Tree Tensor Networks Outperform Parameter-Matched Graph Neural Networks on Gene Regulatory Network Dynamics: A Pre-Registered Falsification Study

Gene regulatory networks have hierarchical, modular topology — yet standard graph neural networks treat them as flatgraphs, learning structure implicitly. We ask whether a model that encodes known community structure explicitly, via aTree Tensor Network (TTN) architecture originally developed in quantum many-body physics, outperforms aparameter-matched GNN when both are given the same task: predict expression dynamics in the DREAM4 benchmark.Using a pre-registered protocol with falsification criteria fixed before any data was seen, we find that it does,consistently. The TTN achieves 4–6× lower MSE per parameter and maintains stable long-range correlation (0.91–0.95) across all 10 tested configurations; GNN baselines (both GCN and GAT) degrade as network size increases.These results hold at maximum statistical significance for n=20 paired seeds across every confirmatory and exploratoryconfiguration tested. Beyond prediction, we show that a quantity native to tensor-network structure — a bond-entropyproxy — predicts which gene communities are most sensitive to knockout perturbation, without being trained for this(Spearman ρ=0.66, p=4.4×10−35 with rigorous canonicalization). We also report a pre-registered negative result: thisadvantage vanishes when the TTN’s hierarchy is uninformed, providing mechanistic evidence that the benefit isspecifically about correct structural priors, not tree architectures in general. We interpret this as a principled argumentfor why the field’s default of treating biological networks as flat graphs is leaving structure on the table.