BLOOM: Biologically Informed Flow Matching for Transcriptomic Virtual Embryo Modeling

BLOOM, a biologically informed conditional flow-matching framework, ranks first against 22 time-series forecasting baselines for gene- and cell-expression correlations across three mouse single-cell developmental atlases.

• Why it matters: Virtual embryo modeling needs generative methods that capture how heterogeneous cell populations change over developmental time, including at unmeasured stages.
• What they did: BLOOM integrates temporal population context with developmental and regulatory priors to generate cell-expression distributions at queried stages in either temporal direction, evaluated on three mouse single-cell atlases.
• The result: Measured-stage evaluations recover developmental expression patterns and marker-specific temporal relationships, while unmeasured-stage interpolation and rollouts extend simulated dynamics and enable gene-perturbation hypothesis testing.