Giant transposons promote strain heterogeneity in a major fungal pathogen

Abstract

Fungal infections are difficult to prevent and treat in large part due to strain heterogeneity. However, the genetic mechanisms driving pathogen variation remain poorly understood. Here, we determined the extent to which Starships—giant transposons capable of mobilizing numerous fungal genes—generate genetic and phenotypic variability in the human pathogen Aspergillus fumigatus. We analyzed 519 diverse strains, including 12 newly sequenced with long-read technology, to reveal 20 distinct Starships that are generating genomic heterogeneity over timescales potentially relevant for experimental reproducibility. Starship-mobilized genes encode diverse functions, including biofilm-related virulence factors and biosynthetic gene clusters, and many are differentially expressed during infection and antifungal exposure in a strain-specific manner. These findings support a new model of fungal evolution wherein Starships help generate variation in gene content and expression among fungal strains. Together, our results demonstrate that Starships are a previously hidden mechanism generating genotypic and, in turn, phenotypic heterogeneity in a major human fungal pathogen.

Introduction

Infectious diseases caused by fungi pose a grave threat to human health. The World Health Organization recently coordinated a global effort to prioritize research among fungal pathogens based on unmet research needs and public health importance (1). Among the pathogens deemed most important for research include Aspergillus fumigatus, a globally distributed human pathogen causing disease in an estimated 3 million people yearly (2, 3). Several infectious diseases are caused by A. fumigatus, including invasive pulmonary aspergillosis, which manifests primarily in immunocompromised individuals with mortality rates of up to 88% (4). The treatment of A. fumigatus infections are complicated by the remarkable variation strains display in virulence, resistance to antifungals, metabolism, and other infection-relevant traits (3, 5–8). Strain heterogeneity confounds “one size fits all” therapies and poses a significant challenge for developing efficacious disease management strategies (9). Recent evaluations of the A. fumigatus pangenome have revealed extensive genetic variability underlying variation in clinically-relevant traits such as antifungal resistance and virulence, yet in many cases, the origins of such variation remain unexplained (10–12). Determining the genetic drivers of strain heterogeneity will help accelerate the development of strain-specific diagnostics and targeted therapies.

Mobile genetic elements (MGEs) are ubiquitous among microbial genomes and their activities profoundly shape phenotypic variation (13). MGE transposition generates structural variation that directly impacts gene regulation and function (14–16), and many elements also modulate genome content by acquiring genes as “cargo” and transposing them within and between genomes, facilitating gene gain and loss (17, 18). The ability to generate contiguous genome assemblies with long-read sequencing technologies has dramatically enhanced our ability to find new lineages of MGEs (13, 19). For example, Starships are a recently discovered superfamily of MGEs found across hundreds of filamentous fungal taxa (20–22). Starships are fundamentally different from other fungal MGEs because they are typically 1–2 orders of magnitude larger (~20–700 kb versus 1–8kb) and carry dozens of protein-coding genes encoding fungal phenotypes. In addition to flanking short direct repeats, all Starships possess a “captain” tyrosine site-specific recombinase at their 5’ end that is both necessary and sufficient for transposition (21). Many Starships make key contributions to adaptive phenotypes: for example, Horizon and Sanctuary carry the ToxA virulence factor that facilitates the infection of wheat, and Hephaestus and Mithridate encode resistance against heavy metals and formaldehyde, respectively (23–26). Starships are capable of horizontal transfer, implicating them in the acquisition, dissemination and repeated evolution of diverse traits ranging from host interactions to stress resistance (20, 21, 24–26). Starships may also fail to re-integrate into the genome during a transposition event, leading to the loss of the element, its cargo and its encoded phenotypes (21). Through horizontal transfer and failed re-integration events, Starships contribute directly to the generation of rapid gene gain and loss polymorphisms across individuals. However, we know little about how Starships drive genetic and phenotypic variation in species relevant for human disease.

Here, we conduct the first systematic assessment of Starship activity and expression in a human fungal pathogen to test the hypothesis that these unusual transposons are a source of strain heterogeneity. We reveal that Starships are responsible in part for generating key instances of previously unexplainable variation in genome content and structure. Our interrogation of 507 diverse clinical and environmental strains of A. fumigatus, combined with highly contiguous assemblies of 12 newly sequenced strains, enabled an unprecedented quantification of Starship diversity within a single species. We leveraged the wealth of functional data available for A. fumigatus to draw causal links between Starship-mediated genetic variation and phenotypic heterogeneity in secondary metabolite and biofilm production contributing to pathogen survival and virulence. We analyzed multiple transcriptomic studies and determined that variation in Starship cargo expression arises from strain- and treatment-specific effects. By revealing Starships as a previously unexamined mechanism generating phenotypic variation, our work sheds light on the origins of strain heterogeneity and establishes a predictive framework to decipher the intertwining impacts of transposons on fungal pathogenesis and human health.

Results

Data curation and long-read sequencing

Starships are difficult (but not impossible) to detect in short-read assemblies due to their large size and frequent localization in regions with high repetitive content (20). The wealth of publicly available short-read genomes are still of high value, however, due to the breadth of sampling they provide combined with the still relevant possibility of finding Starships, especially in higher quality short-read assemblies such as those available for A. fumigatus. We therefore deployed a hybrid sampling strategy that leveraged both short- and long-read genome assemblies, combined with short-read mapping to genotype Starship presence/absence, to enable an accurate and precise accounting of how Starships impact strain heterogeneity. We downloaded 507 publicly available short-read A. fumigatus assemblies that span the known genetic diversity of this species isolated from environmental and clinical sources (51.7% clinical, 47.9% environmental, 0.4% unknown isolation source) and used Oxford Nanopore to sequence 12 additional strains with long-read technology (66.7% clinical, 25% environmental, 8.3% unknown), for a combined total of 519 assemblies (10, 11). The 12 isolates selected for long-read sequencing were chosen because they represent commonly utilized laboratory strains, clinical isolates and environmental strains from three major A. fumigatus clades (10). We performed additional short-read Illumina sequencing of the 12 strains to provide additional support and error-correct the long-read assemblies. The long-read assemblies are of reference quality, representing nearly full chromosome assemblies with an L50 range of 4–5 and a N50 range of 1.9–4.7 Mb, and are thus ideal complements to the available short-read assemblies for investigating the connection between Starships and strain heterogeneity (Table S1).

At least 20 active and distinct Starships vary among A. fumigatus strains

We began evaluating the impact of Starships on strain heterogeneity by systematically annotating them in the 519 assemblies using the starfish workflow (22). We identified a total of 787 individual Starships and validated these predictions by manually annotating a subset of 86 elements (Table S2, Table S3, Methods) (10, 11). As expected, more Starships were recovered in total from short-read assemblies, but on average more Starships were recovered from each reference-quality assembly, highlighting the utility of sampling both short- and long-read assemblies (Fig. S1). Importantly, unlike many other fungal transposons, we found that with few exceptions, if a Starship is present in a strain, it is present at most in a single copy. Some other species appear to readily accumulate multiple copies of the same Starship (20), but this is not the case for A. fumigatus.

