## Main

An accurate and complete genome assembly is required to understand the unique aspects of mosquito biology and to develop control strategies to reduce their capacity to spread pathogens1. The Ae. aegypti genome is large (approximately 1.25 Gb) and highly repetitive, and a 2007 genome project (AaegL3)2 was unable to produce a contiguous genome fully anchored to a physical chromosome map3 (Fig. 1a). A more recent assembly, AaegL44, produced chromosome-length scaffolds that made it possible to detect larger-scale syntenic genomic regions in other species but suffered from short contigs (contig N50, 84 kb, meaning that half of the assembly is found on contigs >84 kb) and a correspondingly large number of gaps (31,018; Fig. 1b). Taking advantage of rapid advances in sequencing and assembly technology in the last decade, we used long-read Pacific Biosciences sequencing and Hi-C (a high-throughput sequencing method based on chromosome conformation capture) scaffolding to produce a new reference genome (AaegL5) that is highly contiguous, with a decrease of 93% in the number of contigs, and anchored end-to-end to the three Ae. aegypti chromosomes (Fig. 1 and Extended Data Figs. 1, 2). Using optical mapping and linked-read sequencing, we validated the local structure and predicted structural variants between haplotypes. We generated an improved gene set annotation (AaegL5.0), as assessed by a mean increase in RNA-sequencing (RNA-seq) read alignment of 12%, connections between many gene models that were previously split across multiple contigs, and a roughly twofold increase in the enrichment of assay for transposase-accessible chromatin using sequencing (ATAC-seq) alignments near predicted transcription start sites. We demonstrate the utility of AaegL5 and the AaegL5.0 annotation by investigating a number of scientific questions that could not be addressed with the previous genome annotations.

This project used the Liverpool Aedes Genome Working Group (LVP_AGWG) strain, related to the AaegL3 Liverpool ib12 (LVP_ib12) assembly strain2 (Fig. 1c and Extended Data Fig. 1a). Using flow cytometry, we estimated that the genome size of LVP_AGWG is approximately 1.22 Gb (Fig. 1d and Extended Data Fig. 1b). To generate our primary assembly, we produced 166 Gb of Pacific Biosciences data (around 130× coverage for a 1.28-Gb genome) and assembled the genome using FALCON-Unzip5. This resulted in a total assembly length of 2.05 Gb (contig N50, 0.96 Mb; and NG50, 1.92 Mb, meaning half of the expected genome size found on contigs >1.92 Mb). FALCON-Unzip annotated the resulting contigs as either primary (3,967 contigs; N50, 1.30 Mb; NG50, 1.91 Mb) or haplotigs (3,823 contigs; N50, 193 kb), representing alternative haplotypes present in the approximately 80 male siblings pooled for sequencing (Table 1 and Extended Data Fig. 1e). The primary assembly was longer than expected for a haploid Ae. aegypti genome, as predicted by flow cytometry and prior assemblies, which was consistent with remaining alternative haplotypes that were too divergent to be automatically identified as primary and associated alternative haplotig pairs.

To generate a linear chromosome-scale reference genome assembly, we combined the primary contigs and haplotigs that were generated by FALCON-Unzip to create an assembly comprising 7,790 contigs. We used Hi-C to order and orient these contigs, correct misjoined sections and merge overlaps (Extended Data Fig. 1c–e). We set aside 359 contigs that were shorter than 20 kb and used the Hi-C data to identify 258 misjoined sections, resulting in 8,306 ordered and oriented contigs. This procedure revealed extensive sequence overlap among the contigs, consistent with the assembly of numerous alternative haplotypes. We developed a procedure to merge these alternative haplotypes, removing 5,440 gaps and boosting the contiguity (N50, 5.0 Mb; NG50, 4.6 Mb). This procedure placed 94% of sequenced (non-duplicated) bases onto three chromosome-length scaffolds that correspond to the three Ae. aegypti chromosomes. After scaffolding, we performed gap-filling and polishing using Pacific Biosciences reads. This removed 270 gaps and further increased the contiguity (N50, 11.8 Mb; NG50, 11.8 Mb), resulting in a final 1.279-Gb AaegL5 assembly and a complete mitochondrial genome (Fig. 1e and Table 1). We used Hi-C contact maps to estimate the position of the centromere with a resolution of around 5 Mb: chromosome 1, approximately 150–154 Mb; chromosome 2, around 227–232 Mb, chromosome 3, around 196–201 Mb. There are 229 remaining gaps in the primary assembly, including 173 on the three primary chromosomal scaffolds (Extended Data Fig. 2a and Supplementary Data 1). Analysis of near-universal single-copy orthologues using BUSCO6 revealed a slight increase in complete single-copy orthologues and a reduction in fragmented and missing genes compared to previous assemblies (see Supplementary Methods and Supplementary Discussion). AaegL5 is markedly more contiguous than AaegL3 and AaegL4 assemblies2,4 (Fig. 1a, b, e and Table 1). Using the TEfam, Repbase and de novo identified repeat databases, we found that 65% of AaegL5 was composed of transposable elements and other repetitive sequences (Fig. 1f and Supplementary Data 2, 3).

Complete and correct gene models are essential for studying all aspects of mosquito biology. We used the NCBI RefSeq annotation pipeline to produce annotation version 101 (AaegL5.0; Extended Data Fig. 2b) followed by manual curation of key gene families. AaegL5.0 formed the basis for a comprehensive quantification of transcript abundance in 253 sex-, tissue- and developmental stage-specific RNA-seq libraries (Supplementary Data 48). The AaegL5.0 gene set is considerably more complete and correct than previous versions. Many more genes have high protein coverage when compared to Drosophila melanogaster orthologues (915 more genes with >80% coverage, a 12.5% increase over AaegL3.4; Fig. 1g) and >12% more RNA-seq reads map to the AaegL5.0 gene set annotation than AaegL3.4 (Fig. 1h and Supplementary Data 9). In addition, 1,463 genes that were previously annotated separately as paralogues were collapsed into single gene models and 481 previously fragmented gene models were completed (Supplementary Data 10, 11). For example, sex peptide receptor is represented by a six-exon gene model in AaegL5.0 compared to two partial gene fragments on separate scaffolds in AaegL3.4 (Extended Data Fig. 2c). Genome-wide, we mapped a 1.8-fold higher number of ATAC-seq reads, known to co-localize with promoters and other cis-regulatory elements7, to predicted transcription start sites in AaegL5.0 compared to AaegL3.4, consistent with more complete gene models in AaegL5.0 (Extended Data Fig. 2d).

We next validated the base-level and structural accuracy of the AaegL5 assembly. We estimate the lower bound of base-level accuracy of the assembly to have a quality value of 34.75 (meaning that 99.9665% of bases are correct, see Supplementary Methods and Supplementary Discussion). To develop a fine-scale physical genome map based on AaegL5, we compared the assembly coordinates of 500 bacterial artificial chromosome (BAC) clones containing Ae. aegypti genomic DNA with physical mapping by fluorescence in situ hybridization (FISH) (Extended Data Fig. 2e and Supplementary Data 12). After removing repetitive BAC-end sequences and those with ambiguous FISH signals, 377 out of 387 (97.4%) of probes showed concordance between physical mapping and BAC-end alignment. The 10 remaining discordant signals were not supported by Bionano or 10X analysis, and therefore probably do not reflect misassemblies in AaegL5. The genome coverage of this physical map is 93.5%, compared to 45% of AaegL38, and is one of the most complete genome maps across mosquito species9,10.

