BLOOM turns sparse developmental atlases into a queryable virtual embryo, topping gene and cell correlation on all three mouse atlases

BLOOM infers expression distributions at developmental stages from neighboring observations, and it works in both directions: predicting later stages from earlier observations and reconstructing earlier stages from later ones. The time-series forecasting baselines were evaluated in their earlier-to-later forecasting configuration, whereas BLOOM explicitly adds a later-to-earlier reconstruction pathway. On the prenatal, visual-cortex and endothelial mouse atlases, BLOOM's stage-averaged gene Pearson correlations are 0.532, 0.634 and 0.488, and its stage-averaged cell correlations are 0.427, 0.864 and 0.814, ranking first among 22 time-series forecasting baselines in all three.

Generation first compresses each stage's cell set into a stage representation and then chains those representations in time order, so the model sees population history rather than individual cells. Destructive sequencing yields cross-sectional populations that are not tracked across stages, and cell number and composition vary by stage, so the model must retain both this history and within-stage heterogeneity. A shared cell encoder with permutation-invariant pooling within each stage produces stage representations, a single-layer GRU summarizes their ordered history, and this is combined with the target cell-state embedding and pseudo-trajectory hints to form the initial hidden state.

Biological priors are not post-hoc explanation but the starting point of generation: they set the mean and standard deviation of the Gaussian base, and a flow-matching velocity field pushes samples toward the requested stage's expression distribution. Earlier flow-matching methods mainly regress a velocity field along probability paths, whereas BLOOM encodes developmental age, cell-identity hierarchy, developmental graph, anatomical context, gene programmes and weighted transcription-factor–target networks as conditioning features. Pseudo-trajectories come from transport couplings inferred by moscot between adjacent stages in PCA coordinates, supplying candidate predecessor or successor correspondences; these priors are fused through dedicated encoders and learned gates whose output layers are zero initialized.

Evaluation covers three mouse single-cell atlases: the prenatal atlas, the visual-cortex atlas and the multi-organ endothelial atlas STED-EC, with readouts concentrated on transcriptomic expression distributions, population organization and marker genes. Reverse reconstruction has measured targets to compare against, whereas interpolated and history/future-extrapolated query populations have no independent experimental counterparts in the supplied evaluation and are assessed through their relationship to observed reference states, time-dependent population organization and expression summaries.

Reliable extrapolation still needs explicit assessment of predictive diversity and uncertainty: in the endothelial history example, gene rankings persist strongly while cell-level variability contracts markedly, so smooth population means or stable expression ranks can conceal loss of cellular heterogeneity. Single-gene input edits currently act only on the final source-history frame and do not impose persistent depletion through the generated course, and matched experimental perturbation outcomes are absent; some predicted directions disagree with published experiments, with tissue composition, developmental timing, isoform specificity and intervention persistence as possible contributors.