Representation choice shapes the interpretation of protein conformational dynamics

This paper demonstrates that the choice of molecular representation fundamentally shapes the interpretation of protein conformational dynamics, arguing for a comparative framework and introducing the "Orientation" feature and "ManiProt" library to capture complementary aspects of protein motion that no single representation can fully reveal.

The Gist

Imagine you are trying to understand how a complex, shape-shifting machine works. You have a high-speed video camera recording every tiny movement of its gears, springs, and levers. This is what Molecular Dynamics (MD) simulations do for proteins: they record the atomic-level dance of a protein molecule over time.

However, the problem is that the video is too big and too messy. It's like having a 100-hour movie of a busy city intersection, but you only have a tiny screen. If you try to watch it all at once, you see nothing but a blur. Scientists need to "zoom in" and pick a specific way to look at the data to find the story.

This paper argues that how you choose to zoom in changes the story you tell.

The Core Problem: The "Camera Angle" Trap

The authors say that most scientists usually pick just one way to look at the protein (one "representation"). It's like trying to understand a person's personality only by looking at their shadow.

• If you look at the shadow (Cartesian coordinates), you see the overall shape.
• If you look at the skeleton (torsion angles), you see how the joints bend.
• If you look at the clothing (point clouds), you see the surface texture.

The paper shows that if you only use one "camera angle," you might miss the whole story. A protein might look like it's doing one thing in the shadow, but something completely different in the skeleton view.

The New Solution: "Orientation Features"

The authors introduce a new way to watch the movie called Orientation Features.

The Analogy: The Dancer's Local Compass
Imagine a ballet dancer spinning on stage.

• Old Way (Cartesian): You track the exact X, Y, and Z coordinates of her nose. If she spins, her nose moves in a circle, but it's hard to tell if she's just turning or actually changing her pose.
• New Way (Orientation): Instead of tracking her nose in the room, you attach a tiny, invisible compass to her shoulder. You only watch how that compass rotates relative to her body.

This new method, called Orientation Features, treats every part of the protein like a tiny compass. It ignores the fact that the whole protein might be spinning in the room (which is boring and unimportant) and focuses entirely on how the individual parts are twisting and turning relative to each other.

What They Found: Three Different Stories

The researchers tested this new method on three different types of protein "movies" and found that different camera angles revealed different secrets:

1. The Fast-Folding Protein (The Origami):

   • The Story: A protein folding itself up like a piece of paper.
   • The Discovery: The new method was great at spotting the tiny, subtle twists that happen before the protein fully folds. Other methods missed these "intermediate" steps, thinking the protein was just wobbling randomly. The new compass view saw the specific turns that mattered.

2. The Giant Helicase (The Moving Crane):

   • The Story: A huge protein that acts like a crane, opening and closing its arms to grab DNA.
   • The Discovery: The new method saw that the "arms" of the crane didn't just move in and out; they twisted in a specific way to lock onto their target. Old methods saw the arms moving but missed the crucial twisting motion that made the lock work.

3. The Protein Handshake (The Association):

   • The Story: Two proteins finding each other and sticking together.
   • The Discovery: When proteins meet, they don't always change shape dramatically. Sometimes they just nudge each other. The new method was sensitive enough to detect the tiny rotation of the "fingers" (amino acids) as they locked hands, whereas other methods thought nothing was happening because the overall shape didn't change much.

The Takeaway: Don't Just Pick One Lens

The authors built a free software tool called ManiProt (like a Swiss Army knife for protein data) that lets scientists switch between these different "camera angles" instantly.

The Moral of the Story:
If you want to understand a protein, don't just look at it from one side.

• Sometimes you need to see the shape (where the atoms are).
• Sometimes you need to see the bends (how the joints twist).
• Sometimes you need to see the spin (how the parts rotate).

The paper concludes that no single view is perfect. To get the full picture of how life works at the molecular level, we need to look at the data through multiple lenses simultaneously. The "Orientation" lens is a powerful new addition to the toolbox that helps us see the hidden twists and turns that were previously invisible.

Technical Summary

1. Problem Statement

Molecular Dynamics (MD) simulations generate high-dimensional, atomic-resolution trajectories that are crucial for understanding protein function, folding, and binding. However, extracting interpretable insights from this data is challenging due to:

• High Dimensionality: The data consists of strongly correlated time series where functional motions are often subtle and collective.
• Representation Bias: Standard analysis methods (clustering, dimensionality reduction, kinetic modeling) typically rely on a single representation of molecular motion (e.g., Cartesian coordinates, torsion angles, or pairwise distances).
• Geometric Limitations: Common representations often impose Euclidean assumptions that obscure the nonlinear geometry and rotational constraints inherent to protein conformational space. They may fail to distinguish between global rotations and internal conformational changes or miss specific dynamical features depending on the chosen coordinate system.

The authors argue that the choice of representation is not neutral; it fundamentally shapes the inferred conformational organization, similarity relationships, and apparent transitions.

2. Methodology

A. Introduction of "Orientation Features"

The core methodological contribution is Orientation Features, a geometrically grounded representation for protein backbone dynamics.

