Protein graph neural networks often treat graph construction as a fixed pre-processing detail. This work questions it, and asks which edges does a protein model actually need.
The study investigates topology, geometric inductive biases, and Pareto trade-offs across EC, GO, and Fold3D benchmarks. It introduces Angle Rewiring and a FiLM-style Efficient IEConv variant designed to retain useful structural information while reducing memory and computation.
This work was completed during my internship at Nostrum Biodiscovery.