Publicly available complete genomes.

We searched for “Neisseria gonorrhoeae” in the National Center for Biotechnology Information (NCBI) Genome Datasets (https://www.ncbi.nlm.nih.gov/datasets/genome/) and downloaded all genomes with an assembly level of “chromosome” or “complete” on July 31, 2025. We determined the sequencing technology and assembly method from the associated publications (where available) and associated metadata available on NCBI. In the rare cases of discrepancies between the publication and NCBI, we used the metadata reported in the publication.

Nanopore sequencing and genome assembly.

N. gonorrhoeae isolates were streaked from frozen stocks onto GCB agar (Difco) with Kellogg’s supplement [27] and grown at 37° C with 5% CO₂ overnight. Genomic DNA was extracted using the Invitrogen PureLink Genomic DNA Mini Kit. We prepared the library using the Oxford Nanopore Native Barcoding Kit 24 V14 (SQK-NBD114.24) and performed Nanopore sequencing using a MinION Mk1B device. Basecalling was performed using Dorado v0.8.2 [28] superaccuracy basecalling. Reads were demultiplexed using Dorado v0.8.2 demux and filtered using Filtlong v0.2.1 (https://github.com/rrwick/Filtlong) to keep the 90% highest quality read bases and only reads longer than 1kbp. Genomes were assembled using Autocycler v0.2.1 [29] with 4 read subsets at 25x minimum depth and using the assemblers: Canu v2.2 [30], Flye v2.9.5 [31], miniasm v0.3 [32], NECAT v0.0.1 [33], NextDenovo v2.5.2 [34], and Raven v1.8.3 [35]. Genomes that were fully resolved (one complete genome plus optional plasmids) and had a length of at least 2 Mbp were kept, as we expect the length of the N. gonorrhoeae genome to be 2.2 Mbp.

Comparison to other sequence search methods.

To confirm that we had identified all the opa genes in the genomes, we compared our method to three additional approaches:

1. We performed a search of the conserved region of the FA1090 opa1 sequence between the semivariable and hypervariable 1 regions using BLAST v2.14.1. We set the maximum number of high-scoring segment pairs (local alignments) to keep as 15 for a single query-subject pair and kept all other parameters at the default values. We matched the hits from the BLAST results that were also found by our search algorithm by merging the nearest start positions of the hits from the two search algorithms if they were within 1200 nucleotides. BLAST hits that were not merged were identified as putative additional hits found by BLAST not found by our search algorithm and we manually inspected these sequences.
2. We annotated the genomes using Prokka v1.14.6 [37]. We set the genus to “Neisseria” and species to “gonorrhoeae” and kept all other parameters at their default values. We noticed that all the opa genes identified with our method had the gene name piiC_*, where * was a number. We then checked if there were any additional genes annotated as piiC_* that were not identified by our method.
3. We used Roary v3.13.0 [38] to define the pan-genome and cluster genes. We set the minimum percentage identity for the BLASTp program within Roary to be 90% and kept all other parameters at their default values. We looked at genes that clustered with the opa genes identified by our method to determine if there were any additional genes in these clusters not identified by our method.

Identification of opa pseudogenes.

To identify opa pseudogenes that were missing the start and/or stop codon, we extracted the sequences, including upstream and downstream sequences, of the hits found by BLAST that were not found by our search algorithm and aligned the sequences with the FA1090 opa sequences using MAFFT v7.520 [39]. We visualized the alignments using JalView v2.11.5.1 [40] to identify if each hit was an opa pseudogene compared to the FA1090 opa sequences.

Genomic rearrangements.

Genomic rearrangements were detected using progressiveMauve v2.4.0 [41] by comparing each complete genome (query) with the complete FA1090 genome (reference). We rearranged the opa position in the query genome relative to the reference genome by reordering and flipping regions of the query genome that fell into locally collinear blocks compared to the reference. Regions of the query genome that did not fall into any locally collinear block with the reference genome were not altered.

Generating reference-mapped pseudogenomes.

To generate reference-mapped pseudogenomes, we used short reads generated by, or simulated to be generated by, Illumina sequencing. Sequencing reads from all publicly available N. gonorrhoeae datasets were downloaded from the European Nucleotide Archive (S3 Table and https://github.com/qinqin-yu/gonococcus_opa_diversity/blob/main/opa_diversity_snakemake/input_data/whole_genome_metadata/global_isolates_metadata_met_qc.csv). For the publicly available complete genomes, we used ART v2016.06.05 [42] to simulate paired-end short reads from the MiSeq v3 sequencing system with a read length of 250 bp, 80x coverage, mean DNA fragment size of 600 and standard deviation of 100. Reads were mapped to the NCCP11945 (NC_011035.1) reference genome using BWA-MEM v0.7.17 [43]. Duplicate reads were marked with Picard v3.0.0 (https://broadinstitute.github.io/picard/), and reads were sorted with SAMtools v1.17 [44]. The quality of the mapped reads was assessed using Qualimap’s bamqc v2.2.1 [45]. We used Pilon v1.24 [46] to call variants (minimum mapping quality 20 and minimum coverage of 10). To create pseudogenomes, we replaced the reference allele with high quality variant calls (at least 90% of reads supporting the allele). Positions called as deletions by Pilon were replaced with a gap character, and positions with low coverage or an indeterminate allele were replaced with an N.

