Spectral Map: Deep-Learning CVs for MD

• Spectral Map is an unsupervised deep-learning framework that constructs collective variables from high-dimensional MD data by maximizing separation between slow and fast dynamics.
• It builds a diffusion-map-like Markov transition matrix using anisotropic kernel density estimates to optimize the spectral gap reflecting metastable state separation.
• The method integrates learned CVs into enhanced-sampling techniques, offering robust reaction coordinates for applications like protein folding and conformational changes.

Searching arXiv for the core Spectral Map papers to ground the article in the molecular-dynamics literature and its immediate extensions. arXiv search query: (Gökdemir et al., 2024) OR (Rydzewski, 2024) OR (Rydzewski, 2024) Spectral Map (SM) is an unsupervised deep-learning framework for constructing collective variables (CVs) from molecular dynamics (MD) data by explicitly maximizing the separation between slow and fast dynamical modes in a learned low-dimensional representation. In this usage, SM maps high-dimensional configurations $x$ to low-dimensional coordinates $z=\xi_w(x)$, builds a diffusion-map-like Markov transition matrix in the CV space, and optimizes the mapping so that the spectral gap of that matrix is large; the resulting coordinates are intended to resolve rare, metastable transitions and to be directly usable in CV-based enhanced-sampling methods (Gökdemir et al., 2024, Rydzewski, 2024).

1. Problem setting and conceptual scope

In atomistic MD, many processes of interest, including conformational changes, nucleation, and folding, are rare events: transitions between metastable states are separated by high free-energy barriers and occur on timescales much longer than typical simulation windows. Because the dynamics are high-dimensional and only a few slow modes drive these transitions, a standard strategy is to introduce CVs that provide a low-dimensional representation of the coordinates, resolve metastable states and barriers, and capture slow dynamical processes. The difficulty is that CV choice is typically heuristic and often relies on physical intuition, for example radii of gyration or selected distances, and may therefore fail to capture the dominant slow kinetics (Gökdemir et al., 2024).

SM addresses this CV-selection problem by learning CVs directly from configurations in an unsupervised way. It does not require labeled states, reaction coordinates, or committor values. Its organizing principle is dynamical rather than purely geometric: instead of maximizing variance or reconstructing inputs, it seeks a reduced representation in which slow modes are clearly separated from faster ones through a spectral criterion. This places SM within the broader transfer-operator and manifold-learning tradition, but with the distinctive feature that the embedding itself is trainable rather than fixed (Rydzewski, 2024).

2. Spectral formulation

The theoretical basis of SM is the spectral analysis of Markovian or transfer operators. For stochastic dynamics with transition operator $\mathcal{T}$, eigenvalues close to $1$ correspond to slowly decaying modes associated with metastable structure. If a propagator or transition matrix has eigenvalues

$1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$

then the gap between consecutive eigenvalues encodes timescale separation; in particular, a large gap between $\lambda_{m-1}$ and $\lambda_m$ suggests $m$ well-separated metastable states (Gökdemir et al., 2024).

SM learns a parametric map

$$\mathbf{z}=\xi_w(\mathbf{x})\equiv \{\xi_k(\mathbf{x},w)\}_{k=1}^d,$$

from the high-dimensional configuration space to a $d$-dimensional CV space. On the mapped samples $z=\xi_w(x)$0, it constructs a Gaussian kernel

$z=\xi_w(x)$1

a kernel-density estimate

$z=\xi_w(x)$2

an anisotropic diffusion kernel

$z=\xi_w(x)$3

and a row-normalized transition matrix

$z=\xi_w(x)$4

The anisotropic normalization is inherited from diffusion maps and is intended to encode density information while mitigating bias from nonuniform sampling in CV space (Gökdemir et al., 2024).

The central objective is the spectral gap

$z=\xi_w(x)$5

where $z=\xi_w(x)$6 is the expected number of metastable states in the reduced space. SM trains $z=\xi_w(x)$7 by maximizing

$z=\xi_w(x)$8

Large $z=\xi_w(x)$9 close to $\mathcal{T}$0 indicate slowly decaying modes, while a large $\mathcal{T}$1 indicates clear separation between those modes and the remaining fast dynamics. In the note-based presentation, this construction does not explicitly require time-lagged pairs; the Markovian structure is induced by the kernel in CV space, in a diffusion-map-style manner rather than through explicit lagged transitions $\mathcal{T}$2 (Rydzewski, 2024).