To determine how many different Starships actively generate variability in A. fumigatus, we assigned each Starship element to a family (based on similarity to a reference library of captain tyrosine recombinases), a navis (Latin for “ship”; based on orthology of the A. fumigatus captains) and a haplotype (based on k-mer similarity scores of the cargo sequences; Methods), following Gluck-Thaler and Vogan 2024. For example, Starship Osiris h4 belongs to haplotype 4 within the Osiris navis, which is part of the Enterprise-family. The most reasonable conclusion when observing the same transposon bounded by identical direct repeats at two different loci in two individuals from the same species is that this transposon is presently or recently active (21). Using a threshold that required observing the same navis-haplotype (i.e., Starship) at two or more sites, we identified 20 high-confidence Starships that met these criteria and 34 medium-confidence Starships that did not, revealing a phylogenetically and compositionally diverse set of Starships actively transposing within A. fumigatus (Methods, Table 1, Fig. 1, Table S4, Table S5). For each of the 20 high-confidence elements, we aligned representative copies present in different genomic regions to highlight how precisely and repeatedly element boundaries are conserved across transposition events (Fig. S2).

Figure 1:

At least 20 distinct Starships carrying hundreds of protein-coding genes vary in their presence/absence across Aspergillus fumigatus strains. A) Top: a SNP-based maximum likelihood tree of 220 A. fumigatus strains from Lofgren et al. 2022 (visualized as a cladogram for clarity), color-coded according to phylogenetic clade as defined by Lofgren et al. 2022 (10). Bottom: A heatmap depicting the presence (gray) and absence (white) of 20 high-confidence Starships in each isolate. B) Schematics and sequence alignments of the type elements from 20 high-confidence Starships, where links between schematics represent alignable regions ≥500bp and ≥80% nucleotide sequence identity and arrows represent predicted coding sequences. C) A donut chart summarizing the number of distinct types of Starships by the quality of their prediction. D) A Venn diagram indicating the number of shared and unique high-confidence Starships in the reference strains Af293 and CEA10 (Table S7). E) A scatterplot depicting pairwise comparisons of SNP-based Identity by State (IBS) and Jaccard similarity in high-confidence Starship presence/absence profiles between 150 Clade 1 strains (Table S8).

After using all available genome assembly data to build the Starship library with automated methods, we explored a variety of other approaches for augmenting and validating our predictions, including manual annotation, BLAST-based detection, and short-read mapping (Methods). Certain detection methods were more appropriate for particular analyses, resulting in the partitioning of our findings into the high-confidence, the expanded, and the genotyping datasets. To analyze Starship features, like gene content and pangenome distributions (see below), we limited our analyses to the set of 459 high-confidence elements for which we have end-to-end boundaries and evidence of transposition (high-confidence dataset). For analyses that examine the presence/absence of elements across genomic regions, we used an expanded set of 1818 elements that consists of the 787 high and medium-confidence elements as well as 1031 elements whose presence was detected through BLAST searches, because these methods maximize our ability to correctly predict the presence/absence of an element in a given region (expanded dataset; Table S4, Table S5). For genotyping analyses that focus on detecting the presence/absence of an element regardless of where it is found in the genome, we took a conservative approach by limiting our analysis to the 13 reference quality assemblies plus 466 A. fumigatus strains with publicly available paired-end short-read Illumina data that were used to validate Starship presence/absence by short-read mapping to a reference Starship library (genotyping dataset; Table S21).

Starships distributions are heterogeneous and partially explained by strain relatedness

All high-confidence Starships show polymorphic presence/absence variation, representing a previously unexamined source of genetic heterogeneity across A. fumigatus strains (Fig. 1A, Table S6). For example, the two commonly used reference strains Af293 and CEA10 carry 12 different high-confidence Starships, but share only 4 in common (or only 2, using a more conservative threshold of not counting fragments or degraded elements; Fig. 1D, Table S7). The number of high-confidence Starships per long-read isolate ranges from 0–11 (median = 3, std. dev. = 3.5)(27, 28). To identify the underlying drivers of variation in Starship distributions, we investigated the relationship between Starship repertoires and strain relatedness by testing if phylogenetic signal underlies Starship distributions. We genotyped Starship presence/absence using short-read mapping in the best sampled clade of A. fumigatus with well resolved phylogenetic relationships (Clade 1 sensu Lofgren et al., 2022, n = 150) and found that isolate relatedness is significantly but only weakly correlated with similarity in Starship repertoire (Fig. 1E, Table S8, Spearman’s ρ = 0.29; P < 2.2e⁻¹⁶, using the genotyping dataset). This indicates that Starships may be a source of genetic heterogeneity among even closely related strains. While no sequence-based or phylogenetic method to our knowledge would enable us to unequivocally differentiate horizontal transfer events from incomplete lineage sorting, loss or sexual and parasexual recombination at the within species level, we hypothesize that horizontal transfer within A. fumigatus is at least partially responsible for generating the observed patchy distributions of Starships, since Starship transfer has occurred between fungal species of vastly different taxonomic ranks, and barriers to transfer are likely lower within species (24, 25). The hypothesis of horizontal transfer within species remains an exciting possibility that must be confirmed with laboratory experiments.

Starships may actively transpose in clones of reference laboratory strains

We examined the short-term potential for Starships to introduce unwanted variation into laboratory experiments by comparing lab-specific isolates of the same strain used by different research groups. First, we found that Starship insertions in the reference Af293 genome coincide with 4/8 known putative structural variants identified in other Af293 strains by Collabardini et al. 2022 (Table S9), suggesting that in addition to other mechanisms, Starship transposition and/or loss contributes to structural variation observed among isolates of what should otherwise be clones of the same strain (6). We complemented this literature-based approach with a genomics approach where we compared sequence data among isolates of the same strain. We compared the locations of Starships between the CEA10 assembly we generated here and a publically available CEA10 assembly (BioSample SAMN28487501), and found evidence that at least 1 Starship (Nebuchadnezzar h1) has jumped to a different chromosomal location, from chromosome 6 to chromosome 4 (Fig. S3). While this specific translocation appears robust, other results from these analyses suggest that the strain of CEA10 sequenced here may have been derived from a sample with mixed lineages (See methods). As a result we cannot currently rule out other more exotic explanations for the movement of Nebuchadnezzar h1 in this isolate. To further address the possibility of Starship movement under short timescales relevant to laboratory experiments, we evaluated Starship heterogeneity in three additional strains with both publicly available short-read data and long-read data generated and found that one of the strains (ATCC-46645-lr) showed signs of Starship heterogeneity among isolates used by different research groups. Specifically, a small number of reads (a single mapped read for h2 and three for h1) supported precise excision events for Nebuchadnezzar h1 and h2, suggesting these elements have transposed or have been lost in a subset of the nuclei within one of the sequenced isolates (Fig. S4). This suggests that some but not all Starships appear to be active under laboratory conditions in some strains, which is directly relevant for experimental design in this case as Nebuchadnezzar h1 carries the HAC gene cluster known to impact biofilm-associated virulence (see below) and genes on Starships may be lost through failed transposition events (29). Additionally, one breakpoint of a large-scale reciprocal translocation between chromosomes 1 and 6 among CEA10 isolates is localized precisely within a Starship, suggesting that this structural variant may have been caused by the presence of the Starship, an association that has also been noted in at least one other species (20). Together, Starship-mediated variation among isolates of the same A. fumigatus strain suggests these transposons may cause genomic instability over short-enough timescales to potentially impact routine laboratory work and experimental reproducibility. It is interesting to note that not all strains or closely related strains demonstrate intraspecific variation: for example, similar to Bowyer et al. 2022, we did not detect large-scale variation in Starship presence/absence between CEA10 and its derivative A1163 (30), which have been separated for decades, indicating that much remains to be known about the factors promoting within strain Starship variability.

