A versatile automated pipeline for quantifying virus infectivity by label-free light microscopy and artificial intelligence

Abstract

Virus infectivity is traditionally determined by endpoint titration in cell cultures, and requires complex processing steps and human annotation. Here we developed an artificial intelligence (AI)-powered automated framework for ready detection of virus-induced cytopathic effect (DVICE). DVICE uses the convolutional neural network EfficientNet-B0 and transmitted light microscopy images of infected cell cultures, including coronavirus, influenza virus, rhinovirus, herpes simplex virus, vaccinia virus, and adenovirus. DVICE robustly measures virus-induced cytopathic effects (CPE), as shown by class activation mapping. Leave-one-out cross-validation in different cell types demonstrates high accuracy for different viruses, including SARS-CoV-2 in human saliva. Strikingly, DVICE exhibits virus class specificity, as shown with adenovirus, herpesvirus, rhinovirus, vaccinia virus, and SARS-CoV-2. In sum, DVICE provides unbiased infectivity scores of infectious agents causing CPE, and can be adapted to laboratory diagnostics, drug screening, serum neutralization or clinical samples.

Fig. 1. Workflow for automated readout of viral infection and dataset composition.

A Classical method (top) for infection readout employs crystal violet staining followed by manual annotation of virally induced lesions in a cell monolayer. Our proposed approach (bottom) uses automated image acquisition and AI-based detection of virus-induced cytopathic effect (DVICE). The red overlay indicates areas of network attention. Scale bar 1 mm. B Composition of acquired dataset, indicating the proportions of viruses in the images of infected wells, and cell lines for uninfected wells. C Quantification of cell confluency for different viruses and for uninfected images. Lines show the medians of the distributions, boxes show the quartiles, and whiskers are drawn to the farthest datapoint within 1.5*inter-quartile range (IQR) from the nearest hinge. CoV-229E: n = 342, VACV: n = 1520, HSV: n = 1561, AdV: n = 10,422, RV: n = 5466, CoV-OC43: n = 675, SARS-CoV-2: n = 2160, IAV: n = 722, uninfected: n = 35,743. Source data are provided as a Source Data file.

Fig. 2. Network performance and characteristics.

A Area under the receiver operating characteristic curve (AUROC) for different machine learning algorithms, including support-vector machine (SVM), k-nearest neighbors (k-NN), Gaussian naive Bayes (GNB), decision tree (DT), logistic regression (LR), random forest (RF), and DVICE. A nonparametric Kruskal–Wallis test with Dunn’s correction for multiple testing was performed to evaluate the conventional ML algorithms against DVICE. SVM, k-NN, GNB, DT, LR, RF: n = 5, DVICE: n = 3. Data are presented as means and error bars indicate standard deviations. Adjusted p-value: ****p = 0.0000276414. B Dependency of infection index on virus concentration. The concentration of plaque-forming units (pfu) was obtained from the plate annotation and well position. Pfu values were grouped into 15 bins, and the plotted points indicate the bins’ mean values. Actual values were annotated by human experts, and predicted values were provided by DVICE. log is the logarithm base 10. The theoretical curve is provided by the Poisson distribution. n = 12,640. C Comparison of humanly annotated (actual) and predicted TCID₅₀ values with linear regression line (red). The shaded region shows the 99.9% confidence interval of the regression curve. DVICE achieved a squared Pearson correlation coefficient of $R^2$ = 0.986 (slope 1.00 ± 0.01, p = ${10}^{-120}$). n = 130. D Example images of different viruses and class activation maps (CAMs), indicating regions of network attention for recognition of virus infection. Scale bar 1 mm. Source data are provided as a Source Data file.

Fig. 3. Generalization of DVICE under new experimental settings.

A Time-resolved analysis of A549 cells infected with AdV-C5-IX-FS2A-GFP with quantification of predicted infection index by DVICE, GFP intensity, and cell confluency. pfu plaque forming units, AUROC area under the receiver operating characteristic curve. Data are presented as means. B Leave-one-out cross-validation of DVICE. Images of wells inoculated with the indicated virus were left out during network training. The performance was evaluated on the left-out images. Data are presented as means and error bars indicate standard deviations. n = 3. C Comparison of DVICE’ performance between two microscopes, including ImageXpress Micro Confocal (IXM-C, Molecular Devices) and Cytation 5 (Agilent). n = 480 (216 infected and 264 uninfected images). Data are presented as means and error bars indicate standard deviations. Source data are provided as a Source Data file.

Fig. 4. SARS-CoV-2 infectious particle stability in saliva.

A Saliva and DMEM were spiked with SARS-CoV-2 and incubated at the indicated temperature and for the specified duration. The SARS-CoV-2 concentration after incubation was quantified by TCID₅₀ titration for infectious particles using the Reed–Muench method (bars) and by RT-qPCR for virus genome copies (dots). n = 2. B Sensitivity and specificity values for DVICE classifications. Sensitivity (true positive rate) is defined as $\mathrm{sensitivity}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$, and specificity (true negative rate) as $\mathrm{specificity}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$, where TP = true positive, FN = false negative, TN = true negative, FP = false positive. Data include 3646 images, of which 627 were infected. Plot shows means ± standard deviations, n = 3. Source data are provided as a Source Data file.

Fig. 5. Virus class-specific identification using DVICE.

A Example images of the segmentation maps used for virus class-specific identification. B Confusion matrix, indicating fractions of correct (across the diagonal) and incorrect classifications. Values are normalized to add up to 1 in each row except for cases of rounding errors. C Sensitivity (true positive rate) and specificity (true negative rate) of virus class detection by DVICE. Sensitivity was calculated as $\mathrm{sensitivity}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$ and specificity as $\mathrm{specificity}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$. TP = true positive, FN = false negative, TN = true negative, FP = false positive. Data are presented as means and error bars indicate standard deviations. n = 3. Source data are provided as a Source Data file.