Evaluating discrepancies in dimensionality reduction for time-series single-cell RNA-sequencing data

Abstract

There are various dimensionality reduction techniques for visually inspecting dynamical patterns in time-series single-cell RNA-sequencing (scRNA-seq) data. However, the lack of one-to-one correspondence between cells across time points makes it difficult to uniquely uncover temporal structure in a low-dimensional manifold. The use of different techniques may thus lead to discrepancies in the representation of dynamical patterns. However, The extent of these discrepancies remains unclear. To investigate this, we propose an approach for reasoning about such discrepancies based on synthetic time-series scRNA-seq data generated by variational autoencoders. The synthetic dynamical patterns induced in a low-dimensional manifold reflect biologically plausible temporal patterns, such as dividing cell clusters during a differentiation process. We consider manifolds from different dimensionality reduction techniques, such as principal component analysis, t-distributed stochastic neighbor embedding, uniform manifold approximation, and projection and single-cell variational inference. We illustrate how the proposed approach allows for reasoning about to what extent low-dimensional manifolds, obtained from different techniques, can capture different dynamical patterns. None of these techniques was found to be consistently superior and the results indicate that they may not reliably represent dynamics when used in isolation, underscoring the need to compare multiple perspectives. Thus, the proposed synthetic dynamical pattern approach provides a foundation for guiding future methods development to detect complex patterns in time-series scRNA-seq data.

Keywords: deep learning; dimensionality reduction; evaluation; single-cell RNA-sequencing; synthetic data; time-series data.

Figure 1

The different rows correspond to different dimensionality reduction techniques applied to the dataset: the first row shows PCA, the second t-SNE, the third UMAP, and the last row scVI. The first column displays the entire dataset, grouped by time point. The subsequent columns represent only the cells from four selected time points, grouped by Louvain clusters. Clustering follows the original authors’ specifications. For each method, we computed a joint dimensionality reduction using cells from all time points and split the representation by time point for visualization.

Figure 2

(A) Different popular dimensionality reduction approaches for scRNA-seq data. (B) We train a (modified) VAE on a snapshot scRNA-seq dataset to obtain (1) a two-dimensional representation, which can optionally mimic a t-SNE, UMAP, or PCA manifold via supervised training, and (2) a matching scVI decoder, which can generate single-cell data based on the respective manifold. (C) Next, we introduce artificial dynamic patterns in the obtained manifold by applying vector fields and use the trained VAE to generate corresponding high-dimensional data for each synthetic time point. (D) We finally concatenate the datasets to create a synthetic time-series dataset, apply different dimensionality reduction techniques and compare.

Figure 3

(A and C) Latent representations from the supervised VAE, trained to match the t-SNE (A) and UMAP (C) embeddings of the PBMC dataset (leftmost panel), used as initial time point (t1), and transformed to induce a visual temporal pattern corresponding to the differentiation of the artificially transformed B-cells, shown in panels t2–t4. (B and D) t-SNE applied to the high-dimensional datasets generated by decoding the transformed latent representations in the t-SNE (A) and UMAP (C) manifolds using the supervised VAE. Cell annotations correspond to manually assigned cell types at the original time point (t1). This comparison illustrates how the same transformation is captured differently by t-SNE across these manifolds.

Figure 4

(A and C) Latent representations from the supervised VAE, trained to match the t-SNE (A) and UMAP (C) embeddings of the Zeisel dataset (leftmost panel), used as the initial time point (t1), and transformed to induce a visual temporal pattern, shown in panels t2–t4. These patterns correspond to structural transitions in the artificially transformed endothelial-mural and microglia populations in A, and in the endothelial-mural and interneurons in C. (B and D) PCA applied to the high-dimensional datasets generated by decoding the differently transformed latent representations in the UMAP manifold using the supervised VAE. Cell annotations correspond to manually assigned cell types at the original time point (t1). This comparison highlights how the performance of the same method (PCA) can vary depending on the transformation applied, even if this transformation is done in the same manifold.