## Curation of multi-gene families

Large multi-gene families are very difficult to assemble and correctly annotate, because recently duplicated genes typically share high sequence similarities or can be misclassified as alleles of a single gene. We curated genes in large multi-gene families that encode proteases, G protein-coupled receptors, and chemosensory receptors using the improved AaegL5 genome and AaegL5.0 annotation. Serine proteases mediate immune responses11 and metalloproteases have been linked to vector competence and mosquito–Plasmodium interactions12. Gene models for over 50% of the 404 annotated serine proteases and metalloproteases in AaegL3.4 were improved in AaegL5.0, and we found 49 previously unannotated protease genes (Supplementary Data 13). G protein-coupled receptors are membrane proteins that respond to diverse external and internal sensory stimuli. We provide major corrections to gene models that encode 10 visual opsins and 17 dopamine and serotonin receptors (Extended Data Fig. 2f and Supplementary Data 1416). Three large multi-gene families of insect chemosensory receptors are ligand-gated ion channels: odorant receptors (OR gene family), gustatory receptors (GR gene family) and ionotropic receptors (IR gene family). These collectively allow insects to sense a vast array of chemical cues, including carbon dioxide and human body odours that activate and attract female mosquitoes. We identified 117 odorant receptors, 72 gustatory receptors (encoding 107 transcripts) and 135 ionotropic receptors in the AaegL5 assembly (Fig. 2a, b, Extended Data Fig. 3 and Supplementary Data 1720), inferred new phylogenetic trees for each family to investigate the relationship of these receptors in Ae. aegypti, Anopheles gambiae malaria mosquitoes and D. melanogaster (Extended Data Figs. 46), and revised expression estimates for adult male and female neural tissues using deep RNA-seq13 (Extended Data Fig. 7). Our annotation identified 54 new ionotropic receptor genes (Fig. 2b, Extended Data Fig. 3 and Supplementary Data 17), nearly doubling the known members of this family in Ae. aegypti. We additionally reannotated ionotropic receptors in An. gambiae and found 64 new genes. In Ae. aegypti, chemoreceptors are extensively clustered in tandem arrays (Fig. 2a and Extended Data Fig. 3), in particular on chromosome 3p, in which over a third of all chemoreceptor genes (n = 111) are found within a 109-Mb stretch. Although 71 gustatory receptor genes are scattered across chromosomes 2 and 3, only AaegGr2, a subunit of the carbon-dioxide receptor, is found on chromosome 1. Characterization of the full chemosensory receptor repertoire will enable the development of novel strategies to disrupt mosquito biting behaviour.

## Structure of the sex-determining M locus

Sex determination in Aedes and Culex mosquitoes is governed by a dominant male-determining factor (M factor) at a male-determining locus (M locus) on chromosome 114,15,16. This chromosome is homomorphic between the sexes except for the M/m karyotype, meaning that males are M/m and females are m/m. Despite the recent discovery of the M factor Nix in Ae. aegypti17, which was entirely missing from the previous Ae. aegypti genome assemblies2,4, the full molecular properties of the M locus remain unknown. We aligned AaegL5 (from M/m males) and AaegL4 (from m/m females), and identified a region that contained Nix in AaegL5 at which the assemblies diverged and that may represent the divergent M/m locus (Fig. 3a). A de novo optical map assembly spanned the putative AaegL5 M locus and extended beyond its two borders. We estimated the size of the M locus at approximately 1.5 Mb, including an approximately 163-kb gap between contigs (Fig. 3a, c). We tentatively identified the female m locus as the region in AaegL4 not shared with the M locus-containing chromosome 1, but note that the complete phased structure of the divergent male M locus and corresponding female m locus remain to be determined. Nix contains a single intron of 100 kb, while myo-sex, a gene encoding a myosin heavy chain protein that has previously been shown to be tightly linked to the M locus18, is approximately 300 kb in length. More than 73.7% of the M locus is repetitive: long terminal repeat retrotransposons comprise 29.9% of the M locus compared to 11.7% genome-wide. Chromosomal FISH with Nix- and myo-sex-containing BAC clones19 showed that these genes co-localize to the 1p pericentromeric region (1p11) in only one homologous copy of chromosome 1, supporting the placement of the M locus at this position in AaegL5 (Fig. 3b). We investigated the differentiation between the sex chromosomes (Fig. 3d) using a chromosome quotient method to quantify regions of the genome with a strictly male-specific signal20. A sex-differentiated region in the LVP_AGWG strain extends to an approximately 100-Mb region surrounding the approximately 1.5-Mb M locus. This is consistent with the recent analysis of male–female FST in wild population samples and linkage map intercrosses21 and could be explained by a large region of reduced recombination encompassing the centromere and M locus22. The availability of a more completely assembled mosquito M locus provides opportunities to study the evolution and maintenance of homomorphic sex-determining chromosomes. The sex-determining chromosome of Ae. aegypti may have remained homomorphic at least since the evolutionary divergence between the Aedes and Culex genera more than 50 million years ago. With the more completely assembled M locus, we can investigate how these chromosomes have avoided the proposed eventual progression into heteromorphic sex chromosomes23.

## Structural variation and gene families

Structural variation is associated with the capacity to vector pathogens24. We produced ‘read cloud’ Illumina sequencing libraries of linked reads with long-range (around 80 kb) phasing information from one male and one female mosquito using the 10X Genomics Chromium platform to investigate structural variants, including insertions, deletions, translocations and inversions, in individual mosquitoes. We observed abundant small-scale insertions and deletions (indels; 26 insertions and 81 deletions called, median 42.9 kb) and inversions and/or translocations (29 called) in these two individuals (Extended Data Fig. 8a and Supplementary Data 21). Eight of the inversions and translocations coincided with structural variants seen independently by Hi-C or FISH, suggesting that those variants are relatively common within this population and can be detected by different methods. AaegL5 will provide a foundation for the study of structural variants across Ae. aegypti populations.

Hox genes encode highly conserved transcription factors that specify segment identity along the anterior–posterior body axis of all metazoans25. In most vertebrates, Hox genes are clustered in a co-linear arrangement, although they are often disorganized or split in other animal lineages26. All expected Hox genes are present as a single copy in Ae. aegypti, but we identified a split between labial and proboscipedia placing labial on a separate chromosome (Extended Data Fig. 8b and Supplementary Data 22). We confirmed this in AaegL4, which was generated with Hi-C contact maps from a different Ae. aegypti strain4, and note a similar arrangement in Culex quinquefasciatus, suggesting that it occurred before these two species diverged. Although this is not unprecedented27, a unique feature of this organization is that both labial and proboscipedia appear to be close to telomeres.

Glutathione S-transferases (GSTs) are a large multi-gene family involved in the detoxification of compounds such as insecticides. Increased GST activity has been associated with resistance to multiple classes of insecticides, including organophosphates, pyrethroids and the organochlorine dichlorodiphenyltrichloroethane (DDT)28. Amplification of detoxification genes is one mechanism by which insects can develop insecticide resistance29. We found that three insect-specific GST epsilon (GSTe) genes on chromosome 2, located centrally in the cluster (GSTe2, GSTe5 and GSTe7), are duplicated four times in AaegL5 relative to AaegL3 (Fig. 4a, b and Supplementary Data 23). Short Illumina read coverage and optical maps confirmed the copy number and arrangement of these duplications in AaegL5 (Fig. 4c, d), and analysis of whole-genome sequencing data for four additional laboratory colonies showed variable copy numbers across this gene cluster (Fig. 4d). GSTe2 is a highly efficient metaboliser of DDT30, and it is noteworthy that the cDNA from three GST genes in the quadruplication was detected at higher levels in DDT-resistant Ae. aegypti mosquitoes from southeast Asia31.

## Genome-wide genetic variation

Measurement of genetic variation within and between populations is important for inferring ongoing and historic evolution in a species32. To understand genomic diversity in Ae. aegypti, which spread in the last century from Africa to tropical and subtropical regions around the world, we performed whole-genome resequencing on four laboratory colonies. Chromosomal patterns of nucleotide diversity should correlate with regional differences in meiotic recombination rates33. We observed pronounced declines in genetic diversity near the centre of each chromosome (Extended Data Fig. 9a, b), providing independent corroboration of the estimated position of each centromere by Hi-C (Extended Data Fig. 2a).

To investigate linkage disequilibrium in geographically diverse populations of Ae. aegypti, we first mapped Affymetrix SNP-Chip markers that were designed using AaegL334 to positions on AaegL5. We genotyped 28 individuals from two populations from Amacuzac, Mexico and Lopé National Park, Gabon and calculated the pairwise linkage disequilibrium of single-nucleotide polymorphisms (SNPs) from 1-kb bins both genome-wide and within each chromosome (Extended Data Fig. 9c, d). The maximum linkage disequilibrium in the Mexican population is approximately twice that of the population from Gabon, which probably reflects a recent bottleneck associated with the spread of this species out of Africa.

