Label-free detection of individual virus-infected cells using deep learning

Technical Summary: Label-free detection of individual virus-infected cells using deep learning

Problem Statement
Reliable identification and quantification of virus-infected cells are fundamental to numerous research and medical applications. However, current methodologies face significant limitations. Reporter virus systems often utilize strains that differ from clinical isolates regarding critical biological properties such as cell tropism and immune evasion, potentially skewing results. Conversely, traditional staining approaches require fixation, which precludes live-cell experiments, and often rely on manual counting, a process prone to human bias and inefficiency. There is a distinct need for a method that can accurately identify infected cells without labels, preserving the native state of the cells for dynamic observation.

Methodology
To address these challenges, the authors developed VAIruScope, a deep learning-based pipeline designed for the label-free identification of virus-infected cells using standard light microscopy images. The core of the methodology relies on the automated recognition of cytopathic effects (CPE)—morphological changes in cells caused by viral infection—rather than exogenous markers. The system was trained and tested across diverse cell models and four clinically relevant prototype viruses representing distinct viral families:

• RNA viruses: Influenza A virus.
• DNA viruses: Human cytomegalovirus (HCMV) and Herpes simplex virus-1 (HSV-1).
• Retroviruses: Human immunodeficiency virus-1 (HIV-1).

Key Contributions
The primary contribution of this work is the demonstration of a fully automated, label-free quantification system capable of distinguishing infected from uninfected cells based solely on intrinsic morphological features visible under light microscopy. By leveraging deep learning, the method eliminates the need for reporter genes or chemical stains, thereby avoiding the biases and experimental constraints associated with those techniques.

Results
The VAIruScope pipeline demonstrated high performance across the tested viral models, achieving classification accuracies of up to 96%. To validate the method's ability to detect specific infection markers without labels, the authors performed a proof-of-concept validation using electron microscopy on wild-type HSV-1 infections, confirming the correlation between the deep learning predictions and actual viral presence.

Significance
The paper positions VAIruScope as a tool with the potential to facilitate live-cell imaging studies. By removing the requirement for fixation or labeling, the method enables the investigation of infection dynamics in real-time. The authors frame this as a significant step forward for studying virus-host interactions under conditions that more closely resemble natural infection, free from the artifacts introduced by reporter systems or staining protocols.