drugs whose structures are known but whose interactions are not. It is trained to recover which chemical cluster each drug came from, but only as far as the clusters' chemical adjacency allows. Interaction types are then decoded from pairs of these embeddings by relation-specific operators. The claim is that this removes the label-shift failure of pair-level adaptation without washing out real pharmacological differences between drug families. We test it on the distribution-shift DDI benchmark that explicitly called for domain adaptation.

Core Claim

In pairwise multi-type interaction classification under new-drug shift, adversarial alignment should act on the entity encoder, not on the pair or relation representation. It should also be structured by chemical-cluster topology rather than applied uniformly. Pair-level alignment, even with class conditioning, is entangled with the interaction-type distribution. That distribution shifts and is extremely long-tailed in this domain. Drug-level alignment targets the shift that actually defines the problem, which is new chemistry. A domain graph over chemical clusters keeps the adversary from erasing differences between drug families that are genuinely distinct. We predict three things: gains over source-only training that grow with shift magnitude; no loss, and ideally a gain, on rare interaction types; and robustness to artificially induced changes in interaction-type priors. Pair-level adversaries should show none of these properties.

Where This Sits

Three papers motivate the problem. Benchmarking drug–drug interaction prediction methods: a perspective of distribution changes builds DDI-Ben. It splits DrugBank and TWOSIDES by approval time and chemical cluster, uses a parameter $\gamma$ to control shift magnitude, and shows that standard multi-type predictors collapse as $\gamma$ grows. It names domain adaptation as the missing piece but proposes no method. On the robustness of generalization of drug-drug interaction models documents the same collapse on one-unseen and both-unseen splits, and finds that SMILES augmentation helps only slightly. Rethinking Drug-Drug Interaction Modeling as Generalizable Relation Learning (GenRel-DDI) argues that chemical proximity in embedding space does not track interaction labels. It responds with a relation-centric cross-attention architecture and uses no unlabeled target data. Our method is orthogonal to that architecture and can sit on top of it.

On the adaptation side, three papers supply the method and constrain it. Conditional Adversarial Domain Adaptation (CDAN) conditions the domain discriminator on classifier predictions to avoid aligning class A in one domain with class B in another. It assumes closed label sets with comparable class structure, and long-tailed polypharmacy labels violate that assumption. That makes CDAN our main pair-level competitor rather than our mechanism. Graph-Relational Domain Adaptation (GRDA) replaces the binary domain discriminator with one that reconstructs a known graph over domains. Its encoder then only needs to make neighboring domains indistinguishable. GRDA shows that uniform alignment (DANN, ADDA) can do worse than no adaptation when domains have topology. We reuse its graph discriminator, but we build the domain graph automatically from chemical clusters and apply it to drug entities inside a pairwise predictor. GRDA left both of these open. Bridging Source and Target Domains via Link Prediction for Unsupervised Domain Adaptation on Graphs (CMPGNN) shows that late feature alignment fails under label shift on graphs. We take this as a negative result constraining design. Its own fix, inserting bridge edges, assumes homophily, and that assumption breaks when edges carry opposite meanings such as enhancive versus degressive (TMFUF).

The closest use of adversarial training on DDI is Drug-drug interaction prediction with Wasserstein Adversarial Autoencoder-based knowledge graph embeddings. It uses an adversary to generate hard negative triples within a single domain. That is a different role entirely, with no cross-domain alignment and no molecular structure.

Research Program

Encoder and decoder. The backbone encodes each drug $u$ from its molecular graph into an embedding $z_u = E(u)$, using a GIN encoder as in the robustness study. Interaction scores follow the relation-specific bilinear form popularized by Decagon, $s(u,r,v) = z_u^\top M_r z_v$, where $M_r$ is a learned matrix for interaction type $r$. A second instantiation replaces the decoder with GenRel-DDI's relation trunk, so we can check that the adversary adds value on top of the strongest architecture.

Domain graph. Drugs are grouped into the chemical/approval-era clusters that DDI-Ben already uses to define its splits. We connect clusters $c$ and $c'$ in a graph $A$ when the mean Tanimoto similarity between their members' fingerprints exceeds a threshold. The threshold is chosen once on source-only data and swept in an ablation. Target clusters, meaning new drugs whose structures are available but whose interactions are unlabeled, are nodes of this same graph. This matches how drugs actually arrive: structure is known at approval, while interaction reports accumulate later.

Objective. We train

$$\min_{E,\{M_r\}} \max_{D} \; \mathcal{L}_{\text{type}}(E, M; \mathcal{P}_s) \;-\; \lambda \, \mathcal{L}_{\text{graph}}(D; E, \mathcal{U}_s \cup \mathcal{U}_t),$$

where $\mathcal{L}_{\text{type}}$ is the class-balanced interaction-type loss on labeled source pairs $\mathcal{P}_s$. $\mathcal{L}_{\text{graph}}$ is the cross-entropy of a discriminator $D$ that, given two drug embeddings $z_u, z_v$ drawn from all source and target drugs $\mathcal{U}_s \cup \mathcal{U}_t$, predicts whether their clusters are adjacent in $A$. The weight $\lambda$ controls the strength of adversarial alignment. Training uses a gradient reversal layer, as in standard domain-adversarial training.

The design choice that matters is what the adversary sees. It never sees pair features or interaction predictions. Its objective therefore does not depend on how interaction types are distributed among new drugs. This is the source of the predicted robustness to label shift.

