generalfactbullishPretraining EnsembleEGNN on conformational ensembles dramatically improves molecular property prediction compared to training from scratch (R²=0.477 vs 0.005).Machine Learning29 Jul 2026http://arxiv.org/abs/2607.21561v1