The researchers used human GENCODE version 38 annotations as the reference and applied TOGA, a tool for inferring orthologs from genome alignments, to compare genes across 462 placental mammals. TOGA identified disruptive mutations including frameshifts, premature stop codons, splice-site disruptions and larger deletions. A gene was classified as lost when its intact reading frame was absent in a species and all corresponding transcripts met the criteria for loss. Across the dataset, each gene had a median of 18 loss events, while each species had a median of 1,130 likely gene-loss events. Only six genes had no observed loss across the complete mammalian sample: NAP1L1, ZC3H14, MEIS2, STRIP1, CCT5 and ATP6V0D1.

GISMO was designed as a weighted estimate of the proportion of species in which a gene has been lost. Each loss is weighted according to evolutionary distance from humans, so losses in more distantly related species contribute more heavily. A lower score indicates that a gene is rarely lost and is therefore more constrained across the sampled mammals. The researchers also examined a phylogenetically adjusted version, which was strongly correlated with the regular score, but the nonphylogenetic version showed stronger empirical associations with independent constraint measures. They attributed some of this performance to TOGA’s requirement for multiple disruptions in an open reading frame before classifying a gene as lost, which can reduce the risk of counting a single questionable disruption as a loss event. The metric covered 97.7% of genes in GENCODE version 38, the highest proportion among the conservation scores benchmarked in the study.

The second measure, GISMO-mis, captures a different aspect of constraint. For each gene, the researchers generated a mammalian consensus sequence and counted missense and synonymous differences in each species relative to that consensus. They then calculated the mean missense-to-synonymous ratio across available species. A lower ratio indicates that amino-acid-changing substitutions are relatively depleted compared with synonymous changes, consistent with stronger functional constraint at the protein level. Codon alignments were included when a gene had a single ortholog in at least 75% of queried species; species with more than four predicted orthologs were excluded. The method extends the logic of the missense-to-synonymous substitution ratio by using a multi-species consensus and by handling ambiguous alignments conservatively.

Tests against independent gene sets showed that the most constrained deciles were enriched for genes whose disruption causes early lethality in mice or loss of viability in cell lines. For mouse heterozygous knockouts, the odds ratios were 3.43 for GISMO and 5.78 for GISMO-mis. The corresponding values for homozygous knockouts were 8.21 and 13.1, while essential cell-line genes had odds ratios of 4.34 and 10.70. By contrast, the least constrained deciles contained the greatest proportions of genes classified as nonessential or associated with olfaction. Frequently lost genes also tended to be expressed in fewer tissues in the GTEx dataset, with expression defined as more than 0.3 transcripts per million. The distributions of GISMO scores did not drift as gene-family size increased, indicating no observed association with that feature in the benchmarked data.

Both measures were correlated with human constraint estimates derived from large-scale sequencing data. GISMO correlated with the gnomAD LOEUF score at Spearman’s R of 0.38, while GISMO-mis reached R of 0.56, with both P values below 1 × 10−300. GISMO-mis also correlated with the missense constraint score MOEUF at R of 0.61 and with the predicted missense effect metric AlphaMissense at R of 0.84. Both measures were associated with haploinsufficiency and triplosensitivity scores, while GISMO-mis had stronger correlations than GISMO for each dosage measure. In a UK Biobank copy-number-variant dataset generated from exome sequencing in 197,306 people, the most constrained deciles of both metrics had the fewest deletions and duplications. GISMO showed an even stronger correlation with deletions than pHaplo, a dosage-sensitivity metric derived from lower-resolution microarray data across about one million human samples. These results establish associations with independent measures of constraint, rather than showing that the scores are interchangeable.

A major difference from LOEUF was the reduced sensitivity of the mammalian measures to coding-sequence length. LOEUF was strongly correlated with coding-sequence length at R of −0.50, whereas the correlations for GISMO and GISMO-mis were each −0.12. When the researchers defined intolerance as the most constrained 15% of genes, GISMO identified 2,897 genes with a median coding-sequence length of 1,502 base pairs, and GISMO-mis identified 2,502 with a median of 1,302 base pairs. LOEUF identified 2,968 intolerant genes with a longer median of 2,267 base pairs. Genes considered intolerant by GISMO or GISMO-mis but tolerant by LOEUF were enriched in mouse phenotypes involving sperm immotility, absence of the acrosome and abnormal actin cytoskeleton morphology. The analysis was comparative and computational, so these enrichments identify previously unclassified constrained genes for further study rather than demonstrating human clinical outcomes.

