Seven hundred thirty-seven cats representing 26 recognized cat breeds were included in the study (S1 Table). Seven breeds (Persian, Scottish Fold, Siberian, Bengal, Ragdoll, Birman and Burmese) were represented by 50 samples. To conduct an intraspecific comparison within a single breed, the samples of these seven over represented breeds were randomly divided into two groups (e.g., Bengal 1 and Bengal 2). The breed sample sizes ranged from 10 to 26 cats and twenty of the breeds were represented by 20 or more cats.

As a preparation of the workable SNP dataset, autosomal markers (n = 58,888) were extracted from the genotype dataset by removing the X-chromosome SNPs (n = 2,700) and the unmapped SNPs (n = 684). Approximately 99.7% (n = 58,768) of the autosomal markers passed SNP genotyping call rate of 90% (each SNP genotyped in at least 90% of the samples), thus, only 120 SNPs were excluded across all samples. The average percentage of monomorphic SNPs within each breed was 31.7% ranging from (19.1%–54.8%) (S1 Table). Approximately 12% SNPs (n = 6,955) had a MAF between (0 and 0.05) across all populations and were excluded. Overall, 50709 autosomal SNPs that were accurately mapped to the latest cat genome assembly (felCat 9.0 [51]) were included in downstream analyses.

To understand the genetic relationships among the breeds under study, PCA was performed and the first two components PC1 and PC2 explained ~ 40% of the total genetic variation (Fig 1). A broad overview showed that the first component, PC1, which accounts for 28.9% of the total variation, structured the breeds according to their origin ancestry as a continuum from Western to Eastern (Fig 1a). Additionally, Western breeds and what we refer to as “Middle” breeds clustered more closely than to the Eastern breeds, which forms a distant distinct group. The population structure Western breeds can be summarized by: (1) the Persian family breeds (Persian, British Shorthair, Scottish Fold, and Selkirk Rex) formed a distinct group, (2) Maine Coon breed formed a single distinct group, and (3) individuals from the Munchkin and Siberian were less closely structured and extended to the “Middle” breeds (Fig 1b). The “Middle” breeds were less distinctly structured with the exception of Bengal, which formed a separate group (Fig 1c). Two general groups were detected in the Eastern breeds cluster where Birman separated from the rest (Fig 1d). Individuals of the Bombay breed did not form a unique group and some extended to the “Middle” and “Western” clusters (Fig 1a). The second component, PC2, that explains 9.5% of the total variation enabled separation of Bengal from other “Middle” breeds (Fig 1c) and Birman from Eastern breeds (Fig 1d).

As a calculation of likely signature of selection, SNP-specific di values were calculated for each breed. averaged for SNPs within individual overlapping 500Kb windows (50% overlap between adjacent windows). The 99^th percentile threshold (top 92 windows) were reported as candidate windows under selection (S1 Fig, S2 Table).

Collectively across breeds, individual candidate windows under selection (top 92) from each breed were distributed across the autosomal chromosomes (S1 Fig, S2 Table). This posed challenges such as: (1) despite being on different locations, the number of candidate windows (92) is identical for each breed, which prevents objective comparisons between breeds, (2) the likelihood of false positive windows due to SNP ascertainment bias associated with the design of the SNP array, (3) the uneven distribution of SNPs across windows, and (4) the lack of a systematic method that prioritizes certain candidate windows for gene content investigation. To overcome these challenges, we sought to identify candidate regions under selection rather than to explore individual scattered windows. We define a candidate region under selection as three or more overlapping candidate windows, which combined would be ≥ 1Mb in size (S3 Table).

Using the aforementioned approach, nearly 50% of the candidate windows in each breed were single scattered windows or only two overlapping windows (S2 Fig, S2 Table). The categorization of candidate windows is only for prioritizing which candidate regions to be investigated for gene contents. These candidate windows were not totally dismissed from subsequent analyses but later used when discussing specific breeds (e.g., Persian) or inspecting specific chromosomal locations (e.g., FGF5 on Chr. B1). Abyssinian breed showed the lowest percentage of nonoverlapping candidate windows (50%) whereas Bombay had over 90% of its candidate windows as single or two windows (S2 Fig).

