Pathogen genome databases and human genetic databases fail in the same way: they reflect where sequencing money is spent, not where biology happens. On the pathogen side, public repositories are dominated by isolates from well-funded urban outbreak centers, so under-sequenced regions look genetically depauperate and evolutionarily slow — an "optical illusion" in which wealthy hubs appear to be sources of innovation and poor regions appear to be passive sinks, as documented for SARS-CoV-2 subtype structure by Genetic grouping of SARS-CoV-2 coronavirus sequences using informative subtype markers for pandemic spread visualization. On the host side, over 80% of genome-wide association studies are European-only, so instruments and risk scores trained on them lose power or reverse sign in other ancestries. Both failures share a root cause: continuous geographic variation in both hosts and pathogens is collapsed into categorical bins (a country, a continent) that were defined by administrative and funding boundaries, and those bins are then mistaken for biology.
This proposal argues that the fix should also be shared: keep geography continuous end to end. Local ancestry inference — methods that partition an admixed genome into segments inherited from different ancestral populations — can now output not just labels but continuous latitude–longitude coordinates per chromosomal segment, cheaply and robustly: XGMix, a stacked gradient-boosting model, reached 97.98% accuracy on a three-way continental admixture benchmark of the 1000 Genomes Project (versus 93.00% for the standard RFMix) and still held 93.19% accuracy with 80% of markers masked, in about 30 minutes on a laptop (XGMix: Local-Ancestry Inference with Stacked XGBoost; Addressing Ancestry Disparities in Genomic Medicine: A Geographic-aware Algorithm). Nobody has ever connected these continuous coordinates to pathogen genomic surveillance. We propose to do exactly that: place hosts and pathogen lineages in one shared coordinate space, measure the distance between them, and use that distance both to re-estimate host–pathogen biological compatibility and to repair emergence forecasting in under-sequenced regions.
Core Claim
The core claim is twofold. First, the apparent geographic "specialism" of pathogen lineages is partly a sampling artifact that continuous coordinate modeling can separate from true co-evolutionary compatibility. Second, adding continuous host-ancestry features to epidemic forecasting models measurably improves predictions of whether an arriving variant will establish sustained transmission, precisely in the under-sequenced countries where current models fail — the "zero-start" problem that the current best forecasting architecture explicitly cannot solve.
The empirical hook is Host-pathogen sympatry and differential transmissibility of <i>Mycobacterium tuberculosis complex</i>, which analyzed 2,279 pulmonary tuberculosis index cases and 12,749 close contacts across New York City, Amsterdam, and Hamburg, with pathogen whole-genome sequencing covering nearly 300,000 single-nucleotide substitutions. It found that "specialist" MTBC lineages transmit poorly outside their ancestral host region: an allopatric host–pathogen combination showed a 32% reduction in odds of infection ($\text{OR}_{\text{adj}} = 0.68$), specialists had a lower adjusted incidence rate ratio ($\text{IRR}_{\text{adj}} = 0.86$), and in European-donor macrophages uptake of specialist lineage L6 was 2.9% versus 33.9% for generalist L4 — evidence for genuine cellular co-adaptation. But the study's host ancestry was a categorical country-of-birth proxy, and its own authors flag this as their central limitation: in multi-generation admixed urban cohorts, country of birth misclassifies exposure badly. Meanwhile, the pathogen side of that study is itself contaminated by surveillance bias: the MTBC pangenome analysis of The Mycobacterium tuberculosis complex pangenome is small and shaped by sub-lineage-specific regions of difference shows that public databases are dominated by outbreak lineages L2 and L4 (335 closed genomes, ~3,746 core and ~506 accessory genes), while endemic specialist lineages L5, L6, and L9 are barely sequenced. So the specialist/generalist dichotomy is estimated on a pathogen sample skewed toward exactly the lineages that circulate where sequencing happens. Both distortions — coarse host labels and skewed pathogen sampling — push in the same direction and have never been disentangled.
Research Program
The program has three linked components.
