Tool Combination Best Practices

Last Updated: November 29, 2025

Combining multiple tools is essential for accurate virome analysis. This guide provides evidence-based recommendations for which tool combinations work best for different scenarios.

Why Combine Tools?

Single tools have limitations: - False positives: 10-30% for viral identification tools - False negatives: Miss 20-40% of viruses - Method bias: Different algorithms favor different virus types

Tool combinations provide: - Higher accuracy (consensus predictions) - Better coverage (union of predictions) - Confidence assessment (number of tools agreeing)

Viral Identification Tool Combinations

Recommended Combinations

Gold Standard (High Accuracy, Moderate Coverage)

# VirSorter2 + VIBRANT + geNomad (consensus ≥2)
# Expected: 60-70% of viruses, <5% false positives

# Run all three tools
virsorter run -i contigs.fa --min-score 0.5 -w vs2/ all
VIBRANT_run.py -i contigs.fa -folder vibrant/
genomad end-to-end contigs.fa genomad/ genomad_db/

# Take consensus (≥2 tools agree)
# Implementation in Tutorial 1

Best for: - Publication-quality datasets - Novel virus discovery - Low tolerance for false positives

Performance: - Sensitivity: 60-70% - Specificity: >95% - Runtime: 4-8 hours (1000 contigs, 16 cores)

Balanced (Good Accuracy, Good Coverage)

# VirSorter2 + geNomad (either tool)
# Expected: 75-85% of viruses, ~10% false positives

virsorter run -i contigs.fa --min-score 0.5 -w vs2/ all
genomad end-to-end contigs.fa genomad/ genomad_db/

# Union: any tool predicts viral
cat vs2/final-viral-combined.fa genomad/viral.fna | \
    cd-hit-est -c 0.95 -i - -o viral_union.fa

Best for: - Diversity studies - Exploratory analysis - Moderate sample size (10-50 samples)

Performance: - Sensitivity: 75-85% - Specificity: ~90% - Runtime: 2-4 hours (1000 contigs)

High Sensitivity (Maximum Coverage, Lower Accuracy)

# VirSorter2 (low threshold) + VIBRANT + geNomad + DeepVirFinder
# Expected: >90% of viruses, 15-25% false positives

virsorter run -i contigs.fa --min-score 0.3 -w vs2/ all
VIBRANT_run.py -i contigs.fa -folder vibrant/
genomad end-to-end contigs.fa genomad/ genomad_db/ --min-score 0.7
python dvf.py -i contigs.fa -o dvf/ -l 1000

# Union of all predictions
# MUST use CheckV to filter false positives

Best for: - Highly novel environments - Pilot studies - When missing viruses is worse than false positives

Performance: - Sensitivity: >90% - Specificity: ~75-85% - Requires: CheckV filtering to remove false positives

Tool-Specific Considerations

Tool	Strengths	Weaknesses	Best Parameters
VirSorter2	Broad virus types, well-tested	Slower, some false positives	`--min-score 0.5-0.7`
VIBRANT	Good for phages, built-in annotation	Misses some virus types	Default settings
geNomad	Fast, recent database, plasmid detection	Newer (less validated)	`--min-score 0.8`
DeepVirFinder	Works on short contigs	Needs GPU for speed, higher FP	`--len 1000` minimum

Quality Filtering After Prediction

Always run CheckV:

# After any viral identification
checkv end_to_end viral_contigs.fa checkv_out/ -t 8

# Recommended filters:
# Conservative: Complete + High-quality only
awk -F'\t' '$8 == "Complete" || $8 == "High-quality" {print $1}' \
    checkv_out/quality_summary.tsv > hq_ids.txt

# Balanced: Complete + High + Medium
awk -F'\t' '$8 != "Low-quality" && $8 != "Not-determined" {print $1}' \
    checkv_out/quality_summary.tsv > good_ids.txt

# Also filter by contamination
awk -F'\t' '$8 != "Low-quality" && $10 < 5 {print $1}' \
    checkv_out/quality_summary.tsv > clean_ids.txt

Host Prediction Tool Combinations

Tiered Approach (Recommended)

# Tier 1: CRISPR spacers (highest confidence, ~5-10% of phages)
# Use PILER-CR or minced + BLAST

# Tier 2: iPHoP integrated prediction (~40-60% of phages)
iphop predict --fa_file phages.fa --db_dir iphop_db/ --out_dir iphop/

# Tier 3: Individual methods for remaining phages
# - WIsH (genomic signatures)
# - CHERRY (deep learning)
# - Protein homology (BLAST)

