Computational biology ranks candidate links, such as "this microRNA is involved in this disease," so that wet labs know which ones to test first. These ranking models score very well on standard benchmarks. They also inherit a popularity problem familiar from recommender systems. Diseases that already have many documented associations, such as lung cancer or Alzheimer's, dominate the top of every candidate list. Rarely studied diseases almost never surface. Recommender-systems research treats the analogous problem as provider-side fairness: guaranteeing that sellers or artists (the "providers") get a fair share of visibility in ranked lists. The obvious move is to import those guarantees and give every disease equal visibility.
In biology, equal visibility is the wrong target, because diseases genuinely differ in how many molecular associations remain to be found. This proposal argues that the target should come from evidence that does not depend on how much has been published: genome-wide association studies (GWAS), which measure how much a disease's inherited risk is spread across genes. Polygenic risk scoring is the field that aggregates many small genetic effects into a single liability measure. We use its machinery to estimate, for each disease and each candidate molecule, how much genetic signal lies in that molecule's target genes. We then allocate top-of-list slots across diseases in proportion to this genetic evidence rather than to literature counts. The claim is testable prospectively. A model trained on an old association database, with genetics-anchored allocation, should recover more of the associations that were discovered later for under-studied diseases. It should do so without losing total discoveries, and it should beat both uniform-fairness re-ranking and simple popularity debiasing.
Core Claim
Provider-fair ranking assumes the relevance scores it redistributes are correct. The research file lists this as an open "proxy estimation problem" for fairness methods (Do et al.; FairRec). In disease-association prediction the scores are systematically wrong in a known direction. Models trained on literature-derived graphs confuse "well studied" with "relevant," and performance tracks training frequency, as FusionGDA's own analysis of DisGeNET shows (FusionGDA).
Our claim is that a second, literature-independent signal can set how much exposure each disease deserves. That signal is the disease's polygenic architecture, aggregated over each candidate's target genes in the way pathway-restricted risk scores are built. Exposure allocation anchored to this signal should improve prospective discovery on long-tail diseases more than any of three alternatives: equal-exposure fairness, inverse-degree reweighting, or simply adding the genetic signal as a model feature. What is new is neither fairness in biology nor genetics in link prediction taken separately. It is using polygenic evidence to define the fairness target itself, and testing that target against future discoveries rather than against held-out edges drawn from the same biased snapshot.
Research Program
The primary setting is microRNA–disease association. A microRNA is a short RNA that silences genes, so its disease relevance flows through the genes it targets. This makes the link to genetics concrete. The base scorer is trained on HMDD v2.0, the snapshot used by KATZNCP and STIM. The later databases HMDD v3.2 and dbDEMC 2.0, which KATZNCP used only for anecdotal case studies, serve as a prospective test set of associations discovered after the snapshot.
For each disease $d$ with public GWAS summary statistics, we compute a gene-level association signal. It is corrected for linkage disequilibrium, meaning the correlation between nearby genetic variants that otherwise causes the same signal to be counted twice. This follows the logic of LD-aware shrinkage scores such as LDpred, as used by Yap et al. and Dennis et al.. We then aggregate the signal over each microRNA's validated targets from miRTarBase. This mirrors how pathway-restricted scores subset variants to a biological subsystem, as Barbu et al. did for NETRIN1 signaling and Bustillo et al. for glutamatergic genes. Summing the target-set signal over all candidate microRNAs gives a disease-level estimate of genetic association mass. Subtracting the associations already known gives $M_d$, our estimate of how much remains undiscovered. We name the source of the summary statistics here because the research file lists no Lacuna resource for them.
