Detection of Frog Virus 3 by Integrating RPA-CRISPR/Cas12a-SPM with Deep Learning

Abstract

A fast, easy-to-implement, highly sensitive, and point-of-care (POC) detection system for frog virus 3 (FV3) is proposed. Combining recombinase polymerase amplification (RPA) and CRISPR/Cas12a, a limit of detection (LoD) of 100 aM (60.2 copies/μL) is achieved by optimizing RPA primers and CRISPR RNAs (crRNAs). For POC detection, smartphone microscopy is implemented, and an LoD of 10 aM is achieved in 40 min. The proposed system detects four positive animal-derived samples with a quantitation cycle (Cq) value of quantitative PCR (qPCR) in the range of 13 to 32. In addition, deep learning models are deployed for binary classification (positive or negative samples) and multiclass classification (different concentrations of FV3 and negative samples), achieving 100 and 98.75% accuracy, respectively. Without temperature regulation and expensive equipment, the proposed RPA-CRISPR/Cas12a combined with smartphone readouts and artificial-intelligence-assisted classification showcases the great potential for FV3 detection, specifically POC detection of DNA virus.

Figure 1

Schematic of the RPA-CRISPR/Cas12-SPM-AI detection system. The nucleic acids of animal-derived samples are released by PINDBK. Target DNA of the virus is amplified and recognized specifically by the RPA-CRISPR/Cas12a system. The Cas12a–crRNA complex binds to the target DNA, which triggers the collateral cleavage of Cas12a on the reporter (FQ-labeled ssDNA probe). The fluorophore on the reporter is then released, and the fluorescence is detected by the developed SPM. Three different deep learning models, including AlexNet, DenseNet-121, and EfficientNet-B7, with transfer learning, are used to classify the captured fluorescence images.

Figure 2

Detection of the purified fragments with three different crRNAs by the RPA-CRISPR/Cas12a system on a plate reader. (A) Normalized signal intensity of CRISPR/Cas12a (left) and RPA- CRISPR/Cas12a (right) by crRNA-1 for 10 nM to 100 pM target DNA versus 10 nM control DNA (left) and for 10 pM to 10 fM target DNA versus 10 pM control DNA (right; incubation time: 30 min). (B,C) Similar experiments are performed by crRNA-2 and crRNA-3. (D) Integrated signal intensity of CRISPR/Cas12a (left) and RPA-CRISPR/Cas12a (right) by crRNA-3 for various concentrations of target DNA versus 10 nM control DNA (left) and 10 pM control DNA (right; incubation time: 30 min): F, FV3; I, ISKNV. Data represent the mean ± s.d. from at least three independent experiments. The Student’s two-sample t-test is used for statistical analysis. *p < 0.05, **p < 0.01, and ****p < 0.001. No error bar appears for certain points because it is shorter than the symbol size.

Figure 3

Detection of FV3 with RPA-CRISPR/Cas12a-SPM. (A) Schematic of SPM for fluorescence detection (left). The physical appearance of the assembled device used for fluorescence image collection after the RPA-CRISPR/Cas12 reaction (right). (B) Fluorescence images of RPA-CRISPR reactions with the purified target DNA or control DNA. (C) The statistics of the signal intensity for fluorescence images collected in RPA-CRISPR reactions with various concentrations of purified fragments: F, FV3; I, ISKNV. (D) The SPM images of the positive and negative animal-derived samples in the RPA-CRISPR/Cas12a detection assay. (E) The statistics of the signal intensity of fluorescence images from animal-derived samples: P, positive; N, negative. (F) The fluorescence signal intensity of animal-derived samples and their absolute concentration detected by qPCR. The Student’s two-sample t-test is used for statistical analysis. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, and N.D. indicates no difference.

Figure 4

Evaluation of deep learning models with transfer learning for the classification of fluorescence images. (A,B) Training traces of EfficientNet-B7 show the accuracy and loss changes versus epoch number for binary classification. (C) Confusion matrix analysis of EfficientNet-B7 for binary classification. The values on the diagonal of the confusion matrix indicate the number of correct predictions, and the off-diagonal values are the numbers of samples with incorrect predictions. (D,E) Training traces of AlexNet show the accuracy and loss changes versus epoch number for multiclass classification. (F) Confusion matrix analysis of AlexNet for multiclass classification.