Machine Learning-Driven and Smartphone-Based Fluorescence Detection for CRISPR Diagnostic of SARS-CoV-2

Abstract

Rapid, accurate, and low-cost detection of SARS-CoV-2 is crucial to contain the transmission of COVID-19. Here, we present a cost-effective smartphone-based device coupled with machine learning-driven software that evaluates the fluorescence signals of the CRISPR diagnostic of SARS-CoV-2. The device consists of a three-dimensional (3D)-printed housing and low-cost optic components that allow excitation of fluorescent reporters and selective transmission of the fluorescence emission to a smartphone. Custom software equipped with a binary classification model has been developed to quantify the acquired fluorescence images and determine the presence of the virus. Our detection system has a limit of detection (LoD) of 6.25 RNA copies/μL on laboratory samples and produces a test accuracy of 95% and sensitivity of 97% on 96 nasopharyngeal swab samples with transmissible viral loads. Our quantitative fluorescence score shows a strong correlation with the quantitative reverse transcription polymerase chain reaction (RT-qPCR) Ct values, offering valuable information of the viral load and, therefore, presenting an important advantage over nonquantitative readouts.

Harnessing CRISPR-Cas to Combat COVID-19: From Diagnostics to Therapeutics

“…Images captured with the phone camera are then sent to the cloud where a machine learning algorithm will be used for result interpretation. The LoD of this RT-RPA-DETECTR assay (6.25 copies/µL) was comparable to the RT-LAMP-DETECTR assay (10 copies/µL) developed by Broughtonet al [14], but higher than that of rRT-PCR (1-3.2 copies/µL) [14,36,80]. The PPA and NPA of the RT-LAMP-DETECTR assay relative to rRT-PCR (n = 82; PPA = 95%; NPA = 100%) [14] were also better than the RT-RPA-DETECTR assay (n = 115; PPA = 87%; NPA = 92%) [36].…”

“…The reported LoD for these one-pot assays ranged from 2.5 copies/µL to 45 copies/µL and achieved 97%-100% concordance with rRT-PCR results when tested with clinical specimens (n = 14-100) [54][55][56]. Like Samacoits et al [36], Chen et al [53] also capitalized on 3D printing technology to fabricate a portable instrument for fluorescence imaging with a smartphone camera, but result interpretation was based on visual inspection instead of a cloud-based analysis and the LoD attained was 20 copies/reaction [53].…”

“…The PPA and NPA of the RT-LAMP-DETECTR assay relative to rRT-PCR (n = 82; PPA = 95%; NPA = 100%) [14] were also better than the RT-RPA-DETECTR assay (n = 115; PPA = 87%; NPA = 92%) [36]. Nevertheless, the PPA and NPA of this RT-RPA-DETECTR assay improved to 97% and 93%, respectively, when clinical samples were restricted to those with high viral loads (Ct < 33; n = 96), indicating that the developed model required further fine-tuning on clinical samples with low viral loads [36].…”

“…The typical workflow of a CRISPR-Dx for COVID-19 consists of RNA extraction, reverse transcription (RT), target amplification, Cas assay, and collateral cleavage activity detection as shown in Figure 2A. RNA extraction is firstly carried out to lyse and purify the RNA genome of SARS-CoV-2 from clinical specimens, such as nasopharyngeal swab [35][36][37][38][39] nasal swab [40], oropharyngeal swab [14,41], saliva [42,43], bronchoalveolar lavage [35,39] and sputum [35]. The viral RNA is then converted into complementary DNA through RT followed by a DNA-based amplification technique in a one-step or a two-step approach to generate a large amount of target DNA prior to the Cas assay and collateral cleavage activity detection.…”

“…Other than LFD and fluorescence reader, a low-cost, three-dimensional (3D)-printed, smartphone-based device for fluorescence imaging has been described for the analysis of RT-RPA-DETECTR results [36]. The highly portable device consists of a sample tray with a capacity of eight PCR tubes, a smartphone holder, and an imaging compartment.…”