• Local Coordinate Systems (LCS): For each residue, a local reference frame is constructed using backbone atoms (N, C$\alpha$, C). This frame is an element of the special orthogonal group SO(3).
• Manifold Formulation: A protein with $n$ residues is modeled as a point in the product manifold SO(3)$^n$.
• Tangent Space Projection: To enable linear analysis (like PCA), rotations are mapped from the Lie group SO(3) to its Lie algebra so(3) (isomorphic to $\mathbb{R}^3$) using the logarithmic map. This converts rotation matrices into 3D vectors representing rotation axes and angles.
• Reference Frames: Two variants are proposed:
  1. Orientation: Uses a single fixed reference structure (e.g., experimental PDB).
  2. Orientation $\odot$: Uses the intrinsic mean of the local coordinate systems for each residue across the trajectory. This captures rotational variability relative to the ensemble average, making it robust to global drift.

B. Comparative Framework

The authors compare Orientation Features against three standard representations:

1. Cartesian Coordinates (C$\alpha$): Absolute atomic positions.
2. Torsion Angles: Dihedral angles ($\phi, \psi$), embedded via sine-cosine encoding.
3. Riemannian Pointcloud: A geometric representation based on the spatial distribution of C$\alpha$ atoms.

C. Computational Tool: ManiProt

To facilitate systematic comparison, the authors developed ManiProt, an open-source Python library. It implements efficient computation of these representations, leveraging NumPy and Numba. The library highlights that Orientation features scale linearly with the number of residues and frames, whereas Pointcloud methods scale quadratically with residues due to pairwise distance calculations.

D. Benchmark Systems

The study evaluates these representations across three distinct dynamical regimes using 23 long MD trajectories from D.E. Shaw Research:

1. Fast-Folding Proteins: Small proteins (e.g., Villin, Trp-Cage, BBA) focusing on local secondary structure formation.
2. Large-Scale Domain Motions: The SARS-CoV-2 nsp13 helicase, focusing on coordinated domain rearrangements.
3. Protein-Protein Association: Systems with minimal internal motion (e.g., Barnase-Barstar, Insulin dimer), focusing on subtle interface rearrangements.

3. Key Results

A. Fast-Folding Proteins

• Kinetic Scores: Torsion angles achieved the highest VAMP-2 scores (indicating strong capture of slow kinetic variance), as folding is driven by dihedral transitions. Orientation features yielded intermediate scores, comparable to Pointcloud and C$\alpha$ coordinates.
• Conformational Partitioning: Different representations resolved different structural states.
  • In BBA and $\alpha$3D, Orientation $\odot$ successfully resolved distinct folded and partially folded states that were merged or indistinguishable in C$\alpha$ or torsion-angle representations.
  • Cross-Orientation Analysis revealed localized rotational changes (e.g., near Pro51 in $\alpha$3D) that were not evident in translational cross-correlation matrices.

B. SARS-CoV-2 nsp13 (Multi-Domain Helicase)

• Structural Similarity: Orientation $\odot$, Pointcloud, and C$\alpha coordinates showed strong correlations with structural metrics (RMSD and lDDT). Torsion angles showed weak correlation with structural similarity.
• Cluster Resolution: Orientation $\odot$ provided complementary discriminative information. It identified distinct clusters (specifically clusters 2, 4, and 6) that C$\alpha$ and Pointcloud representations merged.
• Domain Motions: Orientation $\odot$ captured coordinated motions of the ZB, 1B, and RecA2 domains, revealing intermediate states that were obscured in translational representations.

C. Protein-Protein Association

• Dominant Signal: In systems with minimal internal motion, the rotation axis component ($u$) of Orientation features carried the primary discriminative information, while the rotation angle ($\theta$) contributed minimally.
• System Dependency:
  • Barnase-Barstar: Orientation $\odot$ provided the clearest separation between bound and unbound states, with key residues localizing to the binding interface.
  • RNaseHI-SSB: Torsion angles performed best, consistent with a "folding-upon-binding" mechanism where local backbone rearrangements dominate.
  • Insulin Dimer: Pointcloud and C$\alpha$ coordinates performed best for one trajectory, while Orientation $\odot$ differentiated a non-canonical bound state in another.
• Conclusion: No single representation is universally optimal; the best choice depends on the specific physical mechanism of the association.

4. Key Contributions

1. Orientation Features: A novel, rotation-aware encoding of protein backbone dynamics based on Lie group theory (SO(3)) and tangent space projection, offering a compact and geometrically principled alternative to Cartesian or torsion-based coordinates.
2. Demonstration of Representation Bias: Empirical evidence showing that identical simulation data yields divergent interpretations (different clusters, timescales, and transition pathways) depending solely on the chosen representation.
3. ManiProt Library: A scalable, open-source tool for computing and comparing multiple protein representations, enabling systematic, representation-aware analysis.
4. Complementarity: The finding that different representations emphasize distinct, complementary aspects of conformational space (e.g., torsion angles for local folding, Orientation for domain rotations, C$\alpha$ for global shape).

5. Significance

This work fundamentally shifts the paradigm of MD analysis from seeking a "best" representation to adopting a comparative, representation-aware framework.

• Robustness: It warns against over-interpreting results from a single coordinate system, which may introduce artifacts or miss critical dynamical features.
• Mechanistic Insight: By using multiple representations, researchers can disentangle different physical mechanisms (e.g., distinguishing between global rigid-body motion and local rotational fluctuations).
• Future Directions: The authors suggest that while their current analysis uses linear methods (PCA, TICA) for interpretability, Orientation features are compatible with nonlinear deep learning models (e.g., VAMPnets, graph neural networks), potentially unlocking further insights into complex biomolecular dynamics.

In summary, the paper establishes that representation choice is a critical hyperparameter in MD analysis, and Orientation Features provide a powerful, geometrically grounded tool to capture rotational dynamics that are often missed by conventional methods.