## Dengue competence and pyrethroid resistance

To illustrate the value of AaegL5 for mapping quantitative trait loci (QTLs), we used restriction site-associated DNA (RAD) markers to locate QTLs underlying dengue virus (DENV) vector competence. We identified and genotyped RAD markers in the F2 progeny of a laboratory cross between wild Ae. aegypti founders from Thailand35 (Extended Data Fig. 10a). For this population, 197 F2 females had previously been scored for DENV vector competence against four different DENV isolates (two isolates from serotype 1 and two from serotype 3)35. The newly developed linkage map included a total of 255 RAD markers (Fig. 5a) with perfect concordance between genetic distances in centiMorgans (cM) and AaegL5 physical coordinates in Mb (Fig. 5a, c, d). We detected two significant QTLs on chromosome 2 that underlie the likelihood of DENV dissemination from the midgut (that is, systemic infection), an important component of DENV vector competence36. One QTL was associated with a generalist effect across DENV serotypes and isolates, whereas the other was associated with an isolate-specific effect (Fig. 5b, c). QTL mapping powered by AaegL5 will make it possible to understand the genetic basis of Ae. aegypti vector competence for arboviruses.

Pyrethroid insecticides are used to combat mosquitoes, including Ae. aegypti, and emerging resistance to these compounds is a global problem37. Understanding the mechanisms that underlie insecticide targets and resistance in different mosquito populations is critical to combating arboviral pathogens. Many insecticides act on ion channels, and we curated members of the Cys-loop ligand-gated ion channel (Cys-loop LGIC) superfamily in AaegL5. We found 22 subunit-encoding Cys-loop LGICs (Extended Data Fig. 10d and Supplementary Data 24), of which 14 encode nicotinic acetylcholine receptor (nAChR) subunits. nAChRs consist of a core group of subunit-encoding genes (α1–α8 and β1) that are highly conserved between insect species, and at least one divergent subunit38. Whereas D. melanogaster possesses only one divergent nAChR subunit, Ae. aegypti has five. We found that agricultural and veterinary insecticides impaired the motility of Ae. aegypti larvae (Extended Data Fig. 10c), suggesting that these Cys-loop LGIC-targeting compounds have potential as mosquito larvicides. The improved annotation presented here provides a valuable resource for investigating insecticide efficacy.

To demonstrate how a chromosome-scale genome assembly informs genetic mechanisms of insecticide resistance, we performed a genome-wide population genetic screen for SNPs correlating with resistance to deltamethrin in Ae. aegypti collected in Yucatán, Mexico, where pyrethroid-resistant and -susceptible populations co-exist (Fig. 5e). We uncovered an association with non-synonymous changes to three amino acid residues of the voltage-gated sodium channel VGSC, a known target of pyrethroids (Fig. 5f). The gene model for VGSC, a complex locus spanning nearly 500 kb in AaegL5, was incomplete and highly fragmented in AaegL3. SNPs in this region have a lower expected heterozygosity (Hexp) in the resistant compared to the susceptible population, suggesting that they are part of a selective sweep for the resistance phenotype surrounding VGSC (Fig. 5g). Accurately associating SNPs with phenotypes requires a fully assembled genome, and we expect that AaegL5 will be critical to understanding the evolution of insecticide resistance and other important traits.

## Summary

The high-quality genome assembly and annotation described here will enable major advances in mosquito biology, and has already allowed us to carry out a number of experiments that were previously impossible. The highly contiguous AaegL5 genome permitted high-resolution genome-wide analysis of genetic variation and the mapping of loci for DENV vector competence and insecticide resistance. A new appreciation of copy number variation in insecticide-detoxifying GSTe genes and a more complete accounting of Cys-loop LGICs will catalyse the search for new resistance-breaking insecticides. A doubling in the known number of chemosensory ionotropic receptors provides opportunities to link odorants and tastants on human skin to mosquito attraction, a key first step in the development of novel mosquito repellents. ‘Sterile Insect Technique’ and ‘Incompatible Insect Technique’ show great promise to suppress mosquito populations39, but these population suppression methods require that only males are released. A strategy that connects a gene for male determination to a gene drive construct has been proposed to effectively bias the population towards males over multiple generations40, and improved understanding of M locus evolution and the function of its genetic content should facilitate genetic control of mosquitoes that infect many hundreds of millions of people with arboviruses every year1.

## Methods

### Data reporting

No statistical methods were used to predetermine sample size. The experiments were not randomized and the investigators were not blinded to allocation during experiments and outcome assessment.

### Ethics information

The participation of one human subject in blood-feeding mosquitoes was approved and monitored by The Rockefeller University Institutional Review Board (IRB protocol LVO-0652). This subject gave their written and informed consent to participate.

### Mosquito rearing and DNA preparation

Ae. aegypti eggs from a strain labelled ‘LVP_ib12’ were supplied by M.V.S. from a colony maintained at Virginia Tech. We performed a single pair cross between a male and female individual to generate material for Hi-C, Bionano optical mapping, flow cytometry, SNP-Chip analysis of strain variance, paired-end Illumina sequencing and 10X Genomics linked reads (Extended Data Fig. 1a). The same single male was crossed to a single female in two additional generations to generate high-molecular weight (HMW) genomic DNA for Pacific Biosciences long-read sequencing and to establish a colony (LVP_AGWG). Rearing was performed as previously described13 and all animals were offered a human arm as a blood source.

### SNP analysis of mosquito strains

Data were generated as described34, and PCA was performed using LEA 2.0 available for R v.3.4.041,42. The following strains were used: Ae. aegypti LVP_AGWG (samples from the laboratory strain used for the AaegL5 genome assembly, reared as described in Extended Data Fig. 1a by a single pair mating in 2016 from a strain labelled LVP_ib12 maintained at Virginia Tech), Ae. aegypti LVP_ib12 (laboratory strain, LVP_ib12, provided in 2013 by D. Severson, University of Notre Dame), Ae. aegypti LVP_MR4 (laboratory strain labelled LVP_ib12 obtained in 2016 from MR4 at the Centers for Disease Control via BEI Resources catalogue MRA-735), Ae. aegypti Yaounde, Cameroon (field specimens collected in 2014 and provided by B. Kamgang), Ae. aegypti Rockefeller (laboratory strain provided in 2016 by G. Dimopoulos, Johns Hopkins Bloomberg School of Public Health), Ae. aegypti Key West, Florida (field specimens collected in 2016 and provided by W. Tabachnick). Strains used for the linkage disequilibrium data presented in Extended Data Fig. 9c, d were: Ae. aegypti from Amacuzac, Morelos, Mexico (field specimens collected in 2016 and provided by C. Gonzalez Acosta) and Ae. aegypti from La Lope National park forest, Gabon (field specimens collected and provided by S. Xia).

