Sample collection.

We obtained chemosynthetic bivalve samples from hydrothermal vents, cold seeps, and reducing coastal sediments from around the world (S1 Table). Calyptogena magnifica samples were collected from the East Pacific Rise (EPR) hydrothermal vent fields between 1998 and 2004. Calyptogena fausta were collected from the Juan de Fuca (JDF) Ridge hydrothermal vent system in 2004. The single Calyptogena extenta specimen was collected from Monterey Canyon in 1995. Solemya pervernicosa samples were obtained from the Santa Monica sewage outfall in 1992 (as in [15]). Solemya velum were collected from Point Judith, RI, as described in [15]. B. childressi were sampled from the Veatch Canyon cold seep off of New England. B. septemdierum was sampled from the ABE and Tu'i Malila hydrothermal vent sites in Lau Basin. All tissue samples were stored at -80°C until sterile dissection or subsection of previously sterile-dissected gill tissue as described in [15].

DNA extraction and Illumina sequencing.

We extracted DNA from gill samples for each host individual sampled using Qiagen DNeasy kits following the manufacturer’s instructions. We quantified DNA concentrations using a Qubit dsDNA kit and normalized each sample to 10 ng/ul for Illumina library preparations. We produced the majority of our Illumina sequencing libraries for each sample using a Tn5-based protocol for tagmentation followed by dual-indexing PCR using HiFi DNA polymerase (Kappa Bioscience) and custom primer sequences (IDT) designed to uniquely label both i5 and i7 indexes for each sample (Tn5 enzyme was expressed and purified in-house). Indexed samples were pooled and sequenced on single lanes of a Hiseq4000. We sequenced a total of four lanes of Hiseq4000 paired-end 150 bp sequencing across the entire study. Additionally, we obtained a subset of samples for C. magnifica using genomic methods from our previous work ([15], S1 Table). We also used a dataset we have previously collected for S. velum, specifically the population from Point Judith, Rhode Island ([15], S1 Table). The specific library preparation methods and number of read pairs obtained for each sample included in this work are listed in S1 Table.

Nanopore sequencing.

For each de novo genome assembly in this work, we selected a representative from each host population based on DNA quality as determined using an Agilent Tapestation and DNA concentration based on qubit readings. We sequenced each sample on a single minion flow cell using the ligation-based 1D chemistry, SQK-LSK109 kit per ONT instructions with minor adjustments. The end-repair reaction was incubated for 30 minutes each at 20°C and 65°C and the ligation reaction was performed for 30 minutes instead of the recommended 10 minutes. Read counts obtained and mean read lengths are available in S1 Table. We performed basecalling using the Albacore basecaller v2.0.1 and we discarded the subset of reads whose mean quality score was less than 7. These are the set marked nominally as “failed” by the basecaller software.

De novo symbiont genome assembly.

Reference genomes for the B. septemdierum, B. childressi, S. pervernicosa, and C. fausta symbionts were assembled using combined Nanopore and Illumina reads. First, we assembled the Nanopore reads using the long-read assembly program wtdbg2 [69] using the “ont” presets option and setting the parameter -k to 15. Then, we performed two rounds of reference genome improvement by aligning Illumina sequencing reads from the same individual to the resulting unfiltered assembly and polishing with the Pilon software package [70]. We used BWA-MEM [71] to align Illumina reads in each subsequent polishing round.

We assembled the C. extenta symbiont genome from an Illumina library prepared for the single sampled individual using IDBA [72] and SPAdes [73]. While both assemblies were highly contiguous (N50 = 596007 and 604961 bp, respectively), the SPAdes assembly was able to merge two contigs that were split in the IDBA assembly, so the two contig SPAdes assembly was used for downstream analyses. Comparisons of synteny demonstrated that this join was found in other Vesicomyid endosymbiont genomes suggesting it is correct (see below).

Because read mixtures include host genomic DNA, mitochondrial DNA, and genomic DNA from other bacterial species, we then rigorously filtered the resulting contigs to extract only high confidence contigs contributed by bacteria of the study species. To identify symbiont contigs, we called ORFs with Prodigal v2.60 [74] and annotated coding sequences with BLAST [75] as described in [15] (NCBI nr, TrEMBL, and UniProt database accessed on April 7, 2019). We annotated ribosomal RNAs with RNAmmer [76] and transfer RNAs with tRNAscan [77]. Using the taxonomic information encoded in the annotation, we identified contigs that were confidently of symbiont origin. Then, using these contigs, we filtered the remaining contigs by GC content, read coverage, and coding density using custom scripts. Finally, we evaluated the quality of the assemblies with CheckM [78] and by testing for the presence of core bacterial phylogenetic markers [79] (see S2 Table).