Starships mobilize upwards of 16% of accessory genes that differ across strains

A. fumigatus strains differ extensively in the combinations of genes found in their genomes, resulting in a large and diverse pangenome (10, 11). To quantify how much pangenomic variation is attributable to actively transposing Starships, we estimated the total proportion of genomic content present in the 20 high-confidence Starships (Fig. 2, Methods). We examined these Starships in the 13 reference-quality assemblies (Af293 plus the 12 newly sequenced long-read strains) and found that between 0–2.4% of genomic nucleotide sequence (and 0–2.1% of all genes per genome) is mobilized as Starship cargo (Fig. 2A). We then built a pangenome with all 519 strains and extracted all orthogroups (i.e., genes) found in the 13 reference-quality genomes to gain insight into the distribution of Starship-associated gene content (Table S10). Across the 13 isolate pangenome, 2.9% of all genes, which corresponds to 9.7% of all accessory and singleton genes, have at least 1 member carried as Starship cargo (Fig. 2B). Accessory and singleton genes are overrepresented >3 fold in Starships, with ~92% of Starship genes being either accessory or singleton compared with 24.6% of non-Starship associated genes. We determined the upper bounds of this conservative estimate by examining the 13 isolate pangenome in the context of all 54 high- and medium-confidence Starships, and found that 4.8% of all genes, representing 16% of all accessory and singleton genes, have at least 1 member carried as Starship cargo (Fig. S5, Table S11). Drawing parallels from decades of observations of bacterial MGEs (17), we hypothesize that localization on a Starship promotes a gene’s rate of gain and loss through Starship-mediated horizontal transfer and failed re-integration events (21). Together, these data reveal a previously hidden association between Starships and the making of A. fumigatus’ accessory pangenome.

Figure 2:

Starships are enriched in accessory genes whose presence/absence and genomic location vary across Aspergillus fumigatus strains. All panels visualize data derived from 13 reference-quality A. fumigatus assemblies and the 20 high-confidence Starships (n = 459 elements total). A) Box-and-whisker plots summarizing the total percentage of nucleotide sequence and predicted genes carried by Starships per genome. B) Donut charts summarizing the percentages of gene orthogroups in the core, accessory, singleton, and Starship-associated compartments of the A. fumigatus pangenome (Table S10). C) Iceberg plots summarizing the genomic locations of the single best BLASTp hits (≥90% identity, ≥33% query coverage) to the cargo genes from the type elements of all 20 high-confidence Starships (left) and two individual Starships (right; Table S12). Each column in the iceberg plot represents a cargo gene and is color-coded according to the genomic location of hits (full dataset in Fig. S5).

Starships differ in their potential to mediate gain and loss of unique sequence

We further determined the potential for Starships to introduce variation into A. fumigatus strains by testing if Starship genes are uniquely found in those elements or found elsewhere in the genome. We calculated the degree of association between a given cargo gene and a Starship by identifying the genomic locations of best reciprocal BLAST hits to genes in the 20 high-confidence type elements (Fig. 2C, Fig. S6, Table S12, Table S13, Methods). Across the 13 reference-quality assemblies, we found that the vast majority of Starship genes display presence/absence variation between strains (i.e., are accessory or singleton) although several genes are present in conserved regions in these particular strains (while seemingly counter-intuitive, these genes do vary in their presence/absence and are Starship-associated when examining the larger 519 strain population, as expected). We found that many genes are almost always carried as cargo in active Starships when present in a given genome (e.g., the majority of cargo on Tardis h1), while others have weaker associations and are found in both Starships and non-Starship regions across different strains (e.g., Gnosis h1). Variation in the degree to which genes associate with active Starships suggests elements differ in capacity to generate accessory sequence variation among strains, highlighting the importance of investigating Starships at the individual element level.

Starship activity generates structural variation at a genome-wide scale

We next asked where Starship-mediated variation occurs in the genome to better understand the implications of Starship activity for genome organization. We identified the genomic locations where Starships introduce structural variation by sorting all 787 high and medium-confidence elements, along with elements detected by BLASTn, into homologous genomic regions (Table S14, Table S15, Table S16; expanded dataset; Methods). This enabled us to genotype individuals in the 519 strain population for segregating Starship insertions that are polymorphic across strains (Fig. 3, Fig. S7; Methods). Across all strains and chromosomes, we found 79 regions that contain at least 1 segregating “empty” insertion site, for a total of 154 sites distributed across all eight chromosomes (a single region can have >1 insertion site). The average number of empty sites per genome is 44.5 (range = 9–63, std. dev. = 13.88), indicating that each strain harbors dozens of sites with structural variation introduced by Starships. For example, the Af293 reference strain has a total of 6 full length Starships with annotated boundaries and 2 Starship fragments, and a total of 56 segregating empty sites (Fig. 3A). We found a significant linear relationship between the number of genomic regions a given Starship is present in (a proxy for transposition activity) and the total copy number of that element in the 519 strain population, suggesting active Starships contribute more to strain heterogeneity compared with less active elements (Fig. 3C; y = 14.7x - 1.24; P = 6.5e⁻⁴; R² adj = 0.46; expanded dataset). Predicted Starship boundaries are precisely conserved across copies present in different genomic regions, suggesting transposition, and not translocation or segmental duplication, is the major mechanism responsible for generating variation in Starship location (Fig. 3D).

Figure 3:

Starships and their insertion sites are distributed across all major chromosomes in the Aspergillus fumigatus genome. A) A Circos plot summarizing all inserted Starships (in red) and all genomic regions containing either an empty insertion site or a fragmented Starship (in black, along perimeter and labeled with the r prefix) in the 8 chromosomes of the A. fumigatus reference strain Af293 (Table S14). All genomic regions contain a Starship insertion in some other individual from the 519 strain population. B) Barcharts summarizing the genotypes of segregating genomic regions associated with the four most active high-confidence elements in the 519 strain population (Table S16; full dataset in Fig. S7). If an isolate did not have any Starships within a given region, it was assigned either an “empty” or “fragmented” genotype (Methods). C) A scatterplot summarizing the relationship between the number of genomic regions containing a given Starship and the total number of copies of that Starship in the 519 strain population, where each point represents one of the 20 high-confidence Starships. A line derived from a linear regression model is superimposed, with shaded 95% confidence intervals drawn in gray. D) Alignments of Tardis h1 copies +/− 50kb of flanking sequence across 14 genomic regions. Links between schematics represent alignable regions ≥1000bp and ≥90% nucleotide sequence identity.

Given that Starships are mobilized by a tyrosine site-specific recombinase (21), we attempted to predict each Starship’s target site to gain insight into the genomic sequences that are susceptible to Starship insertions. Among high-confidence Starships, 19/20 have identifiable direct repeats (DRs) ranging from 1–12 bp in length and 18/20 have terminal inverted repeats (TIRs; Table 1, Table S3). DRs typically reflect a portion of the element’s target site, while TIRs are predicted to facilitate transposition (31). While precise target site motifs must be confirmed experimentally, the target site TTACA(N₅)AAT that we recovered for Nebuchadnezzar elements, which belong to the Hephaestus-family, resembles the canonical Hephaestus target site TTAC(N₇)A, demonstrating the utility of our approach as a first step towards understanding what sequence features promote the gain and loss of Starship-mediated variation (21).

