Continuous Directional Statistics for Molecular Conformational Sampling

26 concepts15 papers1 proposal

To model the three-dimensional shapes of proteins and RNA, we must track the rotation of their molecular backbones and side chains. These rotations are measured as dihedral angles ($\phi, \psi, \chi$). Traditionally, structural biology has simplified this continuous landscape by breaking it into discrete pieces. Algorithms rely on fixed fragment libraries for backbones or binned "rotamer" libraries for side chains.

While computationally convenient, discretization introduces severe artifacts. If an actual molecular conformation falls on the border between two arbitrary bins, discrete models suffer from artificial energy jumps and edge effects. More importantly, discretization creates a fundamental resolution limit, blinding models to subtle structural nuances, novel folds, and continuous conformational transitions.

To overcome these limitations, a powerful paradigm has emerged: modeling macromolecular structures using continuous directional statistics on periodic manifolds.

The Geometry of Periodic Manifolds

Why not just use standard continuous tools, like a Gaussian distribution, to model dihedral angles? The answer lies in topology. Dihedral angles are periodic: $0^\circ$ is exactly the same as $360^\circ$. If a probability density is centered near this boundary, a standard Euclidean distribution will treat $-179^\circ$ and $+179^\circ$ as opposite ends of the universe, failing to let the density flow smoothly across the wrap-around point.

To solve this, we must model conformations on their native periodic manifolds. A single angle lives on a circle ($S^1$). A pair of backbone angles ($\phi, \psi$) lives on a two-dimensional flat torus ($T^2 = S^1 \times S^1$). Instead of standard statistics, this paradigm leverages probability distributions explicitly designed for non-Euclidean, periodic spaces:

  • The von Mises distribution: The circular analogue of the normal distribution, which naturally preserves continuity across the $360^\circ$ boundary.
  • Bivariate von Mises & Fisher-Bingham distributions: Multi-dimensional extensions used to model joint distributions on tori or spherical surfaces, capturing the precise spread, stretch, and correlation of orientations.
  • Wrapped Normal distributions: Distributions formed by wrapping a standard normal distribution around a circle or torus, which are highly effective for modeling continuous diffusion processes.

In A generative, probabilistic model of local protein structure, Kanti V. Mardia et al. introduced a framework using the bivariate von Mises distribution on a torus to accurately capture local protein backbones without binning artifacts. Similarly, John T. Kent et al. utilized the 5-parameter Fisher-Bingham distribution in Sampling Realistic Protein Conformations Using Local Structural Bias to sample backbone orientations continuously on a spherical surface. By using these smooth, boundary-free distributions, models can map the true, continuous spread and stretch of molecular shapes.

Continuous Conditioning Beyond Rotamers

The power of continuous directional statistics becomes exceptionally clear when modeling protein side chains. Side-chain flexibility ($\chi$ angles) is tightly constrained by the local backbone configuration. Traditional rotamer libraries handle this by binning the backbone angles into a rigid grid and calculating discrete side-chain probabilities for each grid cell. This suffers from severe data sparsity: many grid cells have few or no experimental examples.

In Beyond rotamers: a generative, probabilistic model of side chains in proteins, Tim Harder demonstrated a solution: continuous conditioning. Instead of a rigid grid, the model conditions side-chain von Mises distributions on the exact, continuous values of the backbone angles. This allows the model to smoothly interpolate across the entire conformational landscape, capturing fine-grained structural coupling while completely bypassing data sparsity issues.

Building on this, Piyumi R. Amarasinghe et al. utilized information-theoretic minimum message length (MML) inference in Getting 'ϕψχal' with proteins: minimum message length inference of joint distributions of backbone and sidechain dihedral angles to learn joint continuous distributions of both backbone and sidechain angles simultaneously. This approach provides a highly concise and physically realistic representation of the full conformational space, allowing researchers to sample favorable conformations that fall entirely between the bins of standard libraries.

Unbiased Sampling and Generative Modeling

Operating in a continuous probabilistic space fundamentally changes how we sample and simulate structures. When structure prediction algorithms use discrete fragment libraries as proposal mechanisms, they introduce unknown statistical biases and unphysical constraints.

By contrast, using explicit continuous distributions as a proposal mechanism enables mathematically rigorous, unbiased sampling in molecular simulations. For example, in A Probabilistic Model of RNA Conformational Space, Jes Frellsen et al. developed BARNACLE, a dynamic Bayesian network (DBN) that uses continuous von Mises distributions to sample RNA structures without fragment libraries, ensuring compliance with the strict requirements of statistical physics. To make these architectures accessible, Martin Paluszewski et al. developed Mocapy++ in Mocapy++ - A toolkit for inference and learning in dynamic Bayesian networks, an open-source toolkit designed specifically for parameter learning and inference in DBNs incorporating directional statistics.

Continuous angular representations also unlock new generative capabilities:

Mathematical Tools for the Flat Torus

As structural biology shifts toward continuous angular representations, researchers have had to develop new statistical tools to analyze and compare these periodic distributions. Standard density estimation and goodness-of-fit tests fail on non-Euclidean manifolds.

To address density estimation, Seyed Morteza Najibi et al. introduced the use of trigonometric splines in Protein Structure Classification and Loop Modeling Using Multiple Ramachandran Distributions, which naturally respect the circular boundaries of Ramachandran plots without artificial constraints.

When analyzing how subtle sequence variations affect local backbone shapes—such as the discovery by Ailie Marx et al. in Codon-specific Ramachandran plots show amino acid backbone conformation depends on identity of the translated codon that synonymous codons correlate with distinct backbone dihedral distributions—researchers require rigorous comparison metrics. To solve this, Javier González-Delgado et al. adapted optimal transport theory to periodic spaces in Two-sample goodness-of-fit tests on the flat torus based on Wasserstein distance and their relevance to structural biology, creating Wasserstein-distance-based tests specifically optimized for the flat torus.

Broader Horizons of Directional Modeling

The core insight of this research direction—that continuous directional modeling outperforms discrete approximations—extends well beyond macromolecular backbones. For instance, in FragBuilder: an efficient Python library to setup quantum chemistry calculations on peptides models, Jan H. Jensen et al. leveraged continuous sampling from Ramachandran distributions to generate highly realistic peptide models for quantum chemistry.

Outside of structural biology, the exact same mathematical machinery solves analogous problems where orientation is inherently continuous:

Whether modeling the fold of a protein, the path of a neural axon, or the orientation of an object in space, treating angular space as a continuous, periodic manifold eliminates the artificial boundaries of the discrete world, revealing the true underlying geometry of the data.

Key Foundational Work

A generative, probabilistic model of local protein structure (2008) — contributed 3 concept elements to this direction.
Sampling Realistic Protein Conformations Using Local Structural Bias (2006) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_continuous_conformational_sampling_cb2a9a52,
  title = {continuous conformational sampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-conformational-sampling/rdr_cb2a9a52a6864cb9b5048af14ea62add},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_continuous_conformational_sampling_cb2a9a52_v1,
  title = {continuous conformational sampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-conformational-sampling/rdr_cb2a9a52a6864cb9b5048af14ea62add/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake