Sequence-Based Prediction of Protein Solubility

Federico Agostini, Michele Vendruscolo, Gian Gaetano TartagliaJournal of Molecular Biology 2011 · DOI

TL;DR

This paper tackles the challenge of predicting protein solubility from amino acid sequences, a crucial factor in understanding protein aggregation and its link to diseases. They developed a novel method using physicochemical properties and machine learning that correlates well with experimental solubility, opening doors for large-scale proteome analysis.

What Problem Does This Paper Address?

Protein aggregation is a significant biological process, often implicated in debilitating diseases like Alzheimer's and Parkinson's. Understanding why some proteins aggregate while others remain soluble is key to tackling these conditions. While protein sequences are known to encode both folding and aggregation tendencies, predicting protein solubility directly from sequence has been challenging due to a lack of comprehensive proteome-level data. Previous studies focused on aggregation rates, but this paper aims to bridge the gap by investigating the relationship between protein solubility and aggregation propensity, considering the thermodynamic stability of proteins.

Methods

The authors developed a sequence-based predictor for protein solubility. Their approach involves:

  • Physicochemical Property Profiling: They calculated profiles of 28 physicochemical properties for each amino acid sequence. These profiles were generated using a sliding window of seven amino acids along the protein sequence.
  • Support Vector Machine (SVM) for Feature Selection: To identify the most discriminative properties, they employed an SVM. This involved splitting the dataset into "head" (highly soluble) and "tail" (least soluble) sets and iteratively selecting properties that best distinguished between these groups. This process was repeated for eight cycles, leading to the identification of six key properties.
  • SVM for Solubility Prediction: The selected physicochemical properties were then used to train an SVM model to predict protein solubility.
  • Cross-Validation: The model's performance was evaluated using a rigorous 10-fold cross-validation process.

Datasets

The primary dataset used in this study was derived from a previous experimental measurement of protein solubility for approximately 70% of the Escherichia coli proteome. This dataset, originally published by Niwa et al. (19), provided experimental solubility values. Protein identifiers were obtained from the Ensembl Bacteria Database. The dataset was further divided into three subsets: "head set" (most soluble), "tail set" (least soluble), and a remaining set of proteins.

Key Insights and Ideas

  • Solubility and Aggregation Propensity are Linked: A significant finding is the strong correlation between experimentally measured protein solubility and aggregation propensity calculated from the folded state. This suggests that the thermodynamic stability of a protein's native state is a crucial determinant of its solubility.
  • Thermodynamics and Kinetics Interplay: The correlation implies a link between the thermodynamics (stability of native vs. aggregate states) and kinetics (energy barriers between states) of protein aggregation.
  • Key Physicochemical Properties Identified: The study successfully identified specific physicochemical properties (coil/disorder, hydrophobicity, hydrophilicity, β-turn, and α-helix propensities) that are most predictive of protein solubility. Notably, disorder and coil propensities were found to be particularly important discriminators.
  • Proteome-Level Analysis Enabled: By developing a sequence-based predictor, the authors pave the way for large-scale, computational studies of protein solubility across entire proteomes, which was previously hindered by data scarcity.

Key Results

  • The developed method shows a strong correlation between predicted and experimental protein solubility scores.
  • The aggregation propensity calculated from the folded state (using the Zyggregator method) showed a high correlation with experimental solubility (Fig. 1a).
  • The CCSOL method (developed in this work) also demonstrated a good correlation with experimental solubility (Fig. 1b).
  • Maximal expression level scores, derived from a different approach, also showed a correlation with experimental solubility (Fig. 1c).
  • The SVM-based predictor identified six key properties that are highly predictive of protein solubility.

Limitations and Future Work

  • Data Scarcity: While the study leverages a significant dataset, the authors acknowledge that proteome-level data on protein solubility is generally scarce, which has historically been a challenge for such analyses.
  • Focus on E. coli: The study primarily uses data from E. coli. While this is a well-studied model organism, extending the predictor to other organisms might require further validation and potentially adaptation.
  • Future Work: The authors suggest that their method, when combined with existing tools for predicting aggregation rates, will enable proteome-level studies to further investigate the intricate relationship between the thermodynamics and kinetics of protein aggregation. This could lead to a deeper understanding of protein stability in living systems.

Summary updated

BibTeX
@article{agostini2011sequence,
  author = {Federico Agostini and Michele Vendruscolo and Gian Gaetano Tartaglia},
  title = {{Sequence-Based Prediction of Protein Solubility}},
  journal = {Journal of Molecular Biology},
  year = {2011},
  doi = {10.1016/j.jmb.2011.12.005},
}
Figure