Combining and overlaying the identified candidate regions under selection for all breeds provides an overall look at the genomic landscape of selection in the pedigreed domestic cats (Fig 2). Using the 205 candidate regions (S3 Table) of 26 breeds (not accounting for redundancy), the following regarding the genomic landscape of selection in cats was observed: (1) candidate regions were found variably in all chromosomes (S3a Fig), (2) the largest number of candidate region were observed in Chrs. B4, A1, and B1, respectively whereas the least number were in Chrs. E1, D4, and F1 (S3a Fig), (3) the number of candidate regions were in correlation with the size of the chromosome with the exception of Chr. B4 and E3 (S3b Fig), (4) the majority of candidate regions were composed of three overlapping windows and measured 1Mb in size (S3c and S3d Fig), (5) the average size of candidate regions was 1.5 Mb, and the largest candidate region (10 Mb) was found on Chr. B4 (S1 Fig).

Whilst candidate regions expectedly varied in number and size between chromosomes, variation was also observed between breeds (S1 Table, S2 Fig). The number of candidate regions of selection averaged to ~ 6 regions per breed where the least was found in Bombay (two regions) and the most was in Birman (11 regions) (S3 Table). Similarly, the average combined size of candidate regions per breed was 9.3 Mb and ranged from 2.5 Mb in Bombay to a 16.25 Mb in Ragdoll (S1 Table, S4 Fig). When examining the number and the combined sizes of candidate regions per breed in relation to the ancestral groups (Western, “Middle”, Eastern), we observed that Eastern breeds (excluding Bombay) had larger number of candidate regions per breed (averaged ~ 9 regions) and covering larger regions (averaged ~ 12 Mb) compared to “Middle” (~ 6 regions and 10.4 Mb), and Western breeds (~ 5 regions and 7 Mb).

We adopted the exclusion of single or two windows regions as a strategy to prioritize the importance of candidate regions and to provide conservative estimates of candidate regions under selection for each breed. However, we find it more useful to include all candidate windows when comparing two populations of the same breed (e.g., Bengal 1 and Bengal 2) or between different breeds especially since partial shared windows are often observed (S1 Fig).

Intra-specific breed comparison was conducted on seven breeds representing the three ancestral groups (Western, “Middle”, and Eastern). The shared candidate windows within a breed ranged from 38% for Scottish Fold to over 84% for Burmese (S4 Table). The Western breeds (Persian, Scottish Fold, and Siberian) exhibited lower number of shred candidate windows among the subpopulations (38–53%) while Bengal and Ragdoll as representatives of “Middle” breeds shared (62–64%). Eastern breeds (Birman and Burmese) showed the highest number of shared windows (77–84%). The unique and unshared windows between the two populations within an individual breed were mostly of single scattered windows or two overlapping windows (S4 Table). This independently supports our strategy of defining candidate regions under selection by three or more overlapping windows.

Many individual windows were shared across several breeds (examples in S1 Fig Chrs. A1, B1, and B2). However, more significantly was the presence of shared candidate regions. One example of such regions was located on Chr.B3 and shared among Bombay, Burmese, Peterbald, Oriental, and Siamese and less uniformly with Scottish Fold and Siberian. A second region is located on Chr. E3 and is shared across seven breeds from the three ancestral groups (S1 Fig).

Genetic variants that are associated with phenotypes under selection in cats serve as useful positive controls for our conducted analyses and a display of the validity subsequent reported findings. Our five positive controls were: (1) Cornish Rex and multiple candidate regions surrounding its curly coat causative gene, LPAR6 (S1 Fig-ChrA1), (2) Abyssinian and a candidate region where ASIP resides, which explains the lack the non-agouti of its coat (S1 Fig-Chr.A3), (3) Persian, Norwegian Forest Cat, Siberian, and Turkish Van with candidate regions of different sizes around the longhair gene, FGF5 (S1 Fig-Chr.B1), (4) Devon Rex’s candidate region for the curly coat associated gene, KRT71 (S1 Fig-Chr.B4), and (5) Siamese a TYR gene candidate region, which is responsible for the distinct dark coat coloration at the extremities (S1 Fig-Chr.D1).

Dominantly inherited genetic variants, previously identified for breed-defining phenotypes, were also observed. The candidate regions encompassing or adjacent to such variants were mostly represented by single or two windows. Examples of these regions included: (1) windows surrounding UGDH gene, which is related to dwarfism in Munchkins (S1 Fig-Chr.B1), (2) a single window adjacent to KRT71, which is responsible for the curly coat of Selkirk Rex (S1 Fig-Chr.B4), and (3) a window covering the tail-length causative gene, HES, in Japanese Bobtail (S1 Fig-Chr.E1).

