Fig. 3: Evaluation of iterative registration using semi-synthetic and real data.

a One of the samples described in Fig. 1 was clustered using the SWIFT algorithm. To create semi-synthetic samples with cluster-scale variability, after compensation and background subtraction, values in two channels for one cluster were divided by scaling factors ranging from 1 (no change) to 10 in 8 steps. For semi-synthetic samples with channel-scale variability, all cells were scaled by the same eight increments. The increments are expressed as magnitudes, i.e., numbers increase to indicate larger deviations between non-scaled and scaled values. The data were then ArcSinh transformed (see axes). All samples were then assigned to the original template. Cells in the specified cluster were assigned correctly (green) or incorrectly (magenta). Gray contours represent all cells. b Three cluster- and channel-scaled samples were registered to the corresponding original, non-scaled samples by ICR, NDCR, or NDCR+ICR. Registered samples were then assigned to the original cluster templates. For three clusters (one from each sample, Cluster One is shown in Fig. 3a), the capture of cells by the correct cluster was measured for increasing magnitudes of deviation. c From SDY420, five samples from the same subject were collected and frozen, then analyzed by CyTOF on different days. One sample was randomly selected as the reference and clustered using the SWIFT algorithm, and then the remaining samples were registered to the resulting template. Registration was repeated with varying numbers of iterations of NDCR or ICR or both and either partial or full position updates per iteration. All registered samples were then assigned to the reference SWIFT cluster template. The RMSE of each registered sample relative to the RMSE of its non-registered counterpart are shown.