Positive selection at core genes may underlie niche adaptation in Fusobacterium animalis

Bacterial genome sequences were retrieved from the National Center for Biotechnology Information (NCBI) database. The list of genomes analyzed in this study and their accession ID is available as Table S1. Genome sequences and their annotations were used as input to run the PPanGGOLiN tool with default settings [17]. PPanGGOLiN generates pangenomes by using a graphical model and clusters proteins in gene families. We then identified which gene families included 32 Fa core genes that are highly expressed in CRC, as defined by [14], and we checked the presence/absence of these genes in the genome of the 161 Fusobacteria strains (Table S2).

The phylogenetic tree was constructed using 120 core genes identified by the Genome Taxonomy Database [18] as previously described [9], using IQ-TREE v1.6.12 [19]. The substitution model (GTR + F + I + G4) was selected by the ModelFinder tool implemented in IQ-TREE.

Gene alignments were generated using the GUIDANCE2 suite [20], setting sequence type as codons and using MAFFT [21] as an aligner. GUIDANCE2 also allows the filtering of unreliably aligned positions; we removed codons with a score lower than 0.90 [22].

To account for recombination, we applied the 3SEQ software (v.1.7) [23], which tests all sequence triplets in a given alignment, scanning for mosaic recombination signals. The result is the identification of mosaic regions in which one of the three sequences is the recombinant (child) of the other two (parental). A total of 12 genes were found to be recombinant (Table S3) and the mosaic regions were masked before running the positive selection analyses.

The single-likelihood ancestor counting (SLAC) method was applied to calculate the average nonsynonymous substitution (dN)/synonymous (dS) substitution rate ratio [24].

Sites that show statistical evidence of purifying selection were identified using FUBAR (Fast Unconstrained Bayesian AppRoximation) [25].

To detect the action of positive selection, a codeml analysis, from the PAML (Phylogenetic Analysis by Maximum Likelihood) package, was run. Specifically, site model M8 (positive selection model) that allows a class of sites to evolve with dN/dS > 1 was compared to a model (M7, neutral model) that does not allow dN/dS > 1. To assess statistical significance, twice the difference of the likelihood (ΔlnL) for the models was compared to a χ2 distribution (2 degrees of freedom) [26‐28]. Genes that were significant after FDR correction for all genes analyzed (p value < 0.05) were also tested by comparing the M8 model with another more conservative neutral model (M8a), which allows dN/dS ≤ 1. In this case, twice the difference of the likelihood for the models was compared to a χ2 distribution with 1 degree of freedom (p value < 0.05) [26, 27]. In order to identify specific sites subject to positive selection, we applied four different methods: (1) the Bayes Empirical Bayes (BEB) analysis (with a posterior probability cutoff of 0.90), which calculates the posterior probability that each codon is from the positive selection site class (under M8 model) [29, 30]; (2) FUBAR (with a posterior probability cutoff of 0.90), an approximate hierarchical Bayesian method that generates an unconstrained distribution of selection parameters to estimate the posterior probability of positive diversifying selection at each site in a given alignment [25]; (3) Mixed Effects Model of Evolution (MEME) (with a p value cutoff < 0.1), which allows the distribution of dN/dS to vary from site to site and from branch to branch at a site [31]; (4) Fixed Effects Likelihood (FEL) (with a p value cutoff < 0.1), a maximum-likelihood (ML) approach to infer dN/dS on a per-site basis, assuming that the selection pressure for each site is constant along the entire phylogeny [24]. Again, to be conservative and to limit false positives, only sites detected using at least two methods were considered as positive selection targets.

FUBAR, MEME, FEL, and SLAC analyses were run locally through the HyPhy suite V2.5.29 [32, 33].

