Using Machine Learning to Measure Relatedness Between Genes: A Multi-Features Model

Yan Wang, Sen Yang, Jing Zhao, Wei Du, Yanchun Liang, Cankun Wang, Fengfeng Zhou, Yuan Tian, Qin MaScientific 2019 · DOI

TL;DR

This paper introduces MFR, a novel machine learning model that combines gene expression similarities with prior biological knowledge to more accurately measure the relatedness between gene pairs. This improved relatedness measurement leads to better prediction of gene-gene interactions, construction of more biologically relevant cancer gene networks, and enhanced gene function prediction.

What Problem Does This Paper Address?

Understanding how genes relate to each other is fundamental to deciphering complex biological systems. Researchers often try to measure this "relatedness" using two main approaches:

  1. Gene Expression Similarities: This involves looking at how genes are "turned on" or "turned off" together across different conditions or tissues. While useful for capturing conditional relatedness (how genes behave together under specific circumstances), this method can suffer from high false discovery rates, meaning it might incorrectly identify genes as related.
  2. Prior-Knowledge Based Similarities: This leverages existing biological databases (like pathways, protein interactions, etc.) to infer relatedness. This approach is good for understanding global relatedness but may not accurately reflect how genes interact under specific conditions.

The challenge is that neither approach alone is perfect. Expression similarities are noisy, and prior knowledge is often too general. This paper aims to bridge this gap by developing a method that can accurately measure conditional relatedness by integrating both types of information.

Methods

The core of this work is the Multi-Features Relatedness (MFR) model. MFR is built upon a Support Vector Machine (SVM), a powerful supervised learning algorithm. Here's how it works:

  • Multi-Feature Input: MFR takes 12 different similarity-based features as input for each gene pair. These features are a blend of:
  • Expression Similarities: Including average expression levels, Pearson correlation coefficient (PCC), Spearman rank correlation (SRC), mutual information (MI), partial Pearson correlation (PPC), and conditional mutual information (CMI).
  • Prior-Knowledge Based Similarities: Incorporating Gene Ontology (GO) similarity, subcellular localization similarity, homology similarity, Reactome pathway similarity, and transcriptional regulatory similarity.
  • SVM Model: An SVM with a linear kernel is trained to predict a "target MFR value" (a score between 0 and 1) for each gene pair. This score represents the degree of conditional relatedness. The SVM is trained to maximize the separation between "positive" (interacting) and "negative" (non-interacting) gene pairs.
  • Assessment Criterion: The output of the SVM, the MFR value, serves as the primary metric for assessing conditional relatedness. A higher MFR value indicates a stronger likelihood of interaction.
  • Training Strategy: MFR is trained on a combined dataset that includes both coexpression and prior-knowledge based gene pairs. This simultaneous training allows the model to learn how to balance and integrate these different types of evidence, rather than relying on one exclusively.

Datasets

The study utilizes several publicly available datasets for training and validation:

  • Development/Training Datasets:
    • Coexpression Sub-dataset: Derived from COXPRESdb, containing positive (coexpressed) and negative (discoexpressed) gene pairs.
  • Prior-Knowledge Sub-dataset: Composed of sub-sub datasets from KEGG (pathways), HPRD (protein-protein interactions - PPI), and TRRUST (transcriptional regulation).
  • Verification Datasets:
    • Test Dataset: Used for initial verification.
  • GeneFriends and DIP Datasets: Used for further robustness testing, with GeneFriends focusing on coexpression data and DIP on prior-knowledge data.
  • Cancer Gene Expression Data: RNA-seq data from four cancer types (BLCA, BRCA, COAD, LUAD) were downloaded from the TCGA database for cancer network construction.
  • Gene Function Prediction Data: Genes from KEGG metabolism pathways were used for predicting gene functions.

