Characterization of gene loss, missense and synonymous variation across mammals
To characterize gene loss, missense and synonymous variation, we analyzed orthologs from 462 placental mammalian species. Using human gene annotations as the reference, we inferred mammalian orthologs with the tool to infer orthologs from genome alignments (TOGA)25. TOGA codon alignments identified synonymous, missense and gene-inactivating mutations, including frameshifts, stop codons, splice-site mutations and larger deletions. We defined gene loss as the absence of an intact reading frame in other mammals, although the dataset includes human-specific or primate-specific genes absent outside primates. For example, 12 genes have been lost across all species except humans, which were typically pseudogenes or single-exon genes (Supplementary Table 1).
We found that each gene had a median of 18 gene loss events (Fig. 1a), whereas, across the species, we found a median of 1,130 likely gene loss events per species (Fig. 1d). Here we defined gene loss as disruption of the open reading frame. We identified only 6 genes with no instances of gene loss across the 462 mammalian species: NAP1L1, ZC3H14, MEIS2, STRIP1, CCT5 and ATP6V0D1.
Fig. 1: Characterization of gene loss, missense and synonymous counts across 462 placental mammals.
a, Distribution of gene loss. The dashed vertical line indicates the median gene loss count. b, Mean missense variant counts across genes relative to the mammalian consensus. The dashed vertical line indicates the median number of missense counts relative to the mammalian consensus sequence. c, Mean synonymous variant counts across genes relative to the mammalian consensus. The dashed vertical line indicates the median number of synonymous counts relative to the mammalian consensus sequence. d, Frequency of gene loss per species. The dashed vertical line indicates the median count of gene loss events per species. e, Total missense variant counts per species. The dashed vertical line indicates the median total number of missense counts per species relative to the mammalian consensus sequence. f, Total synonymous variant counts per species. The dashed vertical line indicates the median total number of synonymous counts per species relative to the mammalian consensus sequence.
For each gene, we calculated the mean missense and synonymous count across the mammals. The median missense average was 27.7 (Fig. 1b), whereas the median synonymous average was 53.4 (Fig. 1c). For each species, there was a median of 610,603 missense variants (Figs. 1e) and 1,084,983 synonymous variants (Fig. 1f) across all genes.
Defining GISMO and GISMO-mis
To capture human gene essentiality, we developed GISMO, which estimates the proportion of gene loss across 462 mammals, weighted by evolutionary distance from humans. Lower GISMO values indicate less frequent mammalian gene loss (Supplementary Table 2). We assessed the effects of phylogenetic compared to nonphylogenetic gene loss counts and found that, in general, there was a strong correlation between both (R = 0.95, P < 1 × 10−16; Supplementary Fig. 1). In addition, we assessed the effects of forward and reverse weighing of evolutionary distance and found that upweighing gene loss events that occurred more distant from humans led to empirically stronger results over conservation and constraint metrics (Supplementary Fig. 2). In particular, we found that GISMO and GISMO-mis were more strongly associated with other human constraint metrics than other conservation metrics such as mean coding sequences (CDS) genomic evolutionary rate profiling scores (Supplementary Fig. 2). We released the phylogenetic version of GISMO in Supplementary Table 11, but have moved forward with all analyses with the regular GISMO due to stronger associations with other constraint metrics. We also quantified missense differences relative to a mammalian consensus sequence as a proxy for molecular adaptation and selection. GISMO-mis measures the mean mammalian missense-to-synonymous ratio (Supplementary Table 3). We benchmarked the number of genes with a score compared to other metrics and found that GISMO covered the highest proportion at 97.7% of all genes in GENCODE v38 (Supplementary Table 4).
We evaluated both metrics against other measures of biological and functional importance of genes. The most constrained GISMO and GISMO-mis deciles had the highest enrichment of genes that cause lethality in mouse heterozygous knockouts (odds ratio (OR)GISMO = 3.43 (95% confidence interval (CI) 2.70–4.33), P = 1.68 × 10−21; ORGISMO-mis = 5.78 (4.61–7.21), P = 2.93 × 10−46), mouse homozygous knockouts (ORGISMO = 8.21 (95% CI 6.42–9.08), P < 1 × 10−300; ORGISMO-mis = 13.1 (11.70–14.70), P < 1 × 10−300) and considered essential in cell lines (ORGISMO = 4.34 (95% CI 3.64–5.17), P = 3.48 × 10−52; ORGISMO-mis = 10.70 (9.05–12.70), P = 6.03 × 10−152) (Fig. 2a–c). The least constrained deciles have the highest proportion of nonessential and olfactory genes (Fig. 2d,e). We also found that genes that were frequently lost across the evolutionary tree were expressed in fewer tissues across Genotype-Tissue Expression (GTEx; Supplementary Fig. 3). We additionally assessed whether GISMO had any biases toward gene family size. We found that the distributions did not drift as family size increased (Supplementary Fig. 4).
