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

Serological studies are critical for understanding pathogen-specific immune responses and informing public health measures1,2. Here, we evaluate tandem IgM, IgG and IgA responses in a cohort of individuals PCR+ for SARS-CoV-2 RNA (n=105) representing different categories of disease severity, including mild and asymptomatic infections. All PCR+ individuals surveyed were IgG-positive against the virus spike (S) glycoprotein. Elevated Ab levels were associated with hospitalization, with IgA titers, increased circulating IL-6 and strong neutralizing responses indicative of intensive care status. Additional studies of healthy blood donors (n=1,000) and pregnant women (n=900), sampled weekly during the initial outbreak in Stockholm, Sweden (weeks 14-25, 2020), demonstrated that anti-viral IgG titers differed over 1,000-fold between seroconverters, highlighting the need for careful evaluation of assay cut-offs for individual measurements and accurate estimates of seroprevalence (SP). To provide a solution to this, we developed probabilistic machine learning approaches to assign likelihood of past infection without setting an assay cut-off, allowing for more quantitative individual and population-level Ab measures. Using these tools, that considered responses against both S and RBD, we report SARS-CoV-2 S-specific IgG in 6.8% of blood donors and pregnant women two months after the peak of spring COVID-19 deaths, with the SP curve and country death rate following similar trajectories.