Variational Generative Modeling for Forecasting Protein Diffusion from Molecular Dynamics
Year: 2026
Type: conference-paper
Abstract: The diffusion of membrane proteins within lipid bilayers occurs over timescales that are challenging to probe with standard molecular dynamics (MD) simulations. Here, we present a family of variational generative models for diffusion forecasting that learn membrane protein motion directly from MD trajectories conducted under periodic boundary conditions. Existing deep learning approaches often fail to maintain realistic protein diffusion over long prediction windows, producing nonphysical drift. To address this, we cast trajectory forecasting as a variational inference problem by introducing a latent variable z that encodes the underlying physical state inferred from the temporal evolution of atomic coordinates. The base Variational AutoRegressive (VAR) model combines 2D spatial encoding via differentiable Gaussian rasterization, recurrent temporal processing, and an autoregressive GRU decoder conditioned solely on z, generating future configurations as stochastic samples from the learned distribution over physical states rather than as deterministic extensions of prior coordinates. Physical realism is reinforced through periodic wrapping, displacement clipping, and a composite ELBO-based loss that includes steric repulsion and center-of-mass drift regularization. VAR captures short-term and spatial dynamics of membrane protein diffusion accurately but exhibits long-term super-diffusion behavior across independent replicas. We further introduce four architectural extensions to address this: VAR-MuLSTM (multi-rate stacked LSTM encoder), VAR-DLS (decoder-side latent sampling), VAR-GMM (Gaussian mixture latent), and VAR-SDE (per-step latent resampling with non-recurrent decoder). VAR-SDE improves the displacement distributions captured by the Van Hove function, while VAR-GMM mitigates long-time superdiffusive bias, indicating that fully resolving long-term diffusion behavior remains an open challenge. The source code is avaliable at https://github.com/LynaLuo-Lab/GenerativeProteinDiffusion
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: hybrid
OpenAlex ID: W7171560133