Present-State Resolution and Lineage History in Early Mouse Embryogenesis

Developmental-history claims depend on how completely the contemporaneous state has been represented. This study tests that dependence in 3,704 lineage-qualified cells from three E7.5 mouse embryos in the MELA dataset, joining source transcriptomes exactly to source-provided reconstructed lineage trees. The analysis separates whether coarse-state discordance survives molecular refinement, whether the molecular present predicts a disjoint molecular target beyond coarse annotations, and whether reconstructed lineage adds predictive information beyond that present state.

A coarse state defined by source germ layer, cell-cycle phase and donor/host type contains 780,783 unordered pairs with matched coarse state, different reconstructed-lineage summary, and different contemporaneous cell type. At molecular partition size K = 32, 3.1338% of these pairs remain co-clustered, while an annotation-discordant control survives at 3.1353%; a tighter annotation-pair-matched comparison yields no corrected history-specific preference (p = 0.527). The raw witness collapse therefore measures resolution of expression-associated annotation discordance rather than independent evidence for lineage-specific memory.

The principal predictive analysis is fully inductive. For each held-out embryo, gene eligibility, panel construction, normalization, representation fitting, and hyperparameter selection use only the other two embryos. A fixed 64-dimensional signed-hash molecular representation improves on both the training-global-mean and S0-only baselines in every held-out embryo, with relative NSE reductions over S0 of 1.87%, 2.33%, and 2.96%. The molecular-over-S0 ordering is also observed under training-fitted PCA target representations.

Lineage models do not establish an additional family-wise predictive contribution. One prospectively specified joint-metric configuration produces positive finite gains in all three embryos, but the complete 76-comparison family with nested retuning under 999 conditional history permutations yields p = 0.553, and the separate 125-bank discrete history diagnostic has no multiplicity-corrected departures. The data therefore support additional information in the measured molecular present while leaving exact state sufficiency, causal developmental memory, Markovianity, and future-state transition laws unresolved.