Bi-clustering interpretation and prediction of correlation between gene expression and protein abundance

Xiaojun Wang, Lin Teng, Haicang Zhang, Qian Zhou, Xiaoquan Su, Xinping Cui, Dongbo Bu, Xinqi Gong, Ansgar Poetsch, Kang NingbioRxiv 2018 · DOI

TL;DR

This paper introduces a novel bi-clustering approach to identify gene groups with consistent expression-abundance patterns, leading to a more accurate "Cluster-based Linear prediction Model" (CLM) for predicting protein levels from mRNA data. This method offers improved predictive power over global models by accounting for cluster-specific relationships, providing deeper biological insights into gene regulation.

What Problem Does This Paper Address?

The correlation between mRNA (gene expression) and protein abundance in cells is surprisingly weak, often around 0.4. This means that mRNA levels only explain about 40% of the variation in protein levels. This disconnect is due to complex regulatory processes like transcription control and protein degradation. While advancements in high-throughput sequencing (RNA-Seq) and mass spectrometry have provided matching datasets, accurately predicting protein abundance from gene expression remains a challenge. Existing methods often provide only a global correlation or lack biological interpretability. This paper aims to bridge this gap by developing a more nuanced and biologically informed prediction model.

Methods

The core of this work is a bi-clustering approach. Unlike traditional clustering that groups items based on one type of data, bi-clustering simultaneously groups genes and experimental conditions (or in this case, different measurement methods for gene expression and protein abundance) to find patterns of consistent correlation. The authors then interpret these clusters by analyzing various features like mRNA and protein half-lives, and protein 3D structure.

Based on these insights, they developed a Cluster-based Linear prediction Model (CLM). This is a "divide and conquer" strategy where instead of building one model for all genes, CLM builds separate linear models for each identified cluster. This allows for more tailored predictions within each group of genes that share similar expression-abundance behaviors. Feature selection was performed using methods like Random Forest, PCA, LDA, PLS, and mRMR to identify the most predictive features for each cluster.

Datasets

The study utilized several well-established datasets:

  • Mouse liver mitochondrial dataset: This was the primary dataset used for building and testing the CLM model. It contained 442 genes with corresponding gene expression and protein abundance data, along with mRNA and protein half-lives.
  • Mouse brainstem mitochondrial dataset: Used for validation, this dataset had a similar structure to the liver dataset but from a different tissue.
  • Saccharomyces cerevisiae dataset: This dataset included protein abundances measured by ICAT and gene expressions from microarrays.
  • Danio rerio embryonic development dataset: This large dataset (5001 genes) utilized NanoHPLC-ESI-MS/MS for protein abundances and RNA-Seq for gene expressions.

Key Insights and Ideas

  • Bi-clustering reveals hidden patterns: The bi-clustering approach successfully identified distinct groups of genes with consistent correlation patterns between their mRNA and protein levels, patterns that are not apparent when looking at all genes globally.
  • Half-life and structure are key drivers: The interpretation of these clusters highlighted that mRNA half-life, protein half-life, and protein 3D structure are significant factors influencing the correlation between gene expression and protein abundance.
  • "Divide and Conquer" is superior: Building separate prediction models for each cluster (CLM) significantly outperforms a single, global model (GLM) because it accounts for the unique regulatory mechanisms within each gene group.
  • Feature importance varies by species/tissue: The study found that the most important features for predicting protein abundance can differ across species and tissues, suggesting that a one-size-fits-all feature set is not optimal.

Key Results

  • The CLM model achieved an average prediction accuracy of 35.09% on the mouse liver mitochondrial dataset, which was significantly higher than the 27.0% achieved by the global GLM model.
  • CLM demonstrated superior performance in predicting protein abundance for specific genes that were difficult for the GLM model to capture, such as Nags and Surfl.
  • When compared to the MARS model, CLM showed better prediction results in several clusters, indicated by higher Pearson correlation coefficients and lower Error Sum of Squares (SSE). For example, in Cluster 1 of the Danio rerio dataset, CLM achieved R²=0.48 and SSE=12720, compared to MARS's R²=0.37 and SSE=17200.
  • Validation on other datasets (mouse brainstem, S. cerevisiae, Danio rerio) confirmed the general applicability and improved accuracy of the CLM approach over simpler models. For instance, in S. cerevisiae, CLM achieved 49.4% accuracy, outperforming all-gene-based linear correlation (42.4%) and a model directly using features from the mouse liver dataset (45%).

Limitations and Future Work

  • Prediction accuracy is not perfect: While CLM shows significant improvement, the prediction accuracies for mouse genes/proteins were still "far from perfect and quite low," suggesting room for further optimization.
  • Dataset specificity: The authors acknowledge that the specificity of each dataset might contribute to the current prediction limitations.
  • Future directions: The authors suggest that future improvements could come from developing more advanced optimization methods for feature selection and exploring non-linear model tuning. This could lead to even better protein abundance predictions from gene expression data without requiring extensive additional information.

Summary updated

Abstract

Abstract Most organisms’ transcript and protein level only moderately correlate for various reasons, such as regulation of transcription and protein degradation. Better prediction and understanding the correlation between gene expression and protein abundance has been possible by harnessing the matching RNA/protein datasets produced by modern high-throughput RNA-Seq and mass spectrometry methods. In this work, we have utilized some well-studied matching RNA/protein datasets, and explored for the first time a bi-clustering method to cluster genes that have consistent correlation patterns between gene expression and protein abundance. The clustering results have been interpreted from the perspective of both transcriptomic and proteomic features, which show that mRNA half-life, protein half-life and protein structure in concert significantly affect the correlation of gene expression and protein abundance. With these and other carefully selected features, a prediction model based on individual clusters, called Cluster-based Linear prediction Model (CLM), was built and tested on mouse liver mitochondrial, mouse brainstem mitochondrial, Saccharomyces cerevisiae and Danio rerio datasets. CLM could find genes for which protein abundance can be predicted from mRNA data. In summary, based on bi-clustering, feature selection and CLM model, we have established a new and valuable cluster-based protein abundance prediction method.
BibTeX
@misc{wang2018bi,
  author = {Xiaojun Wang and Lin Teng and Haicang Zhang and Qian Zhou and Xiaoquan Su and Xinping Cui and Dongbo Bu and Xinqi Gong and Ansgar Poetsch and Kang Ning},
  title = {{Bi-clustering interpretation and prediction of correlation between gene expression and protein abundance}},
  year = {2018},
  doi = {10.1101/270397},
  howpublished = {bioRxiv},
}
Figure