The study also examined whether the evolutionary measures captured disease-associated genetic signals. Using linkage disequilibrium score regression, the researchers partitioned common-variant heritability across 276 independent UK Biobank traits and several additional disease phenotypes. Heritability showed a strong linear enrichment across the constraint deciles of both GISMO and GISMO-mis and was lowest in the least constrained deciles, with consistent results at 10-, 50- and 100-kilobase flanking windows. Genes in the most constrained deciles also showed stronger associations with rare variants linked to reduced fecundity, including schizophrenia, autism and neurodevelopmental disorders. The findings indicate that deeply conserved genes contribute disproportionately to both common- and rare-variant signals in these phenotypes, but they do not establish that conservation itself causes the associations. The researchers also found that genes with high LOEUF constraint but low GISMO conservation were disproportionately X-linked, raising the possibility of differences related to hemizygosity, sex-biased selection or technical challenges in short-read variant detection; this discordant subset was not enriched for Mendelian disease genes or essential phenotypes when restricted to the X chromosome.

GISMO had a distinct association with recessive disease genes because gene loss across species can reflect selection against homozygous loss-of-function alleles. In a set of 1,183 autosomal recessive genes assembled from OMIM, the most constrained GISMO deciles were enriched with an odds ratio of 3.37 and a 95% confidence interval of 2.91 to 3.89. GISMO-mis did not show significant enrichment for this set, with an odds ratio of 0.91 and P = 0.41. Both measures were enriched for 709 dominant disorder genes: the odds ratio was 9.92 for GISMO and 4.38 for GISMO-mis. In the Deciphering Developmental Disorders study of about 14,000 trios, GISMO distributions for recessive developmental-disorder genes with confident or strong evidence were similar to those of Mendelian recessive genes, whereas genes with limited or moderate evidence were unclear. The authors described GISMO and GISMO-mis as complementary: GISMO emphasizes disruption of an entire gene across species, while GISMO-mis emphasizes amino-acid-level constraint. They recommend using the continuous scores where possible and calibrating categorical thresholds against relevant gene sets such as ClinGen annotations, mouse knockout phenotypes or disease-gene catalogs.

Several limitations constrain the interpretation and future refinement of the results. Mammalian conservation may not reflect recent human selective pressures or human-specific innovations, and a gene tolerated in other mammals may nevertheless be essential in humans. The available species may also introduce survivorship and accessibility biases, so broader taxonomic sampling could improve the metrics. GISMO combines signals associated with heterozygous and homozygous selection and therefore does not distinguish between them by itself; pairing it with LOEUF and GISMO-mis can help separate those relationships. Although the nonphylogenetic version of GISMO does not distinguish multiple independent losses from one ancestral loss, it performed better against independent measures of functional constraint than the phylogenetic alternative, and a phylogenetic version was also released. Recent human-specific duplications showed a slight correlation with the conservation score of the original paralog, but such cases were few. Overall, the analyses support gene loss and missense fixation across mammals as measurable signals of selection and show that these scores can rank gene importance, inform disease-gene prioritization and provide evolutionary evidence alongside variation in modern human populations.

Subject of Research: Evolutionary constraint and gene essentiality across 462 placental mammals

Article Title: The landscape of gene loss and missense variation across the mammalian tree informs on gene essentiality

Article References: Liao, C., Ye, R., Ivankovic, F., Fu, J. M., Walters, R., Lowther, C., Valkanas, E., Churchhouse, C., Samocha, K. E., Lindblad-Toh, K., Karlsson, E., Hiller, M., Talkowski, M. E., & Neale, B. M. (2026). The landscape of gene loss and missense variation across the mammalian tree informs on gene essentiality. Nature Genetics. https://doi.org/10.1038/s41588-026-02783-9

Image Credits: AI Generated

DOI: 10.1038/s41588-026-02783-9

Keywords: comparative genomics, gene essentiality, mammalian evolution, gene loss, missense variation, evolutionary constraint, disease-gene prioritization, rare variants, common-variant heritability, copy-number variation, landscape, gene

The central challenge the researchers set out to address is a familiar one in human genetics. Most of the thousands of genetic variants identified by genome-wide association studies as risk factors for disease do not lie within protein-coding genes. Instead, they sit in regulatory regions of the genome, the stretches of DNA that control when, where, and how strongly genes are switched on. Decades of work have established this pattern, but translating a regulatory variant into a concrete molecular mechanism has remained difficult, because the field lacked an empirical framework for characterizing the downstream consequences of regulatory variation at population scale.

Single-cell molecular quantitative trait locus mapping, a technique that links genetic variants to gene regulation measured in individual cells, has emerged as a promising approach. However, previous efforts suffered from limited statistical power and, crucially, from the inability to measure chromatin state and gene expression simultaneously in the same cells. Without both layers of information, researchers could observe that a variant was associated with changes in gene expression or with changes in chromatin accessibility, but they could not reliably connect the two steps into a single causal chain running from DNA accessibility through to transcript output.