Selective events that accompanied the evolution of Fa were investigated with gammaMap, which uses intra-species variation and inter-species diversity to estimate the distribution of selection coefficient (γ) [34]. The sequences of the 32 genes from fifty-five available Fa genomes, as well as their corresponding reconstructed outgroup sequences, were analyzed. In order to reconstruct the nucleotide sequences of the last common ancestor (LCA), we applied the GRASP (Graphical representation of ancestral sequence predictions) tool, which reconstructs the ancestral sequences at internal nodes of a phylogeny [35]. We used all Fa sequences plus the sequences of F. watanabei, F. paranimalis and F. vincentii. We retrieved the reconstructed sequence from the internal node that separates F. paranimalis from all the Fa sequences.

gammaMap requires configuring prior distributions for some of the parameters. Thus, we selected weakly informative distributions, meaning improper log-uniform distributions for the transition/transversion ratio (k), branch length (T), and for the neutral mutation rate per site (θ) parameter. A uniform distribution was chosen for the probability of adjacent codons to share the same selection coefficient (p). The frequency distribution of non-stop codons was calculated by merging all the 32 genes. To check the effect of prior selection, a second run was performed by using log normal distributions for k, T, and θ. Runs of 100,000 iterations were performed with a thinning interval of ten iterations and a burn-in of 10,000.

gammaMap categorizes selection coefficient into twelve predefined classes: strongly beneficial (100, 50), moderately beneficial (10, 5), weakly beneficial (1), neutral (0), weakly deleterious (− 1), moderately deleterious (− 5, − 10), strongly deleterious (− 50, − 100), and inviable (− 500) [34]. Specifically, the program assigns to each codon a posterior probability for each selection coefficient. Because it is often difficult to infer the relative frequency of similar selection coefficients, individual codons are rarely assigned to one selection class with high reliability (i.e. with high posterior probability). This issue can be overcome by grouping coefficients into larger classes. For instance, for the analysis of Fa sequences, we were interested in identifying changes at codons that conferred a selective advantage. We thus called sites as positively selected if they showed a posterior probability > 0.75 of γ ≥ 1.

We used AlphaFold3 [36], through the AlphaFold Server (https://alphafoldserver.com/), to model the structure of human USP7 in complex to the linear motif USP7_MACH of Fa DnaK, and Fa Prp in complex with the N-terminal extension of Fa L27 (LFNIQLFAHKK). ipTM + pTM scores ≥ 0.9 were used as a confidence metric, as previously suggested [37]. The model of USP7 in complex with the DnaK ELM was superposed to the solved structure of human USP7 in complex with the same motif found in human p53 (PDB IDs: 2foj) and MDM2 (PDB ID: 2fop). For Prp, the model was superposed to that of S. aureus in complex with a product peptide (PDB ID: 7kld).

Homology modeling was performed through the SWISS-MODEL server [38]. Template files were derived from the Protein Data Bank (PDB) (7vmx.1.B, 7vmx.1.A, 6riq.1.C) and used for the modelling of Ef-Tu, Ef-Ts, and MinD, respectively. For Fa ClpX we used the Alphafold DB model of CLPX_FUSNN (gene: clpX, organism: Fusobacterium nucleatum subsp nucleatum) as template, in order to also model the N-terminus. GMQE and QMEANDisCo global scores were used for quality estimation.

3D structures were rendered using PyMOL (The PyMOL Molecular Graphics System, Version 1.8.4.0; Schrödinger, LLC).

Positively selected sites in Fa Prp were analyzed for their effects on binding affinity to L27 using the MutaBind2 tool [39], by comparing amino acid states in Fa Prp with the ones in F. paranimalis. The effect on binding was evaluated in terms of ΔΔG_bind values.

Structural disorder was inferred using the Metapredict tool [40, 41], which applies a deep-learning algorithm based on a consensus score calculated from eight different disorder predictors. Metapredict V2 was run using default parameters.

ELMs were identified using the scan tool of the Eukaryotic Linear Motif resource [42], restricting the search to Homo sapiens and using the default motif probability cutoff of 100.

The study gene set comprised 32 genes that were found to be highly expressed in Fa positive tumors (as defined in a previous study with the exclusion of three genes encoding hypothetical proteins) [14] (Table S3).

We first sought to define the distribution of these genes across Fa strains and in Fusobacteria that are closely related to Fa. To this aim, we used PPanGGOLiN [17] to query the genomes of 161 Fusobacteria (Fig. 1A, Table S1 and S2). We found that all genes are present in most genomes. The most notable exception was mliC that seems to be absent in the F. periodonticum/pseudoperiodonticum lineage.

