Universal CRISPR-based RNA virus detection and inhibition via foundation model

Abstract

CRISPR-based technologies offer promising avenues for RNA virus detection and intracellular inhibition. However, the computational design of optimal guide RNAs (gRNAs) is bottlenecked by data-hungry algorithms that struggle to rapidly adapt to newly discovered Cas effectors and emerging viral variants. Furthermore, existing tools largely rely on static reference genomes, ignoring individual genetic polymorphisms (SNPs) and tissue-specific transcriptomes, which risks severe target-activated collateral toxicity. Here, we present CRISPR-viva, a sequence-to-function foundation model for universal CRISPR-based RNA virus detection and inhibition. Pre-trained on a massive unlabelled RNA corpus, CRISPR-viva captures the universal syntax of RNA targeting. This enables few-shot adaptation, rapidly generating highly efficacious gRNAs for novel CRISPR systems and emerging viral variants using minimal experimental training data. Crucially, CRISPR-viva features a dynamic host-context integration pipeline that rigorously filters off-target candidates based on patient-specific SNPs and tissue expression profiles, effectively preventing host-cell toxicity and diagnostic false-positives. We validated the framework across 8 CRISPR systems, demonstrating its precision through in vitro LbuCas13a-based detection and cell-based Cas13d viral inhibition assays. To showcase its scalable inference capability, we deployed CRISPR-viva across 200 segmented viral genomes and 23 individual-derived human cell types, profiling over 300 million gRNA candidates. Overall, CRISPR-viva shifts the paradigm of gRNA design from isolated, task-specific algorithms to a highly adaptable foundation model, providing an agile and precise computational infrastructure for combatting emerging viral threats.