Watermelon accessions WT2 (P₁, lobed leaf), WT20 (P₂, no-lobed leaf), and WCZ (P₃, lobed leaf) are homozygous inbred lines (Fig. 1) that have been manually self-pollinated for at least five generations. Three accessions were used to construct a segregating population of watermelon with lobed leaves. P₁ and P₂ were planted in a greenhouse to derive WT2 × WT20 F₁ generation by crossing. Then, WT2 × WT20 F₁ was self-pollinated to obtain the F₂ population (WT2 × WT20 F₂). Another F₂ population (WCZ × WT20 F₂) was obtained by crossing WT20 and WCZ in the same way. Backcross population of BC₁P₂ or BC₁P₃ was created by hybridizing between WCZ × WT20 F₁ and parents (WT20 and WCZ).

For segregation analysis, the WCZ × WT20 F₂ population was grown and investigated in three experiments conducted in the spring and autumn of 2020 and spring of 2021, with 80,217, and 992 F₂ individuals, respectively. The WT2 × WT20 F₂ population was grown and investigated in two experiments conducted in spring and autumn of 2021 with 231 and 255 F₂ individuals, respectively. Moreover, 113 BC₁P₂ and 69 BC₁P₃ individuals were planted and investigated in the spring of 2020. Parents with 20 individuals and F₁ with 10 individuals were included in each experiment. The leaf phenotype of each individual was observed and recorded for three times independently during plant growth and development, and the segregation ratio was tested using the chi-squared test (Table S1). All watermelon materials used in this study were sown in nursery trays, and the seedlings were transplanted to greenhouses in the Science and Education Park of Henan Agricultural University, Zhengzhou, China.

Genomic DNA from fresh leaves of the WCZ × WT20 F₂ population in the spring of 2020 was extracted using the CTAB method [42], which was used for sequencing and molecular marker analysis. We constructed and re-sequenced two DNA pools, the lobed leaf pool (L-pool) and the no-lobed leaf pool (N-pool), by mixing equal amounts of DNA from 20 lobed leaf individuals and 20 no-lobed leaf individuals in the WCZ × WT20 F₂ population from the spring of 2020. Using an Illumina Genome Analyzer IIx, a paired-end sequencing library (read length of 100 bp) with approximately 500-bp insert sizes was prepared for sequencing. For the sequencing results in L- and N-pools, all short reads were aligned to reference genome ‘97103’ using BWA software [35,43]. Using SAM tools, we converted alignment files into SAM/BAM files [44]. To increase the accuracy of SNP calling, we applied the SNP-calling filter Coval, as previously developed [45]. All SNP positions were calculated using the SNP-index. The SNP-index is the proportion of the difference between the reads harboring SNPs and the reference sequence. SNP positions with a read depth <5 and an SNP-index <0.4 from two sequences were excluded because they may represent pseudo-SNPs due to sequencing/alignment errors or genomic repeat sequences.

According to the SNP-index, we calculated Δ(SNP-index) to identify candidate intervals for lobed leaf loci [45,46]. The Δ(SNP-index) was calculated by subtracting the SNP-index of the L-pool from that of the N-pool. Therefore, if the SNP-index was 0, all short reads were from the no-lobed leaf line; if the SNP-index was 1, all short read fragments were from the ‘97103’ reference genome. A sliding window analysis with a window size of 1 Mb and an increment of 10 kb was used to calculate the average SNP-index of SNPs located in a given genomic region. The SNP-index graphs and the corresponding Δ(SNP-index) graphs were drawn according to the above.

To generate confidence intervals of the SNP-index value under the null hypothesis of no QTL, we conducted a computer simulation. First, we obtained two bulks of offspring through random sampling according to a given number of individuals. A given number of alleles, according to the read depth, were sampled from each bulk. We then calculated the SNP-index of each bulk and obtained the Δ(SNP-index). For each reading depth, this process was repeated 10,000 times and generated a confidence interval. The interval was plotted for all genomic regions with variable read depth.

