Beyond labels: current sequencing data do not support standardized porcine circovirus 3 (PCV3) genotypes

Abstract

Background

Porcine circovirus 3 (PCV3) has been detected worldwide in both healthy and diseased pigs, raising questions about its molecular diversity and potential genotype structure. Since its discovery, several genotyping schemes have been proposed; however, most of them rely on arbitrary criteria and often produce inconsistent or conflicting classifications.

Results

In this study, all high-quality PCV3 ORF2 sequences available at the GenBank to date were analyzed to evaluate whether consistent and reproducible genotyping criteria can be defined. Phylogenetic trees reconstructed using multiple algorithms and statistical comparisons of topological similarity revealed moderate-to-high incongruence among methods, limited phylogenetic signal, and a general decline in bootstrap support toward deeper nodes. Genetic distance distributions showed complete overlap between and within proposed genotypes, further questioning the biological meaning of such partitions.

Conclusions

Overall, these findings indicate that the current genetic landscape of PCV3 does not support a robust or standardized genotyping framework. Given the low genetic variability of PCV3 and the absence of experimental evidence for genotype-specific differences in antigenicity, epidemiology, or virulence, the definition of formal genotypes appears challenging and premature, and likely unnecessary at this stage.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12917-026-05504-w.

Background

Porcine circovirus 3 (PCV3) belongs to the family Circoviridae (phylum Cressdnaviricota, class Arfiviricetes, order Cirlivirales), which includes two genera, Circovirus and Cyclovirus [41]. As of 19/10/2025, 70 species are classified within the genus Circovirus according to the International Committee on Taxonomy of Viruses (ICTV) (https://ictv.global/taxonomy; accessed 19/10/2025). However, porcine circoviruses are by far the most extensively investigated. PCV3 was first detected by metagenomic approaches in diseased pigs in 2016 [26] and has since been reported worldwide across multiple countries and host species, in association with several clinical conditions as well as in healthy pigs [19, 25, 30]. This has fostered intensive diagnostic, surveillance, and research activities, also encouraged by potential analogies with the history of porcine circovirus 2 (PCV2). Despite its frequent detection in healthy animals, the accumulation of epidemiological, pathological, and clinical evidence has established a pathogenic role for PCV3 and led to the definition of PCV3-associated diseases (PCV3-AD), at least for the systemic (PCV3-SD) and reproductive (PCV3-RD) forms [30]. Interest in the relevance of PCV3 has been accompanied by the progressive accumulation of a substantial number of sequences sampled worldwide. A degree of genetic variability has been demonstrated, although it appears markedly lower than that of PCV2, consistent with the lower evolutionary rate estimated for PCV3 [10, 12, 14]. Nevertheless, as with PCV2, an early trend to classify PCV3 into genetic groups emerged. Numerous classification schemes to define genotypes have been proposed across studies, differing both in nomenclature and criteria used [1–3, 5, 10, 16, 27, 31, 35, 40]. In many cases, no explicit criteria were provided, and grouping relied on arbitrary cluster recognition by visual inspection of phylogenetic trees reconstructed from ad hoc subsets of sequences. While classifying strains into simplified categories can significantly aid studies of viral epidemiology and evolution, such schemes must rely on sound biological and methodological assumptions and be standardized and shared across the research community. Otherwise, misunderstanding, bias and incomparability are gathered, especially when nomenclature is adopted or interpreted by non-specialists unaware of the underlying data and model assumptions.

The present study aimed to evaluate whether the global structure of the PCV3 population, represented by all sequences available to date, supported the definition of consistent, standardized genotyping criteria suitable for molecular epidemiology.

