Generalization of energy-guided pretraining across structural changes and multiple poses

Determine the extent to which models pretrained to predict ensemble-averaged MMPBSA energetics from any trajectory frame generalize across changes in three-dimensional structures by distinguishing and correctly ranking multiple docking poses of the same protein-ligand co-complex.

Background

Recent work framed predicting MMPBSA energetics as a pretraining objective encouraging pose and frame invariance, followed by fine-tuning on downstream tasks.

The authors note that this pretraining step was not studied in detail and explicitly state uncertainty regarding how well such models generalize across structural variations, such as distinguishing multiple poses of the same co-complex.