Fig. 2: The biological and functional spectrum of gene loss and fixed missense differences.
a–e, The percentage of genes present in each decile of GISMO and GISMO-mis across different gene sets (mouse heterozygous lethal (a), mouse lethal (b), essential (c), nonessential (d) and olfactory (e) genes). The first decile indicates the most constrained one. f, Comparison of GISMO and GISMO-mis to human LoF constraint. Violin plots show the distribution of values within each conservation-score decile. Embedded box plots indicate the median and interquartile range (IQR), with whiskers extending to 1.5× the IQR. The white circles indicate the mean. g,h, Comparison of GISMO and GISMO-mis to CNVs in the UK Biobank: duplications (g) and deletions (h). For pHaplo and pTriplo, the score was inverted (1 − pTriplo; 1 − pHaplo) such that the lowest percentiles indicated the most constrained to be consistent with other constraint metrics. A Spearman’s correlation test was done for each constraint and conservation score against the mean deletion or duplication per kilobase. The lines were estimated by a locally estimated scatterplot smoothing curve with a 95% CI.
We compared both metrics against human constraint metrics and found that GISMO (R = 0.38, P < 1 × 10−300) and GISMO-mis (R = 0.56, P < 1 × 10−300) were both strongly associated with LOEUF (Fig. 2f and Supplementary Fig. 5), with GISMO-mis having a stronger association. Both metrics also had strong correlations with missense constraint (MOEUF), which was comparable to LOEUF for GISMO-mis (R = 0.61, P < 1 × 10−300) but lower for GISMO (R = 0.33, P < 1 × 10−300). In addition, we found that both metrics were strongly associated with dosage sensitivity metrics to predict the strength of selection against loss of a gene copy (haploinsufficiency (pHaplo); RGISMO = − 0.32, P < 1 × 10−300; RGISMO-mis = −0.44, P < 1 × 10−300) and duplication of a gene (triplosensitivity (pTriplo); RGISMO = − 0.32, P < 1 × 10−300; RGISMO-mis = −0.50, P < 1 × 10−300). In addition, we compared GISMO-mis with the new gene-level AlphaMissense metric and found a stronger correlation to this metric (R = 0.84, P < 1 × 10−300) than LOEUF or missense constraint. We also compared the recently developed sHet metric against GISMO and GISMO-mis, which was developed to account for shorter gene bias. We found that both metrics had strong correlations, albeit weaker for GISMO, consistent with GISMO capturing homozygous selection (RGISMO = −0.28, P < 1 × 10−300; RGISMO-mis = −0.61, P < 1 × 10−300). Finally, we compared against conservation metrics such as mean genomic evolutionary rate profiling and mean phylogenetic codon substitution frequencies and found that GISMO and GISMO-mis consistently outperformed in terms of association strength with other constraint metrics relative to previous conservation scores (Supplementary Fig. 2).
Next, to assess the relationship between copy number variants (CNVs) and mammalian gene conservation, we leveraged a CNV dataset generated from exome sequencing in the UK Biobank26 (n = 197,306) that enabled detection of individual gene-level variants. The most constrained deciles of both GISMO and GISMO-mis had the lowest number of deletions and duplications. We found that GISMO had an even stronger correlation in the UK Biobank for deletions compared to pHaplo, a dosage-sensitivity metric derived from CNVs aggregated from lower-resolution microarray data across ~1 million human samples27 (Fig. 2g,h).
