Structure-guided taxonomic placement of divergent RNA viruses with ViraClass

ViraClass: Using RNA Virus RdRp Structure to Solve Deep Taxonomic Placement

Scientists have created a new tool called ViraClass. It uses the 3D shape of viral proteins instead of just their genetic code. This allows them to accurately classify new viruses. This works even when viruses are so different that their genetic sequences no longer look similar. It helps organize the vast, unknown world of RNA viruses into a clear family tree.

The Collapse of the Sequence Signal

The explosion of metatranscriptomic sequencing—the large-scale analysis of RNA in environmental or clinical samples—has revealed a sprawling "virosphere." This discovery far outpaces our ability to name its inhabitants. For RNA viruses, the standard way to build a family tree is to look at the RNA-dependent RNA polymerase (RdRp). This is the essential enzyme responsible for replicating the viral genome. However, classifying these viruses above the level of "family" is notoriously difficult.

Current bioinformatics pipelines primarily rely on sequence homology (the degree of similarity between nucleotide or amino acid sequences). But RNA viruses evolve with extreme rapidity. Over deep evolutionary timescales, the primary amino acid sequence of the RdRp undergoes many mutations. This causes the "signal"—the recognizable pattern linking a new virus to its ancestors—to collapse. When sequence identity drops to negligible levels, traditional tools like VITAP or geNomad struggle. They cannot easily assign a virus to a phylum or class. This leaves vast swaths of the virosphere unclassified.

Architecture of a Structural Classifier

The researchers propose that while the "letters" of the protein sequence change rapidly, the physical 3D architecture remains stable. The RdRp possesses a conserved "palm" subdomain. This is a structural core essential for its catalytic function. It persists even when the underlying sequence has drifted beyond recognition. ViraClass exploits this by shifting the taxonomic search from a linear string of letters to a three-dimensional shape.

The framework operates through a deliberate, two-stage hierarchical process [Figure 2a,b]:

1. Stage 1: Structure-Guided Descent. The system identifies the RdRp sequence and predicts its 3D structure using AlphaFold 3. It then performs a structural search against a reference library using tools like Foldseek. Classification proceeds rank-by-rank, starting from the phylum and descending toward the genus. At each step, the algorithm uses distance-weighted $k$-nearest-neighbour voting. The weight for each neighbor is calculated as $w_i = \frac{1}{d_i + \epsilon}$, where $d_i$ is the structural distance and $\epsilon$ is a small constant. To handle different levels of detail, the authors use lDDT (a metric for local distance differences in protein geometry) for higher ranks like phylum and family. For the more granular genus level, they use structure-aligned sequence identity. Crucially, ViraClass employs "calibrated acceptance thresholds." It calculates both the vote confidence ($c$) and the vote margin ($m$). If the evidence is not strong enough, the descent stops. This prevents the system from making overconfident, incorrect assignments.
2. Stage 2: Gene-Sharing Refinement. If Stage 1 leaves a virus unresolved, ViraClass checks for fuller genomic information. If available, it builds a gene-sharing network. It compares the presence of various remote protein families. This allows the tool to "rescue" a placement by looking at the broader protein context of the genome.

For viruses falling outside known reference space, the authors implement rank-calibrated structural clustering. This organizes "dark matter" viruses into coherent, compact groups. These serve as candidate lineages for future taxonomic formalization.

Recovering Signal from Evolutionary Depth

The core finding of the paper is that structural similarity retains taxonomic information long after sequence identity has failed. The authors demonstrate this by comparing how different metrics discriminate between taxonomic ranks [Figure 1c,d]. While sequence identity scores overlap heavily when comparing different phyla, structural measures show a clear, rank-dependent decrease. Specifically, the authors report that lDDT achieves an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.940 at the phylum level. It reaches 0.992 at the family level. This indicates a very high capacity to distinguish between different evolutionary lineages.

In rigorous "clade-holdout" benchmarks, the authors removed entire groups of viruses from the reference set. This tests the system's ability to generalize to unseen relatives. ViraClass significantly outperformed sequence-based baselines in these tests. When entire orders were withheld, ViraClass maintained a phylum-level accuracy of 96%. In contrast, the sequence-based VITAP dropped to just 12% [Figure 3a]. This gap proves that structure is a more robust proxy for deep ancestry.

Even in controversial "boundary cases," the structural signal was revealing. When the authors withheld the class Flasuviricetes, they tested the placement of the Flaviviridae family. ViraClass showed a directional, family-wide shift toward the Pisuviricota phylum [Figure 3c]. This mirrors recent phylogenetic debates. It suggests that ViraClass captures the actual structural tensions inherent in viral evolution.

Limitations of the Structural Lens

Despite its strengths, ViraClass is not a universal solution. Its efficacy is tied to two technical dependencies. First, the system requires high-quality structural models. As shown in the LucaProt analysis, phylum-level placement accuracy correlates with the predicted confidence (pLDDT) of the structure [Figure 4b]. If the initial RdRp identification is flawed, the classification becomes unstable.

Second, the framework depends on the completeness of the reference database. Because it relies on nearest-neighbor voting, the taxonomic landscape is shaped by the existing library. If certain phyla are heavily over-represented, they may exert a disproportionate influence on the voting process. Furthermore, the authors acknowledge a fundamental constraint. Since the tool is calibrated against current ICTV (International Committee on Taxonomy of Viruses) labels, it inherits any historical biases present in the existing sequence-based taxonomy.

The Verdict

ViraClass represents a significant step forward for viral ecology. By converting taxonomy from a search for matching strings to a search for matching shapes, the authors have mapped the deep reaches of the RNA virosphere. The tool is particularly valuable for analyzing fragmented metatranscriptomic data. This is especially true when only a single protein fragment is available. While it requires high-fidelity structural predictions, the framework's ability to organize unclassified "dark matter" into coherent clusters provides a scalable roadmap for future taxonomic revision. The code is reportedly available; see the paper for the canonical link.

Figures from the paper