Methods

Dataset preparation

All PCV3 ORF2 sequences were downloaded from GenBank, and their quality was assessed by verifying the absence of poorly aligned sequences, premature stop codons, and frameshift mutations. Sequence MG372488, formerly classified within the tentative PCV3b genotype [10], was excluded as it was the only highly divergent strain and has not been re-detected in recent years. Moreover, the only closely related sequence available at the time of the initial classification was later withdrawn, suggesting quality issues. The retained sequences were aligned at the codon level using the MUSCLE [6] algorithm implemented in MEGA X [21]. Among the remaining sequences, after pairwise genetic distance estimation, those exceeding the 99th percentile of divergence and lacking ≥ 98% nucleotide identity with at least five other sequences were removed, as such unique divergence was attributed to poor sequence quality. Manual inspection of the affected sequences typically revealed nucleotide patterns consistent with common Sanger sequencing artifacts. The resulting dataset was scanned for recombination using RDP5 [23]. Sequences identified as recombinant by at least three methods with a p-value < 10^–5 after Bonferroni correction were removed. Specifically, RDP, GENECONV, Chimaera, and 3Seq were used for the primary scan, while the full set of available methods was employed for analysis refinement. Absence of recombination occurrence was tested using GARD implemented in datamonkey [20, 42].

Phylogenetic analysis

The final sequence dataset was re-aligned and used to reconstruct phylogenetic trees based on different approaches and software:

• Neighbour joining (NJ) tree with uncorrected genetic distances using MEGAX (NJ p − distance) [21]. Reliability of inferred clades was assessed performing 1000 bootstrap replicates.

• Neighbour joining with corrected distances using MEGAX (NJ Corrected) [21]. The substitution model (TN93 + G + I) was selected based on the Bayesian information criterion calculated using the same software. Reliability of inferred clades was assessed performing 1000 bootstrap replicates.

• Maximum likelihood using the algorithm implemented in MEGA (ML-MEGA) [21]. The substitution model (GTR + G + I) was selected based on the Bayesian information criterion calculated using the same software. Reliability of inferred clades was assessed performing 1000 bootstrap replicates.

• Maximum likelihood using the algorithm implemented in IQTree v2.4.0 [24] selecting as substitution model the one (GTR + G + I) with the lowest BIC calculated using the same software. The reliability of inferred clades was assessed performing 10,000 ultrafast bootstrap replicates.

• Maximum likelihood using the algorithm implemented in RAxML v8.2.10 [37] u sing the GTR + CAT substitution model with 10,000 rapid bootstrap replicates.

Phylogenetic tree comparison

The distance among phylogenetic trees was assessed by calculating different metrics.

Since several topological distances can become computationally demanding for very large trees, we adopted a repeated subsampling scheme. A total of 20 subsets of 400 tree tips were randomly selected to ensure computation feasibility and allow the estimation of the effect of sampling on estimated statistics by calculating the relative standard deviation. To promote a reproducible and well-spread subset across the topology, we constructed a pectinate (ladderized) guide topology using PectinateTree function of the Quartet R package [29] and then pruned every tree in the forest to this same subset. Consequently, all metrics computed within a replicate are strictly comparable (identical taxon set), and replicate averages/variability provide a mean estimate and uncertainty (standard deviation, sd) for topological agreement. When computations were possible on the whole tree, no subsampling was performed.

Quartet distance analysis relies on decomposing trees into quartets (sets of four taxa). Given multiple trees with identical tip sets, the ManyToManyQuartetAgreement function tallies, for all tree pairs simultaneously, the numbers of quartet configurations that agree, conflict, or are uninformative. From these counts, the Quartet Divergence, a quartet-based topological distance (0 = identical topologies; 1 = maximally discordant with respect to the available quartets), was derived. Additionally, for each pair of trees and on the same subsets, the following metrics were also reported: 1) DoNotConflict: fraction of quartets not contradicting the shared relationships; 2) ExplicitlyAgree: fraction of quartets identical between the two trees; and 3) Marczewski–Steinhaus: also measuring the topological dissimilarity between two trees based on the proportion of discordant quartets, ranging from 0 (identical trees) to 1 (completely different).

As with many-to-many Quartet Distance, each metric was summarized by its replicate mean and standard deviation per tree pair, thus providing both a point estimate and uncertainty due to taxon choice.