The whole genome sequence of watermelon ‘97103’ was released to the Cucurbit Genomics Database (http://www.icugi.org) in 2013 [35]; 32,869 SSR markers distributed across the watermelon genome were then developed according to the ‘97103’ genome [47]. To confirm the candidate regions from BSA-seq, SSR markers on chromosome 4 were screened to analyze the polymorphism between the parents WT2 and WT20. Moreover, we constructed two DNA pools in the WT2 × WT20 F₂ population from the spring of 2021, the L1-pool and the N1-pool, by mixing equal amounts of DNA from 20 lobed leaf individuals and 20 no-lobed leaf individuals. The SSR markers, which were polymorphic between WT2 and WT20, were used to screen the polymorphism between two DNA pools. Finally, nine polymorphic markers were used for linkage analysis and primarily mapped through 157 WT2 × WT20 F₂ individuals. Then, eight Indel, five dCAPS, and two SNP markers were developed during the interval of primary mapping according to the re-sequencing data for fine-mapping in the WT2 × WT20 F₂ population, which contained a total of 486 individuals from the spring and autumn of 2021 (Table S2). The Indel and dCAPS markers were designed based on the Primer3 software (http://bioinfo.ut.ee/primer3-0.4.0/) and dCAPS Finder 2.0 [48], respectively. To further narrow the candidate interval, the WCZ × WT20 F₂ population with a total of 1289 individuals was used for linkage analysis and fine-mapping with the same methods as above. The PCR reaction system, reaction procedure, restriction enzyme digestion reactions, and subsequent gel electrophoresis were conducted as described by Dou et al. [40]. Linkage analysis was performed using the Kosambi mapping function of JoinMap 4.0, with the threshold logarithm 10 of the odds (LOD) score of 5.0. The important markers used for linkage analysis in this study are listed in Table S2.

The predicted genes in the candidate region were retrieved from the ‘97103’ watermelon reference genome (http://cucurbitgenomics.org/). Gene-specific primers were designed according to the DNA and CDS sequence of the candidate gene. The DNA and CDS sequence of candidate genes were amplified from young leaves of three parents, WCZ, WT20, and WT2; then, PCR products were sequenced and aligned between the ‘97103’ reference genome and the parents (WCZ, WT20, and WT2) for detecting sequence differences. The function of the candidate gene was predicted based on NCBI (https://blast.ncbi.nlm.nih.gov/Blast.cgi) and the Cucurbitaceae Genome Database (http://cucurbitgenomics.org/). To verify the candidate gene, the co-segregation marker was tested on 100 F₂ individuals (randomly selected in different years) and 94 lobed leaf watermelon germplasm resources (Table S3), which were derived from Guo et al. [49].

To check the expression of candidate gene Clnl, the different tissues (root, stem, male flower, female flower, young leaf, mutant leaf) of watermelon WT2 and WT20 were used to extract RNA for investigating the expression pattern of the candidate gene. Total RNA was extracted using the plant RNA purification kit (Omega, Norcross, USA) according to the manufacturer’s protocol. The concentration and quality of the RNA samples were determined and assayed using NanoDrop 2000 (Thermo) and 1.5% agarose gel, respectively. cDNA was synthesized using reverse transcriptase M-MLV (RNase H-) based on the instructions of the manufacturer (Takara, Kyoto, Japan). The primers of genes for real-time PCR were designed for amplifying products of 100–250 bp using Primer 5.0 online software without any interference with the conserved region. The Actin primer [50] and gene primers for real-time PCR are listed in Table S4. The reaction of qRT-PCR was carried out in accordance with a previous report [40]. The experiments were performed with three biological and three technical replicates, and the relative expression pattern was calculated using the 2^{− ΔΔCT} method [51].

The amino acid sequence of Clnl was retrieved from the Cucurbitaceae Genome Database. A total of 102 homologs of Clnl were identified from A. thaliana, tomato (Solanum lycopersicum), rice (Oryza sativa. L.), and Cucurbitaceae crops (cucumber ‘Chinese Long’ v2, melo ‘DHL92’, watermelon ‘97103’ v1). All information on the 102 homologs of Clnl used for phylogenetic analysis is listed in Table S5. According to the sequence alignments, the phylogenetic tree was constructed based on MEGA7.0 software [52]. Neighbor-joining (NJ) was generated with the model of Poisson correction, 1,000 bootstrap replicates, and pairwise alignment [38].

To study the transcriptional difference of watermelon in leaf development, young leaves of WT2 and WT20 at the three-leaf stage were used for RNA-seq by Biomarker Technologies Co., Ltd. (Beijing, China). Total RNA was extracted from fresh samples according to the CTAB method [53]. The sample contained 1.0 g, and three biological replicates were performed for each sample. RNA concentrations and quality were determined and assayed using NanoDrop 2000 (Thermo) and 1.5% agarose gel, respectively. The RNA-seq library was used to prepare the sample libraries according to the constructed flow path, and then the RNA was sequenced on an Illumina HiSeq2000 platform. Clean reads, obtained by filtering the contaminant sequences and low-quality reads, were aligned to watermelon reference genome ‘97103’ according to HISAT2 [54]. The transcript data were assembled using the software Stringtie [55]. The heatmap was generated using TBtools. The RNA-seq data are available with accession number PRJNA483539 on NCBI, and the detailed statistics and different expression genes of RNA-seq are listed in Tables S6 and S7.