Allocation is posed as a welfare problem over diseases. Let $E_d$ be the total position-discounted exposure that disease $d$ receives across all microRNAs' top-$k$ lists. Let $T_d = k \cdot n \cdot M_d / \sum_{d'} M_{d'}$ be its genetics-anchored target, where $n$ is the number of microRNAs queried. Let $s_{md}$ be the base model's score for microRNA $m$ and disease $d$. We choose rankings $\pi$ to maximize
$$\sum_{m,d} s_{md}\, e_{md}(\pi) \;+\; \gamma \sum_d \psi\!\left(E_d(\pi)/T_d\right),$$
where $e_{md}$ is the exposure that pair $(m,d)$ receives, $\psi$ is a concave function that rewards lifting the most under-served diseases first, and $\gamma$ trades relevance against fairness.
This is the Lorenz-dominance welfare formulation of Do et al., which we reuse together with its Frank-Wolfe solver. The one change is that exposure is normalized by a target $T_d$ instead of being compared as raw exposure. We use a concave welfare function rather than hard quotas because Do et al. showed hard equality constraints "level down": they destroy utility without helping the worst-off. We also track the consumer side, here the microRNAs whose lists are reshuffled. Leonhardt et al. showed that fairness costs concentrate on niche users: recommendation disparity rose from 0.2% to 20.1% under greedy diversification.
The secondary setting is gene–disease prediction on DisGeNET-EVAL with FusionGDA as the scorer. There, the genetic prior attaches directly to genes rather than through microRNA targets. This checks whether the effect survives when the prior is less indirect.
Baselines
The naive reference point is the base scorer's unmodified ranking. The main scorer is KATZNCP, which counts graph paths of all lengths with decaying weight and projects similarity profiles onto the association matrix. It reports a leave-one-out AUC of 0.9325 on HMDD v2.0 (KATZNCP). We also run the heterogeneous graph transformer plus gradient-boosting model HG-LGBM (HG-LGBM). It reports AUC 0.9757 ± 0.0032 on the microbe–disease benchmark HMDAD but no result on HMDD. For gene–disease, the reference is FusionGDA's DisGeNET-EVAL AUC of 0.9371 and AUPR of 0.9389 (FusionGDA). No reported result found on Lacuna for any of these models under a temporal HMDD v2.0→v3.2 split or stratified by disease degree. We therefore rerun them ourselves, and the cited AUCs serve only as a sanity check that our reimplementations are faithful.
The strongest fairness competitors are three uniform-target methods run on the same base scores. The first is the Lorenz welfare method with equal targets (Do et al.). The second is FairRec's round-robin allocation, which guarantees each provider a minimum share of exposure and keeps each consumer's list within one item of what they would envy in another's (FairRec). The third is FairSort, which re-ranks by pushing under-exposed providers up and over-exposed ones down, and binary-searches the strength of that push to keep ranking quality above a threshold (FairSort). FairSort reports NDCG kept within 0.95–1.0 on commercial data. The uniform Lorenz variant is the sharpest control because it differs from our method only in the target.
The strongest non-fairness competitor is inverse-degree reweighting, which down-weights popular diseases in the training loss. A second is feature fusion, where the target-set genetic signal is added as an input feature to HG-LGBM's LightGBM stage. The key ablation keeps the entire pipeline but randomly permutes $M_d$ across diseases. This isolates whether the genetic content of the target matters, as opposed to merely having a non-uniform target with the same spread.
Every method receives the same base scores and the same top-$k$ budget. Each gets the same number of hyperparameter trials, spent tuning $\gamma$ or its fairness knob on an inner split of v2.0 only. None ever sees v3.2 during tuning. Our method beats the field if two conditions hold together. It must recover significantly more v3.2-new associations in the bottom degree tercile of diseases than the best competitor. And its total count of v3.2-new hits at $k$ must stay within the confidence interval of the unmodified ranking.
Why not just…?
Why not just give every disease equal exposure with existing fairness re-rankers? Equal exposure spends discovery slots on diseases with little remaining biology. It can therefore lower total confirmed hits while inflating visibility for diseases that have nothing left to find, which is exactly the leveling-down pattern Do et al. documented. The comparison between uniform-target and genetics-target Lorenz allocation, with everything else fixed, settles this directly. If uniform targets match ours on long-tail prospective hits, the genetic anchor is unnecessary and we will report that.
