Mixture models for spherical data with applications to protein bioinformatics

Kanti V. Mardia, Stuart Barber, Philippa M. Burdett, John T. Kent, Thomas HamelryckarXiv 2021 · DOI · arXiv

TL;DR

This paper introduces a robust way to model data that lives on a sphere—like the directions of hydrogen bonds in proteins—by using "Kent distributions." By switching from traditional approximations to exact mathematical calculations, the authors fixed a common glitch in the EM algorithm, allowing them to precisely cluster and identify the stable geometric patterns that hold proteins together.

What Problem Does This Paper Address?

Many datasets don't fit on a flat map; they exist on the surface of a sphere (think of wind directions, planetary orbits, or molecular bonds). The Kent distribution is the spherical equivalent of a bell curve, but it’s "stretchy," allowing for elliptical shapes on the sphere's surface.

While Kent distributions are powerful, they are mathematically "heavy." To save on computing power, researchers previously used a bivariate normal approximation to estimate their parameters. However, when you try to use these approximations within a mixture model (a model that assumes the data is made of several different groups), the standard tool for the job—the Expectation-Maximization (EM) algorithm—breaks. Specifically, it loses its "monotonicity," meaning it can stop getting better at each step, leading to unreliable results.

The authors also tackle a massive problem in biology: protein folding. Hydrogen bonds are the "glue" of proteins, but identifying which bonds are structurally important versus which are just "noise" or outliers is a major challenge in bioinformatics.

Methods

The researchers implemented two key technical shifts:

  1. Exact Maximum Likelihood Estimation (MLE): Instead of using the "shortcut" approximation for the Kent distribution's normalizing constant, they used the exact infinite series involving modified Bessel functions. This ensures the EM algorithm behaves predictably and converges to a solid answer.
  2. Directional Parameterization: They reduced the complex 3D coordinates of four atoms (12 variables) into a simpler set of four variables: one distance ($\delta$) and three angles ($\chi, \psi, \theta$). By treating these angles as coordinates on a unit sphere, they could apply spherical statistics to biological data.

They used the Akaike Information Criterion (AIC) to automatically determine the "optimal" number of clusters in a dataset, preventing the model from over-fitting or under-fitting the data.

Datasets

  • Simulated Data: The authors created synthetic datasets with known clusters (4 and 5 components) to prove that their exact MLE approach could correctly identify the number of groups and their parameters where previous methods might fail.
  • Protein Hydrogen Bond Data: A massive dataset of 11,653 potential hydrogen bonds inferred from protein structures. The study specifically focused on a subset of 3,627 bonds found within $\alpha$-helices (the spiral-shaped building blocks of proteins).

Key Insights and Ideas

  • The EM Glitch Fix: The most significant technical "aha" is that the approximation used in previous literature was the culprit for unstable models. By doing the harder math (exact MLE), the EM algorithm becomes a reliable tool for spherical clustering.
  • Biological "Ground Truth" via Statistics: When the authors ran their model on protein data, they didn't tell the model which bonds were "good." However, the model naturally clustered the bonds into groups that matched their sequence separation ($\Delta L$).
  • Separating Signal from Noise: In $\alpha$-helices, the model identified one massive, tight cluster representing the "standard" stable bonds (where the residues are 4 apart in the sequence) and smaller, more "spread out" clusters for unusual or potentially spurious bonds.

Key Results

  • High Accuracy: In simulations, the model achieved a near-zero misclassification rate as clusters became more distinct.
  • Protein Clustering: For the $\alpha$-helix data, the AIC picked a model with three Kent components plus a uniform "background noise" component.
  • Validation: The statistical clusters showed an Adjusted Rand Index of 0.879 when compared to biological categories. In plain English: the math-driven clusters almost perfectly matched the biological reality of how these bonds are formed.
  • Robustness: The model successfully identified that bonds with a sequence separation of $\Delta L = 4$ are the most stable and geometrically consistent, while $\Delta L = 2$ or $3$ bonds are much more varied (higher "ovalness" or lower concentration).

Limitations and Future Work

  • Initial Guesses: The EM algorithm is sensitive to where you start. The authors had to pick initial mean directions by looking at "Schmidt-net" projections (flat maps of the sphere) and picking points by eye. A more automated way to pick starting points would be a logical next step.
  • Computational Cost: While more accurate, the exact MLE is more computationally intensive than the approximation.
  • Expanding the Scope: This paper focused heavily on $\alpha$-helices. The authors suggest that applying this to other structures, like $\beta$-sheets and loops, could reveal even more about the "hidden" rules of protein architecture.

Summary updated

Abstract

Finite mixture models are fitted to spherical data. Kent distributions are used for the components of the mixture because they allow considerable flexibility. Previous work on such mixtures has used an approximate maximum likelihood estimator for the parameters of a single component. However, the approximation causes problems when using the EM algorithm to estimate the parameters in a mixture model. Hence the exact maximum likelihood estimator is used here for the individual components. This paper is motivated by a challenging prize problem in structural bioinformatics of how proteins fold. It is known that hydrogen bonds play a key role in the folding of a protein. We explore this hydrogen bond geometry using a data set describing bonds between two amino acids in proteins. An appropriate coordinate system to represent the hydrogen bond geometry is proposed, with each bond represented as a point on a sphere. We fit mixtures of Kent distributions to different subsets of the hydrogen bond data to gain insight into how the secondary structure elements bond together, since the distribution of hydrogen bonds depends on which secondary structure elements are involved.
BibTeX
@misc{mardia2021mixture,
  author = {Kanti V. Mardia and Stuart Barber and Philippa M. Burdett and John T. Kent and Thomas Hamelryck},
  title = {{Mixture models for spherical data with applications to protein bioinformatics}},
  year = {2021},
  eprint = {2104.13140},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2104.13140},
}
Figure