Antibody responses to SARS-CoV-2 train machine learning to assign likelihood of past infection during virus emergence in Sweden

Serology is critical for understanding pathogen-specific immune responses, but is fraught with difficulty, not least because the strength of antibody (Ab) response varies greatly between individuals and mild infections generally generate lower Ab titers (1-3). We used robust IgM, IgG and IgA Ab tests to evaluate anti-SARS-CoV-2 responses in individuals PCR+ for virus RNA (n=105) representing different categories of disease severity, including mild cases. All PCR+ individuals in the study became IgG-positive against pre-fusion trimers of the virus spike (S) glycoprotein, but titers varied greatly. Elevated IgA, IL-6 and neutralizing responses were present in intensive care patients. Additionally, blood donors and pregnant women (n=2,900) sampled throughout the first wave of the pandemic in Stockholm, Sweden, further demonstrated that anti-S IgG titers differed several orders of magnitude between individuals, with an increase of low titer values present in the population at later time points (4,5). To improve upon current methods to identify low titers and extend the utility of individual measures (6,7), we used our PCR+ individual data to train machine learning algorithms to assign likelihood of past infection. Using these tools that assigned probability to individual responses against S and the receptor binding domain (RBD), we report SARS-CoV-2-specific IgG in 13.7% of healthy donors five months after the peak of spring COVID-19 deaths, when mortality and ICU occupancy in the country due to the virus were at low levels. These data further our understanding of antibody responses to the virus and provide solutions to problems in serology data analysis.