Fig. 4: Benchmarking CONCORD on C. elegans and C. briggsae embryogenesis atlas.

a, UMAPs from CONCORD and other integration methods, colored by inferred embryo time and species. Zoomed-in UMAPs for scVI and CONCORD (hcl) show approximately matched regions, colored by lineage and species. b, Global 2D and 3D CONCORD (hcl) embeddings colored by cell type and inferred embryo time. c, Overlap between expert-curated cell type and lineage annotations. A histogram shows lineage annotations concentrated in early-stage cells and cell type annotations predominantly in late-stage cells. Integration performance was evaluated separately for early-stage cells (lineage labels) and late-stage cells (cell type labels) using probing classifiers. d, Global 3D UMAPs of CONCORD, scVI and Harmony, highlighting cells mapped to the lineage subtree that give rise to ASE, ASJ and AUA neurons. For each method, the most representative view was selected. e, Heat map showing the top 50 most variable latent dimensions in the ASE, ASJ and AUA neuron subset for scVI, Harmony and CONCORD (hcl). Expression of gcy-5 and gcy-14 is overlaid on a zoomed UMAP recomputed from the CONCORD latent space. f, Lineage purity and average lineage distance computed across 2,000 randomly selected kNN neighborhoods for each method. For each randomly sampled anchor cell, we retrieve its k-nearest neighbors in the embedding and compare their lineage relationships to the lineage graph. Purity is the fraction of neighbors assigned to the same lineage as the anchor; average lineage distance is the mean hop distance on the lineage tree from the anchor to its neighbors. Box plots show the median (center line), quartiles (box limits), 1.5× the interquartile range (whiskers) and outliers (points). g, Zoomed-in UMAPs for mesoderm (excluding pharynx), highlighting major input lineages and cell types. Each lineage is represented by its cluster medoid; edges connect parental lineages to daughters following the lineage tree. h, Zoomed-in UMAPs for pharynx, annotated by cell type and broad input lineages. Selected lineage paths to pm1/2, pm3–pm5 and pm7 are highlighted. i, scVI and CONCORD were trained on the combined C. elegans data from Packer et al.⁴⁹ and our newly collected batch and then used to project the full atlas including C. elegans and C. briggsae data from Large et al.⁴⁸. Resulting UMAPs are colored by species and integration performance was evaluated with probing classifiers. Acc., accuracy; annot., annotation; avg., average.