A genomic distance can predict maladaptation—but only under the right geometry

A genomic offset can be a useful warning signal for future fitness loss without being a direct measurement of it. That distinction is the central result of Olivier François’s theoretical paper: squared genomic offset tracks average log-fitness loss under a specific set of assumptions, but the relationship may fail for the genetic architecture of any one real population. The result matters because conservation decisions increasingly need to anticipate which populations may struggle as climates shift. A genomic map can point toward adaptive mismatch; it cannot, by itself, tell us how much fitness a population will lose.

The paper also explains why predictions sometimes work strikingly well. One route is biological averaging: adaptation spread across many effective traits makes the consequences of any one genetic architecture resemble the average. Another is geometric: even a small number of traits can predict well if the environmental directions that matter most to the genomic offset also dominate the fitness response. The same logic shows why prediction can falter when direct environmental effects—such as plastic responses not captured by genetic associations—pull fitness in a different direction.

This is a theory paper, not an empirical test with field measurements. It does not establish that a particular species is at risk or estimate a real population’s fitness decline. Instead, it clarifies what genomic offset can reasonably mean, what kinds of evidence are needed to validate it, and why a high correlation is not proof that researchers have identified the traits or mechanisms behind adaptation.

The Science

Genotype–environment association, or GEA, asks whether genetic variants systematically differ along environmental gradients. A genomic offset takes those associations and asks how far a population’s predicted genetic profile would be displaced when its environment changes. Imagine a population moving from its current climate toward a warmer, drier one: a GEA model can translate that environmental change into a weighted genetic distance. François calls the squared version of that distance genomic offset.

The intuition is appealing. If genetic variation has been shaped by environmental conditions, then a large shift in the conditions associated with a population’s genetic profile might signal that it is poorly matched to its new surroundings. But distance is not fitness. Whether mismatch harms survival or reproduction depends on what genes do, which traits selection acts on, how traits interact, and whether organisms can respond directly to the environment.

To make those links explicit, the paper models adaptation through multiple latent traits. “Latent” means the traits are conceptual dimensions in the model, not necessarily named or measured characteristics such as body size or flowering time. Each gene-associated environmental response contributes to these traits, and stabilizing selection penalizes deviations from the local optimum. In other words, fitness falls as the organism’s trait profile moves away from the combination favored in its destination environment.

The central fitness quantity is negative log-relative fitness: the log-scale loss compared with the maximum fitness available in that environment. This is not the same as absolute fitness, nor necessarily the same as the log of average fitness across individuals. The distinction matters because variation among organisms can make an average of log losses differ from a log of average fitness.

The model separates two questions that are easy to blur. First: across many possible genetic architectures consistent with the GEA data, what is the expected fitness loss? Second: for one actual population, with its particular genetic effects, how well does offset predict loss? François calls the first “marginal” and the second “realized” or “quenched.” A clean relationship for an ensemble does not guarantee a clean relationship for one population.

The assumptions are important. The framework treats GEA coefficients as fixed and genetic effects on latent traits as uncertain, centered, and independent across loci. It also allows direct environmental effects on trait mismatch, which could include plasticity or other routes that do not pass through the modeled genetic associations. The paper’s exact average relationship requires that these direct effects be absent. Its concentration results additionally use Gaussian assumptions about allelic effects.

The paper’s first conceptual figure should be read as a map of the distinction: environmental change feeds into genomic offset, while fitness depends on latent traits, genetic effects, direct environmental effects, and selection. The offset is a measure of genetically informed environmental mismatch—not a direct reading of a phenotype or fitness.

What They Found

The first result is reassuring, but conditional. When direct environmental effects are absent, squared genomic offset is linearly related to expected negative log-relative-fitness loss across genetic architectures. The slope depends on the covariance of genetic effects and the strength and structure of selection. The relationship is exact for that expectation under the model, rather than merely a pattern that emerges in a particular simulation.

Yet the average can conceal substantial variation. A realized population’s genetic effects may produce a fitness-loss matrix that differs from the GEA metric. Therefore, a high expected relationship cannot be treated as a guarantee for every population. The paper’s distinction is especially consequential for applied forecasts: validation against an average or simulated ensemble may not establish prediction for a focal wild population.

A key determinant is effective adaptive dimensionality, written as $K_{\mathrm{eff}}$. It is not simply the number of traits someone measures or hypothesizes. It counts how many trait dimensions meaningfully contribute after accounting for genetic-effect covariance and selection. Several traits may amount to roughly one effective dimension if one dominates or if correlations and unequal selection weight concentrate the response.

Under the paper’s Gaussian model, the relative variance in log-fitness loss across genetic architectures equals $2/K_{\mathrm{eff}}$. If many dimensions contribute comparably, individual architectures cluster more tightly around the average prediction. If one dimension dominates, they do not. Crucially, adding more loci alone does not solve this problem: the result depends on effective trait dimensionality, not the number of genetic markers.

This chart can show the paper’s reported formula for relative variance against effective dimensionality, using selected integer values calculated directly from $2/K_{\mathrm{eff}}$. It is a visualization of the theoretical relationship, not observed biological data. At one effective dimension, relative variance is 2; at two it is 1; at four it is 0.5; and at eight it is 0.25. The trend illustrates why a diversified adaptive response can make predictions more stable.