# Combine with confidence weighting:
# - CRISPR match: 95% confidence
# - iPHoP + CRISPR: 90% confidence
# - iPHoP >90 score: 70% confidence
# - ≥2 methods agree: 60% confidence
# - Single method: 30% confidence

Implementation:

name="__codelineno-5-1" href="#__codelineno-5-1">def assign_host_confidence(predictions): """Assign confidence based on prediction method(s)""" confidence_map = { 'CRISPR': 0.95, 'iPHoP_high': 0.70, # iPHoP score >90 'iPHoP_medium': 0.50, # iPHoP score 70-90 'WIsH': 0.40, 'CHERRY': 0.40, 'Homology': 0.45, 'Consensus_2': 0.60, # ≥2 methods agree 'Consensus_3': 0.75, # ≥3 methods agree } # Logic to determine confidence if 'CRISPR' in predictions['methods']: return confidence_map['CRISPR'] elif len(predictions['methods']) >= 3: return confidence_map['Consensus_3'] elif len(predictions['methods']) >= 2: return confidence_map['Consensus_2'] # ... etc

When to Use Each Host Prediction Tool

Tool	Use When	Avoid When	Typical Success Rate
CRISPR	Bacterial MAGs available	Limited bacterial data	5-10% (very high confidence)
iPHoP	Any phages	Time/resource limited	40-60% (medium-high confidence)
WIsH	Bacterial genomes available	Only viral databases	30-50% (medium confidence)
CHERRY	Any phages, especially novel	Need genus-level precision	35-55% (medium confidence)
VirHostMatcher	Related hosts in database	Completely novel phages	25-45% (low-medium confidence)

Assembly Tool Combinations

Single Assembly vs Co-Assembly vs Hybrid

Scenario 1: Few Samples (1-5)

# Use individual assemblies per sample
for sample in sample1 sample2 sample3; do
    metaspades.py --metaviral \
        -1 ${sample}_R1.fq -2 ${sample}_R2.fq \
        -o assembly/${sample}
done

# Dereplicate across samples
cat assembly/*/contigs.fasta | \
    cd-hit-est -c 0.95 -aS 0.85 -o derep_contigs.fa

Scenario 2: Many Similar Samples (10-50)

# Co-assemble all samples together
cat sample*_R1.fq.gz > all_R1.fq.gz
cat sample*_R2.fq.gz > all_R2.fq.gz

metaspades.py --metaviral \
    -1 all_R1.fq.gz -2 all_R2.fq.gz \
    -o co_assembly/ -t 32 -m 250

# Better for shared viruses, misses rare sample-specific viruses

Scenario 3: Complex Design (Recommended for Most Studies)

# Hybrid approach:
# 1. Co-assemble within groups (timepoints, treatments, etc.)
# 2. Individual assemblies for each sample
# 3. Combine and dereplicate

# Co-assembly per timepoint
metaspades.py --metaviral -1 T1_all_R1.fq -2 T1_all_R2.fq -o T1_coasm/
metaspades.py --metaviral -1 T2_all_R1.fq -2 T2_all_R2.fq -o T2_coasm/

# Individual assemblies
for sample in T1_R1 T1_R2 T2_R1 T2_R2; do
    metaspades.py --metaviral -1 ${sample}_R1.fq -2 ${sample}_R2.fq -o ${sample}_asm/
done

# Combine and dereplicate
cat *_coasm/contigs.fasta *_asm/contigs.fasta | \
    cd-hit-est -c 0.95 -aS 0.85 -M 64000 -T 16 -o final_derep.fa

Annotation Tool Combinations

Phage Annotation Pipeline

# Step 1: Pharokka (fast, phage-specific)
pharokka.py -i phages.fa -o pharokka/ -t 8 -d pharokka_db/

# Step 2: DRAMv (metabolic annotation)
DRAM-v.py annotate -i phages.fa -o dramv/ --threads 8
DRAM-v.py distill -i dramv/annotations.tsv -o dramv_distill/

# Step 3: PHROGs (for additional functional annotation)
hmmsearch --tblout phrogs_hits.txt -E 1e-5 --cpu 8 \
    PHROGs.hmm pharokka/proteins.faa

# Combine annotations (take best from each tool)
python3 combine_annotations.py \
    --pharokka pharokka/*.gff \
    --dramv dramv/annotations.tsv \
    --phrogs phrogs_hits.txt \
    --output combined_annotations.tsv

