#34 · ML / AI methods

Deep generative models learn conformational distributions beyond single static structures

The dominant thread is learning distributions over molecular states rather than single static structures. The dissertation applies VAE and scattering-based embeddings to DNA reaction kinetics and generates synthetic cryo-EM data with Struc2mapGAN, alongside a cryo-EM benchmark review. A separate preprint compares DDPM diffusion and rectified-flow generative models for conformational sampling on a 2D potential, the Trp-cage system, and α-synuclein dihedral distributions, providing a direct test of sampler behavior on protein conformational landscapes. A further preprint extends biologically-informed neural networks to 2D+t reaction-diffusion systems, combining equation learning with symbolic regression. That item addresses spatiotemporal dynamics rather than conformational ensembles and is included for methodological context. What is distinctive here is the head-to-head comparison of generative sampler classes on shared protein benchmarks.