Prediction of virus-host infectious association by supervised learning methods

Jie Ren, Mengge Zhang, Lianping Yang, Jie Ren 0006, Nathan A. Ahlgren, Jed A. Fuhrman, Fengzhu SunBMC Bioinformatics 2017 · DOI

TL;DR

This paper introduces a machine learning approach to predict which viruses infect specific bacterial hosts, a crucial step in understanding microbial communities. By analyzing viral sequence "word frequencies," they found that a Random Forest model with relative word frequencies achieves high accuracy, outperforming previous methods and enabling predictions even from short viral fragments.

What Problem Does This Paper Address?

Microbial communities are teeming with viruses, and understanding which viruses infect which hosts is fundamental to grasping their ecological roles, impact on host physiology, and involvement in diseases. However, a significant challenge arises from metagenomic studies, which often yield vast amounts of viral DNA sequences (contigs) without direct host information. While some methods exist to infer virus-host associations based on sequence similarity or "word patterns," this area remains significantly understudied. The authors aim to develop more accurate and efficient computational methods to predict these associations, particularly from fragmented viral sequences common in metagenomic data.

Methods

The core of this research lies in applying supervised machine learning to predict virus-host infectious associations. The authors explored several key methodological components:

  • Feature Engineering: They represented viral sequences as feature vectors based on "word frequencies" (k-mers). Four different representations were tested, including simple relative word frequency and normalized versions that account for expected word counts. The length of these words (k) was varied (4, 6, 8) to find optimal patterns.
  • Machine Learning Models: Five common supervised learning algorithms were evaluated: Logistic Regression, Support Vector Machine (SVM) with RBF kernel, Random Forest (RF), Gaussian Naive Bayes, and Bernoulli Naive Bayes.
  • Model Training and Evaluation: For each of nine bacterial host genera with sufficient viral data, models were trained on existing viral genomes and tested on newly discovered viruses. The primary evaluation metric was the Area Under the Receiver Operating Characteristic Curve (AUC), a standard measure of classification performance.
  • Handling Metagenomic Data: The study also investigated the performance of their best model on short viral contigs (1kb, 3kb, 5kb) and whole genomes, simulating real-world metagenomic data, including the introduction of sequencing errors.
  • Estimating Reliability in Viral Tagging: A maximum likelihood approach was developed to estimate the fraction of true infectious associations from viral tagging experiments, which can suffer from false positives and negatives.

Datasets

The primary dataset used was a collection of 1,426 completely sequenced viral genomes with known host information, obtained from the NCBI viral genome database. The study focused on nine bacterial host genera that had at least 45 identified infecting viruses to ensure sufficient data for training machine learning models. For evaluating predictions on metagenomic data, viral contigs of various lengths (1kb, 3kb, 5kb, and whole genomes) were generated from these complete viral genomes, with simulated sequencing errors.

Key Insights and Ideas

  • Random Forest Reigns Supreme: The Random Forest (RF) algorithm consistently outperformed other machine learning methods across various feature representations and word lengths. This suggests RF's robustness in capturing complex patterns in viral sequences for host prediction.
  • Relative Word Frequency is Sufficient: Surprisingly, the simplest feature representation—relative word frequency—performed as well as, or even better than, more complex normalized frequency measures. This highlights the power of basic sequence composition analysis.
  • Optimal Word Length Matters: The choice of word length (k) significantly impacts prediction accuracy. For most host genera, k=6 provided the best balance, capturing sufficient sequence information without introducing excessive noise.
  • Short Contigs are Predictable: Even short viral contigs (as small as 3-5 kbps) can be reliably used to predict virus-host associations, making the approach highly applicable to fragmented metagenomic data. This is a crucial finding for analyzing real-world environmental samples.
  • Host-Specific Patterns: The performance variation across different host genera could be partially explained by the taxonomic composition of the viruses infecting them and the similarity of their word patterns.

Key Results

  • High Prediction Accuracy: For the nine host genera, the Random Forest model using relative word frequencies achieved AUC scores consistently above 0.85, with many exceeding 0.98 when k=6.
  • Outperforming Existing Methods: The RF approach with word frequency vectors demonstrated higher accuracy than previous methods based on Manhattan and d½ dissimilarity measures for predicting hosts of viral contigs.
  • Viral Tagging Reliability: The developed maximum likelihood method estimated that approximately 95% of identified T4-like viruses in a viral tagging experiment likely infect Synechococcus. However, for non-T4-like viruses and general contigs, the estimated infection fraction was much lower (around 29-30%), highlighting the potential for false positives in such experiments.

Limitations and Future Work

  • Data Dependency: The machine learning approach requires a relatively large number of known virus-host associations for training. This limits its applicability to hosts for which many infecting viruses have been identified.
  • Feature Representation: While four feature representations were explored, other methods (e.g., allowing mismatches in word counting) might yield even better results.
  • Model Specificity: Only one version of each machine learning algorithm was tested (e.g., RBF kernel for SVM). Exploring other kernel types or algorithm variations could improve performance.

The authors suggest that future work could focus on expanding the number of host genera studied, exploring more sophisticated feature representations, and investigating alternative machine learning models. The current study clearly demonstrates the potential of machine learning for predicting virus-host associations, with significant implications for microbial ecology and virology.

Summary updated

BibTeX
@article{ren2017prediction,
  author = {Jie Ren and Mengge Zhang and Lianping Yang and Jie Ren and Nathan A. Ahlgren and Jed A. Fuhrman and Fengzhu Sun},
  title = {{Prediction of virus-host infectious association by supervised learning methods}},
  journal = {BMC Bioinformatics},
  year = {2017},
  doi = {10.1186/s12859-017-1473-7},
}
Figure