Extended Data Fig. 9: DEBIAS-M improves cross-study prediction of melanoma immunotherapy response and works within a single study.

a, Box and swarm plots of auROCs, each evaluating the cross-study generalization performance models using gut microbiome data to predict immunotherapy response in melanoma patients (defined as 12-month progression-free survival⁵²). Each auROC is calculated on a held-out study. See Supplementary Table 6 for information on studies and sample sizes. ‘Log10 Linear’ denotes the pipeline used by Lee et al.⁵². Preprocessing with DEBIAS-M shows a consistent albeit small improvement across all studies, with a particularly strong effect for one study. b, Box and swarm plots of auROCS, each representing the aggregate accuracy over the held-out points for a logistic regression model evaluated under 5-fold cross validation. ‘Cervical carcinoma’, ‘Colorectal cancer’ and ‘HIV’ correspond to the benchmark datasets from Fig. 2d, Fig. 2b, and Fig. 2a, respectively, with information on studies and samples sizes in Supplementary Tables 5, 2 and 1. p, one-sided Wilcoxon signed-rank test comparing between linear and DEBIAS-M single-batch across both datasets. Box, IQR; line, median; whiskers, nearest point to 1.5*IQR.