🧬 biology

Information-Geometric Audits of Macromolecular Ensembles: A Prescriptive Protocol for Latent Dimension Selection

This paper introduces an information-geometric framework that reframes latent dimension selection for macromolecular ensembles as a thermodynamic optimization problem, deriving a three-stage protocol to identify the optimal dimensionality by quantifying compression costs and geometric distortions, thereby outperforming existing statistical heuristics in recovering physically meaningful representations.

Original authors: Hesam Mozafarynia

Published 2026-06-29

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hesam Mozafarynia

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Problem: Too Much Noise, Not Enough Signal

Imagine you are trying to describe the movement of a complex machine, like a clock with thousands of tiny gears, springs, and screws. If you tried to track the position of every single screw (there might be 10,000 of them), you would drown in data. Most of those screws are just vibrating slightly because of heat; they aren't actually driving the clock's main function.

In the world of biology, scientists study huge molecules (like proteins) that act like these machines. They have thousands of atoms moving around. The goal is to find the "secret sauce"—the few key movements that actually matter for how the molecule works (like folding or unlocking a door). This process is called Dimensionality Reduction. It's like trying to summarize a 1,000-page book into a 2-page summary.

The Catch: Scientists have been guessing how many pages to keep in that summary. They use rules of thumb like, "Keep enough pages to cover 95% of the words" or "Stop when the story stops changing." But the authors of this paper say these rules are flawed. They might keep too many pages (including the boring noise) or too few (missing the plot twists).

The Solution: A "Thermodynamic Audit"

The authors propose a new, three-step protocol to figure out the exact right number of pages to keep. They treat this not just as a math problem, but as a physics problem. They ask: "What is the physical cost of throwing away information?"

Here is how their three-step "audit" works:

Step 1: The "Thermodynamic Bill" (Global Cost)

Imagine you are compressing a file. Every time you delete a piece of data, you are essentially "erasing" information. In physics, erasing information costs energy (think of it as a tax you have to pay to the universe).

The Analogy: Think of the molecule as a messy room. Cleaning it (compressing it) costs energy. If you throw away a toy that was actually important, you pay a huge energy penalty later because you can't find it when you need it. If you throw away a dust bunny, the penalty is tiny.
The Test: The authors calculate a "Compression Free Energy" bill. They check: "If we reduce the molecule to d dimensions, how much energy (or 'information tax') do we lose?"
The Result: They look for an "elbow" in the graph. At first, adding more dimensions saves a lot of energy (because you are keeping the important stuff). Suddenly, the bill stops dropping. That "elbow" tells you: "Okay, we've kept all the important stuff; anything after this is just dust bunnies."

Step 2: The "Curvature Map" (Where is the distortion?)

Sometimes, the total "bill" looks okay, but the map is still wrong. Imagine you are folding a flat map of the world into a tiny ball. You might get the total area right, but you've squashed the continents together so that Africa is touching North America. That's a geometric disaster.

The Analogy: Think of the molecule's shape as a landscape with hills (stable states) and valleys (transitions). When you compress the data, you are flattening this landscape.
The Test: The authors use a tool called Riemannian Curvature to look at the map. They are looking for "hot spots"—places where the map is being stretched or twisted too much.
The Result: If the map has a sharp, fiery red spot where two different valleys are being squashed into one, the compression failed. Even if the total energy bill looked okay, the map is broken. This step tells you where the compression is going wrong.

Step 3: The "Sweet Spot" (The Phase Transition)

Now you have a list of candidates. Some are too small (missing the plot), some are too big (full of noise). You need the perfect middle ground.

The Analogy: Imagine tuning a radio. You turn the dial slowly. At first, you hear static. Then, suddenly, the music comes in crystal clear. Then, if you keep turning, you hear the next station and the music gets fuzzy again. You want to stop exactly where the music is clearest.
The Test: The authors look at the "Rate-Distortion Curve." This is a graph showing the trade-off between "how much we compressed" vs. "how much the picture got blurry."
The Result: They look for the sharpest bend in the curve. This is the "Phase Transition." It's the exact moment where you stop filtering out harmless noise and start throwing away the actual story.

The Proof: The "Müller-Brown" Test

To prove their method works, they tested it on a mathematical model called the Müller-Brown potential.

The Setup: They created a fake molecule with a known "true" answer. They knew for a fact that this molecule only needed 2 dimensions to describe its main movements.
The Challenge: They added 48 extra dimensions of pure random noise (like static on a TV) to trick the computer.
The Old Way: Standard methods (like looking at variance or gaps in the data) got confused by the noise. They suggested keeping 7 to 9 dimensions, thinking the noise was important.
The New Way: The authors' three-step audit ignored the noise, spotted the geometric distortions, and correctly identified that 2 dimensions was the perfect answer.

Why This Matters

This paper doesn't just give a number; it gives a physical reason for that number.

Old Way: "We kept 9 dimensions because the graph looked like it stopped changing."
New Way: "We kept 2 dimensions because: (1) It pays the lowest energy tax, (2) It doesn't squash the important valleys together, and (3) It sits exactly at the point where we stop throwing away the story."

The authors also found that different compression tools (like PCA, tICA, and VAEs) distort the map in different ways. Their method can tell you which tool is best for a specific job, not just by how well it reconstructs the image, but by how faithfully it preserves the molecule's physical "soul."

In short: They turned a guessing game into a precise physics experiment, ensuring that when scientists simplify complex molecules, they don't accidentally throw away the most important parts.

Technical Summary: Information-Geometric Audits of Macromolecular Ensembles

Problem Statement
The computational analysis of macromolecular ensembles relies on dimensionality reduction (DR) to extract low-dimensional latent representations ( $z \in \mathbb{R}^d$ ) from high-dimensional configuration spaces ( $x \in \mathbb{R}^{3N}$ ). A critical, unresolved question in the field is determining the physically necessary number of latent dimensions ( $d$ ). Current selection heuristics—such as explained variance (PCA), spectral gaps (tICA), or cross-validated kinetic accuracy (VAMP/GMRQ)—are calibrated against statistical proxies or specific kinetic fidelities. These methods fail to account for the thermodynamic cost of information lost during compression. Consequently, they may select dimensions that capture irrelevant thermal noise or, conversely, discard macroscopic collective variables, leading to qualitatively incorrect predictions of folding rates, allosteric communication, and free energy landscapes. There is a lack of a principled framework that connects dimension selection directly to the thermodynamics and geometry of the molecular ensemble.

Methodology: A Three-Stage Prescriptive Protocol
The authors propose an information-geometric framework that reframes latent dimension selection as a physical problem solvable via a three-stage protocol. This approach integrates information geometry and stochastic thermodynamics to evaluate candidate dimensionalities based on global thermodynamic cost, local geometric distortion, and rate-distortion optimization.