To complement quartet-based measures with a split-based perspective, the normalized Robinson–Foulds distance and Steel–Penny index (maximum expected agreement) were computed with the Quartet [29] and phangorn [32] library on the same subsets used for the quartet calculations. Robinson–Foulds distance measures the topological difference between two phylogenetic trees by counting the number of splits (bipartitions) present in one tree but not in the other; lower values indicate higher similarity. Steel–Penny index quantifies how much two trees agree beyond what would be expected by chance, providing a normalized measure of topological similarity ranging from 0 (random agreement) to 1 (identical trees). The mean and standard deviation across replicates for each tree pair were calculated. Also, the topological dissimilarity using the Jaccard–Robinson–Foulds (JRF) distance, a “soft” variant of Robinson–Foulds distance that gives partial credit to near-matching clades, was assessed [36].

The Clustering Information Distance (CID), defined as the variation of information (in bits) between the sets of non-trivial clades of two trees on the same taxa, was calculated for topological dissimilarity using the TreeDist package in R [36]. It is 0 for identical topologies and increases as clades disagree; unlike Robinson–Foulds, it weights disagreements by clade size (i.e. misplacing one tip costs little; breaking a deep clade has a higher cost). A summary of principles, range, and rule of thumb herein applied for metrics interpretation is reported in Table 1.

Table 1.

Summary of the main statistics used in this study: principles, range, rules of thumb, and interpretation of results

Metric (normalized)	Range	Interpretation (rule-of-thumb)	Notes
Robinson–Foulds (RF)	0–1	< 0.1 ≈ nearly identical; 0.1–0.5 = moderate differences; > 0.5 = largely incongruent	Based on shared and unique bipartitions (splits). Sensitive to minor topological changes when many taxa are present
Jaccard–Robinson–Foulds (JRF)	0–1	< 0.1 ≈ very similar; 0.1–0.25 = moderate; > 0.4 = clearly different	Weighted version of RF that accounts for partial similarity between splits; reduces penalty for small topological rearrangements
Quartet distance	0–1	< 0.05 ≈ almost identical; 0.05–0.15 = local differences; 0.15–0.4 = moderate; > 0.4 = strongly incongruent	Compares all possible quartets of taxa; robust to branch-length differences
Clustering Information Distance (CID)	0–1	< 0.1 ≈ very similar; 0.1–0.5 = moderate differences; > 0.5 = strongly divergent structures	Entropy-based; measures how much information must be “exchanged” to make the trees equivalent
Marczewski–Steinhaus distance	0–1	< 0.1 ≈ minimal differences; 0.1–0.3 = moderate; > 0.4 = markedly different	Normalized distance based on differing quartet sets; mathematically related to quartet distance
Steel–Penny index (similarity)	0–1	> 0.9 ≈ nearly identical; 0.7–0.9 = good agreement; 0.5–0.7 = moderate; < 0.5 = low or random agreement	Measures the expected topological agreement between trees beyond chance. Higher values indicate greater congruence

Tree topology testing

To evaluate the fit of inferred trees to the data, the log-likelihood of all trees was estimated based on the provided alignment using IQTree. The RELL approximation was used to assess if some topologies had a significantly lower likelihood calculating several tree topology tests. More in detail, the bootstrap proportion (BP), the Kishino-Hasegawa test (KH) [18], the Shimodaira-Hasegawa (SH) test [34] as well as their weighted version (the weighted KH and weighted SH), the expected likelihood weights (ELW) [38] and approximately unbiased (AU) test [33], were calculated performing 10,000 RELL replicates.

Visual tree comparison

To provide a graphical representation, the IQTree topology was used to propose a tentative genotype partition based on clustering in the unrooted tree. Using subjective, tree-guided criteria, PCV3 sequences were assigned to three major clusters, whereas intermediate lineages were labelled “unclassified.” The same genotype labels were then mapped onto trees inferred with alternative software and methods to visually assess topological differences.

