Founder fish sampling and fish housing.

Founder fish for inbreeding were first-generation lab-reared guppies from the Lower Guanapo River (Twin Bridge North West Trinidad, PS 91100 77800) where they are neither endangered nor protected. The specimens were kindly donated by Dr. David Reznick, UC Riverside, in 2009. Fish collection and export of these fish was approved by the Ministry of Agriculture, Land and Marine Resources, Republic of Trinidad and Tobago, conforming to their legislation. Since their collection progeny was kept and bred at the Max Planck Institute for Developmental Biology, Tübingen, according to German legislation. The facility was approved by the Regierungspräsidium Tübingen, registration number 35/9185.46. Fish required for preparation of DNA were anesthetized with a lethal dose of MS222 before being stored in 95% ethanol.

Genome sequencing.

DNA from a 5^th-generation female was used to prepare Illumina paired-end libraries with insert sizes of 240 to 460 bp and DNA from female offspring of the same lineage at generations 6 to 8 (six individuals total) was used to construct Roche/Illumina hybrid mate-pair libraries of 3 to 20 kb length; for details see [36]. Briefly, mate-pair libraries were prepared by ligating a circularization adaptor (Roche) and fragments were selected. The fragments were then circularized by Cre Recombinase, following the Roche protocol for 454 sequencing. Linear DNA was digested before circularized DNA was fragmented. Illumina paired-end standard adaptors were ligated onto these fragments, mate-pairs (of 180 to 480 bp length) were amplified and sequenced from both ends on the Illumina GA II platform. Long-jump 40 kb fosmid libraries were constructed from a single female offspring at generation 8 using the Nx 40 kb mate-pair cloning kit from Lucigen (Middleton, USA). Fosmid clones were amplified, fosmid DNA extracted in bulk, digested with BfaI and end fragments of 8–9 kb size (vector including insert ends) were selected. After recircularization, and digestion of linear DNA, mate-pairs were amplified by PCR and sequenced on the Illumina HiSeq2000 platform. See S1 Table for details about library sizes and sequencing yield.

Read filtering and quality trimming.

All genomic libraries were retrieved and converted to FASTQ format using the import and convert commands in shore version 0.7.1 [37]. To remove PCR duplicates for a non-random fragment representation, each library was scanned with the filterPCRdupl.pl script (Version 1.01) included in ConDeTri version 2.0 [38]. For paired-end libraries, the first 50 bp of both reads of a pair were compared, and for mate-pair libraries the first 35 bp.

Mate-pair libraries were screened for the following adapter sequences:

5’–TCGTATAACTTCGTATAATGTATGCTATACGAAGTTATTACG– 3’

5’–CGTAATAACTTCGTATAGCATACATTATACGAAGTTATACGA– 3’

Screening was conducted using cutadapt version 1.1 [39], with an error rate of 0.15, an overlap of 6 bp, minimum read length of 35 bp, and matching wildcards. The screening resulted in two sets of mate-pairs: one set that contained parts of the adapter sequences and the other set without any adapter sequence. All other libraries were filtered for low quality bases using ConDeTri version 2.0 (using default parameters for paired-end libraries and with parameters–rmN–hq 20 –minLen 25 for mate-pair libraries).

Estimating mate-pair library insert sizes.

To make use of mate-pair library information even in the absence of adapter sequences in the reads we first prepared a draft assembly. For this assembly SOAPdenovo version 1.05 [40] (kmer size 27, -d 1, -D 2, -F) was used. In detail, all paired-end libraries were used for contig building, but only mate-pair libraries with adapter sequence were included; these mate-pair reads were added stepwise according to their insert size (smallest to longest) for eight rounds of scaffolding. Mate-pair reads without adapter sequence were mapped to the assembly using bwa version 0.6.1 [41] with default parameters. Mate-pairs that mapped within 1 kb distance of each other were excluded from further assembly steps to prevent possible paired-end contamination in the data. Insert sizes for the remaining reads were estimated per library based on distances between the reads.

Fosmid 40 kb insert library.

The fosmid 40 kb library was treated slightly differently from the mate-pair libraries. First, phiX sequences were removed by mapping the library to the phiX genome sequence using bwa with default parameters. Then, reads were filtered for PCR duplicates and low quality bases, similar to the mate-pair libraries. Finally, reads were trimmed from the 3’ end to keep 50 bp per read due to the low quality towards the 3’ end of the reads. Insert sizes of the filtered and trimmed reads were estimated as described above (using the SOAPdenovo assembly).

Genome assembly.

