Single-cell sequencing only yields snapshots: a latent cell-state model fills in gene expression at unobserved timepoints
Synopsis
Single-cell RNA sequencing destroys each cell as it measures it and samples only discrete timepoints, so one cell's expression profile cannot be followed continuously. The Latent Cell-State Model approximates cell-state transitions as a Markov chain and adds a low-dimensional continuous latent variable to generate expression counts given cell state, time and gene. On the pancreatic development and Waddington-OT reprogramming datasets, maximum mean discrepancy between generated and real samples at held-out timepoints was 0.011 and 0.006 (pancreatic days 3 and 6) and 0.030 and 0.018 (reprogramming days 3.5 and 4), lower than the three compared methods.
Interpretation
Trajectory reconstruction can now return each cell's hidden state together with the gene expression under that state, rather than only interpolated expression values. Earlier trajectory methods either interpolate between adjacent observed timepoints without modelling continuous-time dynamics, or learn a global velocity field from start to end while giving less attention to the expression distribution at each observed timepoint; neither captures the hidden cell state that determines the expression profile. Holding out days 3 and 6 of stage 5 in the pancreatic dataset and days 3.5 and 4 in Waddington-OT, maximum mean discrepancy between generated and real samples was 0.011 and 0.006 for pancreatic and 0.030 and 0.018 for reprogramming, below the scNODE, scDiffusion and ARTEMIS comparators.
The model writes cell evolution as a discrete state chain that only moves forward, and lets state, time and a low-dimensional latent variable jointly determine each gene's expression count. State transitions are parameterized by an upper-triangular Markov transition matrix that zeroes backward transitions, and the initial state distribution is given by a softmax and shared across all cells. Gene expression is modelled as a negative binomial to accommodate the sparsity and overdispersion of single-cell counts, with the mean decomposed into total cell count times relative gene abundance; because the latent integral is intractable, training maximizes a variational lower bound, with networks outputting mean and dispersion parameters.
Learned cell states can be annotated one by one against known marker genes, grounding the generated high-dimensional profiles in readable biological identities. On the pancreatic data, 12 states were annotated as pancreatic progenitor-like, early and established endocrine induction, a differentiating endocrine intermediate, SC-EC-like, alpha-like and beta-like identities; on the reprogramming data, 3 states were annotated as matrix-associated stromal-like, a plastic intermediate and a contractile stromal-like state. Annotation used each state's top 20 differentially expressed genes (Wilcoxon rank-sum test) plus mean expression and expressing-cell fraction of canonical markers such as SOX9, PDX1, NEUROG3, PAX4, INS, GCG, Col6a1 and Tagln, broadly concordant with a model-based analysis of the expression distribution.
Perspective
The demonstrated scope is two transcriptomic time-series datasets: stage 5 of pancreatic differentiation from day 0 to day 7 (holding out days 3 and 6) and Waddington-OT mouse embryonic fibroblast reprogramming from day 0 to day 8 (holding out days 3.5 and 4), evaluated by maximum mean discrepancy between generated and real samples in a 50-dimensional principal-component space and by UMAP visualization. The authors propose extending the framework to multiple molecular layers to capture evolutionary dynamics beyond the transcriptome.
This is a preprint that has not been peer reviewed. The model uses 12 states on the pancreatic data and 3 on the reprogramming data, both chosen by ablation on mean validation ELBO; evaluation relies mainly on distribution-level maximum mean discrepancy and two-dimensional UMAP, and state annotation depends on known marker genes and reference data. The authors note the current work focuses on scRNA-seq time-series data.