Optimizing an emperical scoring function for transmembrane protein structure determination.

Malin M. Young, Kenneth L. Sale, Genetha A. Gray, Tamara G. Kolda2003 · DOI

TL;DR

This paper tackles the challenging problem of determining the 3D structure of transmembrane proteins, which are crucial for cell function and drug targets. The authors frame this as an optimization problem, using a scoring function called "Bundler" to evaluate potential structures, and compare two derivative-free optimization algorithms: Simulated Annealing (SA) and Asynchronous Parallel Pattern Search (APPS). APPS emerges as a more efficient and effective method for this specific task.

What Problem Does This Paper Address?

Transmembrane proteins, embedded within cell membranes, play vital roles in cell signaling, transport, and recognition. Understanding their 3D structure is key to deciphering their function and developing targeted drugs. However, experimentally determining these structures is difficult, making computational approaches essential. This paper focuses on a computational method that models transmembrane proteins as bundles of alpha-helices and aims to find the optimal arrangement of these helices. The challenge lies in efficiently searching the vast conformational space of possible helix arrangements and finding the one that best matches experimental data and known structural principles.

Methods

The core of the problem is framed as a bound-constrained nonlinear optimization problem. The objective function, "Bundler," is an empirical scoring function that quantifies how well a proposed protein structure fits experimental data (like distance constraints between atoms) and known helix-packing rules. A lower Bundler score indicates a better-fitting structure.

The authors compare two derivative-free optimization algorithms:

  • Simulated Annealing (SA): A global optimization method inspired by the annealing process in metallurgy. It allows for "bad" steps to escape local minima, aiming to find the global optimum. However, it can be computationally expensive and sensitive to parameter choices.
  • Asynchronous Parallel Pattern Search (APPS): A method that explores the search space by evaluating points in a predefined pattern. It's well-suited for problems where the objective function is noisy or its derivative is unavailable. APPS is designed to leverage parallel computing, allowing for simultaneous evaluation of function values at different points, which can significantly speed up the process.

Datasets

The primary dataset used for evaluation is the experimental distance constraints for dark-adapted rhodopsin, a well-studied transmembrane protein with a known 3D structure. This allows the authors to use Root Mean Square Deviation (RMSD) as a metric to compare their computationally derived structures against the known true structure. They also utilize a set of 16 known transmembrane protein structures to derive parameters for the Bundler scoring function.

Key Insights and Ideas

  • Bundler as a Bridge: The Bundler scoring function is a clever way to integrate diverse experimental data (distance constraints) and biophysical principles (helix packing) into a single objective function for optimization. This moves beyond purely theoretical models.
  • Efficiency Matters: For complex biological problems like protein structure determination, computational efficiency is paramount. The paper highlights that even a good algorithm can be impractical if it takes too long.
  • Global vs. Local Optimization Trade-offs: While SA is a global optimizer, its computational cost and parameter sensitivity make it less ideal for this specific problem compared to APPS. APPS, being a local optimizer, can get stuck in local minima, but this is mitigated by using multiple starting points.
  • Parameter Tuning is a Hurdle: The sensitivity of SA to its parameters (like cooling schedules) is a significant drawback, requiring extensive fine-tuning. APPS, on the other hand, requires minimal parameter tuning, making it more user-friendly.

Key Results

  • APPS Outperforms SA: In their numerical experiments with rhodopsin, APPS consistently produced structures with lower Bundler scores and lower RMSD values compared to SA, especially when SA was run with fewer temperature cycles.
  • Computational Efficiency: APPS was significantly more computationally efficient than SA. For instance, in one test, APPS required approximately 32,458 function evaluations and 17 minutes on 86 processors, while SA required 81,800 function evaluations and 61 hours on a single processor.
  • Parameter Sensitivity of SA: Varying SA's parameters (number of temperature cycles and initial temperature) led to drastically different results, demonstrating its instability and the difficulty in finding optimal settings.
  • APPS's Robustness: APPS, with minimal tuning, achieved comparable or better results than SA, highlighting its robustness and ease of use.

Limitations and Future Work

  • Bundler's Noise Sensitivity: The authors acknowledge that the Bundler scoring function incorporates noisy experimental data, and currently lacks a regularization term to prevent overfitting to this noise. This means achieving a Bundler score of zero is not necessarily the goal, and small score variations might not be meaningful.
  • Future Optimization Methods: The paper suggests that future work could involve investigating alternative optimization methods and making minor modifications to the Bundler function itself.
  • Scaling to Larger Problems: While APPS is efficient, the paper notes that future projects will involve optimizing thousands of structures, emphasizing the need for continued focus on computational time and efficiency.
  • Local Minima Mitigation: The authors suggest that using multiple starting points can overcome the limitation of APPS being a local optimization method.

Summary updated

Abstract

We examine the problem of transmembrane protein structure determination. Like many other questions that arise in biological research, this problem cannot be addressed by traditional laboratory experimentation alone. An approach that integrates experiment and computation is required. We investigate a procedure which states the transmembrane protein structure determination problem as a bound constrained optimization problem using a special empirical scoring function, called Bundler, as the objective function. In this paper, we describe the optimization problem and some of its mathematical properties. We compare and contrast results obtained using two different derivative free optimization algorithms.
BibTeX
@misc{young2003optimizing,
  author = {Malin M. Young and Kenneth L. Sale and Genetha A. Gray and Tamara G. Kolda},
  title = {{Optimizing an emperical scoring function for transmembrane protein structure determination.}},
  year = {2003},
  doi = {10.2172/918349},
}
Figure