The first component builds the shared coordinate space. We run XGMix's continuous regression mode on host genomes to assign every individual an ancestry coordinate $a_i$ (latitude and longitude of ancestral origin, aggregated across chromosomal segments). Independently, we assign each pathogen lineage an endemic coordinate $p_\ell$, estimated from the geographic provenance of reference genomes — for MTBC, lineage labels and regions of difference from the 335-genome pangenome collection; for viral variants, the country-level prevalence trajectories used by the forecasting benchmark below. The key object is the sympatry distance $d(a_i, p_\ell)$, the great-circle distance between a host's ancestry coordinate and a pathogen lineage's endemic coordinate. We then re-fit the transmission model of the tuberculosis study with $d$ replacing the categorical sympatry/allopatry indicator:
$$\text{logit}\,P(\text{contact } j \text{ infected by case } i) = \alpha + \gamma^\top C_j + \beta_1\, d(a_j, p_{\ell(i)}) + \beta_2\, \mathbb{1}[\text{specialist lineage}] \cdot d(a_j, p_{\ell(i)}),$$
where $C_j$ are the same confounders the original study adjusted for (age, sex, smear positivity, homelessness, drug use, setting). If the co-evolution story is real, infection odds should fall monotonically with $d$ — including within the "allopatric" bin, and including for admixed individuals whose coordinates sit between ancestral poles, who should show intermediate susceptibility. The categorical design could never see this gradient; the Bats of a feather work on volant vertebrates shows the general danger: sampling discontinuities manufacture discrete "breaks" out of continuous clines. The same lens applies to the HIV case, where resistance-prediction tools calibrated on Western subtype B fail globally because subtype C follows distinct mutational paths to lopinavir resistance (Comparing mutational pathways to lopinavir resistance in HIV-1 subtypes B versus C) — a categorical-regime error that a continuous representation of viral and host geography would expose.
The second component quantifies the sampling artifact itself. We artificially downsample the over-represented generalist lineages (L2, L4) in the pangenome and surveillance collections to the sequencing density of L5/L6/L9, re-run lineage classification and terminal-branch-length analyses, and measure how much apparent "specialism" (long terminal branches, restricted geographic range, low diversity) is recovered in generalists purely from thinning. This yields a calibration curve: observed specialist signal as a function of sampling density. Any claim of true co-evolution must then survive correction by this curve. This directly operationalizes the open problem of separating biological gaps from collection voids.
The third component carries the coordinates into forecasting. The strongest current model for predicting when a viral variant arrives in a country is FA-GCN, a graph neural network over 87 countries and 36 variants whose inputs are epidemiological dynamics (log-prevalence ratios, global growth rates) and mobility/policy edges weighted by the Oxford COVID-19 Government Response Tracker; it achieved a Median Median Absolute Error of 1.27, beating a physics-informed regularized model (2.24) and a pure dynamics model (1.86) (Dynamics-Based Feature Augmentation of Graph Neural Networks for Variant Emergence Prediction). FA-GCN treats host populations as interchangeable. We augment it with per-country, per-pair ancestry features: the distribution of continuous host coordinates in each country (from 1000 Genomes–trained inference), and the sympatry distance between each variant's origin coordinate and each country's ancestry distribution. This is motivated by the Joint host-pathogen genomic analysis identifies hepatitis B virus mutations associated with human NTCP and HLA class I variation finding that host population genetics (the East-Asian-enriched NTCP S267F variant, HLA class I alleles) directly selects compensatory viral mutations — host ancestry is not noise in forecasting, it is a mechanism. The decisive question is whether sympatry features reduce error, and whether the reduction concentrates in low-sequencing countries where the zero-start problem bites.
A methodological contrast we will test rather than assert: Multi-ancestry polygenic risk scores using phylogenetic regularization (DendroPRS) shares effect sizes across human populations along a bifurcating phylogenetic tree and beat pooled and single-population baselines (BMI prediction correlation 0.14 in aggregated non-White UK Biobank groups versus 0.10 and 0.03). DendroPRS is the closest existing method for cross-population information sharing, but its tree assumption breaks precisely in the admixed urban cohorts where surveillance happens — the authors themselves note that reticulate gene flow requires graphs, not trees. Our continuous-coordinate representation is the admixture-native alternative, and we will run both as competing host-structure models on the transmission regression.
