Paper-based point-of-care sensors for multiplexed quantification of biomarkers using deep learning

Introduction: Point-of-care (POC) testing is an attractive diagnostics approach owing to its low cost, easy accessibility and short turnaround times. Paper-based assays form an important class of POC sensors, which usually come in a lateral flow assay format (LFA), where samples flow horizontally across multiple paper layers. Despite their widespread use, these sensors often lack high sensitivity, provide limited multiplexing due to the in-line geometry and have high sensor-to-sensor variations. To overcome some of these limitations, we developed a fluorescence-based vertical flow assay (fxVFA) processed by a hand-held, cost-effective optical reader along with machine learning (Fig. 1a). Unlike conventional LFAs, our fxVFA maintains the vertical flow of the injected samples and supports a high level of signal multiplexing, allowing for parallel detection of multiple biomarkers with improved inter-sensor repeatability [1]. Furthermore, the fluorescence readout modality allows us to achieve higher sensitivity than colorimetric-based detection approaches that are typically employed in LFAs. We applied the fxVFA platform for multiplexed quantification of three biomarkers in human serum samples, i.e., myoglobin, creatine kinase-MB (CK-MB), and heart-type fatty acid binding protein (FABP) [1]. We used deep learning to optimize the fxVFA design and improve its biomarker quantification accuracy, which helped us demonstrate sub-ng/mL limit-of-detection for all three analytes with minimal cross-reactivity between the antigens (Fig. 1b-d). The optimized fxVFA was blindly tested on 16 serum samples, and output concentrations had a high correlation with the gold standard measurements, showing >0.9 coefficients of determination (R2) for all three biomarkers of interest (Fig. 1f-h).

Materials and

Methods: The fxVFA consists of multiple stacked paper layers, which include a sensing membrane with a 2D array of immunoreaction spots, interacting with the target analytes. We functionalized the sensing membrane with capture antibodies specific to the analytes of interest to enable specificity toward the three biomarkers. In addition to the capture antibodies, we also added positive and negative controls for internal calibration of the assay signals. In total, the sensing membrane contained 17 immunoreaction spots, including 6 testing spots, 4 positive controls and 7 negative controls (Fig. 1e). We used conjugated polymer nanoparticles (CPNs) as our fluorescent probes and conjugated them with the detection antibodies through streptavidin-biotin interaction. The fxVFA was operated in less than 15 min using 50 µL of serum sample and after the assay was complete, a custom-designed smartphone phone-based fluorescence reader captured an image of the activated sensing membrane. Following the image capture, an automated image analysis code segmented the individual immunoreaction spots from the sensing membrane to calculate fluorescence signals from the assay as average pixel intensities of the spots, generating 17 fluorescent intensity values per assay. These values were finally processed by neural network models, which quantified the biomarkers of interest. We separately optimized the architectures of these inference models for the three biomarkers, converging to three independent neural networks. The optimized assay and deep learning models were blindly tested on 16 clinical serum samples that had never been used before.

Results, Conclusions, and Discussions: Before testing the fxVFA on clinical serum samples, we first tested its response to buffer-spiked samples and showed a high correlation between the captured fluorescent signals and the concentrations of the spiked antigens (R2 > 0.98) for all three biomarkers of interest. The assay also showed limits-of-detection (LODs) of 0.52 ng/mL, 0.3 ng/mL and 0.49 ng/mL for Myoglobin, CK-MB and FABP, respectively, which are >10x lower than the corresponding clinical cutoff values for these biomarkers [1] (Fig. 1b-d). To test the biomarker quantification capabilities of our fxVFA using serum, we used clinical samples and generated a dataset comprising a total of 46 unique serum samples and 125 individually activated fxVFA cartridges. We used a portion of these samples (training/validation sets) to computationally optimize the assay design. We performed this optimization independently for the three biomarkers and converged to three separate assay designs with 7 spots for Myoglobin, 10 spots for CK-MB and 5 spots for FABP that served as inputs to each neural network used for the corresponding biomarker concentration inference. Optimized neural network models were trained on 30 samples from the training set and were blindly tested on the remaining 16 new samples never used during the training phase. fxVFA predicted concentrations had a high correlation with the gold standard measurements obtained using standard ELISA tests with an R2 value of >0.9 for all three biomarkers. Additionally, our fxVFA had high repeatability in triplicate measurements with an average coefficient of variation of < 15 % for the three biomarkers (Fig. 1e-h). This computationally optimized assay outperformed assays with different configurations, including those with a higher number of detection spots and those with other machine-learning inference models (e.g., linear regression) [1].
In conclusion, we developed a multiplexed fluorescence-based vertical flow assay (fxVFA) for the quantification of biomarkers from human serum samples. fxVFA predicted concentrations in clinical samples had a strong correlation with ground truth measurements obtained by standard ELISA tests. Competitive quantification performance, along with its low-cost design, rapid operation and hand-held footprint, make our fxVFA a promising POC sensor for highly sensitive multiplexed quantification of biomarkers from biological samples.

Acknowledgements (Optional):