Generally, DRs >6 bp are associated with targeting of the 5S rDNA gene (e.g. Osiris), presumably because this is a relatively stable target site (22). In contrast to this, Tardis has a highly conserved 9 – 11 bp DR that does not correspond to the 5S gene. A k-mer analysis shows that the 10 bp consensus motif CTACGGAGTA is strongly overrepresented in the genome of Af293 (>99.99th percentile of k-mers; Fig. S8), indicating that this motif may represent some other type of highly conserved genomic sequence, such as a transcription factor binding site. Although inconclusive, this motif does overlap with predicted motifs associated with C6 zinc cluster factors in A. nidulans (https://jaspar.elixir.no/matrix/UN0291.2/) (32). This DR sequence is further conserved among other Eurotiomycetes (22). This result highlights the fact that studying Starship elements in detail can reveal other important aspects of a species’ biology, and characterizing the Tardis motif should be of interest for future research.

Starships modulate variation at an idiomorphic biosynthetic gene cluster

Several genomic regions harbor multiple types of Starships, raising the possibility that strain heterogeneity arises in part from the formation of Starship insertion hotspots. We investigated the genomic region with the highest density of Starship insertions to define the upper bounds of Starship-mediated structural variation at a single locus (Fig. 4C, Methods). This region spans an average of 498.76 Kb (range = 158.07–781.54 kb; std. dev. = 83.02 kb) across the n = 210 strains for which we could detect it, and ranges from position 584,521–1,108,409 on Chromosome 3 in the Af293 reference assembly (representing 12.84% of the entire chromosome). By supplementing starfish’s automated genotyping with manual annotation of nine strains with distinct alleles at this region, we found this region contains at least 7 distinct Starships with identifiable DRs and 1 degraded Starship that range from 47.52–151.62 kb long and are inserted into 6 independent segregating sites (Fig. 4C). The majority (75%) of the Starships in this region are inserted into 5S rDNA coding sequences, which effectively fragments that copy of the 5S rDNA gene. Total 5S rDNA copy number varies between 28–33 in the 13 reference quality assemblies (median = 31, std. dev. = 1.4; Table S17), but between 6–8 intact copies are typically present at this single locus, representing a ~10-fold enrichment of the 5S rDNA sequence relative to background expectations given the length of this region (using 5S rDNA frequencies in the Af293 genome). Thus, the enrichment of 5S rDNA at this locus potentiates Starship-mediated variation and the generation of a Starship hotspot.

Figure 4:

Starships mobilize adaptive traits and generate allelic diversity among Aspergillus fumigatus strains. Schematics and alignments of: A) Starship Lamia h2 and h4, which carry the biosynthetic gene cluster (BGC) encoding the polyketide secondary metabolite Fumigermin (42), shown inserted at 3 independent sites. B) Starship Gnosis h2 carrying the BGC encoding the terpene secondary metabolite Fumihopaside A (35), shown inserted at 2 independent sites. C) a large region on Chromosome 3 previously identified as an idiomorphic BGC (33) containing multiple segregating Starship insertions and various combinations of putative BGCs, including a a non-reducing polyketide synthase (NR-PKS) BGC carried by Starship Osiris h4. Starships from the Osiris navis specifically insert in 5S rDNA sequence and are predicted to fragment it. D) Eight Starships in the reference strains Af293 and CEA10 that all carry homologs of biofilm architecture factor A (bafA)(29). Starship Nebuchadnezzar h1 (Neb. h1) carries bafA as part of the H_AAC (hrmA-associated gene cluster), while Starship navis10-var35 carries bafB as part of H_BAC and Starship Osiris h3 carries bafC as part of H_CAC (29). Only a portion of Starship Lamia h4 is visualized for figure legibility. All data for A-D was collected from the 519 A. fumigatus strain population (Table S5). Links between schematics represent alignable regions ≥5000bp and ≥95% nucleotide sequence identity and arrows represent predicted coding sequences. Abbreviations: polyketide synthase (PKS); non-ribosomal peptide synthetase (NRPS); highly reducing (HR).

We were surprised to find that in addition to multiple segregating Starships, the Chromosome 3 region contains an idiomorphic biosynthetic gene cluster (BGC) that was previously noted to be a recombination hotspot (Cluster 10 sensu Lind et al. 2017) (33, 34). The idiomorphic BGC locus is polymorphic for upwards of 4 smaller BGC modules and is upstream of a fifth BGC encoding the biosynthesis of Fusarinine C (BGC8 sensu Bignell et al. 2016) that itself is embedded within a larger stretch of sequence containing NRPS and KR-PKS core genes. While the idiomorphic modules were not predicted by antiSMASH, manual inspection revealed that each contains a different core SM biosynthesis gene and numerous other genes involved in metabolic processes, suggesting they are part of a larger BGC or represent different cryptic BGCs. The NR-PKS BGC module at this locus is carried by Starship Osiris h4, and its presence/absence is directly associated with the presence/absence of the element, implicating Starships as a mechanism generating idiomorphic BGCs. Furthermore, some BGC modules are disrupted by the insertion of a Starship (e.g., the NRPS-like BGC module). Together, these findings support the assertion that Starship activity helps generate selectable variation in BGC genotypes. The implication of this variation for the expression and diversification of natural products warrants further investigation.

Starships encode clinically-relevant phenotypes and are enriched in strains isolated from clinical and environmental sources

Starships encode genes with diverse functions typically associated with fungal fitness, including carbohydrate active enzymes, BGCs and metal detoxification genes (Table 1, Table S5). To investigate broad trends in how Starships might contribute to phenotypic heterogeneity, we compared the predicted cargo functions among the 20 high-confidence Starships (Table S18; high-confidence dataset; Methods). As expected, Starship types differ from each other in their predicted functional content, but functional diversity also varies to some extent within elements from the same type, indicating that both inter- and intra-specific Starship variation may contribute to functional heterogeneity among A. fumigatus strains. In this sense, different copies of the “same” Starship in different individuals have much greater potential to differ from each other compared with much smaller transposons. We found that 49.7% of COG-annotated genes are “Poorly Categorized”; 28.8% belong to the “Metabolism” category, 14.7% belong to “Cellular Processes and Signaling” and 6.8% belong to “Information Storage and Processing”.

Given the importance of metabolic processes for diverse fungal traits, we examined granular classifications within the Metabolism category and found that Starship types differ specifically in their contributions to metabolic heterogeneity (Fig. S9). For example, nearly all 75 copies of Gnosis h1 carry at least 1 gene with a predicted role in “Coenzyme transport and metabolism,” and none carry genes with predicted roles in “Lipid transport and metabolism”, while the exact opposite is true for the 119 copies of Tardis h1. The most frequently mobilized COG metabolism categories are “Carbohydrate transport and metabolism” (present in 371 individual elements belonging to 13 high-confidence Starships), “Secondary metabolite transport and metabolism” (present in 329 elements belonging to 16 high-confidence Starship types) and “Energy production and conversion” (present in 281 elements belonging to 9 high-confidence Starship types), each of which has known associations with clinically-relevant pathogen phenotypes. We found no evidence that any metabolism-associated COG category was enriched in Starships compared with background frequencies in Af293 (data not shown).

