Cluster-informed downsampling preserves MRD. (A) Extent of downsampling of B-ALL MRD files after processing through our pipeline, based on number of events per file (top) and file size (bottom). Bars and whiskers show means and SDs, respectively. (B) Effect of cluster-informed downsampling on total (blue), MRD (red), and benign (gray) absolute number of events detected on pipeline exports, relative to true number of events (estimated per upsampling factor correction). Data are plotted by level of MRD involvement or by specific DNN class. Bars and whiskers show means and SDs, respectively. (C) Actual percentage of MRD events detected on downsampled pipeline exports (no upsampling factor correction), compared with MRD percentages on the corresponding original files. PCs, plasma cells; PDC, plasmacytoid dendritic cells.