We must be honest about one data dependency: the tuberculosis municipal cohorts published country-of-birth labels, and we do not know from the available material whether host genotypes exist for those individuals. The plan is to pursue the harmonized cohorts directly; if host genotyping is unavailable or ethically restricted, we validate the full pipeline semi-synthetically — grafting 1000 Genomes admixture simulations (the Wright-Fisheradmixture mosaics already used to benchmark XGMix) onto the real transmission network topology — and reserve the fully real analysis for cohorts where paired host–pathogen sequencing exists, such as the 567-individual chronic hepatitis B exome-plus-viral-genome cohort used by the G2G study.
Baselines
The do-nothing reference for the transmission component is the published country-of-birth sympatry analysis itself: the same cohort, same generalized estimating equation and quasi-Poisson structure, same confounder set, with ancestry entered as categorical bins (Host-pathogen sympatry and differential transmissibility of <i>Mycobacterium tuberculosis complex</i>). Its reported allopatry penalty is $\text{OR}_{\text{adj}} = 0.68$ and specialist $\text{IRR}_{\text{adj}} = 0.86$; our continuous model beats it if it reproduces the direction of these effects, predicts infection odds at the individual-contact level with better calibration (area under the precision–recall curve on held-out transmission pairs), and — uniquely — recovers a monotone distance gradient and intermediate risk for admixed individuals that the categorical model cannot express by construction. The strongest structured competitor on the host side is a DendroPRS-style discrete-population effect-sharing model (Multi-ancestry polygenic risk scores using phylogenetic regularization), run with the same tuning budget; it will be fitted to the same contacts and lineages, with effect offsets shared along the host population tree. Beating it means lower out-of-sample transmission-prediction error in admixed individuals specifically, since tree-based offsets must assign each admixed person to one population.
For the forecasting component, the baseline to beat is FA-GCN exactly as published: the same 87-country, 36-variant benchmark, the same train/test splits, the same OxCGRT-weighted mobility edges and dynamic features, the same two-hour-scale compute and the same hyperparameter search budget, with our ancestry features as the only addition (Dynamics-Based Feature Augmentation of Graph Neural Networks for Variant Emergence Prediction). Its reported figure is a Median Median Absolute Error of 1.27 (with 1.86 for the pure dynamics model and 2.24 for the physics-informed variant); we beat it by lowering MedMedAE on the identical benchmark, and decisively by lowering it on the pre-registered subset of low-sequencing countries, where the authors concede the zero-start failure. The key ablation isolates our central ingredient: FA-GCN augmented with categorical continental-region features (one-hot region indicators per country) versus FA-GCN augmented with continuous ancestry distributions. If the continuous version wins, the gain is attributable to fine-scale and admixed structure, not merely to adding any ancestry information.
Why not just…?
Why not just stratify everything into three to five continental bins? Because categorical bins demonstrably destroy the signal we are after. The Kenyan genomes analyzed by the geographic-aware XGMix work split bimodally into local Kenyan ancestry and Cameroon-associated ancestry tracing the Bantu expansion — a real biological structure invisible under a single "African" label (Addressing Ancestry Disparities in Genomic Medicine: A Geographic-aware Algorithm). Coarse categories also fail clinically: European-derived diagnostic criteria under-detect monogenic diabetes in non-White groups, requiring ethnicity-specific thresholds (Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing). And stratification wastes power, since separate per-bin models cannot share information — the failure DendroPRS was built to fix. Still, this alternative is cheap and must be tested empirically; that is exactly what the categorical-versus-continuous ablation does.
Why not just pool all cohorts and adjust with principal-component covariates? Because pooling with linear covariates forces a single average effect and actively cancels real biology. In the DendroPRS evaluation, the pooled model collapsed BMI prediction correlation in non-White UK Biobank groups to 0.03, worse than even single-population models at 0.10, while tree-structured sharing reached 0.14; pooling also masked flip-flop variants like the MSRA polymorphism with opposite BMI effects in European and Chinese cohorts (Multi-ancestry polygenic risk scores using phylogenetic regularization). Our allopatry effect is precisely an interaction — host ancestry times pathogen lineage — and no linear covariate adjustment can represent it.