Selecting representative global genomes.

To select representative global genomes from the publicly available N. gonorrhoeae isolates, we included isolates with genomes meeting the following quality control filters: at least 80% of reads mapped to the N. gonorrhoeae reference genome, the coverage of reads mapped to the N. gonorrhoeae reference genome was > 40X, fewer than 12% of the sites in the N. gonorrhoeae reference genome were unable to be confidently called using our assembly pipeline, and the de novo assembly length was 1.75 Mbp-2.5 Mbp. This yielded 21,653 genomes which were then clustered using PopPUNK v2.6.0 [47]. The isolate in each PopPUNK cluster with the fewest contigs in its de novo assembly was chosen as the representative isolate from that cluster, resulting in 737 representative genomes (S4 Table).

Recombination-masked phylogenies.

Recombination-masked phylogenetic trees were created using the reference-mapped pseudogenomes in Gubbins v3.3.4 [48] using the GTR substitution model, the RAxML Next Generation tree builder [49], a maximum of 20 iterations for the phylogeny of complete genomes, and a maximum of 5 iterations for the phylogeny of representative and complete genomes.

Comparison of within- and between-isolate opa diversity.

To compare within- and between-isolate opa diversity, we first translated the opa sequences downstream of the coding repeats. We excluded all sequences with premature stop codons due to frameshift mutations downstream of the coding repeats or nonsense mutations. We then aligned the amino acid sequences using MAFFT v7.520 [39] with the default parameters. We then calculated pairwise distances in the alignment for all pairs of opa sequences from different isolates (between-isolate comparison) and all pairs of opa sequences from the same isolate (within-isolate comparison).

Similar opa within the same isolate.

To identify clusters of similar opa sequences from the same strain, we used the NetworkX Python package (https://github.com/networkx/networkx) to create graphs where the nodes are the opa sequences and the edges connect opa sequences with less than 5% difference in amino acid sequence alignment. The clusters of similar opa sequences were identified as the connected components in the graph. The percentage of isolates with at least one pair of similar opa was determined by finding the isolates that contained at least one opa sequence with less than 5% difference in amino acid sequence alignment with another opa sequence in the same strain.

Comparison of opa sequence distance and genomic sequence distance.

We calculated the opa sequence distance and the genomic distance between pairs of genomes. To facilitate quantitative comparisons between pairs of genomes, only pairs of genomes with the same number of opa genes were kept. The opa sequence distance was calculated using the alignment generated above and pairing the opa sequences between the two genomes that had the fewest number of segregating sites using a greedy algorithm (i.e., find the two closest opa sequences, then find two next closest opa sequences, etc. If there are identical matches, then one is chosen at random to be matched.). The total opa sequence distance was calculated by summing the distances between each pair of opa sequences and dividing by the number of opa genes (where a distance of 0 indicates the opa repertoire in the two genomes are identical and a distance of 1 indicates that all sites in all opa genes in the two genomes are different).

The pairwise genomic sequence distance was calculated as the number of SNPs between pairs of pseudogenomes divided by the length of the pseudogenome using pairsnp v0.3.1 (https://github.com/gtonkinhill/pairsnp). Missing sites (indicated by N) were excluded in the calculation of the SNP distance. Sites that were missing in some isolates but present in the two isolates being compared were included.

Identifying opa genes.

We downloaded complete genomes (“chromosome” or “complete” assembly levels) from NCBI RefSeq on February 13, 2025 using the search term “Neisseria” and filtered out all N. gonorrhoeae genomes (S5 Table). Genomes were rotated using the rotate program v1.0 [36] to match the first 90 bases of dnaA from FA1090 (NC_002946.2), allowing for up to 5 mismatches.

If this sequence was not found, then the unrotated genome was used for downstream analysis. This occurred in 60/198 (30%) genomes. We identified opa genes using the custom script that we used for N. gonorrhoeae, which looks for the coding repeats and the unique conserved sequence near the stop codon. To look for additional opa that were missed by our script, we BLASTed the conserved regions of N. gonorrhoeae FA1090 opa1.