Host mitochondrial genome assembly.

Mitochondrial genomes were assembled from Nanopore reads, which were subsequently corrected with Illumina data, as described above, or they were assembled from Illumina reads directly. As the different samples contained different mitochondrial coverage, higher short read coverage was often better than lower long read coverage for recovering these genomes. We assembled mitochondrial genomes for C. fausta and B. septemdierum with IDBA [72] using Illumina data from two of the highest depth-of-coverage samples (C. fausta 31 and B. septemdierum 231, respectively). The complexity of the B. septemdierum data prevented IDBA from finishing within a week, so we first removed low coverage nuclear kmers (<10x) with Quake [80]. The C. extenta mitochondrial genome was assembled along with the symbiont genome using SPAdes [73], as described above. We were able to use the Nanopore-based assembly from Bathymodiolus childressi for the mitochondrial genome. Lastly, the complete mitochondrial genome for S. pervernicosa was available from [15], so we did not reassemble it here.

After assembly, we identified the mitochondrial scaffold by blasting the full set of scaffolds against a database containing the currently available set of 19 bivalve mitochondrial genomes. Then, we annotated the mitochondrial genome with MITOS [81]. For mitochondrial genomes lacking conserved genes, we repeated genome assembly and mitochondrial genome identification and annotation with a different sample to verify we obtained the full sequence. See Section 1 of S1 Text for a description of the host species identification verification process.

Short read alignment.

After producing endosymbiont and host mitochondrial genome assemblies for each host/endosymbiont species, we aligned short read Illumina data from each individual to a reference genome consisting of both of these genomes. Genomes assembled previously for the C. magnifica symbiont ([11], accession NC_008610.1) and mitochondrian ([82], accession NC_028724.1), and the S. velum symbiont ([13], accession NZ_JRAA00000000.1) and mitochondrian ([83], accession NC_017612.1) were used as references for these populations. We used the BWA mem software package [71], and we then sorted and removed duplicate reads using the samtools software package [84]. After this, we performed indel realignment for each sample separately using the “IndelRealigner” function within the Genome Analysis Toolkit (GATK) software package [85].

Genotyping and variant filtration for each host individual.

We called consensus genotypes for each individual jointly using the GATK “UnifiedGenotyper” option and we ran the program with otherwise default parameters except we required that it output all sites rather than just all variable positions. We filtered variant sites using the vcftools software package [86] largely following the GATK best practices as we have done in our previous work [15]. Briefly, we required that each site have a minimum quality/depth ratio of 2, a maximum Fisher’s strand value of 60, a minimum nominal genotype quality of 20 and a maximum number of reads with mapping quality zero at a putatively variant site of 5. For analyses of within host individual variation for fixation index (Fst) calculations, we also obtained a multiple pileup file using samtools and filtered sites that were not retained after applying these filters. See Section 1 of S1 Text for an estimate of the consensus genotyping error rate.

Within-host diversity analysis.

We called within-host SNP and indel variants for endosymbionts and mitochondria using the method from [87]. Briefly, we created mpileup files from BWA bam alignment files for all individuals from each host species using SAMTools [84]. Then, we called variants and calculated pairwise diversity using the perl script from [87], which only considers sites within one standard deviation of the average genome coverage, filters SNPs around indels, and requires an alternate allele count in excess of the cumulative binomial probability of sequencing error at that site. As very closely related sister taxa have not been sampled for most of these bacterial genomes, ancestral/derived alleles could not be identified and we could not plot unfolded allele frequency spectra. Instead, folded allele frequency spectra were calculated for minor alleles and plotted in R.

No heteroplasmy was detected within the mitochondrial within-host populations (see S7 Table), suggesting that these bivalves do not experience double uniparental mitochondrial inheritance. This is important given our expectations regarding mitochondrial-symbiont co-divergence under strict vertical transmission.

Population genealogy inference.

We produced multiple fasta sequence files for each population for the host mitochondrial genomes and for the concatenated symbiont genomes from the set of filtered consensus genotype calls. We then used the phylogenetic software package RAxML [88] using a GTR+G model and 1,000 bootstrap replicates to estimate the phylogenetic relationships among samples and to quantify uncertainty in our phylogenetic relationships. Using FigTree, we rooted the trees by their midpoints and created cladograms for topological comparisons.