Bild vergrößern

The widespread distribution of these genes across Fusobacteria allows the investigation of their evolutionary patterns. We thus retrieved sequence information for the reference genomes of the Fusobacteria species most closely related to Fa (as reported in Fig. 1A). Only in a minority of cases some orthologous genes were not present in all reference genomes (see Table S3). Using alignments of the genes and after accounting for recombination (see Methods), we first calculated the average non-synonymous substitution/synonymous substitution rate ratio (dN/dS) using the single-likelihood ancestor counting (SLAC) method [24]. The dN/dS metric is commonly used to measure the selective pressure acting on coding sequences: dN/dS < 1 indicates purifying selection, dN/dS around 1 is indicative of neutrality, and dN/dS > 1 reflects positive diversifying selection [43]. For all the studied genes, dN/dS was much lower than 1, indicating a major role for purifying selection in shaping their genetic diversity (Fig. 1B). To gain more detailed insight into the strength of selection, we used FUBAR (Fast Unconstrained Bayesian AppRoximation) to calculate the fraction of sites that show statistical evidence of purifying selection [25]. We obtained variable fractions among genes, ranging from ~ 10% to more than 80% (Fig. 1B). As expected, this measure showed an inverse relationship with dN/dS and confirmed that most of these genes are evolutionarily constrained, possibly as a result of their often-essential functions.

A major effect of purifying selection is not incompatible with positive selection acting on specific sites or domains [43]. To test for evidence of positive selection, we applied the likelihood ratio tests (LRTs) implemented in the codeml program [27]. Specifically, we compared models that allow (NSsite model M8, positive selection model) or disallow (NSsite models M7, null model) a class of codons to evolve with dN/dS > 1. After false discovery rate (FDR) correction for multiple tests, nine genes showed a significant LRT (p value < 0.05), with a better fit of model M8 vs M7 (Table S3). These genes were further tested with a more conservative M8 vs M8a model comparison. Five of them retained a significant LRT (p value < 0.05) and were considered targets of positive selection (Table 1, genes marked in red in Fig. 1B). MliC, which was the gene with the highest dN/dS, was not among them. Together with the comparatively lower fraction of sites targeted by purifying selection, this suggests that this gene experiences a weaker constraint compared to the others we analyzed (Fig. 1B).

The positively selected genes have different functions. Inspection of the Virulence factor database (VFDB, https://www.mgc.ac.cn/VFs) [44] indicated that the protein products of two of these genes (elongation factor Tu, Ef-Tu, encoded by tuf, and septum site-determining protein MinD) are known virulence factors, whereas amino acid replacements in Ef-Tu and the RNA polymerase subunit beta'(encoded by rpoC) are associated with antibiotic resistance according to the Comprehensive Antibiotic Resistance Database (CARD Comprehensive Antibiotic Resistance Database, https://card.mcmaster.ca/) [45].

To identify specific codons subject to positive selection, we applied four methods: the Bayes Empirical Bayes (BEB) analysis from model M8, FUBAR, FEL (fixed effects likelihood) and MEME (Mixed Effects Model of Evolution) [24, 25, 30, 31]. To be conservative, only sites detected using at least two methods were considered targets of positive selection (Table S3). We detected few selected sites, indicating that positive selection is highly localized and indeed it occurs in a background of very strong constraint for some of these genes (e.g., rpoC, clpX, and minD) (Table 1; Fig. 1B).

