Episode 7 - MSMs with deeptime

Previous All episodes

Table of contents

Duration

Objectives

Content

Guide scripts to prepare data

In this episode we reuse OpenMM Application Layer scripts to build consistent data:

These outputs are processed with deeptime to stay consistent with PyEMMA.

Operator foundations

The Koopman operator acts on an observable $g(\mathbf{x})$ by propagating it in time

\[(\mathcal{K}_\tau g)(\mathbf{x}) = \mathbb{E}[g(\mathbf{x}_{t+\tau}) | \mathbf{x}_t = \mathbf{x}],\]

which in practice is approximated with matrices $K_{ij}$ relating microstates. Diagonalizing $K$ yields eigenvalues $\lambda_k$ and eigenfunctions $\psi_k$, allowing reconstruction of the stationary state

\[\rho(\mathbf{x}) \approx \sum_k \lambda_k \psi_k(\mathbf{x}).\]

deeptime complements this view by estimating the transfer matrix and computing spectral projections to distinguish macrostates.

Deeptime showcase

Guided demo

This episode now follows the Deeptime-managed Alanine dipeptide example (Ala2 Example <https://deeptime-ml.github.io/latest/notebooks/examples/ala2-example.html>__). We imported that notebook into the repository (see the iframe below) so you can open it locally in the same environment we use for the course. It fetches the heavy-atom positions and backbone dihedrals from mdshare, runs the same Torch-let Koopman/TICA flow, and highlights how the spectral models compare with PyEMMA’s timelines.

Exercise

Key points

Notebooks and scripts

Previous All episodes