### Flow cytometry

Genome size was estimated by flow cytometry as described43, except that the propidium iodide was added at a concentration of 25 μl mg−1, not 50 μl mg−1, and samples were stained in the cold and dark for 24 h to allow the stain to fully saturate the sample. In brief, nuclei were isolated by placing a single frozen head of an adult sample along with a single frozen head of an adult Drosophila virilis female standard from a strain with 1C = 328 Mb into 1 ml of Galbraith buffer (4.26 g MgCl2, 8.84 g sodium citrate, 4.2 g 3-[N-morpholino] propane sulfonic acid (MOPS), 1 ml Triton X-100 and 1 mg boiled RNase A in 1 l of ddH2O, adjusted to pH 7.2 with HCl and filtered through a 0.22-μm filter)44 and grinding with 15 strokes of the A pestle at a rate of 3 strokes per 2 s. The resultant ground mixture was filtered through a 60-μm nylon filter (Spectrum Labs). Samples were stained with 25 μg of propidium iodide and held in the cold (4 °C) and dark for 24 h at which time the relative red fluorescence of the 2C nuclei of the standard and sample were determined using a Beckman Coulter CytoFlex flow cytometer with excitation at 488 nm. At least 2,000 nuclei were scored under each 2C peak and all scored peaks had a coefficient of variation of 2.5 or less43,44. Average channel numbers for sample and standard 2C peaks were scored using CytExpert software version 1.2.8.0 supplied with the CytoFlex flow cytometer. Significant differences among strains were determined using Proc GLM in SAS with both a Tukey and a Sheffé option. Significance levels were the same with either option. Genome size was determined as the ratio of the mean channel number of the 2C sample peak divided by the mean channel number of the 2C D. virilis standard peak times 328 Mb, where 328 Mb is the amount of DNA in a gamete of the standard. The following species/strains were used: Ae. mascarensis (collected by A. Bheecarry on Mauritius in December 2014. Colonized and maintained by J.R.P.), Ae. aegypti Ho Chi Minh City F13 (provided by D. J. Gubler, Duke-National University of Singapore as F1 eggs from females collected in Ho Chi Minh City in Vietnam, between August and September 2013. Colonized and maintained for 13 generations by A.G.-S.), Ae. aegypti Rockefeller (laboratory strain provided by D. Severson, Notre Dame), Ae. aegypti LVP_AGWG (reared as described in Extended Data Fig. 1a from a strain labelled LVP_ib12 maintained by M.V.S. at Virginia Tech), Ae. aegypti New Orleans F8 (collected by D. Wesson in New Orleans 2014, colonized and maintained by J.R.P. through 8 generations of single pair mating), Ae. aegypti Uganda 49-ib-G5 (derived by C.S.M. through 5 generations of full-sibling mating of the U49 colony established from eggs collected by J.-P. Mutebi in Entebbe, Uganda in March 2015).

### Pacific Biosciences library construction, sequencing and assembly

#### HMW DNA extraction for Pacific Biosciences sequencing

HMW DNA extraction for Pacific Biosciences sequencing was performed using the Qiagen MagAttract Kit (67563) following the manufacturer’s protocol with approximately 80 male sibling pupae in batches of 25 mg.

#### SMRTbell library construction and sequencing

Three libraries were constructed using the SMRTbell Template Prep Kit 1.0 (Pacific Biosciences). In brief, genomic DNA (gDNA) was mechanically sheared to 60 kb using the Megaruptor system (Diagenode) followed by DNA damage repair and DNA end repair. Universal blunt hairpin adapters were then ligated onto the gDNA molecules after which non-SMRTbell molecules were removed with exonuclease. Pulse-field gels were run to assess the quality of the SMRTbell libraries. Two libraries were size-selected using SageELF (Sage Science) at 30 kb and 20 kb, the third library was size-selected at 20 kb using BluePippin (Sage Science). Prior to sequencing, another DNA-damage repair step was performed and quality was assessed with pulse-field gel electrophoresis. A total of 177 SMRT cells were run on the RS II using P6-C4 chemistry and 6 h videos.

#### Contig assembly and polishing

A diploid contig assembly was carried out using FALCON v.0.4.0 followed by the FALCON-Unzip module (revision 74eefabdcc4849a8cef24d1a1bbb27d953247bd7)5. The resulting assembly contains primary contigs, a partially phased haploid representation of the genome and haplotigs, which represent phased alternative alleles for a subset of the genome. Two rounds of contig polishing were performed. For the first round, as part of the FALCON-Unzip pipeline, primary contigs and secondary haplotigs were polished using haplotype-phased reads and the Quiver consensus caller45. For the second round of polishing we used the ‘resequencing’ pipeline in SMRT Link v.3.1, with primary contigs and haplotigs concatenated into a single reference. Resequencing maps all raw reads to the combined assembly reference with BLASR (v.3.1.0)46, followed by consensus calling with Arrow (https://github.com/PacificBiosciences/GenomicConsensus)46.

### Hi-C sample preparation and analysis

#### Library preparation

In brief, insect tissue was crosslinked and homogenized. The nuclei were then extracted and permeabilized, and libraries were prepared using a modified version of the in situ Hi-C protocol that we optimized for insect tissue47. Separate libraries were prepared for samples derived from three individual male pupae. The resulting libraries were sequenced to yield 118 million, 249 million and 114 million reads (coverage: 120×) and these were processed using Juicer48.

#### Hi-C approach

Using the results of FALCON-Unzip as input, we used Hi-C to correct misjoins, to order and orient contigs, and to merge overlaps (Extended Data Fig. 1c–e). The Hi-C based assembly procedure that we used is described in detail in the Supplementary Methods and Supplementary Discussion. Notably, both primary contigs and haplotigs were used as input. This was essential, because Hi-C data identified genomic loci in which the corresponding sequence was absent in the primary FALCON-Unzip contigs, and present only in the haplotigs; the loci would have led to gaps, instead of contiguous sequence, if the haplotigs were excluded from the Hi-C assembly process (Extended Data Fig. 1e).

#### Hi-C scaffolding

We set aside 359 FALCON-Unzip contigs shorter than 20 kb, because such contigs are more difficult to accurately assemble using Hi-C. To generate chromosome-length scaffolds, we used the Hi-C maps and the remaining contigs as inputs to the previously described algorithms4. Note that both primary contigs and haplotigs were used as input. We performed quality control, manual polishing and validation of the scaffolding results using Assembly Tools49. This produced three chromosome-length scaffolds. Notably, the contig N50 decreased slightly, to 929,392 bp, because of the splitting of misjoined contigs.

#### Hi-C alternative haplotype merging

Examination of the initial chromosome-length scaffolds using Assembly Tools49 revealed that extensive undercollapsed heterozygosity was present. In fact, most genomic intervals were repeated, with variations, on two or more unmerged contigs. This suggested that the levels of undercollapsed heterozygosity were unusually high, and that the true genome length was far shorter than either the total length of the Pacific Biosciences contigs (2,047 Mb), or the initial chromosome-length scaffolds (1,973 Mb). Possible factors that could have contributed to the unusually high rate of undercollapsed heterozygosity seen in the FALCON-Unzip Pacific Biosciences contigs relative to prior contig sets for Ae. aegypti generated using Sanger sequencing (AaegL3)2, include high heterozygosity levels in the species and incomplete inbreeding in the samples that we sequenced. The merge algorithm described previously4 detects and merges draft contigs that overlap one another owing to undercollapsed heterozygosity. Because undercollapsed heterozygosity does not affect most loci in a typical draft assembly, the default parameters are relatively stringent. We adopted more permissive parameters for AaegL5 to accommodate the exceptionally high levels of undercollapsed heterozygosity, but found that the results would occasionally merge contigs that did not overlap. To avoid these false positives, we developed a procedure to manually identify and ‘whitelist’ regions of the genome containing no overlap, based on both Hi-C maps and LASTZ alignments (Extended Data Fig. 1c, Supplementary Methods and Supplementary Discussion). We then reran the merge step, using the whitelist as an additional input. Finally, we performed quality control of the results using Assembly Tools49, which confirmed the absence of the undercollapsed heterozygosity that we had previously observed. The resulting assembly contained three chromosome-length scaffolds (310 Mb, 473 Mb and 409 Mb), which spanned 94% of the merged sequence length. The assembly also contained 2,364 small scaffolds, which spanned the remaining 6% (Table 1). These lengths were consistent with the results of flow cytometry and the lengths obtained in prior assemblies. Notably, the merging of overlapping contigs using the above procedure frequently eliminated gaps, and thus greatly increased the contig N50, from 929,392 to 4,997,917 bp.

### Final gap-filling and polishing

#### Scaffolded assembly polishing

Following scaffolding and de-duplication, we performed a final round of arrow polishing. PBJelly50 from PBSuite version 15.8.24 was used for gapfilling of the de-duplicated HiC assembly (see ‘Protocol.xml’ in Supplementary Methods and Supplementary Discussion). After PBJelly, the liftover file was used to translate the renamed scaffolds to their original identifiers. For this final polishing step (run with SMRT Link v3.1 resequencing), the reference sequence included the scaffolded, gap-filled reference, as well as all contigs and contig fragments not included in the final scaffolds (https://github.com/skingan/AaegL5_FinalPolish). This reduces the likelihood that reads map to the wrong haplotype, by providing both haplotypes as targets for read mapping. For submission to NCBI, two scaffolds identified as mitochondrial in origin were removed (see below), and all remaining gaps on scaffolds were standardized to a length of 100 Ns to indicate a gap of unknown size. The assembly quality value was estimated using independent Illumina sequencing data from a single individual male pupa (library H2NJHADXY_1/2). Reads were aligned with BWA-MEM v.0.7.12-r103951. FreeBayes v.1.1.0-50-g61527c5-dirty52 was used to call SNPs and short indels with the parameters -C 2 -0 -O -q 20 -z 0.10 -E 0 -X -u -p 2 -F 0.6. Any SNP and short indels showing heterozygosity (for example, 0/1 genotype) were excluded. The quality value was estimated at 34.75 using the PHRED formula with SNPs as the numerator (597,798) and number of bases with at least threefold coverage as the denominator, including alternate alleles (1,782,885,792).