To overcome these limitations, the consortium performed paired single-nucleus RNA sequencing and single-nucleus assay for transposase-accessible chromatin sequencing on peripheral blood mononuclear cells from FinnGen donors, capturing roughly ten million nuclei. The resulting atlas identified 51,083 cis-expression quantitative trait loci affecting 20,829 genes, 338,100 cis-chromatin accessibility quantitative trait loci affecting 210,584 open chromatin peaks, 119,094 putative causal variants, and 593,765 statistical links between regulatory peaks and the genes they control. These numbers represent a resource of extraordinary depth, spanning the major classes of immune cells found in blood and providing cell-type-resolved maps of genetic regulation.

A key analytical insight came from classifying variants according to whether they completed the full regulatory cascade. Variants that altered chromatin accessibility and were also linked, through a peak-gene connection, to altered gene expression showed twice the rate of colocalization with disease associations compared with variants that affected chromatin alone. In other words, the variants most likely to be genuinely causal for disease are those whose effects can be traced through the entire chain, from open chromatin to enhancer activity to target gene expression. This finding provides a practical filter for prioritizing candidate causal variants in complex disease loci.

To validate these statistical findings experimentally, the team turned to massively parallel reporter assays, a high-throughput technique that tests the regulatory activity of thousands of DNA sequences simultaneously. The assays confirmed the regulatory effects of 10,428 fine-mapped molecular quantitative trait loci, providing direct experimental support for a substantial fraction of the atlas’s predictions and demonstrating that the computational framework captures biologically real regulatory variation rather than statistical artifacts.

Perhaps the most conceptually significant discovery concerns genes under strong evolutionary constraint, those whose sequences change rarely because mutations are harmful. Previous studies had noted a paradox: disease variants preferentially target constrained genes, yet constrained genes appear depleted of detectable expression quantitative trait loci. The new atlas resolves this contradiction by revealing a phenomenon the authors call multilayered regulatory buffering. At constrained genes, chromatin accessibility changes caused by variants occur with normal effect sizes, but their transmission to gene expression is attenuated because these genes are regulated through weaker and more numerous enhancer-gene links. The redundancy of many weak regulatory connections dampens the expression consequence of any single regulatory perturbation.

Crucially, the reporter assay experiments confirmed that this buffering operates downstream of the regulatory element itself. Constraint acts at the interface between chromatin and expression rather than on the intrinsic activity of the cis-regulatory DNA. This distinction matters because it explains why the paradox exists in the first place: standard expression quantitative trait locus studies measure only the final output, and buffering at the chromatin-to-expression step suppresses that output signal even when the underlying chromatin effects are intact. The finding reconciles apparently conflicting observations across the field and suggests that disease risk at constrained genes may accumulate through subtle, distributed effects that individual expression studies systematically miss.

The atlas also delivers concrete mechanistic hypotheses for specific diseases. At autoimmune loci, the researchers traced complete regulatory cascades at TICAM1, an adaptor protein in Toll-like receptor signaling, and RHOH, a small GTPase involved in T cell receptor signaling, both linked to autoimmune hypothyroidism. They similarly dissected Finnish-enriched variants at TNRC18 and IL21R, the latter a receptor for an interleukin with established roles in T cell biology. Beyond blood-related traits, the framework explained associations in other tissues, including loci connected to Alzheimer’s disease such as PILRB and TYROBP, genes involved in microglial signaling, illustrating that regulatory mechanisms mapped in immune cells can illuminate neurological and other complex diseases.

The resource is being made broadly available to the research community. Summary statistics for chromatin accessibility quantitative trait loci, expression quantitative trait loci, and peak-gene links are publicly accessible and browsable through an interactive web interface, and the analysis pipelines, including the CASCADE classification framework and supporting software packages, have been released on GitHub. Individual-level data remain protected under Finnish and European data regulations but can be accessed by approved researchers through established application channels. As the field moves toward functional interpretation of the ever-growing catalog of disease-associated variants, atlases of this kind, which connect genetic variation to regulatory architecture across millions of cells and hundreds of donors, are likely to become foundational references for drug target discovery and precision medicine in immunology and beyond.

Subject of Research: Population-scale single-cell multiomic mapping of regulatory genetic variation in human immune cells

Article Title: Population-scale immune multiome atlas reveals regulatory disease mechanisms

Article References: Kanai, M., Delorey, T. M., Honkanen, J., Rodosthenous, R. S., Juvila, J., Murphy, S., Teixeira-Soldano, I., Hwang, H. S., Karjalainen, J., Halonen, J., Panagiotaropoulou, G., Zhang, Y., McCabe, C., Chen, E., Nanki, K., Yoshida, T., Liu, K., Glean, M., Mehrotra, N., … Xavier, R. J. (2026). Population-scale immune multiome atlas reveals regulatory disease mechanisms. Nature. https://doi.org/10.1038/s41586-026-11078-2

Image Credits: AI Generated

DOI: 10.1038/s41586-026-11078-2

Keywords: single-cell multiomics, chromatin accessibility, gene expression, quantitative trait loci, FinnGen, immune cells, regulatory variants, enhancer-gene links, autoimmune disease, genetic fine-mapping, reporter assays, evolutionary constraint