Predicting viral sensitivity to antibodies using genetic sequences and antibody similarities
Fig 3
Accuracy of IC50 predictions assessed by multiple statistical metrics.
We compared the performance of models that learn the relationship between neutralization and genetic sequences using only sequence alignments with the standard 20 amino acids and gaps (dashed lines) to models that incorporate both sequence alignments and alignment-free features (AFFs). Specifically, we used loop length, net charge, and the number of PNGSs in variable loop regions (solid lines) for the AFFs. To assess accuracy, we employed multiple metrics: (a) Pearson’s R, (b) Spearman’s , (c) regression slope, (d) concordance correlation coefficient, and (e) MSE. Error bars represent 95% confidence intervals, estimated by randomly resampling the training and evaluation sets ten times. The concordance correlation coefficient (panel d) is computed as
where r,
, and
denote the Pearson correlation coefficient, mean, and variance of one dataset (x-axis), respectively. Overall, models that include alignment-free features achieve virtually the same prediction accuracy across a wide range of matrix ranks and across all evaluated metrics.