Estimated causal-effect prior in cell-cell signaling enables out-of-distribution prediction of left-ventricular function across development

Ventricular morphogenesis arises from coordinated ligand-receptor signaling across cells, yet no predictive framework links this signaling state to ventricular function in a way that generalizes across developmental stages. This gap blocks the construction of cardiac digital twins capable of forecasting ventricular trajectories under unseen perturbation. Single-cell profiling of 3,498 cells from the C57BL/6 mouse left ventricle between embryonic day (E) 12.5 and E18.5 identified 117 ligands and 165 receptors active in the developing ventricle, of which 43 ligand and 42 receptor features became newly active and 18 ligand and 20 receptor features fell silent across the window. A model trained on an early-stage subset must therefore predict from a feature distribution containing signaling components never seen during training. Across thirteen benchmarked architectures, forward extrapolation underperformed the reverse direction by a mean ΔR² of 0.40. Here, we hypothesized that a prior encoding the estimated causal relationship between ligands and ventricular traits could enable generalization to held-out stages. Double Machine Learning estimated a standardized causal effect θ for each of 7,020 ligand-to-trait connections; permutation-based false discovery rate control, leave-one-timepoint-out stability filtering, and E-value sensitivity analysis prioritized 30 high-confidence edges, dominated by BMP2, periostin (Postn), and collagen-IX (Col9a3) in the forward direction, with signs matching published manipulation phenotypes for BMP2, periostin, and Slit2. Embedded as a sparse linear branch convexly blended with a deep predictor, the causal-estimates prior improved out-of-distribution R² by +0.12 to +0.49 across GRU, LSTM, Transformer, and MLP families, rescued the MLP from near-overfit collapse to R² of 0.97, and was stable across retained edge counts of 30 to 100, indicating that edge identity, not mask size, drives the improvement. These findings establish a foundation for developmental digital twins of the left ventricle combining out-of-distribution predictability with edge-level interpretability, and extend directly to forecasting ventricular trajectories under candidate therapeutic perturbation and into disease windows that cannot be observed directly, including single-ventricle physiology and hypoplastic left heart syndrome.