Assessing sequence relatedness.

We calculated the k-mer distances using MASH v2.3 [50] with a k-mer size of 9 and used the distances as input to create a neighbor joining tree using rapidNJ v2.3.2 [51].

Definition of semivariable, hypervariable 1, and hypervariable 2 region sequences.

We aligned the nucleotide sequences of all opa genes using MAFFT v7.520 [39] with a gap opening penalty of 4 and a gap extension penalty of 1. The nucleotide sequence of FA1090 opa1 was compared to the sequences in Bhat et al. [22] to identify the semivariable, hypervariable 1, and hypervariable 2 region sequences. The variable regions in the other opa sequences were determined based on where they aligned to FA1090 opa1.

Calculation of k-mer distance.

The k-mer distances between sequences in each of the semivariable, hypervariable 1, and hypervariable 2 regions were calculated using MASH v2.3 [50] with a k-mer size of 6 for the semivariable region and 7 for the hypervariable 1 and hypervariable 2 regions. These k-mer sizes were chosen so that the probability of observing a random sequence of length k in a random sequence the length of each region was 0.01. The p-value is the probability of observing a given k-mer distance (or less) under the null hypothesis that both sequences are random collections of k-mers.

Clustering.

The variable region sequences were clustered using the Markov Cluster Algorithm (MCL) v14.137 [52]. We constructed a distance matrix using the -log10 transformed p-value from MASH and values larger than 200 were set to 200. We made these transformations following the approach in TRIBE-MCL, which clusters proteins using -log10 transformed e-values from BLAST [53]. The distance matrix was input into MCL, which then performs multiple rounds of expansion and contraction to amplify large values and reduce small values of the matrix. The number of rounds of expansion and contraction is set by the inflation parameter. We ran the clustering using a range of inflation parameters: 1.4, 2, 4, 6, 8, 10, 12, and 14. We chose the inflation parameter that gave a stable clustering, inspired by TRIBE-MCL. A stable clustering was defined as having a percentage difference in clustering with the next lowest inflation parameter as less than 1% for the semivariable and hypervariable 2 regions. For the hypervariable 1 region, we relaxed the definition for a stable clustering as having less than a 5% difference in clustering with the next lowest inflation parameter due to the large number of clusters. The percentage difference in clustering is calculated as the sum of the number of nodes (sequences) that must be exchanged to transform the two clusterings into each other (also called the split/join distance) divided by two times the number of sequences. The motivation for this approach is that TRIBE-MCL uses convergence of outputs to determine the inflation parameter.

The cluster sequence logos were visualized by aligning the nucleotide sequences in each cluster using MAFFT v7.520 [39] with the default parameters and plotting the sequence logos with Logomaker v0.8 [54].

Dated phylogeny.

Pseudogenomes were generated as described above but with the reference genome being one randomly selected genome from the subtree. Regions of the genome under recombination were detected using Gubbins v3.3.4 [48] (GTR substitution model, RAxML Next Gen tree builder, 20 iterations) and masked. The recombination-masked alignment was used to build a dated phylogeny in BEAST2 v2.7.5 [55], for all isolates that had date of collection information. When the date was given as a year, we set the date to January 1 of that year. We used a GTR substitution model with gamma rate heterogeneity (4 categories), a strict clock model, a coalescent constant population, and 100,000,000 chain length. Chains were inspected to confirm good mixing (effective sample size greater than 200).

Ancestral state reconstruction.

We performed ancestral state reconstruction for each of the loci and variable region cluster types separately on the dated phylogeny using PastML v1.9.49 [56] with the default parameters. Isolates that did not have an intact opa at a locus were not included for the inference at that locus. If there were no changes to the variable region cluster type at a locus then the average rate of cluster type changes was set to 0.

Number, phase, and locations of opa genes.

While most isolates had 11 opa genes, we found a handful of instances of opa loss and gain. The loss events occurred across the phylogeny. In some instances, we saw examples of the gene loss events passed on to progeny. More than 50% of these loss events were explained by opa pseudogenes that were the beginning or end of the gene. We were unable to identify the mechanism for the other loss events, which may have resulted from larger deletions in surrounding regions of the genome, since we could not find remnants of these opa genes by homology search methods. Furthermore, in intact opa sequences, we observed frameshift mutations occurring downstream of the coding repeat sequence leading to a premature stop codon, but we were unable to determine how many of these frameshift mutations were due to sequencing error. Our simulation studies indicated that genome coverages down to 50x are reliable and those down to 21x could also be reliable. The genomes that we sequenced have coverage above this range. We were unable to conclusively determine the coverages of the publicly available complete genomes. While our simulation tests suggest that frameshift mutations due to sequencing error at the coverage levels for the complete genomes that we sequenced should be rare, we could not fully resolve the issue for the observed frameshift mutations. In one isolate, we identified 12 opa genes, which seemed to be due to a recent genomic duplication event that exactly duplicated another opa gene in the genome. The fact that we only observed one isolate with 12 opa suggests that the expansion of opa number is rare.

