Prediction of novel precursor miRNAs using a context-sensitive hidden Markov model (CSHMM)

Sumeet Agarwal, Candida Vaz, Alok Bhattacharya, Ashwin SrinivasanBMC Bioinformatics 2010 · DOI

TL;DR

This paper introduces a novel Context-Sensitive Hidden Markov Model (CSHMM) for identifying microRNAs (miRNAs), a class of small regulatory RNAs. The CSHMM effectively captures the complex structural language of miRNA precursors, outperforming existing methods and enabling the discovery of new miRNAs in human genomes and small RNA sequencing data.

What Problem Does This Paper Address?

MicroRNAs (miRNAs) are crucial regulators of gene expression, and their dysregulation is implicated in various biological processes and diseases. While thousands of human miRNAs have been identified, it's believed many more remain undiscovered. Identifying these novel miRNAs is challenging because they are small, non-coding, and their precursor molecules (pre-miRNAs) have specific, yet complex, secondary structures. Existing computational methods often rely on homology or a large number of features, which can limit their ability to find entirely new miRNA families or miss subtle structural nuances. This paper aims to develop a more accurate and efficient ab initio method for miRNA identification that can discover novel miRNAs without relying on sequence similarity to known ones.

Methods

The core of this work is the development and application of a Context-Sensitive Hidden Markov Model (CSHMM). Unlike traditional Hidden Markov Models (HMMs), CSHMMs incorporate a "memory" mechanism (a stack) between states. This allows them to model dependencies beyond simple adjacent states, which is crucial for representing the palindromic nature and long-range interactions found in miRNA precursor structures.

Specifically, the CSHMM proposed here uses:

  • Pairwise-emission states (P): For emitting upper halves of stems and symmetric bulges.
  • Regular HMM states (S): For emitting asymmetric bulges and loops.
  • Context-sensitive states (C): These states interact with the stack, allowing them to consider previously emitted nucleotides when determining the current emission and transition. This is key for ensuring complementarity in stems and symmetry in bulges.

The model's parameters (transition and emission probabilities) are learned from known human miRNA precursor sequences. For prediction, the CSHMM is used to calculate a likelihood score for a given sequence. A simple classification tree, using only this likelihood score as a feature, is then trained to discriminate between true miRNA precursors and non-miRNA sequences. This approach significantly reduces the feature space compared to methods like miPred, which use dozens of features.

Datasets

The study utilized several datasets:

  • D1: 323 human miRNA precursors (from miRBase).
  • D2: 646 "pseudo-hairpin" sequences from human genic regions, intended to represent non-miRNA sequences that can fold into hairpins.
  • D3: 1,918 non-human miRNA precursors from 40 species.
  • D4: A set of non-coding RNAs (ncRNAs) from Ensembl.
  • D5: Small RNA sequences obtained from normal human leukocytes (via Solexa deep sequencing).

The datasets were carefully curated for training and testing, with efforts made to remove family-related sequences to ensure independent evaluation.

Key Insights and Ideas

  • CSHMMs are well-suited for miRNA structure: The paper demonstrates that the context-sensitive nature of CSHMMs, with their stack mechanism, effectively captures the complex structural requirements of miRNA precursors, which are not fully addressed by simpler models.
  • Simplicity can be powerful: By reducing the prediction problem to a single feature (the CSHMM likelihood score) and using a simple classification tree, the authors achieve high accuracy. This suggests that the CSHMM itself is effectively encoding the most discriminative information.
  • Generative vs. Discriminative Models: The CSHMM is a generative model, meaning it learns the underlying structure of miRNAs. This allows it not only to predict if a sequence is a miRNA but also to predict its most likely secondary structure, a capability not inherent in discriminative models like SVMs.
  • Pipeline for Novel Discovery: Combining the CSHMM with post-prediction filters (EST matches, Drosha cutting sites) proves effective for ab initio discovery of novel miRNAs, even in large genomic regions or from raw sequencing data.

Key Results

  • Performance: The CSHMM-based classifier achieved high accuracy, comparable to or better than existing methods like miPred and CID-miRNA.
  • On human test sets, it showed a predictive accuracy of 93% with 85% sensitivity and 97% specificity.
  • For non-human miRNAs, it achieved 83% sensitivity.
  • It demonstrated good specificity by classifying only 4% of other ncRNAs as miRNAs.
  • Novel miRNA Discovery:
    • Scanning human chromosome 19, the pipeline identified 49 novel miRNAs.
  • Analysis of small RNA sequences from human leukocytes identified 308 novel miRNAs.
  • Structural Prediction: The CSHMM provides a probabilistic distribution over possible secondary structures, offering insights into the structural basis of miRNA recognition.

Limitations and Future Work

  • Limitations: While the CSHMM performs well, the paper notes that the "optimal" classifier threshold was determined by entropy minimization, suggesting that performance can vary with different threshold choices. The study also acknowledges that some stringent filters used by other methods (like miRDeep) might miss valid miRNAs, implying that the CSHMM pipeline might also have some false negatives.
  • Future Work: The authors suggest that the CSHMM is a promising approach for ab initio miRNA discovery, both for individual sequence analysis and for large-scale genome scans. The availability of the CSHMM-based classifier as an online applet [19] facilitates further research and application. The success in identifying novel miRNAs from deep sequencing data highlights its potential for future discoveries.

Summary updated

BibTeX
@article{agarwal2010prediction,
  author = {Sumeet Agarwal and Candida Vaz and Alok Bhattacharya and Ashwin Srinivasan},
  title = {{Prediction of novel precursor miRNAs using a context-sensitive hidden Markov model (CSHMM)}},
  journal = {BMC Bioinformatics},
  year = {2010},
  doi = {10.1186/1471-2105-11-s1-s29},
}

Loading related content...

Figure