When the Graph Knows Little, Repurposing Models Recommend Famous Drugs

Computational drug repurposing searches existing, approved drugs for new uses by predicting missing links in a biomedical knowledge graph, a network of drugs, diseases, proteins and pathways assembled largely from the published literature. Its main appeal is speed. When a new outbreak or an untreated rare disease appears, it can produce a candidate list in days instead of years. But the graph records where scientists have looked, not only what is biologically true. Heavily studied drugs have thousands of edges and neglected ones have a handful. Graph neural networks, which build each node's representation by repeatedly averaging over its neighbours, are known to turn this unevenness into something worse. This proposal makes one testable claim. The less a graph knows about a disease, the more a repurposing model's ranking for that disease collapses into a list of the most-published drugs. The early, data-poor moment when repurposing matters most is therefore exactly when this inherited research bias is amplified most.

We propose to measure this effect as a function of how much disease-specific evidence the graph contains. We will split the measured bias into the part already present in the data (inherited) and the excess the model adds (amplified), and attribute that excess to specific pipeline stages: graph construction, message-passing training, and final ranking. We then test whether a mitigation targeted at the stage that actually produces the excess beats cheap post-hoc reranking. If reranking matches it, we will report that plainly. In that case the useful contribution is the measurement protocol and the evidence that a calibrated reranker suffices.

Core Claim

Three findings motivate the claim. Chen et al. (Phenotype-Conditioned Drug Repurposing for Undiagnosed Rare Disease Patients via Graph Neural Networks and LLM Hybridization) built a relational graph network over PrimeKG fused with a GPT-4o text branch. They state that the network consistently ranks high-degree, heavily studied drugs higher, and that its attention weights track graph density rather than biology. They apply no mitigation. Separately, work on recommender systems (How Graph Convolutions Amplify Popularity Bias for Recommendation?) proves that a node's gradient influence on graph-convolution weights scales as $d_i^{3/2}$ in its degree $d_i$. This pulls embeddings toward popular items, so tail recall falls as depth grows. That paper supplies the mechanism we will test, but it studied only bipartite user–item graphs. Finally, the LLM stereotyping literature (Uncovering Stereotypes in Large Language Models: A Task Complexity-based Approach) shows that models fall back on ingrained statistical defaults when the task outruns the evidence. We borrow this as an analogy rather than a method: a zero-shot disease with few edges is the graph-model version of an overloaded prompt.

Our hypothesis combines these. Let $e$ be the fraction of a disease's neighbourhood edges retained, which simulates "days since emergence." We expect the rank correlation between a drug's predicted score and its degree to rise as $e$ falls, and to rise faster for message-passing models than for non-message-passing embeddings. The direction of the effect comes from the $d^{3/2}$ argument. When the disease-side signal is weak, the drug-side representation dominates the score, and that representation is shaped mostly by gradient flow from hubs.

There is one case we must handle head on: hubs are sometimes right. Dexamethasone, a heavily connected drug, was the top COVID-19 candidate in both TeX-Graph (TeX-Graph: Coupled tensor-matrix knowledge-graph embedding for COVID-19 drug repurposing) and the complemented-network study (Development of complemented comprehensive networks for rapid screening of repurposable drugs applicable to new emerging disease outbreaks). For that reason we do not treat preferring hubs as bias in itself. Following Directional Bias Amplification, which separates bias present in base rates from bias a model adds, we measure amplification against the degree distribution of the true held-out indications.

Let drugs be grouped into degree quantile bins $b$. Let $p_{\text{true}}(b)$ be the share of held-out true indications whose drug falls in bin $b$, and $p_{\text{pred}}(b)$ the share of the model's top-$k$ predictions in bin $b$. We define amplification as

$$\text{Amp}(b, e) = p_{\text{pred}}(b \mid e) - p_{\text{true}}(b),$$

computed at evidence level $e$. A positive value in the top bin means the model over-recommends famous drugs beyond what the ground truth, itself literature-biased, already contains. The held-out labels come from the same biased literature, so this measure is conservative: it detects only excess bias, never the inherited part.

Research Program

The first component is an amplification ledger, which records Amp after each pipeline stage. Graph construction contributes inheritance, which we measure as the degree skew of $p_{\text{true}}$ itself. Training contributes representation amplification. We measure it with a linear probe on frozen drug embeddings: we ask whether a tail drug's true indications are separable from non-indications at all. Ranking contributes decision amplification, the remaining gap between probe-recoverable signal and what actually enters the top-$k$. This split matters because it determines which fix can work. If tail signal survives in the embeddings but is buried at ranking time, a reranker is enough. If the embeddings have collapsed, no reranker can recover it.