3. Representation and optimization

The learned CVs are represented by a standard feed-forward neural network

$\mathcal{T}$3

with trainable parameters $\mathcal{T}$4. The exact architecture is not fixed by the method; the note emphasizes only that it is parametric, differentiable, and suitable for gradient-based optimization with backpropagation. Because the output is a direct parametric map, the trained CVs can be evaluated on new configurations and used online during simulation (Gökdemir et al., 2024).

The training loop is defined operationally. For each epoch, the current network maps all configurations to $\mathcal{T}$5. Within each batch, SM constructs the Gaussian kernel, the density estimate, the anisotropic kernel, and the transition matrix $\mathcal{T}$6; computes the eigenvalues of $\mathcal{T}$7; evaluates the spectral gap $\mathcal{T}$8; and updates $\mathcal{T}$9 to maximize $1$0 through backpropagation. The objective is global at the batch level because it depends on the spectrum of a matrix built from all mapped points in the batch. The note does not specify the optimizer, but explicitly remarks that standard stochastic gradient methods such as Adam are natural choices (Gökdemir et al., 2024).

Three hyperparameters are structurally central. The output dimension $1$1 fixes the number of learned CVs. The integer $1$2 enters directly into the gap definition and therefore encodes the anticipated number of metastable states. The kernel width $1$3 controls the locality of the diffusion kernel and can strongly affect the constructed spectrum. Practical training uses minibatches for scalability; the note states that kernel construction and eigendecomposition scale as $1$4 and $1$5, respectively, for batch size $1$6. A reference implementation is available through PLUMED-NEST as plumID:24.005 (Gökdemir et al., 2024).

4. Relation to neighboring methods

SM sits at the intersection of spectral manifold learning, kinetic dimensionality reduction, and CV optimization. Its closest nonparametric antecedent is diffusion maps: both use Gaussian kernels and anisotropic density correction, and both interpret eigenstructure in terms of slow diffusive modes. The difference is that diffusion maps operate in a fixed input representation and often use leading eigenfunctions as embedding coordinates, whereas SM constructs the kernel in a learned CV space and uses the spectrum only as a scoring function for the trainable map $1$7. This makes SM a parametric, learnable diffusion-map-like method rather than a direct eigencoordinate method (Gökdemir et al., 2024).

The note explicitly compares SM to parametric dimensionality reduction approaches such as reweighted stochastic embedding, which also use neural networks to construct CVs. The distinction is the specific objective: SM maximizes a spectral gap defined from an anisotropic diffusion kernel. It is also related in spirit to spectral gap optimization of order parameters (SGOOP), since both optimize a spectral gap to isolate slow CVs. The difference is that SGOOP starts from a finite set of hand-chosen order parameters, whereas SM maps full configurations directly to CVs and optimizes the mapping end-to-end (Gökdemir et al., 2024).

By contrast, TICA, time-lagged autoencoders, VAMPnets, and state-predictive information bottleneck use time-lagged trajectory pairs and optimize objectives tied more directly to transfer operators at lag time $1$8. SM, in the presentation of the 2024 papers, is lag-time-free at the objective level: it does not require choosing $1$9, and its transition matrix arises from geometric proximity in the learned CV space rather than from explicit transition counts between $1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$0 and $1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$1. This suggests that SM occupies a distinct position between geometry-based spectral clustering and kinetics-oriented representation learning (Rydzewski, 2024).

5. Markovianity, coarse-graining, and reaction-coordinate interpretation

A major extension of the framework introduces kinetic coarse-graining of the learned reduced space and uses it to define transition state ensembles. In this formulation, SM is not only a CV-learning method but also a route to a coarse-grained Markov description. The reduced dynamics are interpreted through an overdamped diffusion picture on a free-energy landscape $1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$2, and a Markov transition matrix in the learned space is propagated and coarse-grained to partition the reduced space kinetically into metastable states and overlap regions associated with transition states (Rydzewski, 2024).

