Point-of-care identification of SARS-CoV-2 variants of concern using a multiplexed antibody microarray and machine learning

Introduction:: As the COVID-19 pandemic progressed, it became increasingly clear that SARS-CoV-2 mutations would significantly impact the course of the epidemic. Viral mutations to the SARS-CoV-2 genome had significant effects on viral infectivity, disease severity, and vaccine effectiveness. Moreover, the performance of many point-of-care tests (POCTs) which relied on the SARS-CoV-2 spike (S) protein for disease detection was markedly diminished for certain SARS-CoV-2 variants. The emergence of new variants of concern (VOC) also made it apparent that there is an urgent epidemiological and clinical need to identify new VOCs, as comprehensively monitoring VOCs can inform both clinical and public health decision-making. Currently, the primary method of conclusively identifying VOCs is by DNA sequencing, which is slow, expensive, and not easily scalable. Unfortunately, POCTs such as lateral flow assays, which have been critical in widespread diagnostic efforts, cannot effectively identify VOCs. A multiplexed POCT capable of rapidly discriminating between SARS-CoV-2 VOCs, as they arise, is hence an urgent public health need.

To fabricate a POCT that can identify different VOCs, we performed a high-throughput combinatorial screen of antibody pairs against SARS-CoV-2 Spike protein (S protein) using a highly multiplexed fluorescence microarray. From this combinatorial screen, a suite of machine learning methods was used to identify minimal subsets of antibody pairs capable of discriminating between four different SARS-CoV-2 VOCs. This panel of cross-reactive S protein antibodies were used to create a rapid, low-cost POCT that can be paired with statistical learning methods to achieve point-of-care variant discrimination at scale.


Materials and Methods:: High-throughput screening data was obtained using the D4 platform that we have previously developed (Joh et al., 2014) as follows. Capture antibodies (cAbs) were inkjet printed on a polymer brush as ~150 µm diameter spots. A fluorescently labeled detection antibody (dAb) was also printed as microspots atop a soluble excipient pad on the polymer brush. Upon addition of a drop of liquid—typically viral transport medium for detection of SARS-CoV-2—the excipient pad dissolves, releasing the printed dAb into solution to bind the analyte, then form a sandwich with the cAbs at their printed spots. This generates a fluorescence signal that can be used to quantify the concentration of the analyte. Using this platform, a combinatorial screen of 29 cAbs and 13 dAbs was performed against the S protein of four SARS-CoV-2 variants (wild-type, delta, omicron, and beta). Dose-response curves for ~400 antibody pairs were measured as a function of S protein concentration.

Four feature selection algorithms were used to quantify importance of antibody pairs for classifying SARS-CoV-2 variants: L2-regularized logistic regression, random forests, partial least squares-discriminant analysis, and statistically equivalent signature. Importance metrics from 25 runs of each model were aggregated to create a composite score that ranked antibody pairs by frequency of inclusion and average importance. This composite score was used to select a final minimal subset, from which predictive models were created to assess overall performance. Performance results were averaged over 25 runs and compared using a two-way ANOVA with a Tukey’s honest significant difference post-hoc test.


Results, Conclusions, and Discussions:: Of the over 400 antibody pairs screened, around 100 were removed due to poor dose-response behavior. Each feature selection method was then used to quantify importance of the remaining pairs within the training set. Based on the results, the most important antibody pairs were used to select increasingly small subsets with which to build predictive models for the classification of SARS-CoV-2 variants. Models trained using subset sizes between 3-15 total features and were tested on held-out data. Performance of these models generally remained constant until fewer than five features were included, demonstrating that smaller subsets could provide similar performance to maximal models. Next, these feature importance metrics were compiled to create an aggregate ranking ,which identified just 15 total cAbs to be most important in over 90% of the models created regardless of the paired dAb. Out of this final subset of 15 antibodies, predictive models were built using 15, 10, and 5 cAbs. The two best performing models, random forest and quadratic discriminant analysis, demonstrated accuracies of 85 ± 0.01% and 82 ± 0.02%, respectively. Compared to models built using all antibody pairs, smaller subsets (15, 10, and 5) showed significantly higher accuracy and F1 score, though smaller subsets generally showed similar performance to each other.

Altogether, we show the development of a portable, highly sensitive, and inexpensive POCT that can discriminate viral variants with high accuracy. This technology represents a scalable method of identifying disease variants at the point-of-care with the goal of supplementing next-generation sequencing in helping to inform clinical and epidemiological decision making. Besides SARS-CoV-2, this platform could be applied to identifying other disease variants such as different serotypes of Dengue or Ebola virus. Moreover, the ability to distinguish highly similar proteins can provide novel opportunities for diagnostic applications, such as parsing patterns of post-translational modifications and diagnosing complex diseases states that do not have any single ideal biomarker. Both these applications include disease states that affect large swaths of the population in both developed and low-to-middle income nations such as cancer and autoimmune disorders.


Acknowledgements (Optional): : Z.Q. Acknowledges NIH grant 1T32GM144291

References (Optional): : Joh, D. Y., Hucknall, A. M., Wei, Q., Mason, K. A., Lund, M. L., Fontes, C. M., ... & Chilkoti, A. (2017). Inkjet-printed point-of-care immunoassay on a nanoscale polymer brush enables subpicomolar detection of analytes in blood. Proceedings of the National Academy of Sciences, 114(34), E7054-E7062.