Why not just debias the scorer with inverse-degree weighting? This is a real threat. It is cheap, and if degree bias is the whole problem it may suffice. Its weakness is that it pushes every rare disease upward by the same logic, with no notion of which rare diseases actually have undiscovered associations. The degree-stratified prospective evaluation will show whether reweighting's gains land on diseases with high $M_d$, as ours should, or scatter randomly. We also run reweighting combined with our allocation to test whether the two are complementary.
Why not just add the polygenic signal as a feature? This is the most serious alternative. Richer inputs are how the field has progressed, from STIM to HG-LGBM to FusionGDA. But a feature changes scores pair by pair and does not control how exposure concentrates across the whole list budget. A model with a genetic feature can still hand most slots to hub diseases, because popularity features dominate the fit. If feature fusion matches our long-tail and total hit rates, the allocation layer adds nothing, and the correct conclusion is "use genetics as a feature." The experiment is designed to reach that conclusion if it is true.
Proposed Experiments
The first experiment is a precondition test, and it can kill the idea. Across HMDD diseases with usable GWAS data, does $M_d$ computed from v2.0 correlate with the number of associations each disease gained by v3.2, after controlling for v2.0 degree? We use a partial rank correlation with a bootstrap interval. If there is no partial correlation, genetics does not predict future discovery, and the allocation hypothesis fails before any ranking is built.
The main experiment trains every scorer on v2.0 and applies each allocation method. It then counts top-$k$ predictions that appear in v3.2 or dbDEMC 2.0 but not in v2.0. Counts are reported overall, by disease-degree tercile, and as a Lorenz curve of per-disease hits, along with the Gini coefficient of exposure. On the consumer side, we report the Leonhardt-style score and recommendation disparity across microRNAs, split by how many targets each microRNA has. This checks that fairness costs do not pile onto niche microRNAs.
The decisive checks are the permuted-$M_d$ ablation, which must underperform the true prior on long-tail hits, and the head-to-head against feature fusion. The gene–disease replication on DisGeNET-EVAL uses degree-stratified AUPR and top-$k$ recall for low-frequency diseases.
One constraint is stated up front. The GWAS behind $M_d$ are mostly of European ancestry, and their portability to other populations is poor (PRS portability; Mexico FH cohort). The prior may therefore under-weight diseases whose burden falls mainly outside European populations. We flag diseases with little non-European GWAS coverage and report results separately for them, rather than claiming the anchor is neutral.
References
- Two-sided fairness in rankings via Lorenz dominance
- FairRec: Two-Sided Fairness for Personalized Recommendations in Two-Sided Platforms
- Heterogeneous biomedical entity representation learning for gene-disease association prediction
- KATZNCP: a miRNA-disease association prediction model integrating KATZ algorithm and network consistency projection
- Predicting Disease Related microRNA Based on Similarity and Topology
- Analysis of common genetic variation and rare CNVs in the Australian Autism Biobank
- Clinical laboratory test-wide association scan of polygenic scores identifies biomarkers of complex disease
- Association of Whole-Genome and NETRIN1 Signaling Pathway–Derived Polygenic Risk Scores for Major Depressive Disorder and White Matter Microstructure in the UK Biobank
- Risk-Conferring Glutamatergic Genes and Brain Glutamate Plus Glutamine in Schizophrenia
- User Fairness in Recommender Systems
- HG-LGBM: A Hybrid Model for Microbiome-Disease Prediction Based on Heterogeneous Networks and Gradient Boosting
- FairSort: Learning to Fair Rank for Personalized Recommendations in Two-Sided Platforms
- Polygenic risk score portability for common diseases across genetically diverse populations
- Analysis of a deeply-phenotyped familial hypercholesterolemia cohort from Mexico shows a role for both rare and common alleles across known dyslipidemia genes and reveals structural variation in a novel locus