What Drives GNN Performance in Tissue Dynamics? Insights from Vertex-Model Simulations

Graph Neural Networks (GNNs) are increasingly applied to model collective cell behaviors during development, yet choosing an appropriate architecture remains difficult. We present an in silico benchmark of 1,615 tissue graphs, each undergoing a single T1 transition (cell-neighbor exchange), and evaluate five architectures for predicting post-T1 cell–cell interface (edge) lengths. The Provably Powerful Graph Network (PPGN) outperformed all others, achieving high accuracy with only 41 training graphs. PPGN learned a bimodal prediction strategy: near the T1 event, where deformations are large and localized, it predicts edge-length changes accurately, but several hops away it reverts to copying pre-T1 lengths, safeguarding accuracy yet risking cumulative bias. Additionally, prediction error increased almost linearly with tissue disorder but decreased with larger datasets, identifying disorder as a practical proxy for reliability. Our benchmark provides a controlled platform for evaluating GNNs in tissue-dynamics during development and offers actionable guidelines for model and dataset design.