Leveraging GISMO to identify shorter intolerant genes
Existing human constraint metrics are biased by gene length because short genes have smaller mutational targets and are often poorly calibrated. The correlation of GISMO and GISMO-mis with CDS length was R = −0.12, P = 3.8 × 10−52 for GISMO, R = −0.12, P = 4.0 × 10−52 for GISMO-mis and several magnitudes lower compared to LOEUF (R = −0.50, P < 1 × 10−300) (Fig. 3a). As LOEUF may be poorly calibrated for short genes, we tested whether GISMO and GISMO-mis identified shorter intolerant genes. For each metric, we defined intolerance as the top 15% most constrained genes. Both GISMO and GISMO-mis identified much shorter intolerant genes (GISMO median CDS = 1,502 bp for 2,897 genes and GISMO-mis median CDS = 1,302 bp for 2,502 genes) than LOEUF-intolerant genes (median CDS = 2,267 bp for 2,968 genes). We additionally found that genes predicted to be intolerant by GISMO and GISMO-mis but not LOEUF tend to be much shorter compared to genes that are intolerant to LOEUF and one of the GISMO metrics (Fig. 3b).
Fig. 3: GISMO and GISMO-mis capture shorter genes relative to human constraint metrics.
a, Correlation of GISMO, GISMO-mis and LOEUF against CDS length using a Spearman’s correlation test. The thicker colored line indicates the linear regression line across all data points. b, Short constrained GISMO and GISMO-mis genes considered tolerant for LOEUF. The colored line within the distributions indicates the median CDS length for that category.
The genes that were considered tolerant for LOEUF (>0.35) and GISMO intolerant were enriched for genes involved in the following pathways or gene sets in mice: sperm immotility (Bonferroni’s adjusted P (Padj) = 2.17 × 10−3, 23 of 46 genes), absent acrosome (Padj = 4.86 × 10−2, 16 of 30 genes) and abnormal actin cytoskeleton morphology (Padj = 4.86 × 10−2, 15 of 27 genes) (Supplementary Table 5).
Identifying potential sex chromosome biases in LOEUF
We also examined genes that appear highly constrained by LOEUF but fall into the least constrained GISMO deciles—that is, genes that are intolerant to heterozygous LoF in humans yet frequently lost across mammals. This discordant set was notably enriched for X-linked genes, raising the possibility that LOEUF may overestimate constraint for some sex-linked loci due to hemizygosity, sex-biased selection or technical challenges in variant detection from short-read sequencing (Supplementary Table 6). When restricting to X-linked genes only, we found no significant enrichment for Mendelian disease genes or essential phenotypes, suggesting that these discrepancies may reflect biological or technical factors specific to human population data rather than true evolutionary indispensability (Supplementary Table 7).
Mammalian gene loss and fixed missense differences capture human disease signals
We next tested whether mammalian gene loss and fixed missense differences prioritize human disease genes. As many disease models are mammalian, conservation may identify genes amenable to modeling in mammals. First, we assessed common variant heritability. For both metrics, we partitioned heritability across 276 independent traits from the UK Biobank and across several disease traits not well captured in the biobank. We found that there was a strong linear enrichment of heritability across the deciles of both metrics, with lowest enrichment in the least constrained deciles (Fig. 4a and Supplementary Table 8). The partitioned heritability results were consistent across flanking thresholds (10 kb, 50 kb and 100 kb). Second, we looked at rare variant association studies of traits with decreased fecundity, such as autism, schizophrenia and neurodevelopmental disorders. We found that both GISMO-constrained and GISMO-mis-constrained genes had a higher association with neurodevelopmental disorders and the metrics were strongly correlated with the association strength (Fig. 4b and Supplementary Figs. 6 and 7).
Fig. 4: GISMO and GISMO-mis capture the disease signal and GISMO captures the recessive signal.
a, Mean enrichment of per-SNV partitioned heritability of 274 traits explained by common variants within 100 kb of genes for GISMO and GISMO-mis deciles. The dashed horizontal line indicates the null expectation of 1. Violin plots show the distribution of values within each conservation-score decile. Embedded box plots indicate the median and IQR, with whiskers extending to 1.5× the IQR. The white circles indicate the mean. b, Mean neurodevelopmental disorder association across GISMO and GISMO-mis deciles. The association reflects P values estimated from the transmission and de novo association test of neurodevelopmental disorder rare variant associations that were later converted to absolute z-scores. Violin plots show the distribution of values within each conservation-score decile. Embedded box plots indicate the median and IQR with whiskers extending to 1.5× the IQR. The white circles indicate the mean. c, The percentage of genes present in each decile of GISMO and GISMO-mis for 1,183 recessive disorder genes aggregated from Online Mendelian Inheritance in Man (OMIM). d, The percentage of genes present in each decile of GISMO and GISMO-mis for 709 dominant disorder genes aggregated from OMIM.
