🧬 genetics

Gene-level evolutionary conservation and human constraint decline at comparable rates across the age of onset of Mendelian disease

This study demonstrates that gene-level evolutionary conservation and human constraint scores decline at comparable rates as the age of onset for Mendelian diseases increases, challenging the notion that constraint metrics lose predictive power faster than conservation scores while highlighting significant metric-dependent variations in performance, particularly for recessive diseases.

Original authors: Mahaboob Ali, A. A., Nelson, E. J. R.

Published 2026-10-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mahaboob Ali, A. A., Nelson, E. J. R.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the vast library of human genetics, scientists have long relied on two types of ancient maps to find the genes responsible for disease. The first map is a deep-time record, showing which parts of our DNA have remained unchanged for millions of years across many different species. If a gene has stayed the same since our ancestors shared a common path with other mammals, it is usually because that gene is vital for life. The second map is a recent snapshot, looking at how much variation exists within human populations today. If a gene rarely shows up with damaging changes in healthy people, it suggests that nature has been actively weeding out those changes because they are harmful. Both maps are used to prioritize which genes to study when a patient is sick, operating on the idea that the more "constrained" a gene is, the more likely it is to be the cause of a serious illness.

However, a fundamental question has lingered about how these maps work for diseases that appear late in life. Evolutionary theory suggests that natural selection is strongest when it protects the ability to reproduce. Once a person has had children, the pressure to keep the body running perfectly for decades afterward weakens. This has led to a common assumption: that the genetic signals for late-onset diseases, such as those appearing in old age, should be much fainter or harder to find than those for childhood diseases. If this were true, the ancient maps would lose their usefulness as the age of disease onset increases, leaving doctors with fewer tools to diagnose conditions that strike later in life.

A team of researchers set out to test this assumption directly. They gathered data on nearly 3,000 rare genetic diseases, known as Mendelian disorders, and organized them by the age at which symptoms typically appear. They then compared how well the deep-time map and the recent human map could distinguish disease-causing genes from healthy ones across this entire age spectrum. They did not just look at whether the maps worked; they specifically measured whether one map faded away faster than the other as the age of onset got older. The results were surprising. The researchers found that both maps lose their ability to identify disease genes at almost exactly the same rate as the age of onset increases. The deep-time record does not fade faster than the recent human record, nor does the human record fade faster than the deep-time one. The difference between their performance is so small that it is statistically indistinguishable from zero.

This finding challenges a popular explanation that had been used to justify why these maps might fail for older patients. Many scientists had believed that early-onset diseases are caused by genes that are essential for development, while late-onset diseases are caused by genes that are less critical. The new analysis showed that this difference in gene type does not actually explain the pattern. When the researchers adjusted their calculations to account for whether a gene was involved in development or had multiple functions, the result did not change. The relationship between the age of onset and the strength of the genetic signal remained the same. The study also revealed that the specific mathematical tool used to measure the "recent human" map matters more than previously thought. Depending on which tool was chosen, the results could swing from showing no difference to showing a clear difference, even though the tools ranked the genes in nearly the same order. This suggests that the conclusion one draws can depend heavily on the specific method used, rather than on a clear biological truth.

Perhaps the most practical discovery concerns how these maps treat recessive diseases, where a person must inherit two faulty copies of a gene to become sick. One of the most common tools used to measure human constraint actually ranks these recessive disease genes as if they were less important than average genes, effectively hiding them from view. In contrast, other tools rank them higher. The researchers traced this discrepancy to the fact that recessive disease genes tend to be longer and have different structural features than other genes. When the researchers compared the tools fairly, matching genes of the same length and structure, all the tools agreed: none of them could reliably identify recessive disease genes better than chance. This means that current methods used in diagnostic labs may be systematically overlooking a large class of genetic disorders simply because of how the genes are built, not because the genes are unimportant.

The study concludes that while we can stratify genetic evidence by the age of disease onset, our current data is not yet precise enough to say that one type of map is definitively better than the other for late-life conditions. There are only a few hundred late-onset Mendelian diseases with clear genetic labels available for study, which limits how finely we can measure these differences. The researchers emphasize that the idea that evolutionary evidence becomes useless for late-onset disease is not supported by their data, but neither is the idea that it remains perfectly strong. The most important takeaway is a warning to scientists and doctors: the choice of which genetic tool to use can change the answer, and for recessive diseases, the standard tools need a specific correction to work correctly. Until more data becomes available, the best approach is to use multiple tools and understand their limitations, rather than relying on a single score to guide the search for the cause of a late-onset genetic illness.

Technical Summary: Gene-level evolutionary conservation and human constraint decline at comparable rates across the age of onset of Mendelian disease

Problem Statement
Gene-level evolutionary scores, including cross-species conservation and human-population constraint metrics, are standard tools for prioritizing disease genes and benchmarking genomic models. Classical evolutionary theory posits that purifying selection weakens as age increases, particularly after reproduction is complete. Consequently, it is hypothesized that these scores should carry diminishing information about diseases with later onset. While this "selection shadow" effect has been demonstrated at the level of allele frequencies, it remains untested for the gene-level scores used in practice. A critical question is whether conservation (reflecting deep-time selection) and constraint (reflecting recent human selection) decay at different rates as disease onset increases, and whether observed gradients are driven by the fitness costs of disease or by the confounding properties of gene classes (e.g., developmental, pleiotropic genes).