While investigating associations between Starships and characterized pathogen phenotypes, we found three notable examples of cargo genes that encode known traits important for pathogen survival and virulence (Table S19, Methods) (10, 11, 33, 35–41). The BGC encoding Fumihopaside A (AFUA_5G00100-AFUA_5G00135) is carried by Starship Gnosis h2 (Fig. 4A). Fumihopaside A is a triterpenoid glycoside that increases fungal spore survival under heat and UV stress exposure (35). Similarly, the BGC encoding the polyketide Fumigermin (AFUA_1G00970-AFUA_1G01010), which inhibits bacterial spore germination, is carried by Lamia h2 and h4 (Fig. 4B) (42). Both the Fumigermin and Fumihopaside A BGCs were previously identified as “mobile gene clusters” based on their presence/absence at different chromosomal locations in A. fumigatus strains, and our results reveal that Starships are the mechanism underpinning their mobility (Cluster 1 and 33, respectively from Lind et al., 2017) (33). Finally, a cluster of three genes (AFUA_5G14900/hrmA, AFUA_5G14910/cgnA, AFUA_5G14915/bafA, collectively referred to as the hrmA-associated cluster or HAC) carried by Nebuchadnezzar h1 increases virulence and low oxygen growth, and its expression modulates colony level morphological changes associated with biofilm development (Fig. 4D) (29, 43). The genes bafB and bafC, which are homologs of bafA, also mediate colony and submerged biofilm morphology and are each carried by up to 5 additional Starships, indicating a sustained association between this gene family and Starships (Fig. 4D, Fig. S10) (29). Mobilization of biosynthetic gene clusters and biofilm-related loci by Starships likely contributes to the rapid evolution of these survival and virulence traits by mediating gene gain and loss through, we speculate, horizontal Starship transfer (which has yet to be experimentally demonstrated) and Starship loss (which has been experimentally demonstrated to occur through failed re-integration events (21)), effectively increasing the capacity of these traits to evolve in populations. We propose that by mobilizing genes contributing to biofilm formation, Starships help drive heterogeneity in clinically-relevant phenotypes at a population level (29).

To identify Starships that could be relevant in either environmental or clinical settings, we tested for an association between the high-confidence Starships and strain isolation source (genotyping dataset; Methods). We found that 5 high-confidence Starships are significantly enriched (P_adj <0.05) in strains from either clinical or environmental isolation sources across the 475/479 genotyped strains for which source data exists (Table S2, Table S20, Table S21, Fig. S11). Lamia h3 (P_adj = 0.013) and Janus h3 (P_adj = 0.034) are enriched in environmental strains. Osiris h4 (P_adj = 0.022), Nebuchadnezzar h1 (P_adj = 0.015; carries bafA, see above) and Janus h2 (P_adj = 0.013) are enriched in clinical strains, providing complementary evidence to recently published analyses of Starship captain enrichment in clinical isolates from several fungal species (44). Although clinical and environmental strains were sampled in roughly equal proportions, we can’t completely rule out unknown biases in our sampling that would contribute to these enrichment patterns. Nevertheless, strains with these 5 Starships were isolated from different countries (between 3–10 countries) and belong to different phylogenetic groups (between 1–3 clades sensu Lofgren et al., 2022 and 1–6 phylogenetic clusters sensu Barber et al., 2021), suggesting these Starships are excellent candidates to test how giant transposons contribute to fitness in environmental and clinical settings.

Starship expression contributes to heterogeneity in a strain-, treatment- and strain by treatment-dependent manner

The revelation of A. fumigatus’ Starships allowed us to test the hypothesis that Starship cargo is expressed under clinically-relevant conditions through re-examination of the RNA-seq datasets available for this pathogen. We analyzed patterns of differential Starship cargo gene expression in 14 transcriptomic studies from the three commonly used reference strains (Af293, CEA10, A1163). In total, 177 samples were included in this meta-analysis. Samples were split into broad treatment categories and corrected for batch effects to allow for comparison of cargo gene expression between studies (Methods). Out of 596 Starship genes total (carried by the high-confidence Starships present in the three reference strains), we identified 459 differentially expressed genes (DEGs; Supplementary Data). We did not find a significant enrichment of DEGs within Starships compared to the genomic background of any strain (Fisher’s Exact Test P-values > 0.05). However, the expression and significance of Starship DEGs vary across Starship naves, fungal strains, and experimental conditions, indicative of pervasive strain-specific, treatment-specific and strain by treatment interactions impacting Starship cargo expression (Fig. 5B & Fig. S14).

Figure 5:

A) A heatmap of transcript abundances (log₁₀TPM) of Starship captain tyrosine recombinase genes (top) and genes within the hrmA-associated cluster (HAC; bottom), collapsed across treatment replicates, for 11 RNAseq studies from A. fumigatus Af293. B) Results from differential expression (DE) tests for select Starships, treatment categories, and strain combinations. Only genes with valid output from DE tests are shown, and only genes with significantly DE are labeled with gene codes. Genes with extreme count outliers, as determined by Cook’s Distance in DESeq2, do not have valid p-values and are labeled with an asterisk (“*”). DE based on a single study are shown as log₂ fold-change (log₂FC) in black with standard error bars, whereas DE genes identified across multiple studies are represented with summarized log₂FC values from a random effects model (REM) and coloured by “sign-consistency”, the number of studies that reported DE in the positive (+1) or negative (−1) direction, centered around 0. Labels in bold font represent DEGs that are significantly DE in more than one study. C) The results of a weighted gene co-expression network analysis (WGCNA) constructed from 14 “antifungal”, “infection”, and “nutrient” RNAseq studies represented using UMAP clustering (87) based on the co-expression eigengene values for all genes in the A. fumigatus Af293 genome. Non-Starship genes present within modules of interest are shown in blue, while all Starship genes are shown in red, with those genes within modules of interest having a black outline. D) Genes within modules that were significantly associated with samples from specific treatment categories or studies were used to construct sub-networks which contained the top 10 edges in the network that were made between any pair of genes or any gene and a Starship captain. The connections between genes (edges) are based on the topological overlap matrix (TOM) for each module, and have been 0–1 scaled. E) Boxplots of eigengene values from WGCNA, akin to a weighted average expression profile, indicate the extent of co-expression (correlation) of the genes present within each module. Pairwise comparisons of module eigengene values determined if a module was significantly associated with samples from specific treatment categories or studies (Fig. S16).

To identify experimental conditions that may impact Starship transposition and the subsequent generation of Starship-mediated heterogeneity, we examined patterns of transcript abundance for captain genes responsible for Starship transposition within the 12 high-confidence Starships from the reference strains Af293 (Fig. 5A), CEA10 and A1163 (Fig. S12). All captain genes have evidence for constitutive gene expression (a minimum median transcript coverage of 1 across the gene body for biological replicates) in at least one experimental condition (Fig. S13). Overall, 8 of the 12 captain genes from high-confidence Starships are differentially expressed in at least one study for one or more reference strains, revealing myriad conditions that create opportunities for Starship transposition (Supplemental information). Differences in experimental conditions and/or strain backgrounds have no consistent positive or negative influence on captain gene expression, indicating that the intricacies of captain expression remain to be elucidated (Fig. 5B).

In order to identify conditions where Starships have potential to impact phenotypes, we examined patterns of differential expression for Starship-mobilized cargo genes. As with captain genes, the expression of cargo genes is context-dependent, within each strain background, with experimental conditions and Starship haplotypes impacting levels of gene expression (Fig. 5B). For example, we compared transcript abundance patterns for genes within the hrmA-asscociated cluster (HAC) between treatment categories and found that the transcription of cgnA and bafA genes within HAC appear to be down-regulated with a select group of infection samples (Fig. 5A & Fig. S12). Together, strain identity, treatment conditions, and non-additive interactions between strains and treatments generate variation in Starship cargo expression.

