🧬 biology

GRP: A Representation-Aware Effective Geometry of Protein State Space

This paper introduces GRP, a general-relativity-inspired framework that models protein state spaces as coordinate-aware geometric manifolds to identify stable invariants linking AI representations to biological constraints, demonstrating that a few dominant modes capture most conformational variance and that structured biological paths exhibit distinct transport properties compared to null models.

Original authors: Chuanyang Liu

Published 2026-09-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Chuanyang Liu

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Proteins are the workhorses of life, tiny molecular machines that fold into intricate shapes to build cells, fight infections, and drive metabolism. For decades, scientists have studied them by looking at their static blueprints: the sequence of amino acids that makes them up, and the single, frozen three-dimensional structure they often adopt. But this view is incomplete. Proteins are not rigid statues; they are dynamic entities that constantly shift, breathe, and explore a vast landscape of possible shapes. They exist as ensembles of many slightly different forms, moving between states to perform their functions. Understanding how a protein's sequence dictates this entire landscape of possibilities, rather than just one snapshot, is the next great frontier in biology. If we can map the rules that govern these shifting shapes, we could predict how proteins behave, how they react to mutations, and how to design new ones from scratch.

A new study introduces a fresh way to look at this complex world, treating proteins not as isolated objects but as travelers moving through a structured space of possibilities. The researcher, Chuanyang Liu, developed a framework called GRP. Instead of trying to find a single perfect shape for a protein, they asked a different question: what are the underlying rules that organize all the shapes a protein can take? To answer this, they used artificial intelligence not as a black box to predict outcomes, but as a powerful microscope. They fed the AI vast amounts of protein data, allowing it to learn a mathematical "map" of protein states. In this map, every possible shape of a protein is a point, and the distance between points represents how different those shapes are. By analyzing the geometry of this map, they discovered that the chaos of protein movement is actually governed by a hidden, orderly structure.

The researcher examined eight different proteins, generating thousands of possible shapes for each one using advanced computer simulations. They then used the AI to translate these shapes into a common mathematical language. What they found was striking: despite the complexity of each protein, the vast majority of their movement could be described by just a handful of dominant patterns. For the proteins they studied, between three and eighteen specific patterns of movement captured 95% of all the variation. It is as if a complex, multi-dimensional dance could be reduced to a few fundamental steps that repeat in different combinations. This suggests that the space of possible protein shapes is not a random cloud, but a tightly organized landscape with a specific, low-dimensional geometry.

The study also explored how proteins react when they are nudged, such as when a single amino acid is changed by a mutation. In the past, scientists might have measured how far a protein moved after such a change. However, the researcher realized that a large movement in a direction where the protein naturally wiggles a lot is less significant than a small movement in a direction where the protein is usually very stiff. To fix this, they created a new way to measure "cost." They compared the size of a mutation's effect against the natural background noise of the protein's movement. If a mutation pushed the protein in a direction it rarely goes, the cost was high. If it pushed the protein along a path it frequently travels, the cost was low. When they tested this against real-world data from deep mutational scanning—where scientists measure how thousands of mutations affect a protein's ability to function—they found a clear link. Mutations that the model flagged as "expensive" or atypical were indeed the ones that tended to hurt the protein's function. This confirms that the mathematical geometry of the AI map reflects real biological constraints.

Perhaps the most compelling finding concerns the paths proteins take when they change from one state to another. The researcher looked at ordered sequences of states, such as a protein growing longer or undergoing a series of mutations, and compared them to random, scrambled sequences of the same length. They measured the "transport cost" of moving between these states, which essentially calculates how much the protein's shape distribution has to shift to get from one point to the next. The results were consistent and strong: the real, biologically structured paths required significantly less transport cost than the random, scrambled paths. In the primary comparison, the structured paths had a transport cost of 11.75, while the random paths averaged 22.90. This means that nature's pathways through the landscape of protein shapes are remarkably efficient, following the smooth, low-energy valleys of the map rather than climbing over random hills.

The study does not claim to have discovered a universal law of physics that governs all proteins in the same way gravity governs falling apples. The author is careful to state that their findings are specific to the representations and models they used, and that the "geometry" they describe is a statistical description of how AI sees protein data, not necessarily a direct physical force. However, the consistency of the results across different proteins, different types of mutations, and different ways of measuring the data suggests that this geometric organization is a real feature of how proteins work. The framework successfully bridges the gap between the abstract world of AI representations and the tangible world of biological function. It shows that by treating proteins as distributions of possibilities rather than single points, we can uncover the hidden rules that organize their behavior.

This work represents a shift in how we think about molecular biology. Instead of cataloging individual structures, the focus moves to the space of possibilities itself. The study suggests that the sequence of a protein acts as a set of constraints that carve out a specific, efficient route through the vast landscape of potential shapes. The AI serves as a tool to visualize and measure this landscape, revealing that the paths nature chooses are not random but follow a principle of efficiency. While the researcher cautions that this is a preliminary step and that more experimental validation is needed, the results offer a powerful new perspective. They provide a testable, mathematical language for describing how proteins move, change, and function, turning the chaotic dance of molecular life into a map that can be read, understood, and potentially navigated.

