Deep models can now read a stretch of DNA and predict how a single letter change outside protein-coding genes will perturb gene regulation. The best of them, such as AlphaGenome — a convolutional-transformer trained on 1-megabase sequence windows to predict gene expression, chromatin accessibility, and other molecular readouts — are trained on both human and mouse genomes, so the same variant can in principle be scored twice: once against the human regulatory program and once against the mouse program. This proposal starts from a simple observation: nobody uses the disagreement between those two predictions, even though it is free to compute and even though the field's most concrete empirical finding about non-coding variants is that their effects flip direction between human and mouse contexts (Context-dependent regulatory variants in Alzheimer's disease). We propose to treat cross-species concordance as a per-variant confidence signal, validate it against the one dataset where human-versus-mouse regulatory effects were measured directly, and then carry it downstream into pathway enrichment — the step where lists of variant-to-gene assignments get summarized into biological mechanisms. At each stage we ask whether the concordance signal earns its keep against simpler ways of ranking and weighting variants, because the literature is full of cases where a simple baseline quietly beats a sophisticated one (Probabilistic modeling methods for cell-free DNA methylation based cancer classification).
The opportunity sits at a genuine hole in the literature. AlphaGenome predicts 5,930 human and 1,128 mouse tracks but never evaluates when and why mouse-track predictions diverge from human-track predictions — its own authors flag condition-specific effects as an open limitation. GeneCompass measures cross-species transfer (a 7.5% zero-shot cross-species cell-typing boost) but only across 17,465 protein-coding orthologs, with no view of non-coding regulatory sequence at all. The Alzheimer's MPRA study measured the human/mouse discordance empirically for 599 variants but did not benchmark predictors against it or push downstream. And the enrichment benchmark of Ma et al. covers only binary differentially-expressed gene lists, never the probabilistic, distance-biased gene sets that variant-effect models actually emit. No existing work connects these pieces; that is what this proposal does.
Core Claim
Let $\Delta_h(v)$ and $\Delta_m(v)$ be a model's predicted effect of non-coding variant $v$ on a matched target-gene expression track in human and in mouse respectively. We define a concordance score
$$ c(v) \;=\; \mathbb{1}\big[\mathrm{sign}(\Delta_h) = \mathrm{sign}(\Delta_m)\big] \cdot \sqrt{|\Delta_h|\,|\Delta_m|}, $$
the sign agreement between species weighted by the geometric mean of the two predicted magnitudes, so that agreement on a strong effect counts more than agreement on noise. The claim has three parts, each falsifiable on its own. First, $c(v)$ separates causal variants from matched decoys better than either species' score alone, because it filters out predictions that are artifacts of one species' regulatory grammar. Second, $c(v)$ predicts which variants will show discordant effects in actual cross-context MPRA experiments — the 599 late-onset Alzheimer's variants assayed in human macrophage states versus mouse brain regions are the test case. Third, when variant-to-gene assignments are weighted by $c(v)$ and fed into a topology-aware enrichment analysis, recovered pathways are more accurate than under unweighted or thresholded lists, because the weighted analysis acknowledges that predicted target sets have structured, non-uniform error.
A crucial design choice, taken from the FusorSV lesson: we will not fuse model scores globally. FusorSV showed that variant callers switch from useless to essential depending on variant class and size bin, and that a context-binned combination recovered 843 validated structural variants that naive unions and intersections missed. TraitGym found the analogous regime structure for variant-effect models: alignment-based scores dominate where purifying selection is strong (CADD reached an AUPRC of 0.85 on Mendelian traits), supervised functional-genomics models win on complex traits, and the largest language model lagged both and failed specifically on enhancer variants. So the concordance machinery will be fitted and evaluated per regulatory context — 5' UTR, promoter, distal enhancer, gene desert — not pooled.
Research Program
The first component builds the concordance benchmark. Starting from the TraitGym variant set (338 Mendelian causal variants and 1,140 complex-trait causal candidates, each with nine locus-matched decoys), we score every variant and decoy with AlphaGenome against matched human and mouse tracks in two modes: at the orthologous mouse locus (the evolutionarily honest comparison) and with the human sequence transplanted into the mouse-track model (which exactly mirrors what an in vivo mouse MPRA does, and is therefore the comparison that measures preclinical fidelity). The output is a genome-wide map of where cross-species regulatory predictions agree and diverge, stratified by regulatory class — something no paper in the file has produced.
