Scaling Graph Neural Networks to Large Proteins 1. This paper introduces the DISPEF dataset, specifically designed for benchmarking graph neural networks (GNNs) on large, biologically relevant proteins. DISPEF contains over 200,000 proteins with implicit solvation free energies, a key challenge for molecular modeling. 2. A major innovation is the introduction of a multiscale architecture called “Schake,” which enhances GNN performance on large proteins by incorporating both short-range and long-range interactions, ensuring computational efficiency and transferability across protein sizes. 3. The Schake model leverages two types of message-passing layers: a more accurate SAKE layer for short-range atomic environments and a more efficient SchNet layer for long-range alpha carbon interactions. This mixed design significantly improves both accuracy and speed. 4. DISPEF provides a diverse chemical environment, including both folded and disordered protein regions, making it an ideal dataset for testing GNNs. The inclusion of many-body solvation free energies as targets pushes the limits of model accuracy and generalization. 5. Benchmark results show that Schake consistently outperforms existing GNNs on energy and force predictions for large proteins, while maintaining superior computational efficiency, reducing inference times by up to 2.7× compared to prior models. 6. A key insight from the study is that increasing the cutoff distance in GNNs improves model transferability to larger proteins, but this needs to be balanced against computational cost, which Schake addresses with its hybrid architecture. 7. This work highlights the importance of datasets like DISPEF for driving future GNN innovations and optimizing models for real-world applications like protein folding and drug discovery, where large proteins are common. 8. The paper provides valuable insights for advancing GNN architectures in computational biology, emphasizing the need for both accuracy and efficiency in handling the vast complexity of biomolecular systems. @binzmit 💻Code: github.com/ZhangGroup-MITChe… 📜Paper: arxiv.org/abs/2410.03921