Distinguish virulent and temperate phage-derived sequences in metavirome data with a deep learning approach

Shu‐Fang Vivienne Wu, Zhencheng Fang, Jie Tan, Mo Li, Chunhui Wang, Qian Guo, Congmin Xu, Xiaoqing Jiang, Huaiqiu ZhubioRxiv 2020 · DOI

TL;DR

This paper introduces DeePhage, a novel deep learning tool that accurately and efficiently distinguishes between virulent and temperate phage DNA sequences in complex metagenomic and metavirome data. By leveraging a convolutional neural network and a unique "one-hot" encoding of DNA, DeePhage offers significant improvements in accuracy and speed over existing methods, opening new avenues for understanding phage roles in microbial communities and disease.

What Problem Does This Paper Address?

Microbial communities are teeming with viruses, particularly bacteriophages (phages), which play crucial roles in shaping these ecosystems. Phages are broadly categorized into virulent (which kill their host cells) and temperate (which can integrate into the host genome or enter a lytic cycle). Understanding the balance between these two types is vital for comprehending microbial community dynamics, host-pathogen interactions, and even developing phage-based therapies. However, analyzing phage sequences from complex, culture-independent samples like metagenomes and metaviromes is challenging. Existing methods struggle with short DNA fragments, lack of universal marker genes in phages, and the difficulty of reconstructing complete phage genomes. Crucially, there was no effective computational tool to directly classify phage fragments as either virulent or temperate from such data.

Methods

The core of DeePhage is a deep learning model, specifically a Convolutional Neural Network (CNN). The authors developed a novel "one-hot" encoding scheme for DNA sequences, representing each nucleotide (A, C, G, T) as a distinct vector. This encoding captures detailed sequence information without requiring manual feature engineering. The CNN architecture comprises several layers:

  • 1D Convolutional Layer: Extracts local patterns (motifs) from the encoded DNA sequences.
  • 1D Max Pooling Layer: Downsamples the feature maps, reducing dimensionality while retaining important information.
  • 1D Global Average Pooling Layer: Further reduces dimensionality by averaging features across channels, integrating global spatial information.
  • Batch Normalization Layers: Stabilize training and accelerate convergence.
  • Dropout Layer: Helps prevent overfitting by randomly dropping units during training.
  • Dense Layers: Perform classification based on the extracted features.

The network outputs a score between 0 and 1, with scores above 0.5 indicating a virulent phage and scores below 0.5 indicating a temperate phage.

Datasets

The authors constructed artificial datasets for training and testing by extracting contigs from well-annotated complete phage genomes. They utilized a dataset of 227 complete phage genomes with lifestyle annotations (79 virulent, 148 temperate), excluding two with ambiguous lifestyles. These artificial contigs were generated using MetaSim and divided into four length groups (100-400 bp, 400-800 bp, 800-1200 bp, and 1200-1800 bp) to mimic various read lengths and assembled contig sizes.

Real metavirome data from bovine rumen was also used for qualitative evaluation.

Key Insights and Ideas

  • Sequence Signatures Hold the Key: The study demonstrates that distinct sequence signatures, detectable by deep learning, differentiate virulent and temperate phages, even in short DNA fragments.
  • "One-Hot" Encoding for DNA: Representing DNA sequences using a "one-hot" encoding scheme is effective for capturing fine-grained sequence information, which is crucial for deep learning models.
  • CNNs Excel at Local Feature Extraction: Convolutional Neural Networks are well-suited for identifying local patterns (motifs) in DNA sequences, which are important for classifying phage lifestyles.
  • Temperate Phages May Mimic Host Signatures: The authors hypothesize that temperate phages, spending more time integrated into host genomes, may adjust their sequence signatures to resemble bacterial chromosomes, leading to their misclassification as bacterial by some models.

Key Results

  • High Accuracy: DeePhage achieved a five-fold cross-validation accuracy of up to 88%, which is approximately 30% higher than the most related method, PHACTS.
  • Improved Performance on Real Data: On real metavirome data, DeePhage annotated 54.4% of reliable contigs, compared to PHACTS' 44.5%.
  • Computational Efficiency: DeePhage is significantly faster than PHACTS, reducing computational time by 810 times on the same machine.
  • Robustness to Fragment Length: DeePhage maintained reliable performance across different contig lengths, with accuracy increasing for longer fragments.
  • Advantage over PHACTS for Short Fragments: When tested on short DNA fragments (simulating metagenomic data), DeePhage achieved accuracies of 75%-88%, while PHACTS performed at near-random levels (around 50%).

Limitations and Future Work

  • Eukaryotic Viruses: DeePhage is currently designed for prokaryotic viruses (phages) and cannot identify eukaryotic viruses. The authors suggest developing a preprocessing module to filter out eukaryotic viruses for more reliable downstream analysis.
  • Ambiguous Predictions: Sequences with scores near 0.5 are considered uncertain. The tool includes an adjustable cutoff to filter these predictions, allowing users to balance reliability and the proportion of uncertain results.
  • Further Exploration of Mechanisms: The authors suggest further exploring the biological mechanisms behind the sequence signature differences that enable DeePhage's classification.
  • Application to Disease Studies: The tool's ability to detect phage transformations is highlighted as a promising avenue for gaining insights into disease conditions, such as inflammatory bowel disease.

Summary updated

Abstract

ABSTRACT Background Prokaryotic viruses referred to as phages can be divided into virulent and temperate phages. Distinguishing virulent and temperate phage-derived sequences in metavirome data is important for their role in interactions with bacterial hosts and regulations of microbial communities. However there is no experimental or computational approach to classify sequences of these two in culture-independent metavirome effectively, we present a new computational method DeePhage, which can directly and rapidly judge each read or contig as a virulent or temperate phage-derived fragment. Findings DeePhage utilizes a “one-hot” encoding form to have an overall and detailed representation of DNA sequences. Sequence signatures are detected via a deep learning algorithm, namely a convolutional neural network to extract valuable local features. DeePhage makes better performance than the most related method PHACTS. The accuracy of DeePhage on five-fold validation reach as high as 88%, nearly 30% higher than PHACTS. Evaluation on real metavirome shows DeePhage annotated 54.4% of reliable contigs while PHACTS annotated 44.5%. While running on the same machine, DeePhage reduces computational time than PHACTS by 810 times. Besides, we proposed a new strategy to explore phage transformations in the microbial community by direct detection of the temperate viral fragments from metagenome and metavirome. The detectable transformation of temperate phages provided us a new insight into the potential treatment for human disease. Conclusions DeePhage is the first tool that can rapidly and efficiently identify two kinds of phage fragments especially for metagenomics analysis with satisfactory performance. DeePhage is freely available via http://cqb.pku.edu.cn/ZhuLab/DeePhage or https://github.com/shufangwu/DeePhage .
BibTeX
@misc{wu2020distinguish,
  author = {Shu‐Fang Vivienne Wu and Zhencheng Fang and Jie Tan and Mo Li and Chunhui Wang and Qian Guo and Congmin Xu and Xiaoqing Jiang and Huaiqiu Zhu},
  title = {{Distinguish virulent and temperate phage-derived sequences in metavirome data with a deep learning approach}},
  year = {2020},
  doi = {10.1101/2020.12.25.424404},
  howpublished = {bioRxiv},
}
Figure