The second component validates the signal against real cross-context measurements. The Alzheimer's study (Context-dependent regulatory variants in Alzheimer's disease) measured 599 variants in four THP-1 macrophage activation states and three mouse brain regions, finding variants with opposite effects across contexts, stronger immune-context effects for common variants, and stronger brain-context effects for rare ones. If predicted discordance between human and mouse tracks anticipates which variants behave this way, the concordance score is measuring something real rather than a modeling quirk; the file reports no quantitative concordance number for that study ("high reproducibility, and machine learning predictions largely agreed with experimental data"), so we will compute one.
The third component carries the signal into enrichment. Deep predictors emit weighted gene sets with heavy proximity bias — a known failure mode, since the osteoporosis foundation-model work exists precisely because nearest-gene assignment misidentifies effector genes (A DNA foundation model predicts osteoporosis risk genes without proximity bias), and since CRISPRi silencing of one Alzheimer's enhancer upregulated genes megabases away while the nearest genes stayed put. Standard overrepresentation analysis treats gene lists as unstructured binary sets; topology-aware methods like NetGSA model pathway wiring but still expect uniform input quality. We extend a topology-based enrichment test (following the regularized-regression approach of Coherent pathway enrichment estimation by modeling inter-pathway dependencies) to accept per-gene weights derived from variant effect times concordance, and we quantify what happens to power and false-positive rates when the input weights are wrong — the direction Ma et al. established for binary inputs but never for predicted ones.
Proposed Experiments
Experiment one re-scores the full TraitGym benchmark with the concordance machinery, per regulatory bin, and asks whether $c(v)$-weighted fusion beats the best single model on both the Mendelian and complex-trait splits. The target to beat is explicit: CADD achieved an AUPRC of 0.85 on Mendelian traits (TraitGym); for complex traits the file attributes superiority to Enformer and Borzoi but reports no number, so we state plainly: no reported result found on Lacuna for a complex-trait AUPRC of any TraitGym model, and we will measure these ourselves under the paper's protocol. A secondary analysis measures how much of $c(v)$'s value is just evolutionary conservation by correlating it with CADD and GPN-MSA scores; if concordance collapses into conservation, its value should appear only on complex traits where conservation signals are weak — a built-in test of redundancy.
Experiment two evaluates concordance against the Alzheimer's MPRA panel. Using the variant-by-context effect matrix from that study, we test whether variants with low predicted concordance are disproportionately the ones showing opposite-direction effects between human immune contexts and mouse brain contexts, and whether the common-versus-rare split it reported tracks the same signal. Metrics are rank correlation between predicted and measured cross-context discordance and a classification AUPRC for predicting "context-flipping" variants, with the in vivo mouse measurements treated as the fidelity reference rather than the other way round — for the reasons in the next section.
Experiment three runs the enrichment benchmark. We generate weighted target gene sets for TraitGym complex-trait variants and the Alzheimer's panel, then evaluate three pipelines on the same KEGG pathway universe used in the Ma et al. setting (TCGA breast cancer with 114 pathways, TCGA prostate cancer with 112 pathways): Fisher's exact overrepresentation on thresholded nearest genes, NetGSA-style topology enrichment on the binary top-k target list, and concordance-weighted topology enrichment. Power and false-positive rates are estimated by sample-label permutation exactly as Ma et al. did, and recovery is scored against known disease-relevant pathway structure — immune and B-cell programs for the Alzheimer's panel, which is where that study located the common-variant effects and where the multiple-sclerosis epigenomics work independently placed non-coding risk loci (Genomic Regions Associated with Multiple Sclerosis Are Active in B Cells). No reported result found on Lacuna for a NetGSA power or precision value on these datasets, so all comparisons are run fresh on identical inputs.
Baselines
The do-nothing reference is the proximity heuristic plus standard overrepresentation analysis: assign each non-coding variant to the nearest transcription start site, threshold the resulting gene list, and run Fisher's exact tests against GO and KEGG. This is the analysis most GWAS papers still run, and it is the cheapest thing a reader could ask for. It fails on documented grounds — linear proximity mischaracterizes true targets, as the distal GSAP and SSH1 upregulation in the Alzheimer's CRISPRi experiment showed — but it must be beaten empirically, on the same variant sets and pathway universe, with false-positive rates estimated by the same permutation procedure.
The strongest existing competitor is two-part, because the proposal spans two tasks. For variant prioritization it is CADD, which achieved an AUPRC of 0.85 on TraitGym's Mendelian split and whose alignment-based scores the same benchmark crowned on complex-trait enhancers' opposite regime (TraitGym); for the enrichment task it is NetGSA, the network-regularized linear model that Ma et al. found superior on small, tightly coordinated metabolic pathways. Both will be run with the same input variant sets, the same KEGG graphs, and the same per-bin hyperparameter tuning budget as the proposed method, with tuning done on held-out chromosomes so that no method sees its test variants in advance. Beating the baseline means a higher AUPRC than 0.85 on the Mendelian split with a statistically separable gain on the complex split, and, on the enrichment side, strictly higher permutation-calibrated power at matched false-positive rates across both TCGA settings.
