Fig. 2.

Overall performance of different somatic mutation callers on 119 replicates in the SEQC2 data set. The NeuSomatic model trained on SEQC2 data achieved consistent superiority over other techniques across diverse sets of replicates of different purities/coverages in WGS, WES, FFPE, AmpliSeq (AMS), and different library preparation data sets. For each replicate, the best F1-score was computed across different approaches. The heatmaps illustrate the absolute difference between the F1-score of any of the network models according to the best F1-score. In each panel, the mean F1-score is shown for each approach across 119 replicates. Here, the SEQC-WGS-GT50-SpikeWGS10 trained models were used for NeuSomatic and NeuSomatic-S