We explored the candidate regions under selection for each breed and examined genes within for relationships to the breed defining phenotypes (S3 Table). However, we limited our discussion to a representative number of candidate regions selected based on their (1) sizes, (2) presence of phenotype related candidate genes, and when applicable (3) occurrence in replica populations (i.e., Persian1 and 2). Below, we provide detailed examples for three breeds while others are reported in Table 1.

Genome-wide analysis of the signatures of selection for the American Curl breed showed distinctly several overlapping windows with high levels of genetic population differentiation on Chr.E3 and F1 (Fig 3a). The size of the first candidate region was 3.75Mb and contained 123 genes among which MCRIP2 representing a possible candidate gene for the backward curled ears. A second smaller candidate region was detected with PBX1 as another candidate gene (Table 1). Similarly, Maine Coon’s genome-wide survey showed two apparent regions on Chrs.A1 and A3 (Fig 3b). The candidate region on Chr.A1 was a 1Mb region and contained four genes (GABRG2, GABRA1, GABRA6, GABRB2) whereas the region on Chr.A3 was 2Mb with two genes, FAM98A and RASGRP3. Lastly, with a focus on the unique coloration of the Turkish Van, which is characterized by an overall white coat with different coloration of parts of the head and tail, we identified two candidate regions (Fig 3c). The regions were located on Chrs.A2 and E3 with the latter harboring CLCN7 as candidate genes, (Table 1).

A shared candidate region of relatively low SNP density was found in Turkish Van, American Curl, Japanese Bobtail, Peterbald, Oriental, and Siamese (S1 Fig, S3 Table). This shared candidate region among diverse breeds was located on Chr. E3 and contained 36 genes. A second shared candidate region was identified specifically in the breeds of slender body shape; Peterbald, Oriental, and Siamese (See Fig 4 for body shape). The Chr. D3 region was 1Mb in size and contained five genes among ADGRD1 as potential candidate for the unique body form (Table 1).

While top 1% candidate windows (n = 92) under selection for each breed were reported (S2 Table), we prioritized the candidate windows into candidate regions based on the presence of multiple overlapping windows in the same genomic location (S3 Table). This ensured reporting of conservative estimates of candidate regions under selection in each breed, allowed objective comparison between breeds, and avoided false positives due to ascertainment bias. On average, cat breeds had 2–12 regions under selection positioned on 2–8 chromosomes and with combined size of 2.5–16.25 Mb. The number and size of candidate regions varied among breeds based on: (1) ancestry (e.g., Western), (2) population size (i.e., popularity), (3) age of the breed-defining trait(s) (e.g., longhair), (4) mode of inheritance of the trait under selection, and (5) breeding practices (e.g., outcrossing). An example of the differences observed based on breed ancestry was the relative low number and small size of candidate regions under selection in Western breeds contrasted by the high number of large size candidate regions in Eastern breeds (S1 Table). This is consistent with previously reported differences in estimated population parameters such as linkage disequilibrium, genetic diversity, and levels of inbreeding [55, 56].

Persian, including Exotics as a short-haired variant, is the most popular and largest purebred cat population. The effect of population size of the Persian breed on the candidate region under selection was evident in the few small sized candidate regions (max = 1.5 Mb). This effect was especially apparent when comparing the two replica populations (Persian 1 and 2) where 50% of the candidate windows were shared (S4 Table). The variability in the detect candidate regions under selection mirrors the variability observed in its signature brachycephaly phenotype and the range of diseases [57]. Nonetheless, two shared candidate regions among the two Persian populations contained promising candidate genes for brachycephaly (Table 1). Both candidate regions contained distal-less homeobox genes (Chr.A2: DLX6 and DLX5—Chr.C1: DLX2), which are involved in the craniofacial development [58, 59]. We suspect that several candidate genomic regions and several genes are responsible for the wide range of brachycephaly phenotypes in Persian cats.