Results

Sequence dataset and genetic distance features

A total of 2,380 complete ORF2 sequences were included, originating from 28 countries and spanning 1996–2025, based on strains for which metadata were available (Supplementary table). A total of 139 (5.84%) of the included sequences originated from samples collected before the first description of PCV3 [26]. No evidence of recombination was detected under the selected parameters using either RDP5 or GARD. Only one sequence (OR757335) showed limited evidence of recombination when analysed with RDP5 and was therefore removed following a conservative approach. Sequences that were highly divergent relative to the overall genetic distance distribution (i.e., strains lacking close relatives) were excluded. Although original chromatograms were unavailable—preventing a direct assessment of primary sequence quality—the observed patterns in many cases suggested low quality, such as long homopolymeric runs, clusters of multiple mutations confined to short regions, and mismatches inconsistent with the overall alignment.

The distribution of pairwise genetic distances was continuous, with most strains showing distances below 4%. Summary statistics were: 5th percentile = 0.004, median = 0.020, 95th percentile = 0.0248, and maximum = 0.0651. The distribution was approximately normal, though slightly right-skewed (skewness = 0.575) and leptokurtic (excess kurtosis = 1.839) (Fig. 1).

Fig. 1.

Distribution of pairwise genetic distances based on ORF2 selected dataset. The normal distribution having the same mean and standard deviation has been plotted

Phylogenetic analysis

On preliminary visual inspection, the phylogenetic analysis revealed some major cluster separated by long branches. However, across tree-reconstruction methods, intermediate branches representing singletons or small clades were consistently present, and no clear threshold could be identified to delineate all major clades under a shared set of criteria (Fig. 2). Furthermore, irrespective of the method applied, even the major clades encompassed single sequences or small subclades that were clearly separated by long branches from the remaining sequences.

Fig. 2.

Comparative visualization of tree topologies reconstructed using different phylogenetic methods. The trees depict the relationships among PCV3 ORF2 sequences inferred by IQTree, RAxML, Maximum Likelihood with MEGA (ML-MEGA), Neighbor-Joining (NJ) with corrected distances, NJ with p-distance approaches. Each colour represents a tentative genotype or sequence cluster (Unclassified, A, B, and C) as initially defined on IQTree

At graphical inspection, trees reconstructed with different methods starting from the same dataset showed incongruences (Fig. 2). Such finding was confirmed by essentially all metrics of phylogenetic dissimilarity, which revealed moderate to high topological distance among trees, depending on the specific comparison (Table 2).

Table 2.

Summary of pairwise topological distances and quartet metrics. Means ± SD are reported; SD from subsampling (m = 400. B = 20) when applicable

Tree1	Tree2	RF	JRF	JRF (sd)	MS norm mean	MS Norm sd	CID mean	CID sd	QD mean	QD sd	DoNotConflict mean	DoNotConflict sd	Explicitly Agree mean	Explicitly Agree sd	Marczewski Steinhaus mean	Marczewski Steinhaus sd	Steel Penny mean	Steel Penny sd
ML MEGA	NJ Corrected	0.773	0.465	0.015	0.716	0.019	0.550	0.012	0.397	0.011	0.603	0.011	0.603	0.011	0.431	0.011	0.603	0.011
RAxML	ML MEGA	0.770	0.472	0.014	0.755	0.024	0.564	0.012	0.372	0.014	0.628	0.014	0.628	0.014	0.458	0.015	0.628	0.014
RAxML	NJ Corrected	0.765	0.439	0.011	0.739	0.030	0.540	0.011	0.472	0.007	0.528	0.007	0.528	0.007	0.358	0.007	0.528	0.007
ML MEGA	NJ p-distance	0.758	0.425	0.013	0.591	0.021	0.472	0.015	0.255	0.010	0.745	0.010	0.745	0.010	0.593	0.013	0.745	0.010
RAxML	NJ p-distance	0.708	0.417	0.014	0.689	0.025	0.513	0.012	0.414	0.010	0.586	0.010	0.586	0.010	0.414	0.010	0.586	0.010
IQtree	ML MEGA	0.691	0.477	0.010	0.670	0.015	0.582	0.011	0.302	0.015	0.698	0.015	0.698	0.015	0.537	0.018	0.698	0.015
IQtree	NJ Corrected	0.681	0.453	0.008	0.736	0.016	0.587	0.011	0.440	0.010	0.560	0.010	0.560	0.010	0.389	0.010	0.560	0.010
IQtree	NJ p-distance	0.635	0.450	0.011	0.635	0.015	0.552	0.013	0.318	0.018	0.682	0.018	0.682	0.018	0.517	0.021	0.682	0.018
IQtree	RAxML	0.627	0.424	0.018	0.601	0.026	0.502	0.015	0.316	0.012	0.684	0.012	0.684	0.012	0.519	0.014	0.684	0.012
NJ Corrected	NJ p-distance	0.624	0.330	0.012	0.534	0.027	0.411	0.015	0.305	0.011	0.695	0.011	0.695	0.011	0.532	0.013	0.695	0.011