Analysis of polymorphic and recombinant sites.

Using the fasta files described above, we filtered sites to only retain biallelic SNPs with a minimum genotype quality of 10. Without indels, these resequenced genomes were already aligned. Then we used the aligned SNP data to calculate Waterson’s theta [89], pi [90], and the proportion of pairwise sites where all 4-gametes, i.e., all pairwise combinations of alleles, are represented. We then binned the 4-gamete sites by the distances between alleles, with bins at 1e1, 1e2, 1e3, 1e4, 1e5, and 1e6 bp, for model fitting (described below).

Whole genome structural alignment.

We generated whole genome alignment plots by first aligning bacterial genome assemblies with MUMmer 3.23 [91], using the nucmer algorithm and default parameters. In addition to the mb-scale genomes we assembled and referenced above, we obtained mb-scale genomes for Bathymodiolus septemdierum str. Myojin knoll ([42], accession GCA_001547755.1), Bathymodiolus thermophilus str. EPR9N ([92], accession GCF_003711265.1), and Vesicomyosicus okutanii ([12], accession NC_009465.1) from NCBI for alignment. The nucmer output was converted into the BTAB format with the MUMmer tool show-coords and was then visualized using a custom Python script. When necessary, some scaffolds in the bacterial assemblies were split into two parts in order to convert a circular genome into a linear alignment plot.

We used the whole genome aligner progressiveMauve [93] to compare genome synteny on the 10s of Kb scale between the S. velum and S. elarraichensis symbionts. First, we reordered the contigs comprising the S. elarraichensis symbiont draft genome [15] by the S. velum symbiont assembly [13] with the reorder contig function in progressiveMauve. Then, we aligned the S. velum symbiont genome pairwise against the S. elarraichensis symbiont’s reordered contigs and the S. pervernicosa symbiont genome we assembled. We plotted the alignment backbone files in the R package genoPlotR [94].

Mobile element analysis.

We identified mobile elements in the endosymbiont genome sequences by BLAST. First, we generated BLAST database files with the makeblastdb command from the ACLAME [95] and ICEberg [96] nucleotide and amino acid databases of transposable, viral, and conjugative elements. Next, we used blastp and blastn to compare endosymbiont amino acid sequences and full genome sequences, respectively, to these databases (cutoff values: minimum alignment length of 50 nucleotides or 50% amino acid query coverage, 90% identity, and e-value 1e-6). Overlapping hits were consolidated into a single mobile element-containing region (S12 Table).

Ortholog identification.

We identified putative orthologous sequences among sets of bacterial genomes by a reciprocal best BLAST approach. To do this, we first performed pairwise blasts between each pair of genomes’ coding sequences with blastn (-best_hit_overhang 0.1 -best_hit_score_edge 0.1 -evalue 1e-6), alternating each sequence as the query/subject. We parsed these results to only retain the best hits with >50% identity and >100 bp alignment lengths. Then, using a custom perl script, we compared hits between all pairs to identify genes with identical reciprocal best hits among all taxa each homologous gene was detected in. We used the resulting matrix of these reciprocal best hits to extract the coding sequences for each ortholog for each species from the genome fasta files for downstream analysis.

dN/dS analysis.

To evaluate the impact of homologous recombination on patterns of natural selection at the molecular level over long periods of time we computed the average fixation rate among endosymbiont lineages at nonsynonymous and synonymous sites (dN/dS). This ratio of values is an approximate measure of the strength of purifying selection under the assumption that most nonsynonymous substitutions are deleterious. For all genes where a single ortholog was found to be shared among all symbiont lineages we began by producing codon aware alignments using MASCE [97]. Then, we compared each orthologous alignment for pairs of symbiont lineages within each group (solemyid, vesicomyid, and bathymodiolin) to estimate dN and dS using the codeml package in the PAML v4.9 framework [98]. We excluded all comparisons for which dS < 0.05 or dS > 2, as values that exceed this range are often thought to yield unreliable estimates of rates of molecular evolution due to low statistical power and saturated substitutions, respectively. We then compared the distributions of dN/dS for each symbiont group comparison using a Wilcoxon test.

Taxon selection.

