About performance gap between reproduced Pronet results and the reported performance in the paper "LEARNING HIERARCHICAL PROTEIN REPRESENTATIONS VIA COMPLETE 3D GRAPH NETWORKS".

[aagdggdgggdd]

Thank you for providing the Pronet code. I have reproduced Pronet experiments multiple times under identical conditions, including using the same graph cards. However, there is a noticeable performance gap (approximately 1.5-5%) on both ProFunct and HomologyTAPE datasets between my reproduced results and those reported. Could you kindly share the optimal hyperparameter settings, rather than the default ones?