The key ablation replaces the concordance weight with three alternatives on otherwise identical pipelines: no weighting (binary list), the model's own predicted effect magnitude (a within-species confidence weight), and the disagreement between two human-context models (an ensemble-variance weight). This isolates whether the species axis specifically — the preclinical-fidelity ingredient — is doing the work, or whether any confidence weighting would do. If the ensemble-variance ablation matches the concordance weight, the cross-species story collapses and the honest conclusion is a cheaper within-species method; we would report that.
Why not just…?
Why not just pick the best model per trait class? TraitGym already tells us CADD wins on Mendelian loci and supervised functional models win on complex ones, so a reviewer will say: apply that rule and skip the concordance machinery. The reply is that trait class is a benchmark convenience, not a locus-level observable — a clinic does not know whether an unsolved variant sits in a Mendelian or complex-trait regime, and TraitGym's own finer-grained finding is that ranking flips by regulatory class (enhancer versus 5' UTR), which is observable per locus but not the same axis. Our experiment one includes a per-bin oracle-selection baseline to test this directly, and the FusorSV result gives reason to expect fusion with a context filter to beat per-context selection, as binned caller fusion beat every individual caller including the per-bin best. If oracle selection wins, we will say so.
Why not just run the mouse experiment? The skeptical position is that in vivo measurement is the gold standard and computation is a hypothesis. But the entire point is that the mouse readout is the thing whose fidelity is in question: the Alzheimer's MPRA authors themselves warn that human regulatory sequences in mouse brains introduce cross-species artifacts from divergent mouse transcription factors, and the broader preclinical literature — divergent dendritic-cell receptors (Xu (2025)), divergent epitope prioritization on the identical SARS-CoV-2 Spike protein (Sung (2022)), poor xenograft-to-clinic correlation (Lin et al. (2024)) — shows the gold standard is itself miscalibrated. We use MPRA data as validation ground truth, not as a competitor, and the proposal's practical payoff is precisely to say which mouse measurements to trust before spending the money.
Why not just scale up a DNA language model? TraitGym tested this: Evo2 improved with parameter count yet still lagged alignment models and failed specifically on enhancer variants, while AlphaGenome achieved its 75%-to-80% auROC eQTL-sign gain with functional-genomics supervision, not raw scale. Scale without a species axis also offers no fidelity measurement at any size, and it leaves the enrichment-input problem untouched. The concordance score rides on already-trained models and costs roughly one extra forward pass per variant.
Decisive Checks
Three results would decide the proposal's fate. If the correlation between concordance and CADD-style conservation scores is high on complex traits where conservation is uninformative, the signal is redundant and the idea dies in experiment one. If predicted discordance fails to anticipate the context-flipping variants in the Alzheimer's panel better than effect magnitude alone, the signal does not measure fidelity and experiment two fails. If the ensemble-variance ablation matches the concordance weight in enrichment recovery, the cross-species ingredient is not the active one and the surviving contribution is a weighted-enrichment method that should say so plainly. Conversely, a per-bin concordance-weighted pipeline that clears CADD's 0.85 Mendelian AUPRC, gains on the complex split, and delivers higher power at controlled false-positive rates in both TCGA settings would establish cross-species disagreement as a usable, zero-cost error bar — for variant interpretation, for enrichment, and for deciding which mouse experiments deserve to run.
References
- Context-dependent regulatory variants in Alzheimer’s disease
- Probabilistic modeling methods for cell-free DNA methylation based cancer classification
- AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model
- GeneCompass: deciphering universal gene regulatory mechanisms with a knowledge-informed cross-species foundation model
- A comparative study of topology-based pathway enrichment analysis methods
- FusorSV: an algorithm for optimally combining data from multiple structural variation detection methods
- Benchmarking DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics
- A DNA foundation model predicts osteoporosis risk genes without proximity bias
- Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression
- Genomic Regions Associated with Multiple Sclerosis Are Active in B Cells
- Engineered virus-hunter vaccine overcomes HBV immune tolerance
- Detailed analysis of antibody responses to SARS-CoV-2 vaccination and infection in macaques
- Oncolytic activity of a coxsackievirus B3 strain in patient-derived cervical squamous cell carcinoma organoids and synergistic effect with paclitaxel