Methodology
The authors conducted a retrospective analysis of 2,885 Mendelian diseases and 5,322 gene-disease pairs derived from Orphanet. The study employed a logistic regression framework to estimate the interaction between disease age of onset and evolutionary scores.

Data Sources: Mendelian associations and onset intervals were sourced from Orphanet. Constraint metrics included shet (selection coefficient against heterozygotes from GeneBayes), LOEUF, pLI, and missense Z (from gnomAD v4.1). Cross-species conservation was derived from 241-way and 100-way phyloP alignments (Zoonomia) and phastCons.
Statistical Model: A logistic regression model was fitted to predict the probability of a gene-disease association. The primary estimand ( $\Gamma$ ) was the contrast between the onset-slope interaction coefficients of a constraint score and a conservation score ( $\Gamma = \gamma_{constraint} - \gamma_{conservation}$ ). A positive $\Gamma$ would indicate conservation loses discriminative power faster than constraint as onset rises.
Controls and Covariates: Disease genes were compared against matched controls (matched on coding sequence length, expected loss-of-function count, and GC content) and an unmatched pool. The model adjusted for covariates including inheritance mode, gene length, and GC content. Crucially, the study tested whether conditioning on gene-class properties (developmental annotation, haploinsufficiency, pleiotropy) attenuated the contrast.
Robustness Checks: The analysis included sensitivity analyses using alternative constraint metrics (LOEUF v2.1.1, missense Z), different conservation aggregates, and the exclusion of somatic mutations. A permutation null was generated by shuffling onset within therapeutic areas.

Key Results

Comparable Decay Rates: The primary contrast between the onset slopes of 241-way phyloP (conservation) and shet (constraint) was 0.23 logit units (97.5% CI −0.10 to 0.56), equivalent to 0.024 AUROC units. The 97.5% confidence interval excludes the hypothesis that constraint decays faster than conservation by more than 0.011 AUROC units. The authors conclude that conservation and constraint do not lose their value at measurably different rates as disease onset rises.
Gene-Class Explanation Rejected: The hypothesis that the observed relationship is an artifact of early-onset genes being developmental, pleiotropic, or haploinsufficient was not supported. Conditioning on these gene-class properties did not attenuate the contrast. Furthermore, haploinsufficiency and trait-level pleiotropy did not vary significantly with onset among disease genes.
Metric Dependence: The conclusion is sensitive to the choice of constraint metric. While shet yielded a non-significant contrast, alternative metrics (LOEUF v4.1, missense Z, LOEUF v2.1.1) yielded larger contrasts that excluded zero. However, paired comparisons showed these estimates were not statistically separable, suggesting that metrics ranking genes similarly can yield opposite conclusions regarding derived onset contrasts.
Recessive Disease Genes: No constraint metric successfully identified recessive disease genes against matched controls (all metrics ranked them below chance, AUROC ~0.45–0.48). While LOEUF ranked recessive genes above chance against an unmatched pool, this was attributed to coding sequence length and expected loss-of-function count rather than constraint. shet ranked recessive genes below the genome average, consistent with its definition as a heterozygous selection coefficient.
Incremental Value: Adding evolutionary scores to a non-evolutionary baseline (expression, tissue specificity, network topology) improved discrimination by 17–23%, but no onset gradient was detected in this incremental value.

Significance and Claims
The paper claims that the theoretical expectation of a sharp decline in the utility of evolutionary scores for late-onset Mendelian diseases is not supported by current genome-wide data, likely due to the limited number of annotated late-onset diseases (283). The study highlights three critical methodological insights:

Metric Sensitivity: Results derived from gene-level evolutionary scores can be highly sensitive to the specific constraint metric chosen, even when metrics rank genes nearly identically. A result reported by one metric may carry a sensitivity not reflected in its confidence interval.
Multifunctionality Bias: The study warns against using annotation-derived covariates (e.g., disease breadth or trait counts) that inadvertently include the outcome in their inputs, which can introduce bias indistinguishable from biological effects.
Pipeline Implications: For disease gene prioritization, particularly for recessive disorders, standard constraint metrics (like shet) may incorrectly penalize disease genes. The authors argue that pipelines targeting recessive diseases require explicit corrections for how priors are applied, rather than simply switching metrics, as loss-of-function depletion measures also fail against matched controls.

The authors conclude that while onset stratification of evolutionary scores is feasible, current annotation coverage is insufficient to resolve fine-grained differences in decay rates, and the "gene-class" explanation for any observed trends is unsupported.

Technical Summary: Gene-level evolutionary conservation and human constraint decline at comparable rates across the age of onset of Mendelian disease

More like this