We next used structural modelling to map the positively selected sites onto the 3D protein structures, and searched the literature for possible functional data. Ef-Tu and Ef-Ts function as translation elongation factors catalyzing the binding of aminoacyl-tRNA to the A-site of the ribosome. However, in many bacterial species, Ef-Tu has evolved moonlighting functions and the protein can either be exposed at the cell surface or secreted. It contributes to adhesion and invasion of host cells, modulation of immune responses, and binding of host secreted molecules [46]. For instance, in streptococci, secreted Ef-Tu promotes cell-adhesion, biofilm development, and periodontitis onset, whereas the surface exposed Ef-Tu of Lactobacillus johnsonii mediates attachment to intestinal epithelial cells and can induce a proinflammatory responses [47, 48]. Mapping of the positively selected sites on the 3D structure of Fa Ef-Tu indicated that all of them are exposed at the surface and may thus mediate attachment or interaction with host factors. Indeed, one of them (H300) is located in the barrel-like adhesion domain (Fig. 1C) [48]. Conversely, Ef-Ts has no known moonlighting functions. In this protein, the single positively selected site is in the coiled-coil domain (Fig. 1C). Interestingly, mutations in this region have been shown to confer resistance against contact-dependent antibacterial tRNase toxins produced by enterohemorrhagic Escherichia coli [49]. Thus, the positively selected site might represent a signature of inter-bacterial conflict.

MinD, with its partners MinC and MinE, are required for the correct placement of the division site and act to spatially regulate cell division [50]. In Neisseria gonorrhoeae and E. coli, min mutants have altered cell shape and size, as well as decreased virulence [51‐54]. The positively selected sites in minD are not predicted to be located at the homodimer interface, nor in the MinC/MinE binding region (Fig. 1D). As in the case of Ef-Tu the sites are surface-exposed and may mediate interaction with other proteins or components. In this respect, it is interesting to note that F. watanabei and Fa, which are closely related but carry different residues at the positively selected MinD sites (Fig. 1D), were shown to differ in cell length and width [10, 15]. This is relevant because bacterial cell shape and size have been shown to modulate colonization of distinct host niches, the susceptibility to host defenses, and, for many pathogens, disease progression [55]. Experimental studies will however be necessary to determine whether the positively selected sites have a role in modulating Fa cell shape and pathogenicity.

We also detected one positively selected site in RpoC. This variant was not associated with resistance as determined by consulting CARD [45]. However, multiple evidence suggests that changes within housekeeping bacterial genes can have antibiotic-independent adaptive effects [56]. In E. coli, mutations in rpoC have been demonstrated to arise in response to various selective pressures, including prolonged resource exhaustion, growth in minimal media or at high temperatures or under acidic conditions, as well as changes in nutrient sources [57]. Such mutations are often antagonistically pleiotropic—for instance they are adaptive under resource exhaustion but reduce exponential growth rates in full media [58]. Thus, an interesting possibility is that the positively selected site in rpoC contributed to the adaptation of Fusobacteria to different niches within the host [9]. Whether it also evolved to facilitate intracellular survival remains to be evaluated.

Finally, we detected a positively selected site in ClpX, which is located on a flexible loop that joins the zinc coordinating and ATP binding domains of the protease (Fig. 1E) (see below).

We next aimed to assess whether adaptive changes emerged specifically on the Fa lineage. To this aim, we retrieved 55 complete or almost-complete Fa genomes and extracted sequence information for the 32 genes under study (Tables S1 and S2). As the outgroup, we reconstructed the sequences of the ancestor of Fa and F. paranimalis (see methods). We next used the gammaMap program [34], which jointly uses intra-species variation and inter-species diversity, to estimate the distribution of fitness effects (i.e., selection coefficients, γ). In practical terms, γ values can be considered a measure of the fitness consequences of new non-synonymous mutations. The method categorizes selection coefficients into twelve predefined classes ranging from − 500 (inviable) to 100 (strongly beneficial). As expected, and in line with the results above, we observed a preponderance of codons evolving under strong purifying selection (− 500 ≤ γ ≤  − 10). The median γ value for most of the genes was lower than − 50, with only a few exceptions, and the highest median (γ = − 1) was observed for prp (Fig. 2A). Thus, these genes also experience intense constraint in the circulating Fa population.

Bild vergrößern

To determine whether any amino acid changes conferred a fitness advantage to Fa, we estimated codon-wise posterior probabilities for each selection coefficient. We called a codon as positively selected if its cumulative posterior probability of γ ≥ 1 was > 0.75. A total of thirty-three sites in seven genes were found to be positively selected (Fig. 2B). For most such sites, the selected derived allele is fixed or at high frequency in the Fa population (Fig. 2B). The largest numbers of positively selected sites were detected in Prp and DnaK (Fig. 2B). As above, we integrated structural information and literature searches, but also motif analysis, to mine possible effects of the positively selected sites.