To construct dated phylogenies for hosts and symbionts, we downloaded related genomes from NCBI. For the host divergence analysis, all of the bivalve mitochondrial genomes available as of early 2020 and four gastropod mitochondrial genomes were downloaded to serve as ingroups and outgroups, respectively (35 total taxa: 31 bivalves and four gastropods; see S4 Table). For the symbiont divergence analysis, bacterial genomes were identified for inclusion in the analysis by BLAST [75]. While residing in a relatively constrained clade of proteobacteria, these chemosynthetic symbionts do not form a monophyletic clade, have free-living relatives, are basal to more derived groups in Gammaproteobacteria, and are currently taxonomically unclassified, so it was necessary to fish out related genomes by identifying sequence homology. To do this, we aligned the nucleotide coding sequences from each one of the seven symbiont genomes we sequenced and/or analyzed against NCBI’s Prokaryotic RefSeq Genomes database with blastn (-best_hit_overhang 0.1 -best_hit_score_edge 0.1 -evalue 1e-6). Based upon the diversity of hits across symbionts and genomes, we selected the top three best hits to each gene as taxa to include in the full genome divergence analysis (59 total taxa: 58 gammaproteobacteria and one alphaproteobacterium outgroup; see S5 Table).

Multiple sequence alignment.

As bivalve mitochondrial genomes exhibit notoriously diverse structural arrangements [99], we used the whole rearrangement-aware genome aligner progressiveMauve [93] to align the molluscan mitochondrial genomes. These alignments were manually inspected and converted to fasta format in Geneious Prime (version 11.0.6+10) [100].

We identified and aligned orthologous proteins among these diverse bacterial genomes with bcgTree [101], then we back-translated the resulting amino acid alignments to nucleic acids with RevTrans [102]. As bcgTree only includes protein-coding genes, we manually extracted and aligned 16S and 23S ribosomal RNA sequences for these taxa with Mafft (using the accurate mafft-linsi setting) [103]. Although recombination clearly occurs frequently in the solemyid and mytilid symbiont populations, we decided against removing recombinant sites because doing so may exacerbate recombination-induced artifacts [104]. Finally, we concatenated these alignments with the nucleotide alignments from bgcTree/RevTrans with a custom perl script and inspected them in Geneious Prime.

Phylogenetic inference and divergence dating.

We first inferred maximum likelihood phylogenies with RAxML (version 8.2.1, with parameters: f a -m GTRGAMMA -N 1000) [88] to verify that the taxa selected were able to resolve the relationships among hosts and among symbionts and free-living bacteria. Both mitochondrial and symbiont phylogenies were well-resolved (S2 Fig).

We inferred Bayesian phylogenies and dated node divergences for host mitochondria in Beast2 [105]. After several rounds of parameter testing to ascertain the speciation model and calibration date distribution that best fit the data (see Section 1 of S1 Text), we selected the Yule model of speciation, with a gamma distributed Hasegawa, Kishino, and Yano (HKY) model of substitution and a relaxed local molecular clock, and we calibrated dates to the base of the ingroup, Bivalvia. We used the fossil-based minimum appearance date for bivalves of 520 million years (first appearance estimated in Fossilworks [106] from fossil data in [107,108]). MCMC chains were run in duplicate until posterior probability convergence, around 8e8 steps for mitochondria. We also performed independent divergence date estimations in RelTime [109,110] and PATHd8 [111] to compare to the Beast2 results (see Section 1 of S1 Text and S6 Table).

Given the consistency in RelTime and Beast2 estimates for mitochondria (S6 Table) and the unreasonably long run times necessary (several weeks) for symbiont dated phylogenies to reach posterior convergence in Beast2, we estimated symbiont divergence times in RelTime. We used the previously estimated divergence date for Gammaproteobacteria of 1.89 billion years (based on calibration to the cyanobacterial-caused atmospheric oxygenation event [112]) with a log-normal distribution to calibrate a relaxed local clock, using a gamma-distributed Tamura-Nei model of substitution, and allowing for invariant sites. All trees were plotted in FigTree.

Simulation setup.

To simultaneously estimate the rates of horizontal transmission, effective homologous recombination rates, and the recombinant tract length, we used an approximate Bayesian computation approach. Here, we define the population-scaled per-base pair recombination rate, rho, to be equal to two times the effective population size times the per-base pair rate of gene conversion (rho = 2*N_er). The recombination tract length, l, is defined as the length of the gene conversion segment in base pairs. We used the bacterial sequential Markovian coalescent simulation framework, FastSimBac [113], to simulate neutral coalescence across a range of input parameters (see Section 2 of S1 Text for model proof). We drew the effective mutation rate from a log-uniform (3e-5,1e-2) distribution, the effective recombination rate from a log-uniform (1e-6,1e-2) distribution and the recombinant tract length from a uniform (1,1e5) distribution. Because the clonal frame cannot be inferred for the Bathymodiolus and Solemya symbiont populations, presumably due to the high levels of recombination relative to other bacteria, we did not supply the program with a fixed precomputed clonal frame for any simulations. In total we performed 100,000 simulations and we used subsets from each simulation to obtain summary statistics to train the variable sample size simulations.

