GIC: A computational method for predicting the essentiality of long noncoding lncRNAs

Pan Zeng, Ji Chen, Yuan Zhou, Jichun Yang, Qinghua CuibioRxiv 2017 · DOI

TL;DR

Researchers have developed GIC, a novel computational method that accurately predicts gene essentiality using only RNA sequence information. This is a significant breakthrough because existing methods fail to work for long noncoding RNAs (lncRNAs), a crucial class of molecules, and GIC fills this gap, offering a powerful new tool for biological research.

What Problem Does This Paper Address?

Identifying "essential genes" – those critical for an organism's survival – is fundamental to understanding biological processes, disease mechanisms, and developing new therapies. While experimental methods like CRISPR/Cas9 are powerful, they are time-consuming and resource-intensive, especially for large-scale studies. Computational methods have emerged as a valuable complement, but they have historically focused on protein-coding genes. A major challenge has been the inability of these methods to predict the essentiality of long noncoding RNAs (lncRNAs). LncRNAs are a vast and increasingly recognized class of RNA molecules with diverse regulatory roles, yet their essentiality remains largely unexplored computationally due to their unique characteristics, such as low sequence conservation and lack of protein-coding potential. This paper aims to bridge this gap by developing a computational method capable of predicting the essentiality of both protein-coding genes and lncRNAs.

Methods

The core of this work is the development of the Gene Importance Calculator (GIC), a computational method based on a logistic regression model. GIC leverages RNA sequence information to predict gene essentiality. The key features used by GIC include:

  • RNA Sequence Length: Based on the observation that longer RNAs tend to encode more conserved proteins.
  • Nucleotide Triplet Frequencies: Analyzing the occurrence of all 64 possible three-nucleotide combinations (e.g., ACT, GCC) within the RNA sequence.
  • Predicted RNA Secondary Structure: Using RNAfold to predict the secondary structure and calculate its minimum free energy (MFE). To account for the fact that longer RNAs naturally have lower MFEs, a normalized MFE (nMFE) was introduced, calculated as MFE divided by sequence length.

To handle the imbalance between essential and non-essential genes in training datasets, GIC employs a strategy of randomly selecting a subset of non-essential genes to achieve a balanced positive-to-negative ratio for training. The output of the logistic regression model, representing the conditional probability of a gene being essential, is defined as the GIC score.

Datasets

The study utilized several datasets for training and evaluation:

  • RNA Sequences: Human (GRCh37/hg19) and mouse (GRCm38/mm10) mRNA sequences from the UCSC Table Browser. Human and mouse lncRNA transcripts were obtained from the NONCODE database (version 4), with sequences longer than 200 nt retained.
  • Essential Genes: Human and mouse essential protein-coding genes were retrieved from the DEG database (version 10).
  • Independent Testing Set: A set of 7 mouse essential lncRNAs and 7 non-essential lncRNAs, annotated using the Mouse Genome Informatics (MGI) database and data from Sauvageau et al.'s assays.
  • CRISPR/Cas9 Scores: Data from Wang et al.'s study on KBM7 cell lines.

Key Insights and Ideas

  • LncRNAs are Predictable: The most significant insight is that lncRNA essentiality can be predicted using only sequence-based features, a feat not achievable with previous methods. This opens up a vast, understudied area of biology for computational exploration.
  • Sequence Features Hold the Key: GIC demonstrates that fundamental sequence properties like length, triplet composition, and predicted secondary structure are sufficient to infer essentiality, even for non-coding RNAs. This suggests underlying principles of RNA function are encoded within their sequences.
  • Bridging the Protein-Coding/Non-Coding Divide: GIC's ability to work on both protein-coding genes and lncRNAs provides a unified framework for essentiality prediction, simplifying research and enabling cross-comparisons.
  • Limitations of Existing Tools: The paper clearly shows that methods designed for protein-coding genes (like Pheg, which relies on CDS sequences) are not suitable for lncRNAs, highlighting the unique challenges and the need for specialized approaches.

Key Results

  • Human Protein-Coding Genes: GIC showed significant correlations with established measures of essentiality, including homolog number, dN/dS ratio, and protein interaction network degree. It also demonstrated that essential genes are significantly enriched in higher GIC score bins.
  • Mouse Genes: Similar positive correlations and enrichment patterns were observed for mouse genes.
  • CRISPR/Cas9 Comparison: GIC outperformed CRISPR/Cas9 scores in predicting essentiality in a general context, suggesting that condition-specific CRISPR/Cas9 scores may not always translate to broader essentiality predictions.
  • LncRNA Prediction: On an independent mouse lncRNA dataset, GIC achieved a high Area Under the Curve (AUC) of 0.918, significantly outperforming Pheg (AUCs of 0.531 and 0.653).
  • Experimental Validation: A case study using human VSMCs showed that silencing GIC-predicted essential genes (but CRISPR/Cas9 non-essential) led to significant reductions in cell viability and increased apoptosis, validating GIC's predictions. Conversely, silencing GIC-predicted non-essential genes (but CRISPR/Cas9 essential) had less pronounced effects.

Limitations and Future Work

  • miRNA Inapplicability: The authors found that GIC is not suitable for microRNAs (miRNAs) due to their very small sequence size, which limits the utility of nucleotide triplet features derived from protein-coding genes. This suggests that a different approach might be needed for miRNA essentiality prediction.
  • Future Directions: The paper highlights the potential of GIC as a tool for identifying and characterizing novel genes and lncRNAs with important functions, which could significantly advance our understanding of disease pathogenesis, diagnosis, and therapy. The authors also suggest the need for developing new computational methods specifically for predicting essential miRNAs.

Summary updated

Abstract

ABSTRACT Measuring the essentiality of genes is critically important in biology and medicine. Some bioinformatic methods have been developed for this issue but none of them can be applied to long noncoding RNAs (lncRNAs), one big class of biological molecules. Here we developed a computational method, GIC (Gene Importance Calculator), which can predict the essentiality of both protein-coding genes and lncRNAs based on RNA sequence information. For identifying the essentiality of protein-coding genes, GIC is competitive with well-established computational scores. More important, GIC showed a high performance for predicting the essentiality of lncRNAs. In an independent mouse lncRNA dataset, GIC achieved an exciting performance (AUC=0.918). In contrast, the traditional computational methods are not applicable to lncRNAs. As a public web server, GIC is freely available at http://www.cuilab.cn/gic/ .
BibTeX
@misc{zeng2017gic,
  author = {Pan Zeng and Ji Chen and Yuan Zhou and Jichun Yang and Qinghua Cui},
  title = {{GIC: A computational method for predicting the essentiality of long noncoding lncRNAs}},
  year = {2017},
  doi = {10.1101/177923},
  howpublished = {bioRxiv},
}
Figure