The final genome assembly was built with Allpaths-lg version 43668 [42] using default parameters. Due to a quality correction step in the Allpaths-lg pipeline, we used PCR filtered libraries for the paired-end and mate-pair libraries, and PCR-filtered and quality-trimmed fosmid libraries. Overlapping paired-end libraries were used for contig building (insert sizes of 240 and 270 bp, library ID 1–5). Longer insert paired-end libraries (insert size 460 bp, library ID 6–8), mate-pair libraries (ID 9–19), and the fosmid library (ID 20) were incorporated in the scaffold building process. Allpaths-lg was used to estimate the heterozygosity rate of the reference genome sample.

After assembly, all paired-end reads were mapped back to the genome assembly using bwa version 0.6.2 to estimate the proportion of the genome that was not covered by paired-end reads. Only a very small proportion of the genome assembly was not covered by paired-end reads (224,382 bp or 0.03%). Additionally, the assembly was screened for regions of runs of Ns. These regions occur during the scaffolding process and denote that sequence information between two contigs might be missing.

Contamination removal.

To exclude cross contamination in the assembly from other organisms, we aligned the final genome assembly against the NCBI nt database (blastn version 2.2.21, e-value cutoff 10⁻⁵ [43]), reporting the best hit only. Scaffolds with hits only against non-vertebrate organisms were excluded from the assembly. This strategy excluded eight scaffolds with combined length of 27,088 bp (0.004% of the assembly) from the final assembly. Additionally, a more stringent contamination screen for adaptors was performed by NCBI resulting in 1,260 bp of potential adaptor sequences that were removed from the assembly. Visual inspection did not reveal any signs of mis-assemblies in these regions.

Genetics map integration.

Scaffolds were anchored on linkage groups using a genetic linkage map built from 5,493 RAD-seq markers (for details about the linkage map see S1 File). Markers were aligned against the assembled genome (blastn version 2.2.27+, e-value < 10⁻²⁰ [44]) and only markers with unique hits were used to anchor scaffolds using the method described in [1]. Adjacent scaffolds were separated by a character string of 100 Ns. Scaffolds that could not be reliably anchored to one of the linkage groups were grouped into LG Un.

Assembly validation.

Guppy expressed sequence tags (ESTs), were downloaded from the NCBI EST database (accessed 2013-09-25) and blasted (blastn version 2.2.27+, e-value < 10⁻¹⁰ [44]) against the draft genome assembly. Additionally, a set of 454 transcriptome sequences [45] was downloaded (http://www.bio.fsu.edu/kahughes/Databases.html) and blasted (blastn, e-value < 10⁻¹⁰) against the draft genome sequence. See S1 File for further assembly validation steps.

Repeat content.

A guppy specific repeat library was built using the draft guppy genome assembly and RepeatModeler version open-1.0.5 [46] in combination with ab-blast version 2.2.6 [47] (default parameters used for both programs). The resulting repeat-library was applied to identify and mask repeats in the draft assembly using RepeatMasker version open-3.3.0 [48], with ab-blast used as the search engine (default parameters).

Gene set annotation.

Genes were annotated using ‘The NCBI Eukaryotic Genome Annotation Pipeline’ (http://www.ncbi.nlm.nih.gov/genome/annotation_euk/process/, accessed 2014-09-05). This automated pipeline annotated genes, transcripts and proteins on the draft genome assembly (NCBI Poecilia reticulata Annotation Release 100).

Genes with potential functions in pigment pattern development, vision, growth and sex differentiation.

To identify protein coding genes with putative functions in pigment pattern development, vision, growth and sex differentiation, reciprocal blast searches were performed using a sample of 624 published coding sequences (S13 Table), mostly from related fish species. Resulting high-scoring segment pairs (HSPs) were manually scrutinized for e-value (< 10⁻²⁰), percent identity (>90%), length, and reading frame. Further, predicted gene models from scaffolds localized on LG12 were searched against NCBI nt and nr databases as well as Ensembl (version 71) databases of medaka, stickleback and platyfish. To resolve positions of LWS-1 (A180), LWS-2 (P180 and LWS-3 (S180) in the cluster on scaffold 43, all exons were scrutinized for best e-value (≤ 10⁻⁴⁰), percent identity and coding strand. Further, we aligned a Cumaná genomic BAC (GenBank HM540108.1) to genomic scaffold 43. A region of 32,500 bp length in the reconstructed BAC sequence corresponds to the opsin gene cluster that was aligned to about 38,300 bp on scaffold 43 of our assembly (96 to 98% identity, excluding few gaps). Reciprocal blast alignment revealed stretches of N in the Allpaths-lg assembly (up to 4,500 bp) as the main reason for length discrepancy. We inspected this region by eye but did not find evidence that points towards a mis-assembly in this region. A potential explanation for the length discrepancy is wrongly estimated insert sizes of the mate pair libraries for this particular region. Another possible explanation is a length difference for this particular region between the Cumaná and Guanapo strains.