Analysis of beneficial changes in ClpX indicated that one of the sites corresponds to the one that was targeted by positive selection in the extended Fusobacterial phylogeny and the other is flanking (Fig. 1E). Thus, the same loop in ClpX has been a target of selection over different time-frames—i.e. during the emergence of the Fusobacterial species and more recently, as the Fa population diverged from the common ancestor with closely related Fusobacteria. ClpX is the ATP-dependent component of the Clp protease, which determines substrate specificity [59]. In many bacteria, the protease controls virulence, stress tolerance, and biofilm formation [59, 60]. More recently, ClpX was also reported to have a role in phage biology, as it directly interacts with the phage repressor CI and performs an essential role in prophage induction by abolishing the ability of CI to repress lytic phage genes [61]. We therefore speculate that the positively selected sites in ClpX modulate the specificity of the protease towards endogenous or phage proteins. Unlike the observations for ClpX, the beneficial changes in Ef-Ts do not map to the coiled-coil region, nor to the surface involved in the interaction with Ef-Tu (Fig. 1C).

Next, we moved to the analysis of selected sites in the two major targets, Prp and DnaK. Prp is essential for the assembly and maturation of the ribosome, as it removes an N-terminal extension from a precursor of ribosomal protein L27. The protein acts as a dimer. We thus modeled the structure of Fa Prp in complex with the N-terminal extension of Fa L27 using AlphaFold3 [36] (Fig. 2C). The generated complex was then superimposed to the crystal structure of the orthologous protein from Staphylococcus aureus, which was crystallized in complex with a peptide substrate (Fig. 2C) [62], showing a perfect overlap between complexes. Mapping of the positively selected sites indicated that three of them project into the binding pocket with a distance of less than 3 Å from the L27 fragment. An additional selected site is at the homodimer interface (Fig. 2C). To evaluate the effect of the positive sites on binding efficiency, we mutated in silico these sites by recapitulating the amino acid states found in F. paranimalis. We then estimated changes in binding affinity (Table S4): as expected the F. paranimalis mutations were not predicted to be deleterious, but they diminished binding affinity (ΔΔG_bind > 0) (Table S4). Overall, this analysis suggests that Fa Prp may bind the L27 N-terminal tail more efficiently than the protein encoded by F. paranimalis. Whether these changes also modulate the cleavage efficiency of L27 remains to be evaluated. Furthermore, an interesting possibility is that the positively selected sites modulate the specificity of the protease for other substrates. Recent evidence in Streptococcus mutans indicates that a Prp homolog can cleave cellular enzymes, eventually modulating carbohydrate metabolism, biofilm formation, and interaction with other microorganisms (e.g., Candida albicans) [63].

The most interesting target of selection identified in this study was DnaK, not only because it accumulated the largest number of beneficial changes, but also for its role in tumor biology. Indeed, the DnaK chaperone encoded by Mycoplasma fermentans has been demonstrated to promote malignant transformation in vivo and in vitro [64‐66], and both the M. fermentans and F. nucleatum DnaK proteins reduce the effectiveness of anticancer drugs such as cisplatin and 5-Fluorouracil [67]. Experiments with M. fermentans showed that DnaK can be internalized by uninfected human cells and that the protein localizes to several cellular compartments, including the cytoplasm, perinuclear membrane, and nucleus [66]. The DnaK protein of F. nucleatum was found to be secreted [68], suggesting that, irrespective of whether intracellular infection occurs, the protein may gain access to human cells. In the M. fermentans model, DnaK was shown to bind several host proteins, including human USP10 (ubiquitin-specific protease 10), thus reducing p53 stability and anti-cancer functions [66, 69].

