DMGT: Composable Relational Learning from Heterogeneous Developmental Snapshots

Developmental trajectories are often inferred from destructive cellular snapshots, where individual experiments provide only partial and heterogeneous evidence of cellular relationships, such as transcriptional similarity, developmental time, lineage ancestry, and spatial proximity. Existing methods typically assume a fixed supervision signal or a predefined notion of cellular relatedness, limiting their ability to integrate complementary developmental evidence across experimental settings. We introduce DMGT (Developmental Memory Graph Transformer), a relation-composable representation-learning framework that models developmental observations as directed multi-relational graphs and learns shared representations through relation-aware message passing and attention. DMGT does not require a fixed relation set, but it composes the developmental relations available in each experiment, enabling a common model to perform fate prediction, lineage-consistent cross-time correspondence, and molecular-state modeling across heterogeneous datasets. We evaluate whether these representations capture transferable developmental structure using clone-disjoint splits and hard-negative controls that reduce lineage memorization and expression-similarity shortcuts. Across lineage-traced differentiation and reprogramming systems, controlled ablations show that temporal and lineage relations provide complementary information beyond transcriptional state, with their relative utility depending on the developmental task. We further apply the same relational formulation to spatiotemporal mouse and human embryonic development, incorporating spatial and temporal relations when direct lineage supervision is unavailable. These results demonstrate that developmental representations can be learned from heterogeneous and partially observed relational evidence within a unified, relation-composable framework.