#### Identification of mitochondrial contigs

During the submission process for this genome, two contigs were identified as mitochondrial in origin and were removed from the genomic assembly, manually circularized, and submitted separately. The mitochondrial genome is available as GenBank accession number MF194022.1, RefSeq accession number NC_035159.1.

### Bionano optical mapping

#### HMW DNA extraction

HMW DNA extraction was performed using the Bionano Animal Tissue DNA Isolation Kit (RE-013-10), with a few protocol modifications. A single-cell suspension was made as follows. First, 47 mg of frozen male pupae was fixed in 2% v/v formaldehyde in Homogenization Buffer from the kit (Bionano 20278), for 2 min on ice. Then, the pupae were roughly homogenized by blending for 2 s, using a rotor–stator tissue homogenizer (TissueRuptor, Qiagen 9001271). After another 2 min fixation, the tissue was finely homogenized by running the rotor–stator for 10 s. Subsequently, the homogenate was filtered with a 100-μm nylon filter, fixed with ethanol for 30 min on ice, spun down, and washed with more Homogenization Buffer (to remove residual formaldehyde). The final pellet was resuspended in Homogenization Buffer. A single agarose plug was made using the resuspended cells, using the CHEF Mammalian Genomic DNA Plug Kit (BioRad 170-3591), following the manufacturer’s instructions. The plug was incubated with Lysis Buffer (Bionano 20270) and Puregene Proteinase K (Qiagen 1588920) overnight at 50 °C, then again the following morning for 2 h (using new buffer and Proteinase K). The plug was washed, melted and solubilized with GELase (Epicentre G09200). The purified DNA was subjected to 4 h of drop dialysis (Millipore, VCWP04700) and quantified using the Quant-iT PicoGreen dsDNA Assay Kit (Invitrogen/Molecular Probes P11496).

#### DNA labelling

DNA was labelled according to commercial protocols using the DNA Labelling Kit NLRS (RE-012-10, Bionano Genomics). Specifically, 300 ng of purified genomic DNA was nicked with 7 U nicking endonuclease Nt.BspQI (New England BioLabs, NEB) at 37 °C for 2 h in NEBuffer3. The nicked DNA was labelled with a fluorescent-dUTP nucleotide analogue using Taq polymerase (NEB) for 1 h at 72 °C. After labelling, the nicks were ligated with Taq ligase (NEB) in the presence of dNTPs. The backbone of fluorescently labelled DNA was counterstained with YOYO-1 (Invitrogen).

#### Data collection

The DNA was loaded onto the nanochannel array of Bionano Genomics IrysChip by electrophoresis of DNA. Linearized DNA molecules were then imaged automatically followed by repeated cycles of DNA loading using the Bionano Genomics Irys system. The DNA-molecule backbones (YOYO-1 stained) and locations of fluorescent labels along each molecule were detected using the in-house-generated software package, IrysView. The set of label locations of each DNA molecule defines an individual single-molecule map. After filtering data using normal parameters (molecule reads with length greater than 150 kb, a minimum of 8 labels and standard filters for label and backbone signals), a total of 299 Gb and 259 Gb of data were collected from Nt.BspQI and Nb.BssSI samples, respectively.

#### De novo genome map assembly

De novo assembly was performed with non-haplotype aware settings (optArguments_nonhaplotype_noES_irys.xml) and pre-release version of Bionano Solve3.1 (Pipeline version 6703 and RefAligner version 6851). On the basis of the overlap–layout–Consensus paradigm, pairwise comparisons of all DNA molecules were performed to create an overlap graph, which was then used to create the initial consensus genome maps. By realigning molecules to the genome maps (RefineB P = 10 × 10−11) and by using only the best match molecules, a refinement step was performed to refine the label positions on the genome maps and to remove chimeric joins. Next, during an extension step, the software aligned molecules to genome maps (extension, P = 10 × 10−11), and extended the maps based on the molecules aligning past the map ends. Overlapping genome maps were then merged using a merge P-value cut-off of 10 P = 10 × 10−15. These extension and merge steps were repeated five times before a final refinement was applied to ‘finish’ all genome maps (refine final, P = 10 × 10−11). Two genome map de novo assemblies, one with nickase Nt.BspQI and the other with nickase Nb.BssSI, were constructed. Alignments between the constructed de novo genome assemblies and the L5 assembly were performed using a dynamic programming approach with a scoring function and a P-value cutoff of P = 10 × 10−12.

### Transposable element identification

#### Identification of known transposon elements

We first identified known transposable elements using RepeatMasker (version 3.2.6)53 against the mosquito TEfam (https://tefam.biochem.vt.edu/tefam/, data downloaded July 2017), a manually curated mosquito transposable-elements database. We then ran RepeatMasker using the TEfam database and Repbase transposable-elements library (version 10.05). RepeatMasker was set to default parameters with the -s (slow search) flag and NCBI/RMblast program (v.2.2.28).

#### De novo repeat family identification

We searched for repeat families and consensus sequences using the de novo repeat prediction tool RepeatModeler (version 1.0.8)54 using default parameters with RECON (version 1.07) and RepeatScout (1.0.5) as core programs. Consensus sequences were generated and classified for each repeat family. Then RepeatMasker was run on the genome sequences, using the RepeatModeler consensus sequence as the library.

#### Tandem repeats

We also predicted tandem repeats in the whole genome and in the repeatmasked genome using Tandem Repeat Finder55. Long tandem copies were identified using the ‘Match=2, Mismatch=7, Delta=7, PM=80, PI=10, Minscore=50 MaxPeriod=500’ parameters. Simple repeats, satellites and low complexity repeats were found using ‘Match=2, Mismatch=7, Delta=7, PM=80, PI=10, Minscore=50, and MaxPeriod=12’ parameters.

A file representing the coordinates of all identified repeat and transposable-element structures in AaegL5 can be found at https://github.com/VosshallLab/AGWG-AaegL5.

### Generation of RefSeq gene set annotation

The AaegL5 assembly was deposited at NCBI in June 2017 and annotated using the NCBI RefSeq Eukaryotic gene annotation pipeline56. Evidence to support the gene predictions came from over 9 billion Illumina RNA-seq reads, 67,000 Pacific Biosciences IsoSeq transcripts, 300,000 expressed sequence tags and well-supported proteins from D. melanogaster and other insects. Annotation Release 101 was made public in July 2017, and specific gene families were subjected to manual annotation and curation. Detailed descriptions of the manual annotation and curation of multigene families (Hox genes, proteases, opsins and biogenic amine receptors, chemosensory receptors and LGICs) can be found in the Supplementary Methods and Supplementary Discussion. See also https://www.ncbi.nlm.nih.gov/genome/annotation_euk/Aedes_aegypti/101/.

### Alignment of RNA-seq data to AaegL5 and quantification of gene expression

Published RNA-seq reads13,57 and unpublished RNA-seq reads from tissue-specific libraries produced by Verily Life Sciences were mapped to the RefSeq assembly GCF_002204515.2_AaegL5.0 with STAR aligner (v.2.5.3a)58 using the two-pass approach. Reads were first aligned in the absence of gene annotations using the following parameters: --outFilterType BySJout; --alignIntronMax 1000000; --alignMatesGapMax 1000000; --outFilterMismatchNmax 999; --outFilterMismatchNoverReadLmax 0.04; --clip3pNbases 1; --outSAMstrandField intronMotif; --outSAMattrIHstart 0; --outFilterMultimapNmax 20; --outSAMattributes NH HI AS NM MD; --outSAMattrRGline; --outSAMtype BAM SortedByCoordinate. Splice junctions identified during the first pass mapping of individual libraries were combined and supplied to STAR using the --sjdbFileChrStartEnd option for the second pass. Reads mapping to gene models defined by the NCBI annotation pipeline (GCF_002204515.2_AaegL5.0_genomic.gff) were quantified using featureCounts59 with default parameters. Count data were transformed to transcripts per million values using a custom Perl script. Details on libraries, alignment statistics and gene expression estimates (expressed in transcripts per million) are provided as Supplementary Data 48.

