A controlled in silico benchmark for GNN prediction of tissue dynamics

Graph Neural Networks (GNNs) are promising tools for predicting tissue dynamics, but choosing the right architecture remains difficult because experimental datasets are limited, noisy, and system-specific. We introduce a controlled in silico benchmark for comparing GNN architectures on a vertex-model task: predicting relaxed cell–cell interface lengths after a cell neighbor exchange. Because simulated data provide known ground truth, we can vary tissue geometry, mechanics, perturbation complexity, dataset size, and input features independently. Provably Powerful Graph Networks (PPGN) and Principal Neighborhood Aggregation (PNA) were most sample-efficient when pre-event edge lengths were provided, whereas performance dropped sharply with topology alone. We also revealed a predict-or-copy strategy, whereby predictions far from the exchange copied pre-event lengths instead of predicting long-range changes. Prediction was harder in disordered tissues, suggesting hexagonality as a simple indicator of difficulty. These results provide a reproducible testbed for diagnosing feature dependence, copying behavior, and geometric consistency in tissue-remodeling prediction.