Earth's biogeochemical cycles are driven by microorganisms, with archaea playing increasingly recognized crucial roles. Archaea are abundant in various ecosystems, such as the mesopelagic ocean (where they are primary ammonia oxidizers influencing greenhouse gas emissions) and wetlands/permafrost soils (where methanogenic Euryarchaeota contribute significantly to methane production). Understanding the viruses infecting archaea is vital for accurate ecological assessments and climate modeling. While our knowledge of bacterial viruses (phages) is extensive thanks to metagenomic sequencing and analytical platforms, archaeal viruses are severely underrepresented in global studies. This underrepresentation is partly due to the lack of robust high-throughput methods to distinguish archaeal viruses from bacterial viruses in large datasets. This study aims to address this gap by developing a significantly improved tool for identifying archaeal viruses.

Existing methods for identifying archaeal viruses often involve manual curation of gene sharing networks, phylogenetic analysis, sequence homology comparisons, and functional/taxonomic annotations. These methods are time-consuming and not scalable for large datasets. Previous efforts like MArVD utilized text-based approaches with limitations in scalability and flexibility. This necessitates the development of a more sophisticated tool capable of handling the growing volume of viral sequence data and improving the accuracy and efficiency of archaeal virus identification.

The study involved the creation of three datasets: a reference database (OcAVdb), a training dataset, and a benchmarking dataset. OcAVdb comprises manually curated marine archaeal viruses. The training dataset included archaeal viruses and phages from various environments (marine, hypersaline, hot springs), obtained from public databases and previous studies. The benchmarking dataset contained archaeal viruses and phages from IMG/VR-db v2.0 and Tara Oceans GOV2.0 datasets, allowing for independent evaluation of MArVD2's performance. The archaeal viruses in these datasets were validated through vConTACT2 network analysis and manual inspection of functional annotations (DRAMv). MArVD2 uses a random forest machine learning algorithm trained on the combined genomic features (27 features were initially considered, reduced to 8 based on Gini importance index). The model's performance was evaluated using various metrics (TPR, ACC, SPEC, MCC, FDR, AUROC, AUPRC) and sensitivity analyses were conducted to assess the impact of dataset size, sequence length, and microbial contamination. The original MArVD tool was also recreated to provide a comparison point.

MArVD2 demonstrated significantly improved performance compared to its predecessor, MArVD. Using a prediction probability threshold of 80%, MArVD2 correctly classified 85% of verified archaeal viruses in the benchmarking dataset, with a false detection rate below 2%. The model's accuracy was robust even with varying dataset sizes. However, performance decreased with shorter sequence lengths (<10 kb) and increased microbial contamination. The ideal contig size for optimal performance was determined to be above 10 kb, with a recommended probability threshold of 0.80 for minimizing false positives. The study also revealed that the OcAVdb database may not fully represent all archaeal viruses, highlighting areas for future database expansion.

MArVD2 represents a significant advancement in archaeal virus identification, offering improved scalability, usability, and accuracy compared to previous methods. The high accuracy and robustness of the model, especially at the suggested 0.80 probability threshold, make it a valuable tool for analyzing large viral datasets. The findings address the need for efficient and accurate identification of archaeal viruses in non-extreme environments. The improved ability to distinguish archaeal viruses from other viruses will enhance our understanding of archaeal viral ecology and their impact on various ecosystems. The results highlight the importance of using high-quality, well-annotated data and considering the limitations of sequence length and microbial contamination when applying the tool.

MArVD2 provides a powerful and versatile tool for identifying archaeal viruses in large-scale datasets. Its machine learning approach significantly improves upon previous methods, offering enhanced accuracy, scalability, and user-friendliness. Future research could focus on expanding the reference database, incorporating additional genomic features, and further refining the model to improve its performance with shorter sequences and higher levels of microbial contamination. The continuous improvement and wider application of MArVD2 will be crucial for advancing our understanding of the archaeal virosphere and its role in global biogeochemical cycles.

The model's performance is affected by shorter contig lengths and microbial contamination. Optimal performance is achieved with contigs longer than 10 kb and minimal microbial contamination. The training dataset, while extensive, might not fully capture the diversity of all archaeal viruses. The reliance on existing viral identification tools (like VirSorter) as a preprocessing step introduces potential biases. Further curation and expansion of the reference database are necessary to minimize these limitations and improve the accuracy of MArVD2 across a wider range of viral diversity.