A comparison between the Turkish Van and the Bengal breeds provides an illustration to the age the breed and its selected phenotypes on the numbers and sizes of candidate regions under selection. The Turkish Van is an old breed and belongs to a Mediterranean ancestral cluster [16, 44] with signature coat color pattern named after the breed, “Van”. This coat color appears as an overall white color with different colors randomly covering parts of the head and tail (Fig 3c). On the other hand, the Bengal breed is a comparatively younger breed that resulted from an interspecies hybridization [6] with a marking phenotype of patterned, soft, and glittered coat. Reminiscent of the old age of the Turkish Van breed and its phenotype, several small sized candidate regions were identified, most notably on Chr. E3 (Fig 3c) with CLCN7 as a candidate gene for the hypopigmented appearance of the “Van” coat color [60]. By contrast, much fewer and considerably larger candidate regions were localized for the Bengal breed and two signifying a recent selection on Chrs.B4 (~ 5.5 Mb) and D2 (1.75 Mb). A focus on the candidate region on Chr. D2, which had a manageable number of genes, suggested FGFR2 as a candidate gene for the softness of the Bengal’s coat [61].

The effect of the mode of inheritance of the breed-defining phenotype also has effects on the number and size of candidate regions. As an illustration, we refer to the regions of the Scottish Fold breed, which is defined by the dominantly-inherited folded ears [38] and compare it to American Curl, which characterized by the recessively-inherited backward curled ears (Fig 3a) [62]. Using two separate populations of the Scottish Fold breed (SFOLD1 and 2), we observed: (1) an averaged maximum candidate region size of 2 Mb, (2) significantly low number of shared windows between the two populations (S4 Table), and (3) lack of any candidate windows near the phenotype causing gene, TRPV1 on Chr.D3. Conversely, the America Curl breed exhibited larger candidate regions two of which are indicative of selection of a recessive trait (Fig 3a). The largest candidate region (3.75 Mb) was located on Chr.E1 and among the 123 genes in the region, MCRIP2 represents a likely candidate for curled ears. MCRIP2 is a paralog of FGFRL1, which high localized expression of in cartilaginous tissue [63, 64]. A second candidate region also possessed a candidate gene for the curled ears, PBX1, and is implicated with ear malformations [65].

Lastly, breeding practices during the formation of a cat breed has apparent manifestations on the candidate regions under selection. We use the Abyssinian breed as an example of high inbreeding and the Bombay as an outcrossed breed. While Abyssinian is known to be an inbred breed [16, 44, 55], Bombay is documented to be a hybrid breed between the Eastern Burmese and the Western American Shorthair (Fig 1, S1 Table). The signs of inbreeding of Abyssinian contributed to eight candidate regions with a combined size of 15.5 Mb whereas the outcrossing of Bombay allowed detecting as two candidate regions with a sum of 2.5 Mb.

We examined the robustness of our analysis using known phenotype-related genes and inspected their colocalization in candidate regions under selection (e.g. Cornish Rex [46], Devon Rex [32], Abyssinian [25], Siamese [22], and in longhaired breeds [66, 67]). However, beyond the aforementioned known breed-defining variants, we suggested a number of candidate genes that resided in identified regions of different breeds and might be responsible for phenotypes with still unknown causative genes (Table 1).

The Maine Coon breed is known for its large size (Fig 3b) [68] and its candidate region on Chr.A1 might be related to the large body size. The four genes within the region, GABRG2, GABRA1, GABRA6, and GABRB2, are known to alter the tastebuds and enhance food craving [69, 70]. On the other extreme of body-form, we have detected a shared candidate region across Siamese, Oriental, and Peterbald on Chr.D3. The three Eastern breeds exhibit thin-tubular body (Fig 4). The shared candidate region highlighted ADGRD1 as a potential candidate gene for its association with metabolism and body-weight [71, 72].

Although several genes responsible for the curly coat in cat breeds have been identified [32, 33, 35, 46], the genetic variant responsible for the curled coat of LaPerm is still unknown. We have identified 3.75 Mb candidate region in LaPerm with likely involvement of EPGN and EGER to the unique curly coat of the breed [73]. Examining the Birman via its two populations and the highly shared candidate windows showed several windows on Chr.B1 where KIT is located. The white coat that is located on Birman cats’ feet and referred to as “gloves”, has been linked to KIT (Fig 4) [18]. However, a large candidate region was observed on Chr.B4 with KITLG as a potential second player in the formation of the “gloves” phenotype.

