TL;DR
Estimating the long-term kinetics of complex systems (like protein folding) often fails because we don't know which physical features to track. This paper introduces a simple, robust clustering method based on the Kolmogorov-Smirnov (KS) test to "discover" hidden states from noisy data. By using these discovered states as new features, researchers can estimate transition timescales much more accurately—sometimes by an order of magnitude—without needing expert-level intuition about the system's coordinates.
What Problem Does This Paper Address?
In fields like biophysics and chemistry, we want to understand "slow" dynamics—the rare events like a protein finally folding or a chemical bond breaking. The Koopman operator is a mathematical tool that helps us find these slow modes by looking at how observables (like distances between atoms) evolve over time.
However, there’s a catch: the accuracy of a Koopman-based model (like EDMD or TICA) depends entirely on the "basis functions" or observables you choose. If you pick the wrong features, the model will significantly underestimate how long these slow processes actually take. Identifying the "right" features a priori is incredibly difficult for high-dimensional systems. While Hidden Markov Models (HMMs) can help, they are often computationally expensive and difficult to tune.
Methods
The authors propose a procedure called KS clustering to infer "surrogate" observables that act as indicator functions for hidden metastable states.
- Time Partitioning: Divide a long trajectory of observables into small time intervals ($\Delta t$).
- Statistical Comparison: For every pair of intervals, use the two-sample Kolmogorov-Smirnov (KS) test to measure the distance between their distributions. The KS test looks at the maximum difference between two cumulative distribution functions (CDFs).
- Clustering: Use the resulting distance matrix to perform hierarchical agglomerative clustering. This groups time intervals that are statistically similar, effectively identifying when the system is in the same "hidden" state.
- Surrogate Features: Create new "indicator" time series (1 if the system is in cluster $i$, 0 otherwise).
- Augmented EDMD: Feed these surrogate features back into standard Koopman estimation tools (like EDMD).
This approach is "linear" in its final step but captures non-linear state information through the initial clustering.
Datasets
The authors validated their method on two types of data:
- Synthetic Two-State HMM: An analytically solvable model with Gaussian noise. This allowed them to test the method in "clean" vs. "noisy" scenarios where distributions of states overlap significantly.
- Protein Molecular Dynamics (MD): Three well-known protein systems of varying complexity:
- NTL9 (Ribosomal Protein L9)
- BBA (Beta-beta-alpha fold)
- Trp-Cage
- These datasets involved trajectories of at least 200 $\mu$s, using torsional angles and RMSD as base observables.
Key Insights and Ideas
- Statistics over Time vs. Single Points: Standard clustering (like k-means) looks at individual data points. If the distributions of two states overlap (e.g., a "unimodal" distribution in the observable space), k-means fails. KS clustering looks at the distribution over a window of time, which remains distinct even when individual points are ambiguous.
- The "Variational" Rescue: Koopman estimates are variational, meaning they usually provide a lower bound on timescales. By adding indicator functions from KS clustering, the authors "tighten" this bound, bringing the estimated timescales much closer to the true physical values.
- Robustness to Poor Descriptors: Even if you start with a "bad" descriptor (like a single RMSD value that doesn't capture the full folding process), the KS clustering can often extract enough statistical signal to recover the correct slow timescales.
Key Results
- Order of Magnitude Improvement: For the protein systems, the KS-augmented method produced timescale estimates up to 10 times longer (and thus more accurate) than standard EDMD using only raw physical features.
- Unimodal Success: In the synthetic model where state distributions overlapped so much they appeared as a single peak (unimodal), standard methods failed completely. KS clustering, however, correctly identified the two hidden states and their transition times.
- Convergence at Short Lagtimes: Traditional methods often require very long "lagtimes" (the $\tau$ in the Koopman operator) to converge to the right answer. The KS-clustering approach yields accurate results even at very short lagtimes, making it more data-efficient.
Limitations and Future Work
- Loss of Interpretability: The "surrogate" indicator functions are labels for clusters, not explicit physical formulas (like "the distance between atom A and B"). This makes it harder to explain mechanistically what the discovered states represent without further post-hoc analysis.
- Parameter Sensitivity: The method requires choosing the interval length ($\Delta t$) and the number of clusters ($M$). The authors suggest $\Delta t \simeq \tau_M/10$ (where $\tau_M$ is the fastest slow mode), but this still requires some initial estimation.
- Future Directions: The authors suggest that these "labeled" time series could be used to train neural networks or other models to find more interpretable physical coordinates that correspond to the discovered hidden states.
Summary updated