### Identification of ‘collapsed’ and ‘merged’ gene models from AaegL3.5 to AaegL5.0

VectorBase annotation AaegL3.5 was compared to NCBI Ae. aegypti annotation release 101 on AaegL5.0 using custom code developed at NCBI as part of NCBI’s eukaryotic genome annotation pipeline. First, assembly–assembly alignments were generated for AaegL3 (GCA_000004015.3) × AaegL5.0 (GCF_002204515.2) as part of NCBI’s Remap coordinate remapping service, as described at https://www.ncbi.nlm.nih.gov/genome/tools/remap/docs/alignments. The alignments are publicly available in NCBI’s Genome Data Viewer (https://www.ncbi.nlm.nih.gov/genome/gdv/), the Remap interface, and by FTP in either ASN.1 or GFF3 format (ftp://ftp.ncbi.nlm.nih.gov/pub/remap/Aedes_aegypti/2.1/). Alignments are categorized as either ‘first pass’ (reciprocity = 3) or ‘second pass’ (reciprocity = 1 or 2). First pass alignments are reciprocal best alignments, and are used to identify regions on the two assemblies that can be considered equivalent. Second pass alignments are cases in which two regions of one assembly have their best alignment to the same region on the other assembly. These are interpreted to represent regions in which two paralogous regions in AaegL3 have been collapsed into a single region in AaegL5, or vice versa.

For comparing the two annotations, both annotations were converted to ASN.1 format and compared using an internal NCBI program that identifies regions of overlap between gene, mRNA and coding sequence (CDS) features projected through the assembly–assembly alignments. The comparison was performed twice, first using only the first pass alignments, and again using only the second pass alignments corresponding to regions in which duplication in the AaegL3 assembly had been collapsed. Gene features were compared, requiring at least some overlapping CDS in both the old and new annotation to avoid noise from overlapping genes and comparisons between coding and non-coding genes. AaegL5.0 genes that matched to two or more VectorBase AaegL3.5 genes were identified. Matches were further classified as collapsed paralogues if one or more of the matches was through the second pass alignments, or as improvements due to increased contiguity or annotation refinement if the matches were through first pass alignments (for example, two AaegL3.5 genes represent the 5′ and 3′ ends of a single gene on AaegL5.0, such as sex peptide receptor. Detailed lists of merged genes are in Supplementary Data 10, 11.

### Comparison of alignment to AaegL3.4 and AaegL5.0

The sequences comprising transcripts from the AaegL5.0 gene set annotation were extracted from coordinates provided in GCF_002204515.2_AaegL5.0_genomic.gtf. Sequences corresponding to AaegL3.4 gene set annotations were downloaded from Vectorbase (https://www.vectorbase.org/download/aedes-aegypti-liverpooltranscriptsaaegl34fagz). Salmon (v.0.8.2)60 indices were generated with default parameters, and all libraries described in Supplementary Data 4 were mapped to both AaegL3.4 and AaegL5 sequences using ‘quant’ mode with default parameters. Mapping results are presented as Supplementary Data 9 and Fig. 1h.

### M locus analysis

#### Aligning chromosome assemblies and Bionano scaffolds

The boundaries of the M locus were identified by comparing the current AaegL5 assembly and the AaegL4 assembly4 using a program called LAST68 (data not shown). To overcome the challenges of repetitive hits, both AaegL5 and AaegL4 assemblies were twice repeat-masked53 against a combined repeat library of TEfam-annotated transposable elements (https://tefam.biochem.vt.edu/tefam/)2 and a RepeatModeler output54 from the Anopheles 16 Genomes project69. The masked sequences were then compared using BLASTn64 and we then set a filter for downstream analysis to include only alignment with ≥98% identity over 1,000 bp. After the identification of the approximate boundaries of the M locus (and m locus), which contains two male-specific genes, myo-sex18 and Nix17, we zoomed in by performing the same analysis on regions of the M locus and m locus plus 2 Mb flanking regions without repeatmasking. In this and subsequent analyses, only alignment with ≥98% identity over 500 bp were included. Consequently, approximate coordinates of the M locus and m locus were obtained on chromosome 1 of the AaegL5 and AaegL4 assemblies, respectively. Super-scaffold_63 in the Bionano optical map assembly was identified by BLASTN64 that spans the entire M locus and extends beyond its two borders.

#### Chromosome quotient analysis

The chromosome quotient (CQ)20 was calculated for each 1,000-bp window across all AaegL5 chromosomes. To calculate the CQ, Illumina reads were generated from two paired sibling female and male sequencing libraries. To generate libraries for CQ analysis, we performed two separate crosses of a single LVP_AGWG male to 10 virgin females. Eggs from this cross were hatched, and virgin male and female adults collected within 12 h of eclosion to verify their non-mated status. We generated genomic DNA from five males and five females from each of these crosses. Sheared genomic DNA was used to generate libraries for Illumina sequencing with the Illumina TruSeq Nano kit and sequencing performed on one lane of 150-bp paired-end sequencing on an Illumina NextSeq 500 in high-output mode.

For a given sequence Si of a 1,000-bp window, $${{\rm{CQ}}}_{{S}_{i}}={F}_{{S}_{i}}/{M}_{{S}_{i}}$$, where $${F}_{{S}_{i}}$$ is the number of female Illumina reads aligned to Si, and $${M}_{{S}_{i}}$$ is the number of male Illumina reads aligned to Si. Normalization was not necessary for these datasets because the mean and median CQs of the autosomes (chromosomes 2 and 3) are all near 1. A CQ value lower than the 0.05 indicates that the sequences within the corresponding 1,000-bp window had at least 20-fold more hits to the male Illumina data than to the female Illumina data. Not every 1,000-bp window produces a CQ value because many were completely masked by RepeatMasker53. To ensure that each CQ value represents a meaningful data point obtained with sufficient alignments, only sequences with more than 20 male hits were included in the calculation. The CQ values were then plotted against the chromosome location of the 1,000-bp window (Fig. 3d). Under these conditions, there is not a single 1,000-bp fragment on chromosomes 2 and 3 that showed CQ = 0.05 or lower.

#### Chromosome FISH

Slides of mitotic chromosomes were prepared from imaginal discs of fourth instar larvae following published protocols3,70,71. BAC clones were obtained from the University of Liverpool19 or from a previously described BAC library72. BACs were plated on agar plates (Thermo Fisher) and a single bacterial colony was used to grow an overnight bacterial culture in LB broth plates (Thermo Fisher) at 37 °C. DNA from the BACs was extracted using Sigma PhasePrep TM BAC DNA Kit (Sigma-Aldrich, NA-0100). BAC DNA for hybridization was labelled by nick translation with Cy3-, Cy5-dUTP (Enzo Life Sciences) or Fluorescein 12-dUTP (Thermo Fisher). Chromosomes were counterstained with DAPI in Prolong Gold Antifade (Thermo Fisher). Slides were analysed using a Zeiss LSM 880 Laser Scanning Microscope at 1,000× magnification. We note that localization of the M-locus to 1p11 is supported by both FISH and genomic analyses, but is contrary to a previously published placement at 1q2117.