Analysis. Beyond accuracy, we will measure three properties of the learned embeddings. The first is how well new drugs' nearest source neighbors in embedding space predict their true interaction profiles. The second is whether rare-type clusters in pair space stay separated after training, the over-smoothing failure CMPGNN warns about. The third is how the gain varies with each new cluster's graph distance to the nearest labeled cluster.

Baselines

The naive reference point is source-only training of the same encoder and bilinear decoder with class-balanced loss and no unlabeled target drugs. Every adaptation gain is measured against it, at each $\gamma$ level.

The strongest existing method on entity-disjoint splits is GenRel-DDI. It reports AUPR 57.6 on the MeTDDI S3 split, against a previous best of 53.7 (GenRel-DDI). We will run it with its released configuration on MeTDDI and DDInter, and retrain it on DDI-Ben splits with the same tuning budget as our model: an equal number of hyperparameter trials, the same validation split drawn from source clusters only, and matched GPU hours. Beating it means at least matching 57.6 AUPR on MeTDDI S3 when our adversary is added to the GenRel backbone, plus a statistically significant Macro-F1 gain over seeds on the high-$\gamma$ DDI-Ben settings.

On DDI-Ben, the paper reports that the text-based LLM models TextDDI and DDI-GPT degrade least as $\gamma$ grows. The file gives no specific numbers, so no reported result was found on Lacuna for TextDDI or DDI-GPT Macro-F1 on DDI-Ben. We will rerun them on the benchmark's own code and splits.

The adaptation baselines isolate the key ingredient. Pair-level DANN puts a binary domain adversary on the concatenated pair embedding. Pair-level CDAN+E conditions that adversary on the pair embedding crossed with predicted type probabilities (CDAN). For reference, CDAN+E reported 87.7% versus DANN's 82.2% on Office-31, which is images, not drugs. Drug-level uniform DANN places the adversary on single-drug embeddings with a binary old/new label. Drug-level GRDA with a random graph of the same density checks whether the chemical structure of $A$ matters rather than graph sparsity alone. All variants share the encoder, decoder, $\lambda$ schedule and tuning budget. The comparison between pair-level CDAN+E and drug-level uniform DANN tests where to align. The comparison between drug-level uniform and drug-level chemical-graph alignment tests how to align.

Why not just…?

Why not just use a bigger pretrained molecular model or an LLM? GenRel-DDI reports diminishing returns from scaling to models such as MolT5-340M on entity-disjoint splits, with a smaller relation-centric model winning. DDI-Ben reports that text LLMs, though more robust than MLPs, still degrade as $\gamma$ rises. We include TextDDI and DDI-GPT as baselines. If they match our method at high $\gamma$, the adversarial component is not worth its cost, and we will say so.

Why not just use GenRel-DDI's architecture with no adaptation? It uses no unlabeled new-drug structures, which are available at approval. Our method can be added to GenRel-DDI, so the honest test is GenRel with and without the drug-level adversary. If the adversary adds nothing on top of GenRel, the claim fails for strong backbones.

Why not just apply CDAN to pair embeddings? CDAN conditions on predicted classes. For rare interaction types those predictions are unreliable, and the type mix among new drugs differs from source. We predict that pair-level CDAN+E will improve head classes while hurting the 1–10-instance bucket. MRCGNN and SumGNN both report rare-type Macro-F1 or F1 separately, and we adopt that bucketed reporting. The induced-label-shift experiment below settles this directly.

Why not just pseudo-label target pairs and self-train? This is a real threat. Simple confidence-thresholded self-training on unlabeled old–new pairs might capture most of the gain. Its likely weakness is confirmation bias toward head types, but that is a prediction, not evidence. We will run class-balanced self-training under the same budget, and also combine it with our adversary. If self-training alone matches us, the contribution shrinks to a cheaper regularizer.

Why not just include new drugs in self-supervised pretraining? Contrastive pretraining over source and target molecules also exposes the encoder to new chemistry without any adversary. We will pretrain the GIN encoder this way, as a baseline and as an initialization for our method. If it closes the gap, the adversary is unnecessary.

Why not just augment SMILES? The robustness study found only minor relief on both-unseen splits (On the robustness of generalization of drug-drug interaction models). We include it as a cheap control.

Proposed Experiments

The primary testbed is DDI-Ben. We use its DrugBank multi-class setting (86 types) and its TWOSIDES multi-label setting, on both S1 (known–new pairs) and S2 (new–new pairs), sweeping $\gamma$ across the benchmark's range. Multi-class metrics are accuracy, Macro-F1 and Cohen's kappa, with Macro-F1 additionally broken out by training-frequency bucket. Multi-label metrics are per-type AUROC and AUPR averaged over side effects. The secondary testbed is MeTDDI and DDInter S1–S3, to compare against GenRel-DDI's reported number. All results use five seeds with paired significance tests.

Three checks decide the claim. First, the gain of drug-level graph alignment over source-only should grow with $\gamma$ and be positive in the rare bucket, while pair-level CDAN+E should not improve, or should degrade, the rare bucket. Second, in an induced label shift test, we resample target test pairs to skew interaction-type priors at several strengths. Drug-level methods should be largely flat across this resampling, and pair-level methods should drop. This experiment most directly tests the mechanism. Third, replacing the chemical-cluster graph with a density-matched random graph, or with uniform alignment, should remove most of the gain on high-$\gamma$ splits. That would replicate GRDA's negative-transfer finding in a molecular setting.

If none of the three checks holds, the result is still informative. It would show that the failure DDI-Ben exposes cannot be repaired by aligning unlabeled molecular structure, which would point the field back toward relation-centric architectures or text priors.