Self-Learning Virome Intelligence System (SLVIS): An Unsupervised Deep Learning Framework for Emerging Virus Detection and Genomic Drift Surveillance

Abstract

The rapid emergence of novel and highly mutated viruses poses a significant threat to global health and pandemic preparedness. Traditional reference-dependent genomic surveillance pipelines struggle to detect previously unseen viral genomes and adapt to nonstationary sequencing data. To address these challenges, this paper proposes a Self-Learning Virome Intelligence System (SLVIS) — a hybrid unsupervised learning framework combining a Variational Autoencoder (VAE), Self-Organizing Map (SOM), and Fuzzy Clustering for real-time detection and interpretation of viral novelty in metavirome datasets.

SLVIS performs label-free viral clustering, mutation-tolerant similarity detection, and incremental self-learning on streaming genomic data. The model integrates nonlinear latent encoding with topology-preserving organization and fuzzy membership estimation, generating calibrated novelty and drift alerts for ongoing genomic surveillance. Evaluations on large-scale datasets (Global Ocean Virome, MetaPhage, ZoonoMix) demonstrate superior clustering quality (Silhouette = 0.72, DBI = 0.96) and high novelty detection accuracy (AUPRC = 0.901, AUROC = 0.923) compared to baseline methods (DEC, One-Class SVM, AE).

The framework also achieved sub-two-day detection delay for drift events, accurately flagging emerging viral families before reference database annotation. By coupling interpretability (U-Matrix, motif attribution) and incremental adaptation, SLVIS represents a significant advancement toward autonomous, mutation-resilient, and explainable virome surveillance, aligning with next-generation goals in global biosurveillance and pandemic early-warning systems.