BLOOM: Biologically Informed Flow Matching for Transcriptomic Virtual Embryo Modeling

Abstract

Constructing a virtual embryo requires a generative representation of how heterogeneous cell populations change across developmental time. Here we introduce BLOOM, a biologically informed framework for conditional generation of developmental cell populations. BLOOM integrates temporal population context with developmental and regulatory priors through conditional flow matching to generate cell-expression distributions at queried stages in either temporal direction. Evaluation spans three mouse single-cell atlases covering prenatal development, visual-cortex maturation and multi-organ endothelial specialization. In comparisons with 22 time-series forecasting baselines, BLOOM ranks first for gene- and cell-expression correlations across all three datasets. Measured-stage evaluations support recovery of developmental expression patterns, population organization and marker-specific temporal relationships. Biological analyses connect these predictions to cell-state identities, gene programmes and regulatory associations. At unmeasured stages, interpolation and autoregressive history and future rollouts extend simulated expression dynamics between and beyond measured timepoints. Analyses of marker genes and temporal descriptors identify candidate differences in the timing of muscle-associated, neural and vascular-specialization programmes. Input-based gene perturbations further probe predicted responses of edited genes and associated molecular programmes across time. Together, BLOOM supports transcriptomic virtual embryo modeling by bridging temporal sampling gaps, exploring extended developmental trajectories and prioritizing gene-dependent hypotheses for experimental investigation.