Small non-coding RNAs and transfer RNAs (tRNAs).

Small non-coding RNA loci were annotated using Infernal version 1.1rc1 [49] (e-value threshold 10⁻⁴) in combination with Infernal Rfam database version 1.1. To annotate tRNAs we additionally ran tRNA-Scan version 1.3.1 [50].

Pairwise alignments/Synteny analysis.

The guppy genome was aligned to repeat-masked versions of the medaka (Oryzias latipes) and stickleback (Gasterosteus aculeatus) genomes (Ensembl version 71) using NUCmer version 3.1 from the MUMmer package version 3.23 [51]. Alignments were visualized with Circos plot version 0.67–7 [52].

Three-way alignments.

Coding sequence annotations for the guppy were downloaded from GenBank (Accession GCF_000633615.1) and coding sequences from platy (Xiphophorus maculatus) and medaka (Oryzias latipes) were downloaded from Biomart (Ensembl 70). Three-way 1:1:1 orthology sets were identified using ProteinOrtho version 5.11 (parameter settings: minimum similarity for additional hits 0.8, blastp+). In total, 10,840 1:1:1 orthologs were identified. Next, codon sequences were aligned using prank version 140603 [53] with an empirical codon model.

Substitution rate estimates.

Substitution rates were estimated separately for synonymous (d_S) and nonsynonymous (d_N) substitutions per nucleotide using a maximum likelihood method, implemented in the Codeml program (model = 1, star-like user tree specified according to the phylogeny) of the Paml package v4 [54]. Alignments with d_S > 2 along any branch were excluded to minimize statistical artifacts from short sequences and saturation effects in d_S (no alignment showed an estimated d_N > 2). The final data set comprised 9,111 1:1:1 orthologs with mean d_S estimates of 0.052 (±0.033 s.d.), 0.058 (±0.040 s.d.), and 0.949 (±0.262 s.d.) for the platy, guppy, and medaka branches, respectively. Estimates for average d_N were 0.008 (±0.010 s.d.), 0.008 (±0.013 s.d.) and 0.093 (±0.066 s.d.) for the platy, guppy, and medaka branches, respectively.

Parent-offspring trios were used.

We crossed a Quare female (laboratory strain, originally from the Quare River in North East Trinidad) with an EnUlmBL male (laboratory strain, presumably originally from Venezuela). RAD-seq libraries for the parents and five F₁ individuals were prepared using the restriction enzymes PstI and MseI and 8 unique barcodes for each parent and one unique barcode for each F₁ individual [55]. The RAD-seq libraries, with an approximate insert size of 120–220 bp, were sequenced single-end on an Illumina HiSeq 2000 lane.

The raw reads were obtained from the sequencing platform, converted to FASTQ format and de-multiplexed using shore version 0.8.1. Read mapping was performed separately for each individual with bowtie2 version 2.1.0 [56]. Mapping results were enhanced by local realignment using GATK version 2.4–9 [57]. Single nucleotide polymorphism (SNP) detection was performed using GATK UnifiedGenotyper (default parameters). High quality bases were extracted using the mpileup command, implemented in SAMtools version 0.1.18 [58] (BAQ score cutoff 20).

To detect de novo mutations, SNP calls from the parents were compared with the offspring using only sites that were covered by at least 10 reads. Approximately 16 million bases reached the base quality cutoff and between 0 and 2 new mutations were detected per F₁ individual. None of the de novo mutations were detected in more than one individual.

Sampling and sequencing.

Ten males from a downstream high-predation population were collected in 2011 (PS 91100 77800, Twin Bridges) from the Guanapo drainage in North West Trinidad. Fish were euthanized with MS222 and stored in 95% ethanol.

Paired-end DNA sequencing libraries were prepared according to the "Illumina Paired End Preparation protocol"; using unique barcoded Illumina TruSeq adaptors for each individual. The PCR amplified fragments were size selected on a 2% Low Range Ultra Agarose (Bio-Rad) gel. Libraries were pooled and sequenced on two flowcell lanes with an Illumina HiSeq 2000 instrument, aiming for approximately 10x coverage per individual (101 bp read length).

Data preparation.

Libraries were checked out from the sequencing platform using the shore import command version 0.8.1 to retrieve raw data. Raw reads were converted to fastq files using the shore convert command (see S2 Table for details about sequencing yield per sample).