Key Insights and Ideas

  • Synergy of Data Sources: The most significant insight is that combining gene expression data with prior biological knowledge is crucial for accurately measuring conditional gene relatedness. Neither source alone is sufficient.
  • MFR's Advantage: MFR's ability to integrate these diverse features allows it to overcome the limitations of individual approaches, particularly in identifying gene pairs that are both highly coexpressed and have strong prior-knowledge support.
  • Feature Importance: The study highlights that certain features, like mutual information (MI) and partial Pearson correlation (PPC) from expression data, and GO similarity and subcellular localization similarity from prior knowledge, are particularly important for MFR's performance. This suggests specific biological signals are more indicative of conditional relatedness.
  • Punishing Extremes: MFR's negative weights for average expression levels (exp1, exp2) indicate a "punishment" for genes with exorbitant expression. This makes sense, as extreme expression in one gene might imply a disconnect rather than a strong interaction with another.

Key Results

  • Superior Performance in Interaction Prediction: MFR consistently achieved the highest Area Under the Curve (AUC) values across development, test, and DIP datasets for identifying gene-gene interactions.
  • Improved Precision: MFR showed an average improvement of 1.1% in precision for detecting gene pairs with both high expression and high prior-knowledge similarities, outperforming other linear models and coexpression analysis methods.
  • Robustness: MFR demonstrated strong robustness across various verification datasets, including those relying heavily on coexpression (GeneFriends) or prior-knowledge (DIP) data.
  • Enhanced Biological Applications:
    • Cancer Gene Networks: MFR-based networks led to the prediction of more biologically significant metabolic pathways influenced by glutamine and glutamate metabolism in cancer types compared to other methods.
  • Gene Function Prediction: MFR-based networks enabled more accurate prediction of gene functions using a shortest-path method, outperforming other models.

Limitations and Future Work

  • Prior Knowledge Dependency: While MFR leverages prior knowledge effectively, its performance is still influenced by the quality and completeness of these existing databases. The authors suggest improving MFR by incorporating more available and accurate prior knowledge.
  • Model Complexity: The current SVM-based model is a regression model. The authors propose exploring deep learning models, such as deep belief networks, to potentially further enhance accuracy and robustness by automatically learning complex functions.
  • Feature Engineering: While MFR uses 12 features, there's always room to explore additional or refined features that could capture even more nuanced gene relationships.

Summary updated

Abstract

Measuring conditional relatedness between a pair of genes is a fundamental technique and still a significant challenge in computational biology. Such relatedness can be assessed by gene expression similarities while suffering high false discovery rates. Meanwhile, other types of features, e.g., prior-knowledge based similarities, is only viable for measuring global relatedness. In this paper, we propose a novel machine learning model, named Multi-Features Relatedness (MFR), for accurately measuring conditional relatedness between a pair of genes by incorporating expression similarities with prior-knowledge based similarities in an assessment criterion. MFR is used to predict gene-gene interactions extracted from the COXPRESdb, KEGG, HPRD, and TRRUST databases by the 10-fold cross validation and test verification, and to identify gene-gene interactions collected from the GeneFriends and DIP databases for further verification. The results show that MFR achieves the highest area under curve (AUC) values for identifying gene-gene interactions in the development, test, and DIP datasets. Specifically, it obtains an improvement of 1.1% on average of precision for detecting gene pairs with both high expression similarities and high prior-knowledge based similarities in all datasets, comparing to other linear models and coexpression analysis methods. Regarding cancer gene networks construction and gene function prediction, MFR also obtains the results with more biological significances and higher average prediction accuracy, than other compared models and methods. A website of the MFR model and relevant datasets can be accessed from http://bmbl.sdstate.edu/MFR .
BibTeX
@inproceedings{wang2019using,
  author = {Yan Wang and Sen Yang and Jing Zhao and Wei Du and Yanchun Liang and Cankun Wang and Fengfeng Zhou and Yuan Tian and Qin Ma},
  title = {{Using Machine Learning to Measure Relatedness Between Genes: A Multi-Features Model}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41598-019-40780-7},
}
Figure