Summary statistics.

We selected a set of summary statistics that each incorporate some feature of the overall diversity (relevant for theta), and the overall effective recombination rate per site (rho*l). Specifically, we computed two estimators of theta, Waterson’s theta [89], and pi [90], and we included the proportion of pairs of non-singleton SNPs at various genomic distances where all four possible combinations of alleles are observed in our sample. We placed the divisions between distance bins for pairwise comparisons of sites at 1e1, 1e2, 1e3, 1e4, 1e5, 1e6 base pairs.

Model fitting via random forest regression.

We use the scikit-learn package to perform random forest regression to obtain estimates of each parameter for each endosymbiont population using the 4-gamete sites identified above and a custom Python script available on Github. First, we confirmed that our summary statistics are sufficient to accurately fit our desired population parameters using out-of-bag score during model training and for each population sample size (S9 Table). Although the score is often slightly lower for smaller sample sizes, this approach performs sufficiently well and consistently across samples for our applications here. We additionally obtained confidence intervals for each parameter estimate using the forestsci package [114]. It should be noted that our simulations assume an equilibrium population. If this assumption is violated in a subset of the taxa that we examined, it might affect our parameter estimates. Nonetheless, it is unlikely that the large-scale differences in estimated parameters that we observe among groups, which are consistent with prior expectations, are entirely attributable to this potential bias.

Method validation on existing datasets.

In light of the relatively high recombination rates that we estimate, it is valuable to confirm that our approach for estimating rho and theta performs as expected. We therefore applied our method to the Bacillus cereus dataset [115] that has been studied in similar contexts using several related approaches [51,113,115]. In prior work with this dataset, estimates of the total impact of recombination, rho*l/theta, have varied somewhat, from approximately 3.7 [113] to 35.9 [115] and 229 [51]. Using our method, we obtain a value intermediate to these at 17.1, which indicates a moderate impact of recombination on genome evolution. This result suggests that our method is reliable and does not substantially inflate recombination rate estimates.

Read-backed phasing to confirm recombination estimates.

Because we used consensus symbiont genomes during model fitting, it could be possible that the high estimated rates of recombination in solemyid and bathymodiolin samples are an artefact of differential haplotype coverage across the genomes of genetically diverse symbiont chromosomes. We therefore sought to confirm our recombination rate estimates via comparing the rate of occurrence of all four possible configurations of two proximal alleles using read-backed phasing. Because each read-pair must ultimately derive from a single DNA fragment, when two alleles are observed on the same read or read pair, in the absence of errors, they must reflect an allelic combination that’s observed within a single bacterial chromosome.

We therefore analyzed in aggregate all reads that overlapped two or more consensus alleles for each population and computed the fraction of all four possible sampling configurations. More specifically, we computed the proportion of reads sampled from across all reads in all individuals that contained alleles, AB, Ab, aB, and ab, for two adjacent biallelic sites with alleles A/a and B/b. To reduce the impact of sequencing errors, we recorded a site as containing all four possible configurations when all configurations were present at proportion greater than 0.05 in the total set of read pairs. We further limited comparisons to alleles where we could obtain at least 100 observations of both sites on single read pairs. We excluded the populations of vesicomyid endosymbionts from this analysis because too few polymorphic sites were present within the distances spanned by individual read pairs to confidently infer the frequencies of 4D sites. However, because samples from these populations contain virtually no within-host variation, we have little reason to doubt the accuracy of the consensus genotypes resulting from differential coverage of genetically diverse bacterial haplotypes.

Because we are sampling a much larger number of lineages than we did in analyzing the consensus genome sequences, we would expect if anything is different that we observe higher rates of sites where all four possible allele configurations are present. This is precisely what we found. Specifically, we observe higher proportions of pairs of sites where all four possible allelic combinations are represented at each distance considered and in each population in the read-backed dataset than in the consensus chromosomes (S10 Table). Furthermore, because we have placed conservative cutoffs on the proportions of read pairs required to consider a pair of sites as containing all four alleles, these values are likely underestimates of the true rates. Our observed rates of four-gamete test failures is therefore consistent with our analyses of consensus genome sequences and confirms that recombination must be common within these endosymbiont populations.