Why not just use mobility-only forecasting, as FA-GCN already does? This is the most serious objection, and we treat it as a genuine threat rather than a straw man. Mobility and growth-rate features already achieve MedMedAE 1.27, and for fast airborne viruses host population genetics may matter too little to register. The countervailing evidence is that transmission can fail even when mobility trivially succeeds: in the tuberculosis cohort, specialist strains exposed large numbers of close contacts in hyper-connected transit hubs and still transmitted at 32% lower odds, with macrophage uptake differing tenfold by donor ancestry — a biological filter no mobility matrix captures (Host-pathogen sympatry and differential transmissibility of <i>Mycobacterium tuberculosis complex</i>). The forecasting experiment settles this honestly: if ancestry features do not reduce MedMedAE below 1.27 under matched tuning, the mechanism hypothesis is wrong for that regime, and we will report that.
Why not just reweight the data by inverse sequencing volume? Reweighting corrects estimates of totals but not model misspecification. It cannot resurrect lineages that were never sequenced (L5, L6, L9 are largely absent from public collections, per the MTBC pangenome study), and it amplifies variance on the few remaining representatives. Our downsampling calibration experiment is the complementary tool: it measures how inferred structure — not just abundance — distorts with sampling density, which no reweighting scheme addresses.
Proposed Experiments and Decisive Checks
The experiments run on four datasets from the compiled material: the harmonized tuberculosis surveillance cohort (2,279 index cases, 12,749 contacts, New York/Amsterdam/Hamburg) with pathogen whole-genome sequences; the 335-genome closed MTBC pangenome collection with lineage and region-of-difference annotations; the 1000 Genomes Project panel for training the continuous ancestry model (the same YRI/CHB/IBS benchmark on which XGMix reports 97.98% versus RFMix's 93.00%, plus its Wright–Fisher admixture simulations); and the FA-GCN forecasting benchmark of 87 countries and 36 variants with OxCGRT mobility edges.
Three checks decide the proposal. First, the distance-gradient check: in the re-fit transmission model, infection odds must decline monotonically with sympatry distance, and admixed individuals must show intermediate risk; a null or non-monotone result means the original allopatry finding was confounded or too coarse to generalize. Second, the sampling-artifact check: after thinning generalist lineages to specialist-level coverage, we measure what fraction of specialist signature (terminal branch length, restricted range) is reproduced; this number, currently unquantified anywhere in the literature — no reported result found on Lacuna for sampling-density correction of lineage classification — becomes the bias floor that any co-evolution claim must clear. Third, the forecasting check: FA-GCN plus continuous ancestry features must beat the published MedMedAE of 1.27 on the identical benchmark, with the decisive win coming in the pre-registered low-sequencing-country subset and in zero-start (pre-detection) predictions. As an external sanity anchor for the host-genetics mechanism, we will also verify on the 567-individual hepatitis B cohort that the NTCP S267F frequency structure from gnomAD, expressed in our coordinate space, predicts which host populations select preS1 escape mutations at positions 17, 35, and 51. If all three checks pass, the outcome is a reusable, admixture-native bridge between human ancestry inference and pathogen surveillance — the first pipeline in which the same continuous geography debiases both sides of the host–pathogen system at once.
References
- Genetic grouping of SARS-CoV-2 coronavirus sequences using informative subtype markers for pandemic spread visualization
- XGMix: Local-Ancestry Inference with Stacked XGBoost
- Addressing Ancestry Disparities in Genomic Medicine: A Geographic-aware Algorithm
- Host-pathogen sympatry and differential transmissibility of Mycobacterium tuberculosis complex
- The Mycobacterium tuberculosis complex pangenome is small and shaped by sub-lineage-specific regions of difference
- Bats of a feather: Range characteristics and wing morphology predict phylogeographic breaks in volant vertebrates
- Comparing mutational pathways to lopinavir resistance in HIV-1 subtypes B versus C
- Dynamics-Based Feature Augmentation of Graph Neural Networks for Variant Emergence Prediction
- Joint host-pathogen genomic analysis identifies hepatitis B virus mutations associated with human NTCP and HLA class I variation
- Multi-ancestry polygenic risk scores using phylogenetic regularization
- Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing