Skip to content

Host Prediction Tools

Last Updated: November 29, 2025

Host prediction tools help determine the bacterial hosts of viral sequences, particularly phages. This is crucial for understanding phage-host interactions, designing phage therapy, and interpreting the role of phages in microbial communities.

Reality Check: Host Prediction Accuracy

Host prediction is one of the most challenging problems in virome analysis. Current computational tools have significant limitations:

  • Real-world accuracy: Typically 30-60% at genus level for novel viruses (not the 75-80% often cited from benchmarks on known viruses)
  • Many predictions fail: 20-40% of viral sequences may get no confident prediction
  • Predictions are hypotheses: Always treat them as computational predictions requiring validation, not ground truth
  • Method-dependent: Different tools can give conflicting predictions for the same sequence

Best practices: - Use multiple prediction methods and look for consensus - Higher confidence when multiple lines of evidence agree (sequence homology + CRISPR spacers + co-occurrence data) - CRISPR spacer matches are the "gold standard" but only available for ~1-5% of viruses - Consider biological context (sample environment, known host distributions) - Validate computationally important predictions experimentally when possible

Key Host Prediction Tools

iPHoP

iPHoP (Integrated Prediction of Host and Phage) is a state-of-the-art tool for predicting phage-host interactions at various taxonomic levels.

  • Version: v1.3.3, 2023
  • Installation: conda install -c bioconda iphop
  • GitHub Stars: Hosted on Bitbucket
  • Key Features:
  • Integrates multiple prediction methods
  • Assigns confidence scores to predictions
  • Works at multiple taxonomic levels
  • Pre-trained with a comprehensive database
  • Optimized for metagenomic data

Usage Example:

iphop predict --fa_file input_phages.fasta --out_dir iphop_results

CHERRY

CHERRY (v1.0, 2022) uses deep learning for phage host prediction.

  • GitHub Stars: ⭐ 24
  • Installation: git clone https://github.com/KennthShang/CHERRY.git
  • Key Features:
  • Uses deep learning (CNN and LSTM)
  • Pre-trained with thousands of phage-host pairs
  • Works well with novel phages
  • Can make predictions at different taxonomic levels

How It Works: CHERRY uses a combination of convolutional neural networks and LSTM to learn sequence patterns indicative of phage-host interactions.

VirHostMatcher-Net

VirHostMatcher-Net is a network-based virus-host prediction tool.

  • GitHub Stars: ⭐ 21
  • Key Features:
  • Uses both oligonucleotide frequencies and protein alignment
  • Incorporates network-based information
  • Achieves high accuracy at genus level
  • Works well with novel viruses

WIsH

WIsH (Who Is the Host) predicts phage-host interactions using genome homology.

  • Key Features:
  • Uses Markov models of genomic composition
  • Fast and lightweight
  • Good performance on well-characterized host taxa
  • Works well with complete genomes

Additional Host Prediction Tools

Host Prediction Based on Sequence Similarity

  • HostPhinder: K-mer based phage host prediction
  • PHP: Phage host prediction tool
  • PHPGCA: Similarity graphs for phage-host prediction

Host Prediction Based on CRISPR Spacers

Host Prediction Based on Machine Learning

  • DeepHost: CNN for phage host prediction
  • HostG: Graph convolutional network for phage host prediction
  • PhageHostLearn: Machine learning for phage-host prediction

Comparison Table

Tool Method Taxonomic Level Strengths Limitations
iPHoP Integrated Kingdom to Strain Combines multiple methods, high accuracy Resource intensive
CHERRY Deep Learning Genus, Species Works well with novel phages Needs substantial computing power
VirHostMatcher-Net Network-based Genus Good accuracy at genus level Limited to specific host taxa
WIsH Markov Models Species Fast, lightweight Better with complete genomes
HostPhinder K-mer Species Simple, efficient Limited to known hosts
CrisprOpenDB CRISPR spacers Strain High specificity Limited coverage

Performance Benchmarks

Understanding Benchmark Numbers

The performance metrics below are from controlled benchmark studies on test datasets with known virus-host pairs. Real-world performance on novel environmental viruses is typically 20-40% lower.

Benchmark studies often: - Use viruses with known hosts (easier than novel viruses) - Test on sequences similar to training data - Exclude difficult cases - Report best-case accuracy

For your research: Expect genus-level accuracy of 30-60% on novel viruses, not 75-80%.

Based on published benchmarks (test datasets with known hosts):

  • Genus level prediction (benchmark datasets):
  • VirHostMatcher-Net: ~75-80% (novel viruses: ~40-50%)
  • iPHoP: ~70-75% (novel viruses: ~45-55%)
  • CHERRY: ~65-70% (novel viruses: ~35-45%)

  • Species level prediction (benchmark datasets):

  • iPHoP: ~60-65% (novel viruses: ~30-40%)
  • CHERRY: ~55-60% (novel viruses: ~25-35%)
  • WIsH: ~50-55% (novel viruses: ~25-30%)

Performance depends on: viral genome completeness, taxonomic coverage of reference databases, host diversity in training data, and sequence similarity to known virus-host pairs.

For optimal host prediction results, we recommend a multi-tool approach:

  1. Start with iPHoP:
  2. Comprehensive predictions with confidence scores
  3. Good starting point for most analyses

  4. Complement with specialized tools:

  5. For well-studied hosts: WIsH and HostPhinder
  6. For novel phages: CHERRY and VirHostMatcher-Net
  7. If CRISPR data is available: CrisprOpenDB and SpacePHARER

  8. Consensus approach:

  9. Take predictions agreed upon by multiple methods
  10. Weight predictions by confidence scores when available
  11. Consider biological context (e.g., environment of isolation)

Future Directions

The field of phage host prediction is rapidly evolving:

  • Integration of metagenomic co-occurrence data
  • Improved deep learning models
  • Single-cell and spatial information integration
  • Expanded reference databases

Further Reading