Based on the Variational Approach for Markov Processes (VAMP).
VAMPnets (Variational Approach for Markov Processes networks) are a Deep Learning framework designed to replace traditional Markov State Models (MSM) for molecular kinetics. They map high-dimensional, nonlinear molecular coordinates into a low-dimensional probabilistic state space.
Key Advantages
- End-to-End Learning: It replaces the traditional multi-step pipeline (featurization, dimensionality reduction via TICA, and Markov State Model (MSM) clustering) with a single framework.
- Metastable Decomposition: The output naturally identifies the long-lived chemical states (e.g., folded vs. unfolded protein states) without manual intervention.
- Non-linear Transformations: Unlike TICA, which is a linear projection, VAMPnets can capture highly non-linear kinetic manifolds.
Architecture
The core architecture consists of two parallel Lobes (Twin Neural Networks) sharing identical weights:
- Input Lobe 1: Receives frame at time t (xt).
- Input Lobe 2: Receives frame at time t+τ (xt+τ), where τ is the lag time.
The Transformation (Koopman Operator)
VAMPnets aim to approximate the Koopman Operator, a theoretical operator that governs the evolution of the system.
- Input: Nonlinear molecular coordinates (chaotic, difficult to predict).
- Transformation: The NN acts as a non-linear encoder, warping the data into a Latent Space.
- Output: A probability distribution vector (e.g.,
[0.1, 0.9, 0.0]) representing membership in metastable states. - Result: In this latent space, the dynamics between t and t+τ become approximately linear.
VAMP-2 Loss
The network is trained by maximizing the VAMP-2 Loss.