#31 · ML / AI methods

Quantified imputation of missing values in single-cell RNA-seq under dropout

The common thread is computational imputation of missing values in sparse measurement data, with single-cell RNA-seq as the principal setting. A benchmark of six scRNA-seq imputation methods (MAGIC, IterativeImputer, KNNImputer, Mean, SoftImpute, GAN) on acute myocardial infarction patient PBMC data evaluates performance under 10-30% MCAR dropout. P-CSDI and PH-CSDI extend imputation to time-series scRNA-seq by conditioning diffusion models on pseudotime, reporting roughly 40-60% lower MAE and RMSE than prior imputation approaches. A separate preprint evaluates BRITS and SAITS under block-mask missingness in wearable smartwatch data and extends BRITS with circadian encoding, addressing the same problem outside single-cell biology. What is new is explicit, quantified performance assessment rather than assumed correction.