To gain further insight into Starship-mediated heterogeneity in gene expression and regulation, we constructed weighted gene co-expression networks (WGCNA) which visualize correlations in gene expression. Genes are grouped into modules based on the level of co-expression (Methods). We identified five modules as being significantly associated with “antifungal”-, “infection”-, or “nutrient”-based studies (Fig. 5E). Two major trends in the transcriptional network properties of Starship cargo are apparent. First, multiple Starship-associated genes are integrated into modules containing many other genes from the genome at large, underscoring the possibility that phenotypes emerging from these networks are the product of regulatory interactions between Starship cargo and non-Starship genes (Fig. 5D & Fig. S15). In particular, connections involving captain genes further highlight candidate loci involved in regulatory interactions between Starships and the A. fumigatus genome, and are prime targets for future investigations of the gene networks involved in Starship transposition (Fig. 5E). Second, we also identified modules composed of mainly Starship cargo, indicative of modular networks specific to particular Starships (Table S24). These network analyses paint a nuanced picture of how Starships interact with the broader transcriptional networks of the cell, and implicate the expression of Starship cargo in the generation of transcriptional variation.

Discussion

Phenotypic heterogeneity among fungal pathogen strains poses a major challenge for combating infectious diseases, yet we often know little about the genetic basis of this variation. We hypothesized that a newly discovered group of unusual transposons, the Starships, make important contributions to strain heterogeneity. We tested this hypothesis by systematically characterizing Starship presence/absence and expression in A. fumigatus, an important human fungal pathogen of critically high research priority (1, 45). Our work provides fundamental insight into the mode and tempo of fungal evolution by revealing that strain variation emerges not only from expected genetic mechanisms (e.g., single nucleotide polymorphisms, copy number variation, chromosomal aneuploidies, and other types of structural variants) (10, 11) but from the previously hidden activity of giant transposons.

Although much remains unknown about the fundamentals of Starship biology, Starships in A. fumigatus likely generate variation on a scale approaching MGE-mediated evolution in bacteria, which is a major mechanism driving bacterial adaptation. The median number of Starship-borne genes per A. fumigatus isolate is 0.8% (range = 0 – 2.1%), while between 2.9–4.8% of the pangenome (corresponding to between 9.7–16% of the accessory genome) is Starship-associated. These observations approach analogous measurements of bacterial plasmids in the Enterobacteriaceae, where the median number of plasmid-borne genes per genome is estimated at 3.3% (range = 0 – 16.5%) and where between 12.3 – 21.5% of the pangenome is plasmid-associated (46). Given our requirements that a high-confidence Starship be found in 2 different locations in 2 different strains, our estimates represent a relatively conservative assessment of the mobile fraction of A. fumigatus’ genome. Mobility underpins prokaryotic genome dynamics, yet current models of fungal pathogen physiology and evolution do not take gene mobility through transposition into account. These data highlight the need to understand the population level impacts of Starships across a larger number of plant and animal pathogen species and to integrate these findings into a new predictive framework for fungal pathogenesis. Our Starship atlas thus provides practical insights for elucidating A. fumigatus pathobiology and establishes a roadmap for deciphering the origins of strain heterogeneity across the fungal tree of life.

Materials and Methods

Long-read genome sequencing, assembly and annotation

For DNA extraction strains were grown at 37 °C for 24 hours in shaking liquid cultures using 1% glucose minimal media (47). Biomass was collected by gravity filtration through Miracloth (Millipore). High molecular weight DNA extraction followed the Fungal CTAB DNA extraction protocol (48). Biomass was ground in liquid nitrogen and incubated at 65 °C in a lysis buffer composed of 650 μl of Buffer A (0.35 M sorbitol, 0.1 M Tris-HCl pH 9, 5 mM EDTA pH 8), 650 μl Buffer B (0.2 M Tris-HCl pH 9, 0.05 M EDTA pH8, 2M NaCl, 2% CTAB), 260 μl of Buffer C (5% Sarkosyl), 0.01% PVP, 0.2 mg proteinase K for 30 minutes. Potassium acetate (280 μl of 5M solution) was added and incubated on ice for 5 minutes. DNA was extracted by phenol:chloroform:isoamyl alcohol extraction followed by a secondary extraction with chloroform:isoamyl alcohol. The supernatant was RNAse treated (2.5μg), and precipitated using sodium-acetate (0.3M) and isopropanol. DNA was finally purified using 70% ethanol, dried and rehydrated in TE (pH 9). To preserve long strands of DNA suitable for long-read sequencing samples were gently mixed by inversion and transferred using large bore pipette tips. DNA quality was assessed using NanoDrop and concentration quantified using Qubit dsDNA BR assay kit. Finally DNA fragment size was assessed on 1% agarose gel, looking for large quantities of DNA to remain in the well after 1 hour of running at 100v. DNA was sent to SeqCoast for Oxford Nanopore Sequencing. We used previously sequenced Illumina short read data for polishing the assemblies with Pilon (10). Two strains not previously sequenced were sequenced with Illumina at SeqCoast.

The long-read genome assemblies were assembled with Canu v2.2 and polished with five rounds of Pilon v1.24 wrapped with AAFTF v0.3.0 (49–51). Summary statistics for each assembly were computed with AAFTF and BUSCO v5.7 using eurotiomycetes_odb10 marker set (52). Genome annotation was performed using Funannotate v1.8.10 which trained gene predictors with PASA using RNA-Seq generated for each strain (Puerner et al., forthcoming), and predicted genes de novo in each assembly and produced a consensus gene prediction set for each strain (53, 54).

Functional predictions

We annotated all predicted genes with Pfam domains, InterPro domains and GO terms using InterproScan v5.61–93.0 (-appl Pfam --iprlookup --goterms)(55). We annotated carbohydrate active enzymes using dbcan v3.0.4 (--tools all) and EggNOG orthogroups and COG categories using eggnog-mapper v2.1.7 (--sensmode very-sensitive --tax_scope Fungi)(56, 57). We annotated biosynthetic gene clusters using antiSMASH v6.0.1 (--taxon fungi --clusterhmmer --tigrfam) and 5S rDNA using infernal v1.1.4 (--rfam -E 0.001)(58, 59). On average, 35.7% of genes within Starships have no known PFAM or Interpro domain, GO term, COG category or eggNOG orthology (range = 0–67.6%, std. dev. = 14.8%) but approximately half of all mobilized genes could be assigned to a COG (e.g., 50% of 11,900 genes total, across 459 high-confidence elements). To investigate associations between Starships and pathogen phenotypes, we searched for the presence of 798 genes known to be associated with virulence or infection-relevant phenotypes on Starships (Table S19, Methods)(10, 11, 33, 35–41).

SNP analysis

We calculated pairwise kinship among strains by analyzing a previously generated VCF of all high-quality, filtered, biallelic single nucleotide polymorphisms (SNPs) from the Lofgren et al. 2022 population with plink v1.9 (--distance square ibs flat-missing --double-id --allow-extra-chr; Fig. 1E)(60, 61). We calculated Jaccard similarity in Starship repertoire among all pairwise combinations of strains using genotyping data for the set of 20 high-confidence Starships and the following formula: Starships shared between strain 1 and 2 / (Starships unique to strain 1 + Starships unique to strain 2 + Starships shared between strain 1 and 2).

Pangenome construction

We constructed a pangenome for the 13 reference-quality strains (our 12 long-read assemblies plus Af293) combined with the 506 publicly available short-read assemblies (10, 11) using Orthofinder v2.5.4 (--only-groups -S diamond), and then extracted all orthogroups with a sequence from at least 1 of the 13 reference-quality strains (Fig. 2B)(62). Core orthologs are defined as being present in all 13 strains; accessory orthologs are present in between 2–12 strains; singleton orthologs are present in 1 isolate. We included all 519 strains in the Orthofinder analysis to ensure that ortholog groups are consistent across different population subsets.