Evaluation of tree log-likelihoods indicated that the RAxML topology was the one with the highest fit to the provided alignment. However, other alternatives—particularly the MEGA maximum-likelihood tree—were not statistically worse than the best tree; when differences were detected, the associated p-values were often > 0.01 (Table 3).

Table 3.

Topology likelihood comparison across trees. Reported are the log-likelihood (logL). difference from the best tree (ΔlogL). RELL bootstrap proportion (bp-RELL). and p-values for KH. SH. weighted KH (WKH). weighted SH (WSH). expected likelihood weight (c-ELW). and the approximately unbiased test (AU). Signs “ + ”/“–” are appended to each statistic to indicate, respectively, support comparable to the best-scoring topology (in bold) or significantly worse fit

TREE	LOGL	ΔLOGL	BP-RELL	P-KH	P-SH	P-WKH	P-WSH	C-ELW	P-AU
IQTree	−19,260.407	378.64	0.00136 -	0.00437 -	0.0108 -	0.00437 -	0.014 -	0.00136 -	0.00459 -
RaxML	−18,881.771	0	0.919 +	0.931 +	1 +	0.931 +	0.992 +	0.919 +	0.961 +
ML MEGA	−19,091.449	209.68	0.0669 +	0.0686 +	0.163 +	0.0686 +	0.17 +	0.0669 +	0.0875 +
NJ Corrected	−19,228.134	346.36	0.00205 -	0.00772 -	0.0143 -	0.00772 -	0.0263 -	0.00202 -	0.0102 -
NJ p-distance	−19,155.792	274.02	0.0108 -	0.0212 -	0.0543 +	0.0212 -	0.0674 +	0.0108 -	0.0224 -

Because of the easiness of use and apparently clearer topology, the ML IQTree generated tree was arbitrary selected to define major clades as tentative new genotypes (Fig. 2).

Pairwise genetic distance evaluation showed overlapping among intra and inter genotypes p-distances) (Fig. 3).

Fig. 3.

Density plot of the pairwise genetic distances calculated within and among tentatively defined genotypes based on IQTree (See also Fig. 1)

Phylogenetic analysis also confirmed the limited consistency of the proposed classification scheme. Although certain clades were preserved across different trees, several exceptions were observed. While many discrepancies involved clades with a small number of sequences, notable uncertainty—and corresponding differences in branching patterns—also affected larger and more numerous groups (Fig. 2). These findings are reflected in the often-limited bootstrap support associated with many nodes—particularly deeper ones, farther from the tips and closer to the tree root (Fig. 4).

Fig. 4.

Relationship between node support and distance from the tips across different phylogenetic reconstructions. Scatterplots represent the correlation between bootstrap support (%) and branch length–based distance from the tip for all analyzed trees generated using IQTREE, RAxML, ML-MEGA, Neighbor-Joining (NJ) with corrected distances, and NJ with uncorrected p-distances. The smoothed regression lines highlight the general trend toward decreasing bootstrap values in deeper nodes, indicating lower confidence in ancestral relationships across reconstruction methods. To minimize the impact of highly divergent terminal branches, the tip corresponding to the 95th percentile of root-to-tip distances was used as the starting point for the analysis

Similarly, likelihood mapping visualises the tree-likeness of all potential sequence quartets, estimated by IQTree, revealed approximately 20% of potential quartets to be unresolved (Supplementary figure).

