Extended Data Fig. 2: Benchmarking CONCORD and other dimensionality reduction methods across diverse structures.
From: Revealing a coherent cell-state landscape across single-cell datasets with CONCORD
a, Heatmaps of simulated expression for the three-cluster structure and the corresponding CONCORD latent encoding in hcl or kNN modes. b, Heatmaps of simulated expression for the trajectory-loop structure and the corresponding CONCORD latent encoding in hcl or kNN modes. c, Trustworthiness measured across neighborhood sizes (k) in the three-cluster simulation. In the noise-free reference, within-cluster neighbors are assigned at random, so trustworthiness is < 1. CONCORD (h, k) denotes the hcl and kNN modes, respectively. d, Trustworthiness measured across neighborhood sizes in the complex trajectory-loop simulation. e, Heatmaps of simulated expression for the complex-tree structure shown in Fig. 2g, alongside the corresponding CONCORD latent encodings under a moderate degree of hard-negative enrichment in hcl and kNN modes. f, kNN-graph visualizations of latent spaces from the complex tree simulation, generated using naïve contrastive learning and the hcl and kNN modes of CONCORD with varying degrees of hard-negative enrichment. Zoomed-in views highlight improved resolution of a representative branch achieved through hard-negative sampling. g, Trustworthiness across neighborhood sizes for hcl and kNN modes in the complex-tree simulation, evaluated under varying degrees of hard-negative sampling. An inset for k < 20 highlights improved local neighborhood preservation with hard-negative sampling.
