A Single-tier Multiplexed Lyme Disease Assay Enhanced by Machine Learning

Introduction: Lyme disease (LD) is the most prevalent vector-borne disease in North America and Europe (Fig 1a). Yet, the current two-tier centralized lab tests have significant drawbacks, including long turnaround times, underreporting of cases, often resulting in failure to detect and treat LD in its early stages when treatment is most effective at preventing widespread complications (Fig 1b). In addition, the tests used are either subjective or use epitopes that are limited in sensitivity and specificity leading to undesired outcomes. Rapid, multiplexed and low-cost diagnostic platforms that can assay the repertoire of antibodies formed against pathogens can improve the diagnosis of difficult to detect diseases such as Lyme disease, facilitating early detection and improving patient outcomes. In this work, we report a single-tier rapid detection of Lyme disease using a multiplexed vertical flow assay (xVFA) that employs synthetic peptides as diagnostic targets, complemented by machine learning-assisted disease classification (Fig 1c). Clinical testing from two US repositories, including the Bay Area Lyme Disease Biobank (LDB) and the CDC, shows that the single-tier xVFA matches the performance of standard two-tier testing (STTT). This indicates the potential to replace the cumbersome two-tier centralized lab test with a rapid, single-tier point-of-care assay.

Materials and

Methods: We developed a multiplexed vertical flow assay incorporating nine peptide biomarkers that target various antigenic epitopes conserved across Borrelia burgdorferi genospecies. These biomarkers, which react against IgG and IgM antibodies, are each immobilized in duplicate on a paper-based sensing membrane. The assay also includes positive and negative control spots, resulting in a comprehensive 25-spot immunoassay within a single device (Fig 2a). We screened synthetic peptides to identify the most effective targets using Lyme disease-positive and negative control samples. Following this, we tested clinical samples from the LDB (Fig 2b) to train a machine learning-based diagnostic algorithm, which identified the most critical peptides, optimizing for the highest diagnostic accuracy as indicated by the area under the curve (Fig. 3a, 3b). We validated the performance of the xVFA using samples from the CDC and the LDB (Fig. 3c) comparing the effectiveness of the single-tier xVFA to traditional two-tier centralized lab tests.

Results, Conclusions, and Discussions: Results
We trained a machine learning model using 40 clinical samples classified as Lyme positive or negative by STTT. Signal intensities from each immunoreaction spot as shown in Fig. 2b were used as input into the neural network model, which classified samples as LD positive (prediction value > 0.5) or negative (prediction value ≤ 0.5). Using a cross-validation approach with 120 tests (40 patients × 3 replicates per patient), we optimized the neural network model architecture and selected the peptide combination that achieved the highest diagnostic performance, as represented by the area under the curve in the receiver operating characteristic plot (Fig. 3a). Peptides mod-VlsE-FlaB, Var2FlaB, OppA4 yielded the highest diagnostic performance with a sensitivity and specificity of 81.1% and 96.7% respectively (Fig. 3b). Further, we challenged the performance of the xVFA and the diagnostics model by testing a diverse group of LD, cross-reactive, and healthy samples obtained from the CDC, demonstrating a 100% agreement with the STTT and 96.6% agreement with the modified two-tier testing, indicating near-perfect agreement (Fig. 4). Finally, we validated the xVFA and the machine learning model by testing using 15 LD-positive and 15 LD-negative samples from the LDB, obtaining a 95.5% sensitivity and 100% specificity (Fig. 3c).
Conclusion
In conclusion, the single-tier xVFA offers a rapid and reliable alternative to traditional centralized lab testing. By utilizing synthetic peptides as diagnostic targets, we enhanced both sensitivity and specificity. The incorporation of a machine learning diagnostic model further improved predictive accuracy. Testing clinical samples from various biobanks demonstrated that the xVFA performs comparably to current commercial tests while offering superior specificity against cross-reactive look-alike diseases. The performance of the xVFA LD test across different patient groups could eliminate the need for a second-tier assay, making LD testing straightforward and accessible.

Acknowledgements (Optional): The authors acknowledge Bay Area Lyme Foundation’s Lyme Disease Biobank and the U.S. Centers for Disease Control & Prevention for graciously providing serum samples that were used for training and blinded validation of our platform respectively. We also acknowledge financial support from the National Institute of health (Grant #R44AI150060) and the National Science Foundation PATHS-UP Engineering Research Center (Grant #1648451).