Paired-end reads were mapped to the reference genome using bowtie2 version 2.10, applying the ‘end-to-end’ mapping option (no read-clipping) in the ‘very-sensitive’ mode. Discordant alignments for paired reads were suppressed. Mapping was enhanced for each individual by local realignment as implemented in GATK version 2.4–9 (RealignerTargetCreator, IndelRealigner) and duplicates were marked using Picard version 1.89 (http://picard.sourceforge.net, last accessed 2014-07-09).

SNP calling.

SNPs were called using three different variant calling programs: GATK UnifiedGenotyper, freebayes version 0.9.9 [59], and SAMtools mpileup version 0.1.18. GATK UnifiedGenotyper and SAMtools mpileup were run with standard parameters, freebayes was run with ‘—no-indels—no-mnps—no-complex—use-mapping-quality’ parameters. SNPs called by all three SNP calling approaches (~0.7 million) were selected as input data for base quality recalibration with the default set of covariants (GATK BaseRecalibrator). Base quality of the mapping files was adjusted using GATK PrintReads and the resulting bam files were then used as input for a second round of SNP calling with GATK UnifiedGenotyper (standard parameters). Variants with a quality by depth annotation above 10.0 (QD>10.0) were selected as the training set for quality recalibration and filtering using GATK (~2.1 million). Next, we conducted a third round of variant calling using GATK UnifiedGenotyper. As there was no significant difference in the number (and location) of SNPs between the second and the third round, the third round SNPs were used as the final variant set for downstream analyses. In the final SNP set, the average transition-to-transversion ratio was 1.35. The predicted effect of each SNP was annotated using SnpEff version 3.3h [60].

Pairwise nucleotide diversity estimation.

The genome was divided into non-overlapping windows of 50 kb for each scaffold separately. The last window of a scaffold was disregarded if it was shorter than 50 kb. Scaffolds were ordered and oriented along chromosomes as described above (see section ‘Physical assembly’). Windows with at least 25 kb coverage of unique sequences were retained for downstream analyses; unique sequences were defined as the sites within each window that were neither N nor repeat masked. Additionally, we required that these windows were covered by sequence information from at least 7 individuals (i.e., a missing data threshold of 30%). For each retained window, average pairwise nucleotide diversity (π, 0…1) was computed for non-repetitive sites using compute, as implemented in the analysis package version 0.8.3 (https://github.com/molpopgen/analysis, last accessed 2014-11-11) found in the libsequence library [61] with default parameters.

Demographic inference.

To infer the effective population sizes (N_e) history of the Guanapo high predation population, a coalescent-based hidden Markov model (PSMC) was applied as described in [62]. This method infers ancestral N_e over time, exploiting a probabilistic model of coalescence that accounts for recombination and changes in heterozygosity rates along a single diploid genome. We ran PSMC with standard parameters for each individual using recalibrated bam files (see above for details) on long, repeat masked scaffolds (size >10 kb). To visualize the results, psmc_plot.pl was used, assuming a mutation rate of 4.89x10^-8 bp per generation and a generation time of 0.5 years [20]. Results were plotted for 800 to 25,000 generations ago, which translates to 400 to 12,500 years before present. The lower bound was set because estimates more recent than 800 generations are difficult to predict by the PSMC method [62].

Resequencing and assembly.

From each of the 10 males used for resequencing, reads that did not align to the female genome were extracted using the view command from the SAMtools package. Reads from all individuals were pooled together, resulting in 35 million read pairs. To lower the computational burden, ten million read pairs were randomly extracted from this pooled set of unmapped reads and assembled using Trinity version r20140717 [6]. Contigs shorter than 1,000 bp were discarded and only the longest isoform was kept for each assembled component. Other assemblers (SOAPdenovo, Velvet [63] and ABySS [64]) were tested for the male-specific assembly as well, but resulted in higher fragmentation compared to the Trinity assembly (data not shown).

Annotation.

We used blastn to compare contigs against env-nt and nt databases (blastn version 2.2.30+, e-value < 10⁻⁵, database version 20141104) to screen for contamination. Contigs with hits against env-nt longer than 5% of the contig size were removed. Additionally, contigs with no hits against nt were removed. The remaining contigs were blasted against the female genome (blastn, e-value < 10⁻¹⁰) and all contigs with alignment length longer than one-quarter of the contig length were discarded. The resulting contigs were compared to the nr database (blastx, e-value < 10⁻¹⁰, database version 20151112). Only hits with at least 120 amino acids of target protein coverage, 50% identity and an e-value < 10⁻⁴⁰ were kept.