BLAST analysis

We determined the genomic locations of all best reciprocal blast hits to cargo genes carried by the 20 high-confidence Starships using BLASTp v2.13 (Fig. 1C)(63). For each cargo gene, we retrieved the highest scoring hit with ≥90% identity and ≥90% query coverage in each of the 13 reference-quality assemblies, and determined whether that hit was found in the same Starship, a different Starship, or not in a Starship at all. If no hit was retrieved, that gene was marked as missing.

Structural variant analysis

To evaluate the validity of putative structural rearrangements within our CEA10-lr assembly, we took two approaches. First, we aligned the raw nanopore reads to the CEA10-lr assembly using Mummer 4.0 with the nucmer command and default parameters (64). Alignments were then visually inspected using the Genome Ribbon software (65). This method confirmed that individual reads spanned the junctions of putative rearrangements and confirmed that no misassemblies were present. However, this approach suggested that two alternative chromosomal topologies existed: one that matches the chromosomal assemblage of the reference strain Af293, and another that exhibits a reciprocal translocation between Chromosomes 1 and 6, as reported for other assemblies of CEA10 (30). To resolve this paradox, we then mapped the raw nanopore reads to both the CEA10-lr assembly, as well as a publicly available high-quality assembly (BioSample: SAMN28487501) with the minimap2 v2.24 map-ont command (66). These were then visually inspected using the IGV genome browser (67). Multiple reads support both topologies but genome-wide variation at single nucleotide polymorphisms is extremely low, implying that the CEA10-lr isolate sequenced here is a heterogeneous mix of at least two lineages derived from the same CEA10 ancestor. The heterogeneous mixture was subsequently confirmed by morphological phenotyping, so single conidial strains of the mixed culture will have to be obtained and resequenced. Despite the mixed CEA10 culture, annotations of Starships in our CEA10 assembly are still valid for this strain, as all of its Starships are present at the same loci in the other publicly available CEA10 long-read assembly (30).

To further assess whether Starships transpose frequently enough to generate variation among isolates of the same strain, we evaluated three additional strains that we sequenced with Nanopore long-read technology that also had publicly available short-read data generated by other research groups: ATCC46645-lr, S02–30, and TP9. Short-read sequence data were mapped to our long-read assemblies and visually inspected in IGV to determine if any Starships were deleted/transposed in the public data. Both S02–30 and TP9 had no apparent variation at Starship loci. However, ATCC46645-lr showed polymorphism at three Starships, namely Nebuchadnezzar h1, Nebuchadnezzar h2, and Gnosis h1, indicative of active transposition. Both of the Nebuchadnezzar haplotypes show clean deletions, indicating that they may have jumped out of the genome of some nuclei within a particular isolate, while the deletions around Gnosis (5 in total) were more complex, which may be a signature of genomic instability rather than Starship mobilization. Note that neither S02–30 nor TP9 have Nebuchadnezzar h1.

Starship annotation

We systematically annotated Starships in the 12 newly sequenced long-read genomes plus 507 publicly available A. fumigatus genomes by applying the starfish workflow v1.0.0 (default settings) in conjunction with metaeuk v14.7e284, Mummer v4.0, CNEFinder and BLASTn (Table S2)(22, 63, 64, 68, 69). Briefly, captain tyrosine recombinase genes were de novo predicted with starfish’s Gene Finder Module and full length elements associated with captains were predicted with starfish’s Element Finder Module using pairwise BLAST alignments to find empty and occupied insertion sites. We filtered out all elements that were <15kb in length, which we have found to correspond to indels of captain genes but never to the transposition of a full length Starship (22). We then manually examined alignments between each putative Starship and its corresponding insertion site, and filtered out all “low confidence” poorly supported alignments indicative of a false positive insertion. Poor alignments were observed most often when Starships inserted into smaller transposons; when an inversion breakpoint occurred within a putative Starship, resulting in an over-estimation of Starship length; or when the insertion site was on a very small contig resulting in small flanking region alignments.

We verified and supplemented these automated Starship predictions with manual annotations of the 3 reference strains Af293, CEA10 and A1163, along with a set of 86 additional elements (Table S7, Table S3). Insertion sites and DRs were manually verified by generating alignments of the elements plus the 50kb flanks to a corresponding putative insertion site, as determined by starfish (Table S4). MAFFT v7.4 was used to generate alignments with default parameters (70). The DRs were visually assessed and the insertion regions were examined for the presence of TEs, which could lead to erroneous target site, DR and TIR determinations. For Starships in the three reference genomes, the set of reference Starships was used as a query with BLAST to verify their start and end coordinates, and insertion sites. Additionally, the starfish output was examined for additional elements which did not meet our initial strict cutoffs to attempt to capture the entire Starship repertoire of these strains.

Starship classification

Starships were then grouped into naves (singular: navis, latin for “ship”) by clustering captain sequences in ortholog groups using mmseqs easy-cluster (--min-seq-id 0.5 -c 0.25 --alignment-mode 3 --cov-mode 0 --cluster-reassign)(71). Starship sequences were then grouped into haplotypes using the MCL clustering algorithm in conjunction with sourmash sketch (-p k=510 scaled=100 noabund) that calculated pairwise k-mer similarities over the entire sequence length, as implemented in the commands “starfish sim” and “starfish group”(72, 73). These automated and systematic predictions yielded a core set of 787 individual elements grouped into 54 distinct navis-haplotype combinations.

We identified segregating insertion sites associated with the 787 elements across all 519 strains using the command starfish dereplicate (--restrict --flanking 6 --mismatching 2) in conjunction with the Orthofinder orthogroups file (see above) that was filtered to contain groups absent in at most 517 strains and present in at most 8 copies per isolate. Starfish dereplicate enables the identification of independently segregating insertions by grouping Starships and their insertion sites into homologous genomic regions using conserved combinations of orthogroups between individuals. We genotyped the presence of each navis-haplotype combination within each region for each isolate. If an isolate did not have any Starships within a given region, it was assigned either an “empty” genotype (if the set of orthogroups that define the upstream region flank were adjacent to the set of orthogroups that define the downstream region flank) or a “fragmented” genotype (if additional orthogroups were in between the upstream and downstream region flanks).

To define high-confidence Starships, we used a threshold that required observing the same navis-haplotype in ≥2 independent segregating sites in different strains, which ensured an accurate and precise determination of each element’s boundaries and the sequences therein (22). We found 18 navis-haplotype combinations meeting these criteria. We considered 2 additional navis-haplotype combinations of interest (Osiris h4 and Lamia h4) to be high-confidence after manually annotating their boundaries and finding evidence for flanking direct repeats that are signatures of Starship boundaries. Each high-confidence Starship is represented by a type element (similar in concept to a type specimen for defining a biological species) that constitutes the longest element assigned to that navis-haplotype. The type elements of the reference Starships range in size from 16,357–443,866 bp, and collectively represent 459 individual elements with an average length of 71,515 bp (Table S4, Table S5). The remaining 328 elements have an average length of 50,249 bp and are represented by 34 “medium-confidence” navis-haplotypes that were only observed at a single segregating site. We predicted the target sites of each high-confidence Starship by manually aligning multiple empty insertion site sequences to each other and identifying columns within 25bp of the core motif that had conserved nucleotides in at least 80% of sequences (Table 1).