Discussion

The growing availability of genetic data has greatly advanced our understanding of viral molecular epidemiology, enabling extremely fine levels of characterization and even the identification of single variants. However, such a high-resolution approach is often unnecessary and impractical for many studies, where a certain degree of simplification would provide greater applicability.

The species level represents the highest rank of taxonomic resolution formally recognized by the International Committee on Taxonomy of Viruses (ICTV); in consequence, several classification schemes have been proposed to differentiate viruses below this threshold. This trend has also affected PCV3, for which multiple genotyping systems have been proposed over time [35]. Unfortunately, most of these schemes lack consistent criteria, resulting in a multitude of nomenclatures that not only differ in naming but, more problematically, sometimes classify the same strains into different genotypes depending on the study. A parallel example is represented by PCV2, for which several classification schemes have been proposed over time, largely based on the definition of reference strains and genetic distance thresholds within ORF2 [4, 8, 13]. These frameworks have progressively contributed to the standardization of nomenclature and are now widely accepted for studying the evolution of the pathogen, its variants, and its molecular epidemiology over time [9, 15]. In this context, such classifications have gradually gained recognition and practical adoption, not only within the scientific community but also among field veterinarians. Therefore, nomenclature harmonization on PCV3 genotypes would be also desirable. However, an even more pressing question concerns whether establishing formal PCV3 genotyping criteria is both possible and necessary, since the results of the present study raise several doubts regarding the feasibility and reliability of such a classification based on the current stage of knowledge.

To answer abovementioned questions, all available high-quality ORF2 sequences were included and analysed to obtain a comprehensive depiction of the PCV3 genetic landscape and to assess the robustness of phylogenetic inference. Indeed, selectively analysing only a subset of strains can lead to artificially inflated branch supports, thereby exposing different studies to potentially discordant outcomes depending on the particular reference sequences chosen. PCV3 ORF2 was selected, similarly to what proposed for PCV2, as the target region because it provides the best compromise between length and informativeness [3, 10]. ORF2 is characterized by higher average genetic diversity, likely attributable to stronger immune-driven evolutionary pressures, therefore, a higher phylogenetic signal is expected. While the complete genome could potentially provide more comprehensive information, its use may pose practical challenges. In fact, any classification scheme should not only be accurate but also readily applicable across a wide range of laboratories, including those engaged in routine diagnostic activities, for which whole-genome sequencing may be impractical. Interestingly, PCV3 appears to exhibit a lower tendency toward recombination compared to other circoviruses. Recombination events in circoviruses have been reported to occur predominantly within the intergenic regions rather than in ORFs [22]. In our dataset, no statistically significant evidence of recombination was detected. While the possibility of undetected events cannot be entirely excluded—particularly given the limited genetic variability and high strain similarity that characterize PCV3, which may reduce the sensitivity of detection tools—the overall weak recombination signal suggests that this phenomenon likely played a marginal role in shaping the observed phylogenetic patterns. Therefore, the inferred evolutionary relationships are unlikely to have been substantially distorted by recombination, supporting the overall robustness of our analyses. Highly divergent sequences that did not belong to clades comprising at least five additional strains were excluded due to the high likelihood of sequencing artifacts. Notably, one such highly divergent sequence—previously proposed as a potential new genotype in [10]—was later withdrawn, suggesting sequencing inaccuracies [10]. Although this may not apply to all excluded sequences—something that cannot be verified in the absence of raw data—they can at least be regarded as evolutionary dead ends and therefore currently epidemiologically irrelevant. However, this exclusion from the current analysis does not preclude future developments, as the evidence that non-investigated animal populations may harbour divergent or rare PCVs variants or species [7, 11, 17] —whose future behaviour is difficult to predict—serves as a reminder of the importance of continuous surveillance and regular updating of molecular epidemiology frameworks and classification systems.