Technical Summary: GRP – A Representation-Aware Effective Geometry of Protein State Space

Problem Statement
Proteins are not static structures but multilevel state spaces defined by the coupling of sequence, structure, conformational ensembles, and trajectories. Traditional biophysical approaches often treat these levels in isolation or rely on static, a priori choices of coordinates and potentials (e.g., elastic networks), which may obscure the organization of the system's space of possibilities. There is a need for a theoretical framework that unifies these levels as related components of a single state space and provides a mathematically tractable way to describe how biological constraints shape the geometry of this space. Furthermore, while AI-based molecular representations (e.g., protein language models) encode substantial structural and evolutionary information, their relationship to physical reality is often treated as a "black box" rather than a scientific instrument for discovering invariant relations.

Methodology: The GRP Framework
The author introduces GRP (General Relativity of Protein), a framework inspired by General Relativity that shifts focus from coordinates to relations that remain invariant under changes in representation, scale, or model. GRP treats AI representations as a "lossy mirror space" where biological constraints leave statistical traces. The methodology proceeds through four distinct but connected mathematical constructions:

Local Effective Geometry (Ensembles):
- Protein states are defined as empirical probability measures over observations (conformations, mutations, or learned embeddings).
- For Cartesian ensembles, rigid-body alignment (Kabsch) removes global translation/rotation.
- Covariance matrices are estimated from centered coordinates. The framework analyzes the covariance spectrum (eigenvalues) to define local statistical summaries such as participation ratio, anisotropy, and spectral decay. These metrics describe the organization of variance within the declared representation space, not intrinsic physical dimensions.
Covariance-Normalized Perturbation Geometry:
- To quantify the "atypicality" of a perturbation (e.g., a mutation), the framework calculates a regularized Mahalanobis-type cost.
- Using the Ledoit-Wolf estimator, the inverse of the regularized covariance matrix of the unperturbed background defines a local metric ( $G_S$ ).
- The geometric cost $C_{geo}$ is the squared displacement of the perturbed state projected into this local metric. This normalizes displacements by the background's variability: large displacements along highly variable directions are penalized less than small displacements along constrained directions.
Biological Constraint Fields (Source-to-Geometry):
- The framework models the relationship between biological source variables (sequence length, amino acid composition, physicochemical descriptors) and geometric responses (ensemble organization metrics) as a conditional statistical field.
- This is formulated as a predictive model where biological descriptors predict the organization of the ensemble geometry. The author tests for a "shared" response component across systems versus system-specific deviations, using grouped hold-out validation to ensure transferability.
Ordered State Paths and Transport:
- Ordered sequences of states (e.g., mutation trajectories or chain-length progressions) are represented as paths through the state space.
- The framework calculates the mean adjacent-state Gaussian transport (a surrogate for Wasserstein distance) between consecutive states in a common representation space.
- This statistic is compared against matched null paths (randomized paths with identical endpoints and constraints) to test the hypothesis that structured biological paths occupy more coherent, lower-transport routes through the representation space.

Key Results

Covariance Concentration: Across diverse protein ensembles, conformational variance is concentrated in a limited number of empirical modes. Approximately 3–18 modes capture 95% of the retained variance, depending on the protein and representation.
Chain-Length Scaling: Participation ratios and spectral decay parameters show reproducible scaling with chain length and residue composition (e.g., stronger expansion for PolyG and PolyLys), indicating systematic reorganization of ensemble geometry rather than simple rescaling.
Perturbation Normalization: Covariance normalization improves the directional consistency of perturbations across different protein backgrounds and reduces the coefficient of variation of perturbation costs. While the geometric cost shows a weak negative correlation with experimental deep-mutational scanning (DMS) fitness, it provides information orthogonal to sequence-based baselines.
Predictive Fields: Biological source variables (chain length, composition) predict a substantial portion of ensemble geometry organization ( $R^2 > 0.1$ for 73.6% of responses). The coupling structure between sources and responses is stable across resampling but is not a universal, coordinate-free law; it is representation-dependent.
Low-Transport Paths: Structured protein-state paths exhibit significantly lower mean Gaussian transport compared to matched null paths (e.g., 11.75 vs. 22.90 in primary comparisons). This suggests that biologically ordered paths traverse the representation space more efficiently than random sequences.

Significance and Claims
The paper positions GRP as a preliminary, testable statistical route from AI-observed organization to biological principles, rather than a completed physical theory.

Ontological Shift: The author argues for a shift in molecular biology from studying isolated objects (sequences, static structures) to studying distributions and paths within a constrained state space. A protein is viewed as a structured space of possibilities rather than a point in a pre-existing space.
AI as an Instrument: GRP treats AI representations not as the biological reality itself, but as a scientific instrument (analogous to a telescope) that makes collective structures measurable. The goal is to identify invariants—relations that remain stable across different models, representations, and scales—which can then be translated into testable biological hypotheses.
Modest Scope: The author explicitly states that GRP does not claim to derive physical laws (like Einstein's equations) for proteins, nor does it assert that geometric costs are equivalent to free energies or forces. The framework is "representation-aware," acknowledging that the extracted geometry depends on the chosen representation and sampling protocol.
Future Direction: The work establishes a language for relating sequence variation, ensemble heterogeneity, and mutational response. Its ultimate value lies in generating hypotheses about which geometric relations transfer to biological measurements, pending independent experimental validation.

In summary, GRP provides a mathematical framework to quantify the "effective geometry" of protein state spaces using AI representations, demonstrating that biological constraints manifest as reproducible statistical patterns in covariance, perturbation costs, and path transport.

More like this