Each navis associated with a high-confidence element was assigned a charismatic name (e.g., Tardis) and a new, sequentially named haplotype (h1, h2, etc.) to distinguish it from the automatically generated haplotype codes (var14, var09, etc.). All other 34 navis-haplotype combinations with evidence of only a single segregating site were classified as “medium-confidence”. Medium-confidence Starships with captains not belonging to the high-confidence naves kept their automatically generated navis and haplotype codes (e.g., navis01-var09), while medium-confidence Starships with captains belonging to the high-confidence naves were assigned to that navis but kept their automatically assigned haplotype (e.g., Osiris-var04).

Starship genotyping

Starfish requires a well-resolved empty insertion site in order to annotate the boundaries of a contiguous Starship element. It will therefore not annotate any element whose corresponding empty site is missing, or any element that is partially assembled or not assembled at all. We therefore supplemented starfish output using a combination of BLASTn and short-read mapping to the library of reference Starship sequences to decrease the false negative rate for presence/absence genotyping. First, we recovered full length and partial elements using the set of high-confidence Starships as input to the command “starfish extend”. This command uses BLASTn to align full length Starship elements to the sequence downstream from all tyrosine recombinase genes not affiliated with a full length Starship element.

Separately, we downloaded publicly available paired-end short read Illumina sequencing data for all 466 strains for which these data were available (10, 11) and mapped them to the library of 20 high-confidence Starships using the command “starfish coverage” and the aligner strobealign (74), ensuring that we would correctly genotype the presence/absence of a Starship even if it is not present in an assembly or partially assembled. By mapping short reads directly to a library of known Starship elements, we overcame some of the drawbacks associated with working with short-read assemblies. We considered any Starship with a minimum of 5 mapped reads at each position across 95% of its entire length to be present. We considered a Starship to be present in a given individual if either the main starfish workflow, BLASTn extension, or short-read mapping identified it as present; otherwise we considered it absent.

Meta-Analysis of A. fumigatus RNAseq Datasets

We collected transcriptomes from a total of 434 publicly available paired-end libraries of A. fumigatus strains Af293, A1163, and CEA10/CEA17 (Table S22). We surveyed the available metadata from each BioProject and binned the samples into general categories based on the type of treatment applied in each study. The categories that were the most well represented across multiple BioProjects include exposure to antifungals (“antifungal”), in vivo and in vitro infection experiments (“infection”), or supplemented growth media with specific nutrients (“nutrient”; Table S22, Table S23).

RNAseq reads were retrieved from the NCBI SRA using fasterq-dump from the SRA toolkit (last accessed June 26, 2024: https://github.com/ncbi/sra-tools) and were trimmed for quality and sequencing artifacts using TrimGalore (last accessed June 26, 2024: https://github.com/FelixKrueger/TrimGalore). Transcripts were quantified using salmon quant, which employs a reference-free pseudo-mapping based approach to quantify transcripts. In addition to the pseudo-BAM files created by salmon v1.10.1 (75), we created separate sets of BAM files to assess transcript coverage across Starship genes by mapping to the transcriptome with STAR v2.7.11a (76). We assessed transcript coverage across the core of Starship genes using the bamsignals R package (last accessed June 26, 2024: https://github.com/lamortenera/bamsignals. Genes with a median core-transcript coverage less than 1 were considered to have insufficient evidence of being expressed. We applied an abundance filter on transcript abundances, removing genes represented by fewer than 10 transcripts in at least 3 samples. We corrected for batch effects between BioProjects using CombatSeq (77) and excluded BioProjects with inconsistent/unclear metadata, or those that remained as outliers in the PCA after batch-correction. DESeq2 v1.45.0 (78) was used to perform a series of differential expression (DE) tests using the corrected transcript counts. These DE tests were performed individually for each BioProject, with each test comparing the differences in expression between all control and treatment samples. Differences in treatment level were included as a cofactor in the model, where applicable.

To summarize expression patterns across BioProjects which tested similar conditions, we employed a random effects model (REM) using the R package metaVolcanoR (79). This method identifies differentially expressed genes (DEGs) as genes that are significantly differentially expressed in all studies and accounts for the variation in the expression for each DEG observed in multiple BioProjects. The output from the meta-volcano REM are summarized log2 fold-change, a summary p-value, and confidence intervals for each DEG. In addition, a measure of “sign-consistency” is used to evaluate the consistency of DEG, which is expressed as a count of the number of BioProjects where a DEG was observed with the same directional change (+/−), centered around 0.

We performed weighted gene co-expression network analysis (WGCNA) using PyWGCNA v1.72–5 (80) to construct gene co-expression networks for A. fumigatus reference strain Af293. A single network was constructed using the batch-corrected TPM values from the collection of samples belonging to “antifungal”, “infection”, and “nutrient” treatment categories. PyWGCNA automatically estimates an appropriate soft-power threshold based on the lowest power for fitting scale-free topology and identifies modules of co-expressed genes through hierarchical clustering of the network and performing a dynamic tree cut based on 99% of the dendrogram height range. The topological overlap matrix is then computed and a correlation matrix is constructed to produce the final network. Distributions of TOM scores were generated based on edges in the network between genes within Starships, between Starships, between Starships and the rest of the genome, and between non-Starship genes in the genome. We compared these distributions using a Wilcoxon test and anova, adjusting p-values with the Holm–Bonferroni method. Modules in the network that were significantly associated with samples from a specific treatment category were identified using the pairwise distances between observations using pdist from the python module scipy (81). To highlight the genes within each module that have the most strongly correlated expression profiles, sub-networks were constructed using only the edges 10 highest TOM scores made between either Starship captain genes or any non-Starship gene that was present in the module. The collections of genes within these modules were also tested for functional enrichment using gprofiler2 (82).

Statistical tests and data visualization

Enrichment tests were performed with either the Binomial test (binom.test; alternative = “greater”) or the Fisher’s exact test (fisher.test; alternative = “greater”) implemented in base R, using the p.adjust function (method = Benjamini-Hochberg) to correct for multiple comparisons. We visualized all element alignments and insertion site distributions using circos and gggenomes (83, 84). All other figures were generated in R using ggplot2 (85).

Importance.

No “one size fits all” option exists for treating fungal infections in large part due to genetic and phenotypic variation among strains. Accounting for strain heterogeneity is thus fundamental for developing efficacious treatments and strategies for safeguarding human health. Here, we report significant progress towards achieving this goal by uncovering a previously hidden mechanism generating heterogeneity in the major human fungal pathogen Aspergillus fumigatus: giant transposons called Starships that span dozens of kilobases and mobilize fungal genes as cargo. By conducting the first systematic investigation of these unusual transposons in a single fungal species, we demonstrate their contributions to population-level variation at the genome, pangenome and transcriptome levels. The Starship atlas we developed will not only help account for variation introduced by these elements in laboratory experiments but will serve as a foundational resource for determining how Starships shape clinically-relevant phenotypes, such as antifungal resistance and pathogenicity.

Data availability

Sequencing reads and genomic assemblies for the 12 isolates sequenced with Oxford Nanopore have been deposited at NCBI (Table S1). We downloaded the Af293 reference assembly (accession: GCF_000002655.1) as well as 252 assemblies and annotations from Barber et al. 2021 from NCBI, and 256 assemblies and annotations from Lofgren et al. 2022 from the associated Zenodo repository (61). We subsequently reformatted the contig and gene identifiers to include genome codes generated as part of this study (Table S2). All scripts (including raw data for figure generation), genomic data used for this study, and results from differential expression tests are available through the following Figshare repository DOI: 10.6084/m9.figshare.26049703.