Dynamic Consistency Reveals Predictable Genes in Cross-Cell Type Temporal scRNA-Seq Data

Abstract Understanding how gene expression evolves over time after trauma is central to modeling immune responses, yet single-cell temporal data remain sparse and heterogeneous across cell types. Using a temporal trauma scRNA-seq dataset, we formulate the task of predicting next-time gene expression from earlier observations under a cross-cell-type generalization setting. We introduce the Dynamic Consistency Index (DCI), which quantifies how consistently a gene’s temporal trajectory aligns across cell types, serving as a measure of biological regularity and predictability. High-DCI genes exhibit reproducible temporal dynamics and are markedly easier to model. By integrating DCI-based gene selection with a recurrent neural architecture trained under a Gaussian negative log-likelihood objective, we achieve superior accuracy and well-calibrated uncertainty compared to deterministic baselines. Overall, DCI reliably identifies dynamically consistent genes, and uncertainty-aware recurrent modeling provides a robust framework for capturing cross-cell-type gene-expression evolution.