This extension argues that slow CVs learned by SM closely approach the Markovian limit for an overdamped diffusion. For the FiP35 protein-folding example, a Bayesian test based on the transition-state probability $1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$3 yields $1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$4, close to the Markovian overdamped limit $1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$5. The same study reports that coordinate-dependent diffusion coefficients only slightly affect the constructed free-energy landscapes, with the average discrepancy between $1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$6 and the diffusion-corrected landscape below $1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$7. In that example, a single slow CV learned by SM produces a free-energy barrier of about $1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$8 between folded and unfolded basins and is presented as a physical reaction coordinate for folding (Rydzewski, 2024).

The same work also uses spectral-gap-based analysis to quantify feature importance. Individual structural features, such as pairwise C$1=\lambda_0 \ge \lambda_1 \ge \lambda_2 \ge \dots,$9 distances, can be ranked by the spectral gap they induce when treated as one-dimensional coordinates, and residue-level importance can then be aggregated from those feature-level scores. This provides a mechanism for comparing learned slow CVs with structural descriptors such as the fraction of native contacts. In the FiP35 analysis, the SM CV achieves a larger spectral gap than the native-contact coordinate, reinforcing the claim that SM can separate slow and fast modes more effectively in that system (Rydzewski, 2024).

6. Demonstrations, implementation limits, and terminological breadth

The first SM paper demonstrates the method on three reversible folding systems. For chignolin, trained on $\lambda_{m-1}$0 samples from a $\lambda_{m-1}$1 trajectory, the learned two-dimensional CV space exhibits a folded basin and an unfolded basin separated by a barrier of about $\lambda_{m-1}$2 kJ/mol; for $\lambda_{m-1}$3, the reported spectral gap is about $\lambda_{m-1}$4. For trp-cage, trained on $\lambda_{m-1}$5 samples from a $\lambda_{m-1}$6 trajectory, scanning $\lambda_{m-1}$7 yields the largest spectral gap at $\lambda_{m-1}$8, with $\lambda_{m-1}$9, whereas $\lambda_m$0 and $\lambda_m$1 give $\lambda_m$2 and $\lambda_m$3, respectively. For BBA, also trained on $\lambda_m$4 samples from a $\lambda_m$5 trajectory, the learned CVs reveal folded, unfolded, and misfolded $\lambda_m$6-hairpin states; for $\lambda_m$7, the reported spectral gap is about $\lambda_m$8 (Rydzewski, 2024).

These examples define the practical strengths and limits of the method. SM requires adequate sampling of all relevant metastable states; otherwise the empirical $\lambda_m$9 can be noisy and the optimization may miss states or overemphasize sampled basins. It is sensitive to the expected number of metastable states $m$0, to the kernel scale $m$1, and to the chosen output dimension $m$2. Poor sampling, limited CV dimension, or mis-specified hyperparameters can degrade interpretability or produce unstable spectra. The note also identifies overfitting to finite data and poor extrapolation outside sampled regions as standard deep-learning caveats. At the same time, because the learned CVs are parametric and differentiable, they can be evaluated on the fly and plugged into metadynamics, umbrella sampling, variationally enhanced sampling, and related CV-based enhanced-sampling algorithms (Gökdemir et al., 2024).

The expression “spectral map” is also used in several unrelated technical senses outside molecular simulation. In astronomy it can denote a Self-Organizing Map that clusters $m$3-dimensional stellar spectra on a two-dimensional grid (Mahdi, 2011), a radio spectral index map between $m$4 MHz and $m$5 GHz (Gasperin et al., 2017), or a spectral cube in Spitzer/IRS and SPHEREx mapping pipelines (Donnelly et al., 13 Dec 2025, Cukierman et al., 26 Mar 2026). In matrix and operator theory it can denote the map $m$6 associated with weighted Cauchy matrices, proved to be an involution (Pushnitski et al., 25 Apr 2025). In singularity theory it can refer to the assignment $m$7 of an image-computing spectral sequence to a map $m$8 (Cisneros-Molina et al., 2019). In non-Hermitian quasicrystals it denotes a butterfly spectral map encoding localization and spin alignment (Padhi et al., 8 Jan 2026). In dimer integrable systems it denotes the spectral transform from a dimer cluster Poisson variety to spectral-curve data (George et al., 2022). This suggests that “Spectral Map” is a field-dependent label rather than a single cross-disciplinary object; in current MD literature, however, it has the specific meaning of a spectral-gap-maximizing CV-learning framework (Gökdemir et al., 2024).