Improving recessive disease–gene prioritization
Given that gene loss in a species can reflect selection on homozygous LoF carriers, we hypothesized that GISMO could be used to prioritize recessive genes associated with disease (that is, genes under homozygous selection should be more constrained by GISMO). To assess this, we assembled a list of 1,183 recessive genes and found that there was enrichment among the most constrained deciles (OR = 3.37 (95% CI 2.91–3.89), P = 1.42 × 10−52; Fig. 4c), whereas GISMO-mis expectedly did not have a significant enrichment (OR = 0.91 (0.73–1.13), P = 0.41). We additionally found that both GISMO and GISMO-mis prioritize enrichment in dominant disorder genes in the most constrained deciles (ORGISMO = 9.92 (95% CI 8.07–12.20), P = 1.28 × 10−94; ORGISMO-mis = 4.38 (3.67–5.22), P = 3.46 × 10−52) (Fig. 4d). Next, we sought to test whether GISMO may help with prioritization of recessive inheritance candidate genes in the Deciphering Developmental Disorders study consisting of 14,000 trios. We found that the GISMO distribution for recessive developmental disorder genes with a high confidence (confident or strong) followed similar distributions to Mendelian recessive genes, whereas the recessive genes with limited and moderate evidence were unclear (Supplementary Fig. 8). We also found that the limited and moderate evidence categories prioritized genes with high constraint for LOEUF, a reflection of strong heterozygous selection, which are unlikely gene candidates for recessive disorders. To further investigate divergence between cross-species and human-specific constraint, we examined known autosomal recessive disease genes from the curated set of definitive recessive genes in the Deciphering Developmental Disorders study, which fall into the least constrained GISMO deciles. Although this subset is small, enrichment analysis revealed phenotypic associations related to corneal development, thoracic structure and skeletal morphogenesis (Supplementary Table 7). These findings suggest that some genes may be essential in humans yet dispensable across mammals due to lineage-specific compensation, ecological specialization or relaxed evolutionary pressures.
Recommended usages across mammalian and human genetics
GISMO and GISMO-mis are complementary gene-level conservation metrics derived from placental mammal alignments. GISMO quantifies gene-loss frequency across the phylogeny, whereas GISMO-mis measures the depletion of missense relative to synonymous substitutions. Both can be applied in human and comparative mammalian genomics.
We recommend that both GISMO and GISMO-mis be used as continuous variables whenever possible, to preserve the full information content across the spectrum of conservation. However, for applications requiring categorical classification, decile-based thresholds may be employed. Below we outline cases of recommended use and recommended strategies for interpretation.
In human genetics, the most constrained deciles (for example, deciles 1–2) are enriched for genes that are dosage sensitive, cause early lethality in model organisms or are associated with Mendelian disease. GISMO-mis, in particular, shows strong concordance with human constraint metrics such as LOEUF and AlphaMissense and is informative for polygenic traits when used continuously or in broader decile windows (for example, the top 20–40%). Notably, GISMO retains scoring power in duplicated or poorly mappable genes that are often excluded from short-read human constraint models.
In comparative mammalian genomics, GISMO enables the identification of genes that are broadly retained across mammals, reflecting long-term evolutionary indispensability. Conversely, high GISMO values may highlight lineage-specific dispensability or adaptive gene loss. GISMO-mis complements this by identifying codon-level amino acid conservation, which can inform studies of protein evolution, functional domain constraint or convergent adaptation.
Importantly, conservation-based metrics are agnostic to demographic features of contemporary human populations and can offer more reliable estimates in regions where human constraint models may be biased or underpowered. For instance, we observe that genes on the X chromosome are disproportionately assigned strong constraint scores by LOEUF, even when conservation metrics suggest that they are moderately or weakly constrained (Supplementary Table 9). This discrepancy likely reflects miscalibration of LOEUF in regions of reduced effective population size, skewed mutation rates or sex-specific selection—factors that complicate modeling of constraint from standing human variation. In contrast, GISMO and GISMO-mis show more uniform constraint distributions across autosomes and sex chromosomes, underscoring the value of incorporating deep-time conservation into gene prioritization frameworks.
These metrics may also facilitate phylogenetic profiling, pathway enrichment analyses and cross-species annotation transfer. To support application-specific thresholding, we recommend calibrating cutoffs empirically against known gene sets (for example, ClinGen annotations, mouse knockout phenotypes or disease–gene catalogs). A summary of recommended applications is provided in Table 1, but we emphasize that these are not the only cases.
Table 1 Recommended cases for use in GISMO and GISMO-mis