The single largest candidate region across all breeds was observed in Ragdoll on Chr.A1 (7 Mb) and shared among its two populations (RAG1 and 2). Even though the region contained forty genes, the Lysosomal Associated Membrane protein 1 (LAMP1) represented a candidate gene that may explain multiple aspects related to Ragdoll cats. LAMP1 is associated with Lysosomal Trafficking Regulator (LYST), which causes the multi-system Chediak-Higashi Syndrome. The syndrome manifests as dilution of skin, hair, and iris colors, bleeding diathesis, and recurrent infections (reviewed in [74]). Although the LYST and its severe consequences have been identified and described in cats (cited in [75]), we suspect that the large candidate region on Chr.A1 and LAMP1 more moderately contributes to the phenotypic characteristics of Ragdoll cats (e.g. light coat and eye colors) and explains its mucopolysaccharidoses [76].

We have empirically identified candidate regions under selection in twenty-six cat breeds and due to the nature of the data (i.e., Array SNPs ascertained from different breeds), it is not possible to rigorously account for and distinguish between the effects of demography and population structure from artificial selection. Using whole-genome sequence data can foreseeably provide more definitive confirmations of the candidate regions presented. Furthermore, our suggested candidate genes also await an in-depth molecular investigation to reach satisfying conclusions.

This study does not require an ethics statement.

A previously published SNP dataset using the Illumina Feline-infinium 63K SNP array was used in the current study [55]. The dataset contains 34 domestic cat breeds (n = 1,570). A subset of the genotype dataset was selected by excluding cat breeds with sample size less than 10 individuals, and samples with genotyping rate less than 90%. Sample genotyping rate was determined by using the command (--mind) in plink v1.7 [77] (S1 and S2 Data).

Within each cat breed, only unrelated samples were retained to perform further analyses. The genetic relatedness between individuals within each breed was determined using the statistic (PI_HAT) that measures the proportion of Identity by descent (IBD) of alleles for pairs of individuals and was calculated by the command (--genome) in plink v1.7 [77]. A (PI_HAT) of 0.25 or lower was used to select samples that were unrelated at least to the grandparent level. The final list of samples included in the study and their corresponding breed assignment are shown in S5 Table.

The original datasets consists of ~63K SNP markers [55] and the following markers were excluded from downstream analyses: (1) X-chromosome SNPs, (2) unmapped SNPs, (3) SNPs with genotyping rate less than 90%, and (4) SNPs unmapped to felCat 9.0 [51]. The final list of SNPs included in the study and their updated positions to felCat 9.0 are shown in S6 Table. SNP genotyping rate was determined using (--geno) while percentage of monomorphic SNPs was determined with (--freq) both implemented in plink v1.7 [77].

For each population independently, the following summary statistics were calculated using PLINK v1.7 [77]: (1) the function (--freq) was used to calculate the mean and standard deviation of minor allele frequency (MAF), (2) the mean and standard deviation of observed and expected heterozygosity were obtained using the function (--hardy), and (3) the mean individual inbreeding coefficient (F) in each breed was determined with the command (--het).

Population structure and relationships between different cat breeds (n = 26) were analyzed using Principal Component Analysis (PCA) that is based on the variance-standardized relationship matrix implemented in PLINK v1.9 [78]. With the default setting, (--pca) function was used to generate the first 20 principle components. The first two components of the PCA (PC1, PC2) were used to illustrate the overall population structure of the breeds.

F_st was calculated following Weir and Cockerham method [79], which provides accurate estimates of F_st especially when sample sizes are uneven across breeds or are small [80] and using vcftools v 0.1.13 [81]. Original files (.ped and .map) were first converted into Variant Call Format (vcf) file using (--recode-vcf) command in plink v1.9 [78]. Using vcf files and (--weir-fst-pop) command of vcftools, F_st was calculated pairwise between all breeds and for each SNP.

A simple summary statistic (d_i) that measure locus specific divergence in allele frequencies was previously developed and implemented [15, 46, 48]. A script written in R was used to calculate d_i values according to the equation for each breed independently against all breeds. At each SNP, the di represents the sum of the values calculated above.

Based on SNP distribution and ensure sufficient SNP density in our scans, each chromosome was divided into 500Kb overlapping windows (e.g., window1: 0-500Kb, window2: 250-750Kb). The di estimates were averaged for all SNPs within each window (n = 9108 windows). The di averaged values of each window within each chromosome were visualized via an R Manhattan plot script. The 99^th percentile was used as a threshold for window prioritization and the top 92 windows were considered candidate windows under selection. To avoid false positives and to allow comparison across breed beyond the 99^th percentile threshold, the candidate windows were further categorized into candidate regions if three or more windows overlapped. The resulting candidate regions were visualized across breeds in chromosomal plots based on the latest cat genome assembly (felCat 9.0 [51]).