Optimizing weights of protein energy function to improve ab initio protein structure prediction

Chao Wang, Yi Wei, Juntao Liu, Haicang Zhang, Bin Ling, Shuai Cheng Li, Wei‐Mou Zheng, Dongbo BuarXiv 2013 · DOI · arXiv

Abstract

Predicting protein 3D structure from amino acid sequence remains a significant challenge in computational biology. When structural homologues are unavailable, researchers rely on ab initio approaches using empirical energy functions. A prominent algorithm, Rosetta, generates decoy conformations by assembling fragments of known structures and identifies the native state as the low-energy, highly populated conformation. However, the effectiveness of these functions depends heavily on the relative weights of various energy terms (e.g., van der Waals, hydrogen bonding). This paper proposes a linear programming (LP) algorithm to optimize the weighting of 14 energy terms. By employing a "reverse sampling" process that generates conformations starting from the native state, we aim to enlarge the "basin of attraction" for the native structure. Experimental results on benchmark proteins demonstrate that optimized weights improve both the size of the attraction basin and the quality of predicted native states, while revealing that different protein classes require distinct weighting schemes.

Introduction: The Challenge of Ab Initio Prediction

Predicting a protein's 3D structure from its sequence is a fundamental goal in biology. In the absence of templates, ab initio methods build structures from scratch. Tools like Rosetta utilize complex energy functions to find the most stable "native state." These functions combine multiple physical and chemical forces, but determining the optimal weight for each term is difficult. Incorrect weights can lead search algorithms into "decoy" states—conformations that appear stable but are biologically inaccurate.

Methodology: Reverse Sampling and LP Optimization

The researchers developed a two-step iterative approach to refine these energy weights:

  1. Reverse Sampling: Instead of searching for the native state from a random starting point, the algorithm starts at the known native structure and moves outward. It generates a path of conformations where energy increases slightly at each step until it reaches an "edge point"—a conformation where all neighboring states have lower energy. This defines the current "basin of attraction."
  2. Linear Programming (LP) Optimization: An LP model then adjusts the 14 energy term weights to "cut off" these edge points. By lowering the energy of the edge point relative to its neighbors, the algorithm effectively pushes the boundary of the basin further out.
The energy landscape visualization shows the "reverse sampling" paths (arrows) moving away from the native state (N) to map the basin of attraction.

This process repeats iteratively: sampling the basin, optimizing weights to expand it, and re-sampling until the weights stabilize.

Key Insights and Results

Shaping the Energy Landscape

The core innovation is shifting focus from finding the global minimum to shaping the surrounding landscape. By enlarging the basin of attraction, the algorithm creates a larger "funnel" that guides random starting conformations toward the correct native state. For the protein 1ctf, the radius of this basin grew from approximately 6 Å to nearly 14 Å RMSD.

Class-Specific Physics

The study found that different protein architectures (all-alpha, alpha-beta, etc.) prioritize different energy terms. For example, in all-alpha proteins, local geometrical interactions (the "environment" term) are dominant. Conversely, in alpha-beta proteins, "sheet" and "residue pair" terms are critical for stabilizing complex folds.

Evolution of weights for the 14 energy terms across six optimization iterations, showing how the algorithm converges on a stable weighting scheme.

Performance Gains

Optimized weights led to significant accuracy improvements. For the protein 1ilo, the best predicted structure improved from 2.7 Å RMSD (standard Rosetta) to 1.3 Å RMSD. Furthermore, the "good decoy ratio"—the percentage of generated structures close to the native state—increased from 1% to over 20%.

Conclusion and Future Directions

While the current study is limited by a small sample size of three benchmark proteins, it provides a powerful proof-of-concept. The ultimate goal is to apply this method to a massive library of proteins to establish generalized weighting schemes for specific protein classes. This would allow researchers to apply pre-optimized weights based on a sequence's predicted class, significantly enhancing the reliability of ab initio structure prediction.

Summary updated

Abstract

Predicting protein 3D structure from amino acid sequence remains as a challenge in the field of computational biology. If protein structure homologues are not found, one has to construct 3D structural conformations from the very beginning by the so-called ab initio approach, using some empirical energy functions. A successful algorithm in this category, Rosetta, creates an ensemble of decoy conformations by assembling selected best short fragments of known protein structures and then recognizes the native state as the highly populated one with a very low energy. Typically, an energy function is a combination of a variety of terms characterizing different structural features, say hydrophobic interactions, van der Waals force, hydrogen bonding, etc. It is critical for an energy function to be capable to distinguish native-like conformations from non-native ones and to drive most initial conformations assembled from fragments to a native-like one in a conformation search process. In this paper we propose a linear programming algorithm to optimize weighting of a total of 14 energy terms used in Rosetta. We reverse the Monte Carlo process of Rosetta to approach native-like conformations to a process generating from the native state an ensemble of initial conformations most relevant to the native state. Intuitively, an ideal weighting scheme would result in a large "basin of attraction" of the native structure, which leads to an objective function for the linear programming. We have examined the proposal on several benchmark proteins, and the experimental results suggest that the optimized weights enlarge the attraction basin of the native state and improve the quality of the predicted native states as well. In addition, a comparison of optimal weighting schema for proteins of different classes indicates that in different protein classes energy terms may have different effects.
BibTeX
@misc{wang2013optimizing,
  author = {Chao Wang and Yi Wei and Juntao Liu and Haicang Zhang and Bin Ling and Shuai Cheng Li and Wei‐Mou Zheng and Dongbo Bu},
  title = {{Optimizing weights of protein energy function to improve ab initio protein structure prediction}},
  year = {2013},
  eprint = {1312.6960},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1312.6960},
}

Loading related content...

Figure