Fig. 2

Resulting diagnostic graphs of INFLECT on the Levine32 benchmark dataset. The Unimodality set (\({U}_{i}\)) is plotted versus the number of metaclusters generated through metaclustering of SOM-clusters. Applying the L-function returns the inflection point where fraction of unimodal distribution plateaus for increasing numbers of metaclusters. A Diagnostic graph using hierarchical clustering for metaclustering. The L-function is applied to the fitted curve, resulting in an inflection point of 41 metaclusters. Value for \({U}_{41}\) is determined at 98.17%. B Diagnostic graph comparing results for the entire Levine32 dataset using hierarchical clustering to five smaller subsampled datasets, 75%, 50%, 25%, 10% and 5% of events. Unimodality plateaus and inflection points for dataset sizes of 25–100% are consistent. 10% and 5% sizes display a lower plateau and lower inflection point. Of note is that for these smaller datasets, the fraction of unimodal distributions at higher metacluster numbers is less stable. C Diagnostic graph using ConsensusClusterPlus metaclustering and L-function on the fitted curve. Due to the longer runtimes of ConsensusClusterPlus fewer \({U}_{i}\) were calculated. Resulting inflection point is 33 metaclusters with \({U}_{33}\) at 98.76%. D Marker performance diagnostic plot. For 5 selected markers the fraction of metaclusters which passed the dip test and marker spread test is collected per number of metaclusters. Data is shown in a scatterplot, amount of FlowSOM-metaclusters on the x-axis and percentage of metaclusters that passed the unimodality and interquartile range checks. Some (CD8, CD123, CD16) fail the unimodality tests at lower metacluster numbers, while reaching 100% at a higher metacluster numbers. CD13 and CD47 prove very stable at 100%. CD8 displays the highest variability and does not reach 100%, indicating poorer clustering performance for this marker