LSS Applied to DEShaw simulations of the BBA (PDB: 1FME)

The collab notebook loads the Ca positions of BBA simulations with stride=50 and three pre-trained models for the SRV, MDN, and DDPM. Walkthrough includes analysis of implied timescales, eigenvector correlations with input features, and differences in metastable states. The LSS_BBA notebook provides training code for the stride=1 data originally used for each model.

SRV:

• Trained on 10 x 100_000 trajectories of BBA with lag=50
• Features are all pairwise distances between Ca positions
• Analysis shows correlations to input features and differentiates metastable states

MDN:

• Trained on the 6 leading eigenvectors learned by the SRV with lag=50
• Synthetic trajectories reproduce the thermodynamics and transition kinetics of training data

DDPM:

• Trained to reconstruct the Ca positions conditioned on a coordinate in SRV space
• Uses a 100-step diffusion process guided by coordinate space
• Analysis shows high-quality structures that adhere to the latent conditioning