Analysis of this curated dataset revealed an overall limited phylogenetic signal (≈20% of quartets unresolved) and a decline in bootstrap support toward deeper nodes, indicating low confidence in the reconstruction of ancestral relationships—the very level typically targeted for genotype definition. This was reflected by the presence of frequent dichotomies, even along long branches, leading to singletons or very small clusters. Such tree structure also hampered the initial attempts to define clusters using automatic approaches, as previously performed for other pathogens (data not shown) [28, 39]. A preliminary inspection of the phylogenetic trees revealed clear topological inconsistencies among those reconstructed using different methods on the same dataset. Although a formal evaluation of topological incongruence among phylogenetic trees is inherently challenging—and most available metrics have been criticized for potential biases or limitations—the combination of multiple calculated statistics and their low variability across random tree-tip resampling consistently indicated a moderate-to-high degree of incongruence among tree reconstructions. Altogether, these findings suggest limited confidence and reproducibility in the inferred phylogenetic relationships.

Although clades of interest can be defined through arbitrarily selected reference sequences, this approach proved inconsistent in our setting. As a preliminary observation, pairwise genetic distances—already low overall —showed complete overlap both within and between the proposed genotypes, thereby calling into question the practical meaning of assigning strains to “different” genotypes. Especially, genotypes tentatively identified on the IQTree ML topology exhibited markedly different relationships when mapped onto trees reconstructed with other methods—not only for unclassified sequences or minor clades, but also for larger groups. While log-likelihood comparisons sometimes favoured specific trees for the given alignment, differences were frequently non-significant; when significance was detected, p-values were sometimes modest (often p > 0.01). In contrast, distance-based metrics indicated meaningful topological disparities among trees. Taken together, the small likelihood contrasts, coupled with metric-based evidence of incongruence, point to substantial uncertainty and multiple equally plausible evolutionary pathways—offering limited justification for privileging one reconstruction method over another.

Based on current evidence, we are prone to discourage the tendency to classify PCV3 in multiple genotypes. The difficulty of defining clear, objective criteria and the sensitivity of outcomes to the chosen dataset and inference method represent a hard-to-overcome challenge. Mandating analyses on the complete PCV3 ORF2 dataset with a fixed pipeline might offer some standardization, but it would be impractical and remain vulnerable to future changes. In particular, the continued growth of sequence availability and ongoing improvements in phylogenetic algorithms—even within the same software—are likely to generate inconsistencies across studies performed at different times.

Beyond these technical considerations, it is worth reflecting on the necessity of formal genotyping for PCV3, given its overall low genetic variability and the lack of experimental evidence posing in favour of differences in antigenicity, epidemiology, or virulence. In our view, classification should not be an “exercise in style,” but should respond to the practical needs of researchers and veterinarians. Importantly, the present findings do not preclude the use of author-defined PCV3 clade names based on explicit, even if arbitrary, criteria when this aids the discussion and interpretation of genetic differences within an individual study. However, such study-specific nomenclature should be clearly presented as internal to the work and not intended for generalization unless supported by formal evidence.

Considering the documented dynamism of other porcine circoviruses and our still preliminary understanding of PCV3 epidemiology and pathogenesis, the definition of meaningful genotypes in the future is not excluded; our position is contingent on current data. We emphasize, however, that any proposed classification must be accompanied by a transparent methodological description, quantitative measures of statistical support, and evidence of reproducibility—ideally alongside correlations with epidemiologically and/or biologically relevant traits.

Authors’ contributions

Author Contributions: Conceptualization, GF and JS; methodology, GF; software; GF; formal analysis, GF; investigation, GF and JS; data curation, GF; writing, —review and editing, GF and JS. All authors have read and agreed to the published version of the manuscript.

Funding

No funding was received for the present publication.

Data availability

All obtained sequences were previously submitted to GenBank and are freely available.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.