Background: The COVID-19 pandemic caused by severe acute respiratory syndrome-coronavirus 2 (SARS-CoV-2) has overwhelmed health systems worldwide and highlighted limitations of diagnostic testing. Several types of diagnostic tests including RT-PCR-based assays and antigen detection by lateral flow assays, each with their own strengths and weaknesses, have been developed and deployed in a short time. Methods: Here, we describe an immunoaffinity purification approach followed a by high resolution mass spectrometry-based targeted qualitative assay capable of detecting SARS-CoV-2 viral antigen from nasopharyngeal swab samples. Based on our discovery experiments using purified virus, recombinant viral protein and nasopharyngeal swab samples from COVID-19 positive patients, nucleocapsid protein was selected as a target antigen. We then developed an automated antibody capture-based workflow coupled to targeted high-field asymmetric waveform ion mobility spectrometry (FAIMS) - parallel reaction monitoring (PRM) assay on an Orbitrap Exploris 480 mass spectrometer. An ensemble machine learning-based model for determining COVID-19 positive samples was developed using fragment ion intensities from the PRM data. Findings: The optimized targeted assay, which was used to analyze 88 positive and 88 negative nasopharyngeal swab samples for validation, resulted in 98% (95% CI = 0.922-0.997) (86/88) sensitivity and 100% (95% CI = 0.958-1.000) (88/88) specificity using RT-PCR-based molecular testing as the reference method. Interpretation: Our results demonstrate that direct detection of infectious agents from clinical samples by tandem mass spectrometry-based assays have potential to be deployed as diagnostic assays in clinical laboratories, which has hitherto been limited to analysis of pure microbial cultures. Funding: This study was supported by DBT/Wellcome Trust India Alliance Margdarshi Fellowship grant IA/M/15/1/502023 awarded to AP and the generosity of Eric and Wendy Schmidt.

Antibody was biotinylated using biotinylation kit (ThermoFisher Scientific, San Jose, CA) as per manufacturer’s instructions. Biotinylated antibody (1 µg) was coated on streptavidin MSIA tips (Catalog#991STR11, ThermoFisher Scientific, San Jose, CA) in 0.1% BSA containing 1X PBS on the Versette automated liquid handler (ThermoFisher Scientific, San Jose, CA). Nasopharyngeal swab samples (750 µl) were mixed with zwitterion Z316 at final concentration of 0.002% in 96 well plate and were inactivated at 70°C for 30 minutes. Inactivated samples were subjected to enrichment using mass spectrometry immunoassay (MSIA)-based enrichment using biotinylated antibody, washed two times with 200 µl 1X PBS and eluted in 100 µl of 50% ACN/0.002% Z316 in 0.1% TFA. Sample eluent was mixed with 300 l of rapid trypsin digestion buffer (Promega Corporation, Madison, WI) and subjected to in-solution trypsin digestion (Gold Trypsin, Promega Corporation, Madison, WI) at 70°C for 1 hour on a shaker incubator. The digest was acidified using TFA to a final concentration of 1% TFA. The acidified digests were spiked-in with synthetic isotope labelled heavy peptides and the samples were loaded on EvoTips as per manufacturer’s instructions. Briefly, the C18 EvoTips were activated using 20 µl of 100% acetonitrile followed by equilibration with 20 µl of 0.1% formic acid in water. Activation and equilibration was carried out at 700 x g for 1 minute. The sample was loaded at 500 x g for 5 minute followed by washing using 0.1% formic acid once. At last the tips were loaded with 100 µl of 0.1% formic acid and processed for targeted analysis.

The PRM data were processed using the Skyline software package (36). Peak integration of all plausible fragments of analytes were carried out. The fragment ion intensities were exported from skyline and (natural) log transformed. A supervised machine learning method was used to select the optimal fragments and determine their weights for maximizing the performance of the targeted mass spectrometry assay. All computations were performed in R (version 4.0.1). For this, we utilized an ensemble-based machine learning approach encoded in the Super Learner as described previously. This method was configured to use a generalized linear model via penalized maximum likelihood (glmNET), generalized linear model (glm) and random forest model; all configured to use binomial distribution. A 10-fold cross-validation with a goal to maximize the AUC was instituted during the learning process. An optimal weighted average of the different trained models was computed and considered as final model for an independent validation.