We identified 17 beneficial changes in Fa DnaK. With the exclusion of 18 V, all of them are located in the C-terminal region (Fig. 3A). Because the terminal tail of some DnaK proteins was shown to be intrinsically disordered (i.e., to lack a fixed 3D structure) [70, 71], we used Metapredict V2, a deep-learning-based approach that combines different predictors to generate consensus disorder scores, to test whether this was also the case for the fusobacterial proteins. Results indicated that the ~ 30 C-terminal residues of DnaK from Fa, F. paranimalis, F. watanabei, F. vincentii and the LCA are unstructured (Fig. 3B). In eukaryotes, intrinsically disordered protein regions can accommodate eukaryotic linear motifs (ELMs), short peptide modules that are important for protein regulation, often mediating protein–protein interactions or acting as sites for posttranslational modifications [72‐74]. ELMs are often mimicked by viral and bacterial pathogens to hijack host cell regulatory networks [75, 76]. Indeed, these elements of host molecular mimicry have been studied on a large scale in F. nucleatum using a computational strategy. This enabled the identification of a network of protein interactions between the putative F. nucleatum secretome and human proteins [77]. We thus used an ELM prediction tool to investigate the presence of linear motifs in the disordered C-terminal tails of the fusobacterial DnaK proteins [42]. No ELM was found in the proteins encoded by F. paranimalis, F. watanabei, and F. vincentii, nor in the LCA sequence. Conversely, five ELMs were detected in the Fa DnaK sequence (Fig. 3B; Table 2). Two of them are N-glycosylation sites, which are unlikely to be functional because, in eukaryotic cells, the N-glycan precursor is biosynthesized in the endoplasmic reticulum. Another one is a casein kinase 2 (CK2) phosphorylation motif (Fig. 3B; Table 2). CK2 is a constitutively active kinase frequently over-expressed in cancer cells, where it has multiple roles, including proliferation, cell cycle control, and protection from apoptosis [78]. Notably, in CRC, CK2 was suggested to regulate the epithelial-mesenchymal transition (EMT) and Fa is particularly associated with tumors of the mesenchymal subtype [14, 79]. Although these observations do not necessarily imply causation, they indicate that the high expression of CK2 in malignant cells may indeed result in the phosphorylation of the disordered tail of DnaK, a modification that is known to modulate the functional properties of unstructured regions [80].

Bild vergrößern

The two final ELMs we detected are involved in post-translational modifications, namely ubiquitination and SUMOylation (Fig. 3B; Table 2). The latter is a dynamic, reversible process involving the covalent modification of specific lysine residues on target proteins with SUMO (small ubiquitin-related modifier). SUMO conjugation is an important regulatory mechanism for protein stability, subcellular localization, and protein–protein interactions [81]. Interestingly, SUMOylation (and ubiquitination) have been shown to regulate the nuclear import of specific proteins [82]. For instance, SUMO modification promotes the nuclear import of polo-like kinase and suppresses its ubiquitin-mediated proteasomal degradation [83]. Likewise, caspase-8 SUMOylation is associated with nuclear localization [84]. It is thus possible that SUMOylation of DnaK allows its access to the nuclear compartment. It is also worth noting that ubiquitin-specific protease 7 (USP7) is predicted to target a motif flanking the SUMOylation site. USP7 is a key regulator of the p53-MDM2 pathway and, as in the case of CK2, it is over-expressed in most cancers [85]. Thus, the pharmacological inhibition of USP7 is regarded as a promising therapeutic strategy [86]. Notably, USP7 functions as a SUMO deubiquitinase: by acting on SUMOylated proteins, USP7 counteracts their ubiquitination [87]. It is thus tempting to speculate that SUMOylation facilitates the nuclear import of Fa DnaK where USP7 promotes its stability. Given its potential biological relevance, we used AlphaFold3 to model the interaction between human USP7 and the cognate motif in DnaK. This confirmed that the Fa DnaK ELM peptide binds USP7 in the same region as the p53 and MDM2 motifs. We also observed that two sites found as positively selected (N588 and S590) contribute to this interaction (Fig. 3C).

Clearly, these analyses will require experimental validation. We acknowledge that the lack of in vitro/in vivo data to support our hypotheses represents a limitation of this study. Another limitation is the relatively small number of genes we analyzed. Such genes were selected due to their high expression in Fa positive tumors and their possible role in promoting intracellular survival [14]. It is however possible that additional genes modulate the intratumor survival of Fa and its oncogenic potential.