Watermelon accessions WT2 and WCZ were homozygous inbred lines with the normal lobed phenotype in all mature leaves, while watermelon accession WT20 had a no-lobed phenotype in all mature leaves (Fig. 1). The phenotype of the leaf could be easily distinguished from the seedling period to the maturation period. All three accessions were manually self-pollinated for at least five generations and were genetically stable. To study the genetic regulation of lobed leaves in watermelon, WT2, WT20, and WCZ were used as parents to construct genetic populations. No-lobed leaf watermelon WT20 (P₂) was crossed with WT2 (P₁) and WCZ (P₃) for constructing segregation populations. All F₁ individuals had lobed leaves, suggesting that watermelon lobed leaves are dominant over no-lobed leaves (Fig. 1). Segregation populations were constructed through a cross between WCZ and WT20. WT2 was used for constructing the F₂ population by crossing with WT20. The segregation ratio of lobed/no-lobed leaves among different populations in two years is presented in Table S1. The χ² test showed that the segregation ratio in the F₂ and BC₁ populations were in line with 3:1 and 1:1, respectively, indicating that the no-lobed leaf phenotype in watermelon is a qualitative trait and regulated by a single recessive gene, Clnl (Citrullus lanatus no-lobed leaf).

Two pools (L-pool and N-pool) were constructed from the F₂ population in the spring of 2020 for BSA-seq analysis using 20 lobed leaf individuals and 20 no-lobed leaf individuals, respectively. Two pools were sequenced using the Illumina HiSeq^TM PE150 platform. A total of 36.5 GB data were generated from both pools, with approximately 20× depth and more than 99% coverage for each. All data from L- and N-pools were mapped to the watermelon reference genome ‘97103’ v1 (http://cucurbitgenomics.org/) and 5634 SNPs were detected between the two pools. Each identified SNP was computed using an SNP-index. We calculated the average SNP-index in the interval of 1 Mb using a sliding window of 1 kb. Graphs of the SNP-index with the L- and N-pools were generated by plotting the average SNP-index relative to the position of each sliding window in the watermelon genome reference ‘97103’ v1 assembly. The Δ(SNP-index) graph was drawn and the genome locations were calculated by combining the information on the SNP-index in L- and N-pools (Fig. 2a). Only the interval on chromosome 4 between 19.8 and 23.8 Mb had an average Δ(SNP-index) that was higher than 0.4 with a distinct peak. This suggests that candidate gene Clnl was located at an interval of 19.8–23.8 Mb on chromosome 4 in watermelon (Fig. 2a).

To check the BSA-seq results and further fine-map of the candidate gene of watermelon lobed leaves, two F₂ populations (WT2 × WT20 F₂ and WCZ × WT20 F₂) were used for linkage analysis and fine-mapping. The polymorphism SSR markers were screened on chromosome 4 among the parents according to genome-wide SSR markers identified in the watermelon reference genome [47]. A total of nine SSR markers (Table S2) were clearly polymorphic among the three parents (WT2, WT20, and WCZ) and were used for polymorphic analysis in 157 individuals of the WT2 × WT20 F₂ population. The Clnl gene was mapped to 3.9 cM genetic distance between ClSSR11194 and ClSSR11265 (Fig. 2b), which was consistent with the BSA-seq result.

To further narrow the candidate region of the Clnl gene, four Indel markers (Table S2) with significant polymorphisms were developed between ClSSR11194 and ClSSR11265 makers according to the re-sequencing data of the parents. A total of 486 individuals of the WT2 × WT20 F₂ population were used for fine-mapping. The Clnl gene was mapped between Indel3 and Indel4. Furthermore, another four Indel, five dCAPS, and two SNP markers (Table S2) were developed and used for polymorphic analysis in 486 individuals of the WT2 × WT20 F₂ population. To further narrow the predicted region using more F₂ individuals, 1289 individuals from the WCZ × WT20 F₂ population of three cultivated seasons (spring and autumn of 2020 and spring of 2021) were also used for linkage analysis and fine-mapping based on the above markers. The linkage analysis from two F₂ populations (WT2 × WT20 F₂ and WCZ × WT20 F₂) showed that the Clnl gene was located between SNP1 and dCAPS4 markers, with a genetic distance of 1.4 cM. According to the physical position of the markers, the Clnl gene was fine-mapped to the 61.5 kb region between 21,224,481 and 21,285,957 bp on watermelon chromosome 4 (Fig. 2b).

