Extended Data Fig. 2: MIRA outperforms standard methodology for resolving cell state trajectories using expression data alone.

Benchmarking results comparing MIRA to standard methodology of Seurat PCA + Slingshot in the indicated metrics of cell state trajectory inference using expression data alone. Top row shows ground truth scaffolds, which are computationally synthesized by mixing reads from distinct populations of single cells from a 10X Genomics dataset⁶³ of peripheral blood mononuclear cells (PBMCs). Scaffold difficulty increases from left to right, where more difficult scaffolds contain cell states where mixture components are more similar (increased entropy), making them more difficult to distinguish by the tested lineage inference methodologies. Line plots indicate MIRA (red) versus Seurat PCA + Slingshot (blue) performance in each of the four scaffold difficulties with trials for three different mean read depths (lower read depth further increases the difficulty of solving the topology). For each trial, 5 replicates were tested for each modeling approach. Edge accuracy measures the accuracy of the inferred edges compared to ground truth (dynverse’s edge flip score⁶⁴). Branch F1 score⁶⁴ measures the precision and recall of the inferred branches compared to ground truth. Pseudotime correlation⁶⁴ measures the correlation between inferred versus ground truth pseudotime for each cell. The bottom rows show example UMAPs for MIRA or Seurat PCA + Slingshot for each scaffold difficulty with black edges showing cell state parsing from each algorithm. Cells colored by ground truth branch assignment where blue cells are the origin state. In the line plots above, black outlines indicate the points for the models shown in the example UMAPs.