Recent developments in protein structure prediction, notably those based on AI, showed that protein models can routinely reach unprecedented levels of near-experimental accuracy. In this context, modeling protein motions in the living cell is becoming more central than ever before.
I will present our recent developments in geometric deep learning in general, and for the prediction of the dynamics of macromolecules in particular. I will show architectures that handle arbitrarily shaped volumetric patterns with operations inherently invariant or equivariant to patterns’ positions and orientations in 3D. When benchmarked on diverse volumetric datasets, they demonstrate superior performance over the baselines with significantly reduced parameter counts. I will discuss the applications of our developments to current biological problems and beyond.