Across the population, fewer opa were in frame than expected by chance. Furthermore, near-identical opa genes in the same isolate tended to be out of frame more often than unique opa genes, suggesting selection against expression of redundant opa; we note, however, that the receptor binding characteristics of Opa proteins cannot be predicted from opa sequence alone and there may be additional phenotypic redundancy, i.e., Opa proteins with low sequence homology may also share receptor binding characteristics. We further note that the isolates we sequenced were not specifically selected for Opa expression state during culturing, so the average number of in frame opa genes may be lower than observed in isolates directly obtained from human specimens with no or minimal passaging. Similarly, publicly available genomes may have an over representation of out of frame opa genes due to unknown culturing conditions. Nevertheless, our results suggest mechanisms other than chance alone guide the patterns of phase variation, for example, regulation of changes in repeat number, selection against isolates with many opa alleles in frame, or selection for Opa proteins with specific receptor binding characteristics. While these mechanisms are yet unknown, previous work suggested that promoter strength may play a role in regulating opa phase [26], and this could potentially regulate changes in repeat number. Another potential mechanism is that there may be selection against in-frame opa in certain environments, for example the presence of progesterone [65–67], which has been shown to inhibit the growth of Opa positive isolates in vitro.

Genomic rearrangements shuffle opa position in the genome, but by accounting for them, we were able to assign consistent opa loci. Genomic rearrangements in closely related isolates happened very rarely, but the locus assignment approach can be updated if we detect more examples of genomic rearrangements in closely related isolates.

opa diversity and evolution.

The diversity of opa alleles within isolates was generally high, with the average isolate having 7 distinct opa alleles, but there were some isolates with as few as 3 distinct opa alleles. This suggests that there is some pressure to have diverse opa alleles within an isolate. At the same time, almost all isolates had at least two near-identical opa alleles, possibly as a result of gene conversion, suggesting there may be a balance between maintaining diversity and maintaining function while diversifying.

opa genes are on average 74 times more diverse than the rest of the genome. More closely-related isolates have more similar opa sequence repertoires. In the context of reinfections, our results suggest that when reinfections occur with the same or a closely related isolate, there will be a similar opa repertoire. We speculate that if there is immunological memory to the opa that are expressed, then there will be selection in subsequent reinfections for isolates that have different opa turned on, an idea that is also supported by evidence of variation in opa expression over the course of single infections in human challenge studies [68,69]. When reinfections occur with different isolates, there will likely be different opa alleles expressed.

N. gonorrhoeae opa allelic diversity did not overlap with opa from other Neisseria species. While sequencing more complete genomes may lead to more observations of interspecies mosaicism, the separate clustering of N. gonorrhoeae and N. meningitidis opa in our dataset and in Malorny et al. [9] suggests phylogenetic separation of opa sequences between species. It is possible that interspecies recombination was historically important in shaping the opa repertoire and that the phylogenetic separation of opa between species may reflect ancient divergence. The separate clustering of Neisseria species opa may be due to the different functional constraints of host receptor binding or differences in flanking regions of opa genes from different species that prevent homologous recombination from occurring between species.

opaK varies more rapidly than other loci, but it is not significantly more in frame in our dataset. The differences in rates of evolution by opa locus may be due to variability in the accessibility of each locus for recombination or other diversity generating processes (i.e., structural accessibility). The changes in opaK tend to reflect recombination among existing opa types, suggesting that diversity is mostly generated through shuffling of existing sequences. Bilek et al. [24] suggested a possible connection between opa11 (opaK in this study) variation and the neighboring pilE Pilin expression gene variation, but noted this may not be causal.

Hypervariable region 1 and 2 alleles in N. meningitidis are more correlated than expected due to chance [70]. We showed that this phenomenon also occurs in N. gonorrhoeae, which is consistent with studies showing non-random associations of particular hypervariable region 1 and 2 sequences [71,72]. The non-random association of hypervariable alleles is consistent with predictions of a mathematical model of strong immune selection [73], physical linkage on the chromosome, and selection for particular combinations of hypervariable alleles for functional reasons (i.e., to bind to specific human receptors) [74]. Previous work has shown that swapping the hypervariable region sequences between opa genes does not lead to predictable changes in receptor binding [74], underscoring the importance of functional coordination between these regions and providing a plausible mechanism for the selection of specific combinations.