### Identification and analysis of Ae. aegypti GST and P450 genes and validation of the repeat structure of the GSTe cluster

Genes were initially extracted from the AaegL5.0 genome annotation (NCBI release 101) by text search and filtered to remove ‘off target’ matches (for example, ‘cytochrome P450 reductase’), then predicted protein sequences of a small number of representative transcripts were used to search the protein set using BLASTp, to identify by sequence similarity sequences not captured by the text search (resulting in two additional P450s, no GSTs). For each gene family, predicted protein sequences were used to search the proteins of the AaegL3.4 gene set using BLASTp. All best matches, and additional matches with amino acid identity >90% were tabulated for each gene family (Supplementary Data 23) to identify both closely related paralogues and alleles annotated as paralogues in AaegL3.4. On the basis of a BLASTp search against the AaegL3.4 protein set, the two putative P450 genes not annotated as such in AaegL5.0 (encoding proteins XP_001649103.2 and XP_021694388.1) appear to be incorrect gene models in the AaegL5.0 annotation, which should in fact be two adjacent genes (CYP9J20 and CYP9J21 for XP_001649103.2; CYP6P12 and CYP6BZ1 for XP_021694388.1). Compared to AaegL3.4, which predicts a single copy each of GSTe2, GSTe5 and GSTe7, the NCBI annotation of AaegL5.0 predicts three copies each of GSTe2 and GSTe5, and four copies of GSTe7, arranged in a repeat structure. BLASTn searches revealed one additional copy each of GSTe2 and GSTe5 in the third duplicated unit. Both contain premature termination codons owing to frameshifts, but these could be owing to uncorrected errors in the assembly. Error correction of all duplicated units was not possible owing to the inability to unequivocally align reads to units not ‘anchored’ to adjacent single-copy sequence.

To validate these tandem duplications, two lanes of Illumina whole-genome sequence data from a single pupa of the LVP_AGWG strain (H2NJHADXY) were aligned to a hard-masked version of the AaegL3 reference genome using Bowtie2 v.2.2.473, with ‘--very-fast-local’ alignment parameters, an expected fragment size between 0 and 1,500 bp and relative orientation ‘forward–reverse’ (‘-I 0 -X 1500 –fr’). Aligned reads with a mapping quality less than 10 were removed using Samtools63. ‘featureCounts’, part of the ‘Subread’ v.1.5.0-p2 package74, was used to assign read pairs or reads (‘tags’) aligned to either DNA strand (‘-s 0’) and overlapping the coding regions of a gene by at least 100 bp (‘-t CDS–minOverlap 100’) to genes as an estimate of representation in the genome. Gene-wise tag counts were normalized by calculating the fragments per kilobase of gene length per million mapped reads (FPKM), using the following equation: (tag count/gene length in kb)/(sum of tag counts for all genes in genome/1,000,000).

Median FPKM for all genes in the genome was calculated (48.22), allowing FPKM of GSTe genes to be expressed relative to this. To examine strain differences in coverage at this cluster, we repeated this analysis for the four laboratory colonies analysed in Extended Data Fig. 9a, b. Median FPKM values across all genes ranged from 47.68 to 48.46 and gene-wise FPKM values normalized relative to these medians are plotted in Fig. 4d.

To visualize the sequence identity of the repeat structure in the GSTe cluster (Fig. 4b), we extracted the region spanning the cluster from AaegL5 chromosome 2 (351,597,324–351,719,186 bp) and performed alignment of Pacific Biosciences reads using minimap2 v2.1.1 (minimap2 -DP -k7 -w1 -B2 -r200 -g100 -m1 L5_gst.fa L5_gst.fa)75 and visualized these alignments using D-GENIES v1.2.092 with minimum identity set to 0.15 and ‘Strong Precision’ enabled. To validate this repeat structure, we aligned two de novo optical maps created by Bionano using linearized DNA labelled with Nt.BspQI or Nb.BssSI. Single molecules from both maps span the entire region and the predicted restriction pattern provides support for the repeat structure as presented in AaegL5 (Fig. 4c).

### QTL mapping of DENV vector competence

In theory, a good-quality genome assembly is not necessary for QTL mapping procedures, because it relies on a linkage map that can be generated de novo from empirical recombination fractions. This typically involves three steps: (i) marker selection based on the Mendelian segregation ratios, (ii) marker assignment to linkage groups and (iii) marker ordering within each linkage group. However, if a high-quality reference genome assembly is available, the physical position of each marker can be determined and this prior information greatly facilitates steps (ii) and (iii), as shown below.

To demonstrate the improvement enabled by our new genome, we generated two linkage maps using the same Illumina sequence data that were aligned either to AaegL3 or AaegL5 genome assemblies. Although the initial number of markers was 616 in both cases, the final linkage map was 3.3-fold denser with AaegL5 than with AaegL3, as shown in Extended Data Fig. 10b. The difference in marker density between the two linkage maps is because many markers were filtered out from the AaegL3 data. Because the AaegL3 assembly is highly fragmented (>4,700 scaffolds), the position of each marker within the linkage groups is primarily determined from the recombination fractions. This ordering step is performed by creating a backbone with a subset of informative markers using a two-point algorithm, followed by the positioning of the remaining markers one at a time using a multi-point method. Only markers that are unambiguously positioned are kept in the final linkage map for QTL mapping. We note that AaegL4, which de-duplicated and scaffolded AaegL3 onto chromosomes4, would probably yield a similar improvement in mapping resolution.

Ultimately, AaegL5 has a markedly improved QTL mapping resolution over AaegL3. For instance, we mapped the same QTL underlying systemic DENV dissemination at the extremity of chromosome 2 with both AaegL3 and AaegL5. The 1.5 LOD support interval was much larger for the AaegL3-guided linkage map (0–50 cM, 74% of the linkage group) than for the AaegL5-guided linkage map (0–17 cM, 9% of the linkage group). We present this analysis in Extended Data Fig. 10b.

#### Mosquito crosses

A large F2 intercross was created from a single mating pair of field-collected F0 founders. Wild mosquito eggs were collected in Kamphaeng Phet Province, Thailand in February 2011 as previously described35. In brief, F0 eggs were allowed to hatch in filtered tap water and the larvae were reared until the pupae emerged in individual vials. Ae. aegypti adults were identified by visual inspection and maintained in an insectary under controlled conditions (28 ± 1 °C, 75 ± 5% relative humidity and 12:12-h light:dark cycle) with access to 10% sucrose. The F0 male and female initiating the cross were chosen from different collection sites to avoid creating a parental pair with siblings from the same wild mother76,77. Their F1 offspring were allowed to mass-mate and collectively oviposit to produce the F2 progeny (Extended Data Fig. 10a). A total of 197 females of the F2 progeny were used as a mapping population to generate a linkage map and detect QTLs underlying vector competence for DENV.

#### Vector competence

Four low-passage DENV isolates were used to orally challenge the F2 females as previously described35. In brief, four random groups of females from the F2 progeny were experimentally exposed to two virus isolates belonging to DENV serotype 1 (KDH0026A and KDH0030A) and two virus isolates belonging to DENV serotype 3 (KDH0010A and KDH0014A), respectively. All four virus isolates were derived from human serum specimens collected in 2010 from clinically ill patients who were infected with DENV at the Kamphaeng Phet Provincial Hospital35. Because the viruses were isolated in the laboratory cell culture, informed consent of the patients was not necessary for the present study. Complete viral genome sequences were previously deposited into GenBank (accession numbers HG316481, HG316582, HG316483, and HG316484)35. Phylogenetic analysis assigned the viruses to known viral lineages that were circulating in southeast Asia in the previous years35. Each isolate was amplified twice in C6/36 (Aedes albopictus) cell lines (maintained at AFRIMS in Bangkok, Thailand; used only to grow virus, not explicitly authenticated or checked for mycoplasma contamination) before vector competence assays. Four- to seven-day-old F2 females were starved for 24 h and offered an infectious blood meal for 30 min. Viral titres in the blood meals ranged from 2.0 × 104 to 2.5 × 105 plaque-forming units per ml across all isolates. Fully engorged females were incubated under the conditions described above. Vector competence was scored 14 days after the infectious blood meal according to two conventional phenotypes: (i) midgut infection and (ii) viral dissemination from the midgut. These binary phenotypes were scored based on the presence or absence of infectious particles in body and head homogenates, respectively. Infectious viruses were detected by plaque assay performed in LLC-MK2 (rhesus monkey kidney epithelial) cells as previously described35,78.

