Fig. 2: Illustration of neural network training and evaluation for the prediction of SARS-CoV-2 fold change in neutralization activity.

a SARS-CoV-2 isolates were subjected to therapeutic and vaccine in vitro assays, and resultant neutralization activity fold change ratios between the wild-type and variants were compiled and log10 transformed. b Variant spike protein sequences were VAE encoded and corresponding therapeutics and vaccines were one-hot encoded. Data collected from January 9, 2021, to October 31, 2022, was used as training data (N = 6089), and data collected from November 1, 2022, to June 22, 2023, was used as test data (N = 980). c A neural network model was trained to predict log10 fold change in neutralization activity and estimate the uncertainty (variance) in each prediction. d Comparison between model predictions and actual measurements for the test set, highlighting yellow points for higher prediction uncertainty. e Correlation between prediction error and estimated uncertainty for the test set, with least-squares line of best fit shown.