Fig. 5

Performance comparison of reference-free cell mixture adjustment methods based on simulated data and continuous phenotypes. The samples were randomized into five batches, where random methylation differences were added to each batch. Nine scenarios were investigated with different levels of signal density (0.1%, 1% and 10%) and cell mixture confounding (no, moderate and strong). The “Perfect” method, which adjusts for real cell proportions and batch effects, is included for benchmarking purposes. a-f Performance was evaluated by (a) Genomic inflation factor λ on non-DMPs, (b-c) the observed false discovery rate (FDR) and true positive rate after FDR control (5% level, dashed line), (d-e) the observed family-wise error rate (FWER) and true positive rate after Bonferroni correction (5% level, solid line; 95% CI, dashed lines) and (f) the fraction of cell compositional variability (8 cell types jointly) explained by the components (PC/SV) as quantified by adjusted R². Error bars represent the standard errors. The SmartSVA is the only method that controls the type I error under the nominal level across scenarios and retains power in dense signal scenarios