A second result concerns the geometry of prediction across environmental changes. The paper defines predictive accuracy as the squared correlation between genomic offset and conditional mean log-fitness loss as environmental displacement varies. In whitened environmental coordinates—where differences in the natural scale and correlation of environmental variables have been accounted for—this accuracy is the squared Frobenius cosine between two matrices: the GEA metric and the realized fitness matrix. Put simply, prediction depends on whether the two measures emphasize the same environmental directions, and with similar relative weights.

This gives an alternative route to good prediction. A population might have only one or a few effective adaptive traits and still show strong predictive accuracy if those traits align with the dominant directions of genomic offset. Conversely, if offset spreads its weight across many equally important environmental directions while the fitness response can span only a few, a low-rank response cannot reproduce the entire offset pattern. The paper derives upper and lower bounds from the spectrum of the GEA metric and the rank of the fitness response.

A chart of the theoretical upper bound, ${\sum }_{i=1}^r{\lambda }_i^2/{\sum }_{i=1}^d{\lambda }_i^2$, can illustrate this geometric ceiling as the rank $r$ increases, for a chosen spectrum. But the supplied paper text gives no numerical example spectrum or empirical measurements. A chart with invented eigenvalues would misrepresent the evidence, so the result is best presented through the equation and the paper’s qualitative conclusion rather than fabricated plotted values.

Direct environmental effects complicate the story further. If plasticity or other environmental pathways affect trait mismatch, the complete fitness matrix includes direct effects and interactions between direct and genetic components. These terms disappear from the average interaction under the paper’s centered-effects assumption, but remain in a particular realized architecture. Alignment can preserve prediction; misalignment can weaken it. Therefore, even a robust association between offset and fitness does not identify whether the relationship runs through genetic adaptation, plasticity, or both.

The paper’s geometry result can be visualized as two environmental maps: one weighted by GEA and one by fitness loss. Their overlap—not merely the size of either map—determines predictive accuracy. The rank bound explains why a few adaptive dimensions can suffice when the dominant environmental directions coincide.

Why This Changes Things

The practical contribution is a sharper interpretation of genomic offset. It is best treated as a measure of genetically informed adaptive mismatch: a way of identifying where environmental change may pull conditions away from patterns associated with genetic variation. It is not a direct predictor of realized fitness, and it does not specify which trait will fail.

That distinction can make conservation science more useful, not less. A warning indicator need not be a complete forecast to inform priorities. If genomic offset highlights populations facing unusually large environmental shifts, managers can direct common-garden experiments, demographic monitoring, or assisted gene-flow assessments toward those populations. But decisions with high stakes should combine offset with evidence on survival, reproduction, plasticity, demography, and the environmental variables actually limiting the species.

The framework also helps interpret validation studies. If offset correlates with performance in a provenance experiment, that is meaningful evidence of predictive alignment in that setting. It does not establish that the same relationship will transfer to another population, time period, or environmental regime. Nor does correlation reveal effective adaptive dimensionality: high accuracy could result from many traits averaging together or from a few traits aligned with the dominant environmental gradients.

This is a useful caution against a common inferential leap. Predictive success and biological explanation are different achievements. A model can forecast a response because its geometry happens to match, even if the mechanisms remain unknown. Conversely, a biologically plausible mechanism can fail to yield reliable predictions if direct effects, population structure, or mismatch among environmental dimensions distort the relationship.

For climate adaptation, these distinctions matter because genomic tools can be attractive precisely where direct fitness data are scarce. The temptation is to convert a genetic distance into a crisp vulnerability score. François’s analysis argues for a more disciplined use: regard offset as a hypothesis about adaptive mismatch, then test whether it predicts fitness in the populations and environments where decisions will be made.

A useful evidence chart would compare observed genomic offset with measured fitness across real populations, but the paper provides no empirical dataset or reported population-level values. It would be misleading to manufacture one. The paper’s contribution is theoretical: its measurable findings are equations and conditions, not observed percentages of survival, fitness decline, or species vulnerability.

What’s Next

The next step is empirical discrimination among the mechanisms the theory identifies. Researchers can ask whether prediction improves when adaptation spans many effective dimensions, or whether a few environmental directions explain most of the alignment. That requires more than counting traits: it calls for estimates of genetic covariance, selection structure, and how environmental variation is distributed in the study region.

Common-garden and reciprocal-transplant studies can test whether offset predicts relative performance under changed conditions. Repeated measurements can help distinguish genetic effects from plastic responses, while demographic data can connect performance to population growth. These studies should validate predictions out of sample—across populations or conditions not used to build the GEA model—because a relationship fitted to one setting may not travel.

There are also limits in the theory itself. The allelic-effect distribution is an uncertainty model, not a complete account of how evolution generated genetic architectures. Real genetic effects may depend on GEA coefficients; population structure and demographic history can confound associations; selection may change across environments; and environmental variability may alter residual variation. The model targets mean log-relative-fitness loss, not every individual’s outcome or absolute population persistence.

The paper does not offer a universal calibration from a unit of genomic offset to a percentage drop in survival or reproduction. Its strength is more basic: it identifies why such a conversion might work, why it might fail, and what underlying structure determines the difference. The argument shifts the question from “Is genomic offset predictive?” to “For which population, under what architecture, and along which environmental directions?”

That is a more demanding question—and a more useful one. As climate conditions move beyond the environments in which populations evolved, genomic offset can help flag where genetic mismatch deserves attention. The theory’s message is that the signal becomes decision-worthy when it is treated as a guide to testing and monitoring, not as fitness itself.