#### Genotyping

Mosquito gDNA was extracted using the NucleoSpin 96 Tissue Core Kit (Macherey-Nagel). For the F0 male, it was necessary to perform whole-genome amplification using the Repli-g Mini kit (Qiagen) to obtain a sufficient amount of DNA. F0 parents and females of the F2 progeny were genotyped using a modified version of the original double-digest restriction site-associated DNA (RAD) sequencing protocol79, as previously described80. The final libraries were spiked with 15% PhiX and sequenced on an Illumina NextSeq 500 platform using a 150-cycle paired-end chemistry (Illumina). A previously developed bash script pipeline80 was used to process the raw sequence reads. High-quality reads (Phred scores >25) trimmed to the 140-bp length were aligned to the AaegL5 reference genome (July 2017) using Bowtie v.0.12.762. Parameters for the ungapped alignment included ≤3 mismatches in the seed, suppression of alignments with >1 best reported alignment under a ‘try-hard’ option. Variant and genotype calling was performed from a catalogue of RAD loci created with the ref_map.pl pipeline in Stacks v.1.1981,82. Downstream analyses only used high-quality genotypes at informative markers that were homozygous for alternative alleles in the F0 parents (for example, AA in the F0 male and BB in the F0 female), had a sequencing depth ≥10× and were present in ≥60% of the mapping population.

#### QTL mapping

The newly developed linkage map was used to detect and locate QTLs that underlie the DENV vector competence indices described above. Midgut infection was analysed in all F2 females whereas viral dissemination was analysed only in midgut-infected females. The four different DENV isolates were included as a covariate to detect QTL–isolate interactions. Single QTL detection was performed in the R/qtl environment85 using the expectation–maximization algorithm of the scanone function using a binary trait model. Genome-wide statistical significance was determined by empirical permutation tests, with 1,000 genotype–phenotype permutations of the entire dataset.

### Mapping insecticide resistance and VGSC

The mosquito population Viva Caucel from Yucatán State in Southern Mexico (longitude −89.71827, latitude 20.99827), was collected in 2011 through the Universidad Autónoma de Yucatán. We identified up to 25 larval breeding sites from 3–4 city blocks and collected around 1,000 larvae. Larvae were allowed to eclose, and twice a day we aspirated the adults from the cartons, discarding anything other than Ae. aegypti. Then, 300–400 Ae. aegypti were released into a 2-foot (61-cm) cubic cage where they were allowed to mate for up to five days with ad libitum access to sucrose, after which they were blood fed to collect eggs for the next generation. Then, 390 adult mosquitoes were phenotyped for deltamethrin resistance. We exposed groups of 50 mosquitoes (3–4 days old) to 3 μg of deltamethrin-coated bottles for 1 h. After this time, active mosquitoes were transferred to cardboard cups and placed into an incubator (28 °C and 70% humidity) for 4 h; these mosquitoes were referred to as the resistant group. Immobile mosquitoes were transferred to a second cardboard cup. After 4 h, newly recovered mosquitoes were aspirated, frozen and labelled as recovered; these were excluded from the current study. The mosquitoes that were immobilized and remained inactive at 4 h post-treatment were scored as susceptible. DNA was isolated from individual mosquitoes by the salt extraction method86 and resuspended in 150 μl of TE buffer (10 mM Tris-HCl, 1 mM EDTA pH 8.0). We constructed a total of four gDNA libraries. Two groups were pooled from DNA of 25 individual females that survived 1 h of deltamethrin exposure (resistant replicates 1 and 2). The second set of two libraries was obtained by pooling DNA from 25 females that were immobilized and inactive at 4 h post-treatment (susceptible replicates 1 and 2). Before pooling, DNA from each individual mosquito was quantified using the Quant-IT Pico Green kit (Life Technologies, Thermo Fisher Scientific) and around 40 ng from each individual DNA sample (25 individuals per library) was used for a final DNA pool of 1 μg. Pooled DNA was sheared and fragmented by sonication to obtain fragments between 300 and 500 bp (Covaris). We prepared one library for each of the four DNA pools following the Low Sample protocol from the Illumina TrueSeqDNA PCR-Free Sample preparation guide (Illumina). Because 65% of the Ae. aegypti genome consists of repetitive DNA, we performed an exome-capture hybridization to enrich for coding sequences using custom SeqCap EZ Developer probes (NimbleGen, Roche). Probes covered protein-coding sequences (not including untranslated regions) in the AaegL1.3 genebuild using previously specified exonic coordinates87. In total, 26.7 Mb of the genome (2%) was targeted for enrichment. TruSeq libraries were hybridized to the probes using the xGenLockDown recommendations (Integrated DNA Technologies). The targeted DNA was eluted and amplified (10–15 cycles) before being sequenced on one flow cell of a 100-bp HiSeq Rapid-duo paired-end sequencing run (Illumina) performed by the Centers for Disease Control (Atlanta, GA, USA).

The raw sequence files (FASTQ) for each pair-ended gDNA library were aligned to a custom reference physical map generated from the assembly AaegL5. Nucleotide counts were loaded into a contingency table with four rows corresponding to ‘Alive Rep1’, ‘Alive Rep2’, ‘Dead Rep1’ and ‘Dead Rep2’. The numbers of columns (c) corresponded to the number of alternative nucleotides at a SNP locus. The maximum value for c is 6, corresponding to A, C, G, T, insert or deletion. Three (2 × c) contingency tables were subjected to χ2 analyses (c − 1 degrees of freedom) to determine whether there are significant (P ≤ 0.05) differences between (1) Alive replicates, (2) Dead replicates and (3) Alive versus Dead. If analysis (1) or (2) was significant, then that SNP locus was discarded. Otherwise the third contingency table consisted of two rows corresponding to Alive (sum of replicates 1 and 2), Dead (replicates 1 and 2 summed), and c columns. The χ2 value from the (2 × c) contingency χ2 analysis with (c − 1) degrees of freedom was loaded into Excel to calculate the one-tailed probability of the χ2 distribution probability (P). This value was transformed with −log10(P). The experiment-wise error rate was then calculated following the method of Benjamini and Hochberg88 to lower the number of type I errors (false positives).

### Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.

### Code availability

The overview of the Hi-C workflow, as well as modifications to 3D-DNA associated with AaegL5, is shared on GitHub at https://github.com/theaidenlab/AGWG-merge. The source code and executable version of Juicebox Assembly Tools are available at http://aidenlab.org/assembly. Data files and scripts used for the final polishing of scaffolded, gap-filled assembly are available at https://github.com/skingan/AaegL5_FinalPolish.

### Data availability

All raw data have been deposited at NCBI under the following BioProject accession numbers: PRJNA318737 (primary Pacific Biosciences data, Hi-C sequencing primary data and processed contact maps, whole-genome sequencing data from a single male (Fig. 4d) and pools of male and females (Fig. 3d), Bionano optical mapping data (Figs. 3c, 4c) and 10X linked-read sequences (Extended Data Fig. 8a and Supplementary Data 21)); PRJNA236239 (RNA-seq reads and de novo transcriptome assembly13 (Extended Data Fig. 2c and Supplementary Data 4, 5, 7, 9)); PRJNA209388 (RNA-seq reads for developmental time points57 (Fig. 1h and Supplementary Data 46, 9)); PRJNA419241 (RNA-seq reads from adult reproductive tissues and developmental time points, Verily Life Sciences (Fig. 1h and Supplementary Data 4, 5, 8, 9)); PRJNA393466 (full-length Pacific Biosciences Iso-Seq transcript sequencing); PRJNA418406 (ATAC-seq data from adult female brains at three points in the gonotrophic cycle (Extended Data Fig. 2c, d and data not shown)); PRJNA419379 (whole-genome sequencing data from four colonies (Fig. 4d and Extended Data Fig. 9a, b)); PRJNA399617 (restriction-site-associated DNA-sequencing data (Fig. 5a–d)); PRJNA393171 (exome-sequencing data (Fig. 5e-g)). Intermediate results related to the AaegL5 assembly are also available via GitHub (http://github.com/theaidenlab/AGWG-merge) and have been uploaded to GEO (GSE113256). The Hi-C maps are available via http://aidenlab.org/juicebox. The complete mitochondrial genome is available as Genbank accession MF194022.1, RefSeq accession NC_035159.1. The final genome assembly and annotation are available from the NCBI Assembly Resource under accession GCF_002204515.2.