Stage I: Thermodynamic Screening via Compression Free Energy
The protocol defines the compression free energy, $\Delta F_{\text{compression}}(d) = k_B T D_{KL}(P \parallel Q_d)$ , where $P$ is the true Boltzmann ensemble and $Q_d$ is the $d$ -dimensional reconstruction. This quantity is identified as the Kullback–Leibler divergence between distributions, which, via the generalized Landauer bound, represents the minimum thermodynamic work required to execute the compression.
- Procedure: The marginal thermodynamic gain, $\delta F(d)$ , is calculated for increasing $d$ . A "thermodynamic elbow" is identified where the marginal gain drops below a noise floor (e.g., a tenfold decrease). This filters out dimensions that are either too small (discarding macroscopic modes) or too large (overfitting thermal noise), producing a set of thermodynamically admissible candidates ( $D_I$ ).
Stage II: Geometric Audit via Riemannian Curvature
The scalar $\Delta F_{\text{compression}}$ is a spatial average and cannot reveal where in the latent space the compression fails. To resolve this, the protocol computes the Fisher information metric ( $g_{ij}$ ) on the latent manifold and its associated scalar curvature ( $R(z)$ ).
- Procedure: The curvature field identifies regions where the DR framework is forced to geometrically distort the true energy landscape (e.g., conflating transition states or flattening barriers). Two statistics are computed: the peak curvature ratio ( $\rho$ ) and the hot-spot coverage fraction ( $\eta$ ). Candidates are rejected if curvature is too sharp or extensive, indicating that the dimensionality is topologically insufficient to resolve critical transition corridors. This yields a geometrically admissible set ( $D_{II}$ ).
Stage III: Ranking via Rate-Distortion Optimization
The final stage identifies the optimal dimensionality $d^*$ by locating the critical phase transition on the empirical rate-distortion curve.
- Procedure: The protocol evaluates the trade-off between the compression rate (mutual information $I(X; Z)$ ) and the energy distortion ( $E[|U(x) - U(\hat{x})|^2]$ ). The optimal $d^*$ is defined as the point of maximum curvature (the "elbow") on this curve, representing the transition from filtering uncoordinated thermal noise to actively discarding macroscopic collective variables. A composite physical score ranks the surviving candidates.

Key Contributions

Theoretical Framework: The paper establishes $\Delta F_{\text{compression}}$ as the canonical physical cost measure for latent compression, rigorously linking representation error to the Landauer bound on irreversible information erasure. It further derives the Fisher information metric and Riemannian scalar curvature as the local differential decomposition of this global penalty.
Methodological Protocol: It introduces a method-agnostic, three-stage protocol with explicit estimators for finite simulation data. Unlike previous approaches, it simultaneously enforces thermodynamic fidelity (via $\Delta F$ ) and geometric/topological fidelity (via curvature), rather than optimizing for variance or kinetic rates alone.
Comparative Diagnostics: The framework provides a physically interpretable basis for comparing DR methods (PCA, tICA, VAE) by revealing method-specific distortion profiles (e.g., distributed bias in PCA, timescale prioritization in tICA, and near-isometric mapping in VAE) that are invisible to scalar selection criteria.

Results
The protocol was validated on the high-dimensional embedded Müller–Brown potential, a system with analytically known ground truth $d^* = 2$ (representing two independent transition pathways) embedded in $D=50$ dimensions with 48 Gaussian noise coordinates.

Performance: The protocol unambiguously recovered $d^* = 2$ across three distinct compression regimes: linear (PCA), kinetically motivated (tICA-like), and nonlinear generative (VAE-like).
Comparison with Heuristics: Existing heuristics failed to consistently identify the correct dimensionality. The explained variance criterion selected $d = 7–9$ , and the spectral gap heuristic failed on two of the three methods.
Diagnostic Insights:
- Stage I correctly identified $d=1$ and $d=2$ as thermodynamically significant but could not reject $d=1$ alone.
- Stage II rejected $d=1$ by detecting severe geometric distortion (conflation of two saddle points into a single latent coordinate), evidenced by high curvature ratios and coverage fractions.
- Stage III confirmed $d=2$ as the global optimum, situated at the sharp phase transition between macroscopic information capture and noise absorption.

Significance and Claims
The paper claims to provide the first rigorous equivalence between machine learning representation error and physical thermodynamic dissipation. By establishing a common language rooted in statistical mechanics, the framework allows latent molecular representations to be audited, compared, and improved based on physical principles rather than statistical proxies. The authors assert that this protocol transforms dimension selection from a hyperparameter tuning exercise into a falsifiable physical claim about which degrees of freedom are relevant to the system's biology. The work is presented as a necessary foundation for the analysis of increasingly complex molecular simulations and deep generative models, offering a prescriptive method to ensure that latent representations faithfully preserve the thermodynamic and kinetic structure of the ensemble.