The second component is a stage-targeted mitigation, applied only where the ledger shows excess. Before training, we downweight literature-derived edge types (co-mention and text-mined relations) relative to curated biological edges. We do not prune hubs, because hubs carry shared biology. During training, we rescale each node's loss contribution by $d_i^{-1/2}$. This is meant to cancel the extra degree factor in gradient influence while keeping the usual graph normalisation. After training, we calibrate a drug-degree prior on validation diseases so that $p_{\text{pred}}$ matches $p_{\text{true}}$. This step adapts the base-rate threshold calibration of Directional Bias Amplification from per-class thresholds to ranked lists.

As a secondary measurement, we run the ledger on Chen et al.'s score-level fusion. The aim is to see whether the language-model branch dampens or adds to popularity skew. Their own analysis already notes that the language model may rationalize reporting bias.

Baselines

The naive reference is the unmodified backbone: TxGNN (A foundation model for clinician-centered drug repurposing) on its zero-shot disease split, plus Chen et al.'s fusion model on their 108 held-out diseases. For Chen et al. the file reports MRR 0.325 overall and 0.311 on undiagnosable cases. TxGNN is reported only as relative gains (49.2% on indications over eight baselines), so we will rerun it to obtain absolute numbers. We found no reported tail-stratified or degree-stratified result on Lacuna for any PrimeKG model. Establishing that number is part of the contribution.

The strongest single competitor is degree-calibrated post-hoc reranking, which is our own final stage run alone. It is the cheapest method and the one most likely to match us. On vision and language benchmarks, base-rate calibration alone cut measured amplification from 0.040 to 0.001 (Directional Bias Amplification).

The second competitor is DAP, the post-hoc method from How Graph Convolutions Amplify Popularity Bias for Recommendation?. It subtracts cluster-estimated "popularity gravity" from embeddings layer by layer. On recommendation tail sets it reports +23.96% recall on Gowalla and gains above 50% on Yelp2018. No reported result exists on Lacuna for DAP on biomedical knowledge graphs. We will port it with per-relation clustering.

The third competitor is ComplEx, a non-message-passing embedding, run as in Knowledge Graph Embeddings in the Biomedical Domain: Are They Useful?, which reports HITS@10 0.793 on BioKG. ComplEx tests whether amplification is specific to message passing at all.

All methods share the same backbone checkpoint where applicable, the same splits, and an equal hyperparameter budget of 30 trials, selected on validation MRR subject to a ceiling on Amp. The ablations remove one stage at a time.

Beating the reranker means two things. Tail-bin recall@50 must be higher at matched overall MRR, with the overall-MRR drop under 2% relative to the unmodified backbone. The gain must also be concentrated at low evidence levels $e$.

Why not just…?

Why not just rerank by a degree penalty? This is the real threat. For it to fall short, the probe must show that tail drugs' indication signal is lost inside the embeddings, not merely outranked. A reranker cannot recover signal that is absent from the representation. The ledger measures this directly. If probe accuracy on tail drugs is near the hub level, then reranking is the right answer, and we will recommend it over in-training changes.

Why not just apply DAP? DAP assumes one relation type. PrimeKG and DRKG mix dozens of relation types (107 in DRKG), so a single subtracted popularity vector may erase relation-specific structure that mechanism paths depend on. We test this by measuring mechanism-path quality on DrugMechDB after DAP. If DAP preserves paths and matches our tail recall, it wins on cost.

Why not just prune hubs or mask literature edges? Pruning removes shared protein and pathway infrastructure, which is what zero-shot transfer relies on. Masking can move bias into proxy paths, as the masking paradox in Directional Bias Amplification shows. We include hub pruning as an ablation and expect overall MRR to fall more than tail recall rises.

Why not just reweight the loss by inverse degree? This is a fair single-stage alternative. Our leave-one-out ablation is exactly this test. If training-stage reweighting alone accounts for nearly all the gain, the claim narrows to "fix the gradients" and the extra stages are dropped.

Proposed Experiments

The primary experiment sweeps $e$ from 1.0 down to 0.05 on PrimeKG zero-shot diseases. For each value it reports the degree–score rank correlation, Amp in the top and bottom degree bins, overall MRR, and tail recall@50 for every baseline and ablation. A second experiment is a retrospective COVID-19 check on DRKG. TeX-Graph reports 10 of its top 100 drugs already in clinical trials at the time of the study. We will check whether mitigation keeps Dexamethasone near the top while adding trial-validated low-degree drugs. This is the key guard against debiasing that simply penalizes correct hubs. A third experiment, on RTX-KG2c, measures DrugMechDB mean percentile rank after each mitigation. It tests whether mechanism explanations survive, against the 94.7% reported for KGML-xDTD (Predicting Drug Repurposing Candidates and Their Mechanisms from A Biomedical Knowledge Graph).

Three results would decide the proposal. The hypothesis fails if the degree–score correlation does not rise as $e$ falls. The $d^{3/2}$ mechanism is not the driver if ComplEx amplifies as much as the graph networks do. The stage-targeted pipeline is unnecessary if the post-hoc reranker equals it on low-$e$ tail recall at matched MRR. Each outcome changes what we would recommend to anyone building a rapid-response repurposing system.