Precursor microRNA Identification Using Deep Convolutional Neural Networks

Huỳnh Thị Thanh Bình, Vladimir Golkov, Göktuğ Erce Gürel, Daniel CremersbioRxiv 2018 · DOI

TL;DR

This paper introduces a novel deep learning approach for identifying precursor microRNAs (pre-miRNAs) by representing their predicted secondary structures as a 2D matrix. This method, utilizing a convolutional neural network (CNN), automatically learns relevant features, outperforming existing state-of-the-art methods on multiple benchmark datasets.

What Problem Does This Paper Address?

MicroRNAs (miRNAs) are small non-coding RNA molecules crucial for regulating gene expression, with implications in various diseases. Identifying them is challenging due to their short length. A common workaround is to identify their precursor molecules (pre-miRNAs), which are longer and possess more distinct structural features. However, accurately distinguishing pre-miRNAs from similar non-coding RNAs (pseudo-hairpins) is difficult, exacerbated by class imbalance (many more negative examples than positive ones). Existing methods often rely on handcrafted features, which may not capture all essential characteristics, or use representations that entangle structural information, making it harder for neural networks to learn.

Methods

The core of this work is a deep learning method that leverages a 2D convolutional neural network (CNN). The key innovation lies in the input representation: instead of using traditional one-hot encoding of the RNA sequence alone, the authors propose a multi-channel 2D matrix. This matrix combines:

  • One-hot encoding of the RNA sequence: Each nucleotide (A, C, G, U) is represented as a binary vector.
  • Predicted secondary structure: This is encoded as an L x L pairing matrix, where L is the sequence length. A non-zero value indicates a pair, and a zero indicates no interaction.
  • Minimum free energy (MFE): The MFE of the predicted structure is also incorporated.

This 2D representation is fed into a CNN. The network is designed to handle variable-sized inputs by incorporating a downscaling layer, allowing it to process sequences of different lengths. The authors also compare this to a fixed-sized input approach that uses zero-padding.

Datasets

The authors evaluated their method on three benchmark datasets:

  • Human dataset: Primarily derived from miRBase and NCBI.
  • Cross-species dataset: Also obtained from miRBase and NCBI.
  • New dataset: Collected from various sources including fRNAdb, NON-CODE, and snoRNA-LBME-db.

These datasets are characterized by a significant class imbalance, with many more negative samples than positive ones.

Key Insights and Ideas

  • The power of structural representation: The paper strongly emphasizes that incorporating the predicted secondary structure, particularly through the pairing matrix, is crucial for accurate pre-miRNA identification. Results show a significant performance drop when the pairing matrix is omitted.
  • Automatic feature extraction: By using a 2D CNN on a rich structural representation, the method bypasses the need for manual feature engineering, allowing the network to discover optimal features directly from the data.
  • Handling variable input sizes: The proposed downscaling layer in the CNN architecture effectively addresses the challenge of processing RNA sequences of varying lengths, a common issue in biological sequence analysis.
  • Deep learning's advantage: The study demonstrates that deep learning approaches, especially those that leverage structural information, generally outperform traditional machine learning methods for pre-miRNA identification.

Key Results

The proposed method achieved state-of-the-art results on the human, cross-species, and new datasets, particularly excelling in metrics like F-score, g-mean, AUROC, and AUPR. For instance, on the human dataset, their "Proposed variable-sized (CV)" model achieved an AUPR of 0.922, outperforming most other methods. On the new dataset, they achieved the highest scores in specificity, PPV, F-score, and AUPR, outperforming deepMiRGene in most metrics. The paper also highlights that their variable-sized input CNN performed comparably to the fixed-sized version but required less memory.

Limitations and Future Work

  • Data imbalance: While the authors address class imbalance using weighted cross-entropy loss, it remains a challenge. They suggest that collecting more data would be beneficial for better generalization.
  • Dataset size: The paper notes that a larger dataset is needed for good data generalization, especially given the imbalanced nature of the classes.
  • Future directions: The authors suggest expanding their method to other miRNA-related tasks, such as miRNA target prediction and miRNA function prediction. They also propose exploring larger neural network architectures like AlexNet or ResNet for improved performance.

Summary updated

Abstract

Abstract Precursor microRNA (pre-miRNA) identification is the basis for identifying microRNAs (miRNAs), which have important roles in post-transcriptional regulation of gene expression. In this paper, we propose a deep learning method to identify whether a small non-coding RNA sequence is a pre-miRNA or not. We outperform state-of-the-art methods on three benchmark datasets, namely the human, cross-species, and new datasets. The key of our method is to use a matrix representation of predicted secondary structure as input to a 2D convolutional network. The neural network extracts optimized features automatically instead of using a large number of handcrafted features as most existing methods do. Code and results are available at https://github.com/peace195/miRNA-identification-conv2D .
BibTeX
@misc{binh2018precursor,
  author = {Huỳnh Thị Thanh Bình and Vladimir Golkov and Göktuğ Erce Gürel and Daniel Cremers},
  title = {{Precursor microRNA Identification Using Deep Convolutional Neural Networks}},
  year = {2018},
  doi = {10.1101/414656},
  howpublished = {bioRxiv},
}
Figure