According to the annotation database of the watermelon reference genome ‘97103’ v1 (http://cucurbitgenomics.org/), four functional genes (Cla018360, Cla018361, Cla018362, Cla018363) were annotated in the candidate region. The sequences of these four genes were amplified and aligned among the three parents. The sequences of the three genes were consistent among three parents, while Cla018360 with three exons and total 2387 bp DNA sequence had sequence differences among the three parents. A single-base deletion (A/-) at the 391st base of Cla018360 CDS only existed in the no-lobed leaf watermelon WT20 (Fig. 3). The deletion resulted in premature translation termination in WT20 (Fig. 3b). The amino acid sequence of Cla018360 was 233 aa in WT2, WCZ, and reference genome ‘97103’, while it was only 158 aa in WT20 (Fig. 3c). The function annotation from the watermelon reference genome and BLAST-P search in the Arabidopsis database (https://www.arabidopsis.org/) showed that Cla018360 belongs to the HD-Zip gene family, which has been reported to regulate leaf development [41,56]. Therefore, we further suggested that Cla018360 was the candidate gene controlling the lobed/no-lobed leaf phenotype in watermelon.

To validate the candidate gene, we selected 100 F₂ individuals (randomly selected in different years) for testing for polymorphisms according to the co-segregation marker of Cla018360. The results showed that all 77 lobed leaf individuals were dominant (24 individuals for homogeneous and 53 individuals for heterozygous dominance), and 23 no-lobed leaf individuals were homozygous recessive, which was consistent with the phenotypes. Furthermore, 94 lobed leaf watermelon germplasm resources [44] were used for analysis of the sequence difference of Cla018360. The results showed that there was no mutation of Cla018360 in all 94 accessions (Table S3), which indicated that the normal Cla018360 may regulate watermelon lobed leaves. These results further indicate that the Cla018360 gene is the candidate gene Clnl controlling watermelon lobed/no-lobed leaves, and mutation of Cla018360 CDS may result in a no-lobed leaf phenotype in watermelon.

The expression patterns of Cla018360 were investigated via qRT-PCR in different tissues (root, stem, male flower, female flower, young leaves, and mature leaves) between WT2 and WT20. The specific primers of Cla018360 and the reference gene Actin are listed in Table S4. The results showed that the expression level of Cla018360 in lobed leaf watermelon WT2 was higher than in no-lobed leaf watermelon WT20. The young leaves had the highest expression among all tested tissues, with a significant difference between WT2 and WT20 (Fig. 4). The lobed/no-lobed leaves can be distinguished at the watermelon seedling stage. This corresponded to the expression pattern of young leaves, suggesting that the Cla018360 gene plays an important role in the leaf formation stage.

To investigate the classification and function and to gain insights into the potential function of the HD-Zip gene family, 102 HD-Zip proteins from Arabidopsis thaliana, S.lycopersicum, O. sativa, Cucumis sativus, Cucumis melo, and Citrullus lanatus were identified for phylogenetic analysis (Table S5). The phylogenetic tree was constructed according to amino acid sequences of all 102 HD-Zip proteins, which were divided into four groups (I, II, III, and IV) (Fig. 5). There were similar numbers of genes in the four groups. The gene Cla018360 was clustered into group I according to phylogenetic analysis and had a close genetic relationship with AtHB51, which has been reported to regulate the formation of leaf shape in Arabidopsis [14,56]. These results further suggested that the Cla018360 gene may play an important role in regulating lobed/no-lobed leaf formation in watermelon.

To explore the regulatory pathway by which the Clnl gene regulates the formation of leaf shape in watermelon, according to the expression pattern of Clnl in different tissues, the young leaves from WT2 and WT20 at the three-leaf period were used for RNA-seq. The expression profiles of 333 genes were more than two-fold between WT2 and WT20, of which 115 genes were upregulated, and 218 genes were downregulated in no-lobed leaf watermelon WT20 (Table S7). We selected 23 transcription factors from 333 differential expressed genes to create the heatmap (Fig. 6), of which nine transcription factors were upregulated, and 14 transcription factors were downregulated in no-lobed leaf watermelon WT20. Furthermore, eight genes, which were reported to regulate leaf development, were selected to validate the expression level between lobed leaf watermelon WT2 and no-lobed leaf watermelon WT20 using the qRT-PCR (Fig. 7). The results showed that ClERF5, ClADP, ClCUC2, ClKNAT, and TCP-like genes had higher expression level in WT20 compared with WT2, in conformity with the RNA-seq results. In contrast, the transcript levels of two HP-zip genes and the ClCUC3 gene were more highly expressed in WT2 compared with WT20, similar to the RNA-seq data. These results indicate that the Clnl gene and other leaf development genes together regulate the formation of watermelon leaf shape.