When to use each: - Pharokka: Primary phage annotation (fast, comprehensive) - DRAMv: Auxiliary metabolic genes (AMGs), metabolic pathways - PHROGs: Additional functional categories - PROKKA: If Pharokka fails (less phage-specific)

Taxonomic Classification Combinations

Multi-Method Taxonomy

# Method 1: BLAST (sequence similarity)
blastn -query viral.fa -db nt -outfmt 6 -max_target_seqs 5 > blast.txt

# Method 2: vConTACT2 (protein sharing network)
vcontact2 --raw-proteins proteins.faa --db ProkaryoticViralRefSeq --output vcontact2/

# Method 3: PhaGCN (graph convolutional network)
python PhaTYP.py --contigs viral.fa --threads 8

# Combine:
# - BLAST for known viruses (>80% identity)
# - vConTACT2 for viral clusters (species-level)
# - PhaGCN for novel viruses (family-level)

Decision tree:

For each virus:
├─ BLAST identity >90%? → Use BLAST taxonomy (high confidence)
├─ BLAST identity 70-90%? → Use BLAST family + vConTACT2 genus (medium confidence)
├─ vConTACT2 cluster with references? → Use cluster taxonomy (medium confidence)
└─ No hits? → Use PhaGCN family prediction (low confidence) + mark as novel

Abundance Estimation Combinations

Read Mapping Strategy

# Method 1: BBMap (sensitive, slower)
bbmap.sh in=reads.fq ref=viral.fa out=bbmap.sam covstats=bbmap_cov.txt

# Method 2: Bowtie2 (fast, standard)
bowtie2-build viral.fa viral_idx
bowtie2 -x viral_idx -1 R1.fq -2 R2.fq -S bowtie2.sam

# Method 3: CoverM (batch processing, multiple metrics)
coverm contig --coupled *_R1.fq *_R2.fq --reference viral.fa \
    --methods mean trimmed_mean covered_fraction variance \
    --output-file coverm_abundance.tsv

# Recommendation: Use CoverM for multi-sample studies

Coverage metrics comparison:

Metric	Robust to Outliers	Good for Low Coverage	Best Use
Mean	No	Yes	Even coverage, high depth
Trimmed Mean	Yes	Yes	Recommended for most cases
Median	Yes	No	Very uneven coverage
RPKM	No	No	Cross-sample comparison
Covered Fraction	N/A	Yes	Presence/absence

Statistical Analysis Combinations

Diversity Analysis

# R script combining multiple diversity metrics

library(vegan)
library(phyloseq)

# Alpha diversity (within-sample)
richness <- specnumber(abundance)  # Species richness
shannon <- diversity(abundance, index="shannon")  # Shannon
simpson <- diversity(abundance, index="simpson")  # Simpson

# Beta diversity (between-sample)
bray <- vegdist(abundance, method="bray")  # Bray-Curtis
jaccard <- vegdist(abundance, method="jaccard", binary=TRUE)  # Jaccard

# Ordination
nmds_bray <- metaMDS(abundance, distance="bray", k=2)
pca <- rda(abundance)

# Statistical tests
# PERMANOVA for group differences
adonis2(abundance ~ Treatment, data=metadata, method="bray")

# ANOSIM (alternative)
anosim(abundance, metadata$Treatment, method="bray")

# Recommendation: Use PERMANOVA (more powerful)

Differential Abundance

# Use DESeq2 for count data (recommended)
# Use ANCOM for compositional data (alternative)

library(DESeq2)
library(ALDEx2)

# Method 1: DESeq2 (recommended for most cases)
dds <- DESeqDataSetFromMatrix(
    countData = counts,
    colData = metadata,
    design = ~ Treatment
)
dds <- DESeq(dds)
results <- results(dds, contrast=c("Treatment", "A", "B"))

# Method 2: ALDEx2 (for compositional data)
aldex_result <- aldex(counts, metadata$Treatment, mc.samples=128)

# Recommendation:
# - DESeq2 for most virome studies
# - ALDEx2 if concerned about compositionality
# - Validate with both if results differ substantially

Workflow Integration Examples

Complete Virome Workflow (Recommended Stack)

#!/bin/bash
# Complete virome analysis combining best practices

# 1. QC
fastp -i R1.fq.gz -I R2.fq.gz -o clean_R1.fq -O clean_R2.fq

# 2. Assembly (hybrid approach)
metaspades.py --metaviral -1 clean_R1.fq -2 clean_R2.fq -o assembly/

# 3. Viral ID (consensus of 3 tools)
virsorter run -i assembly/contigs.fa --min-score 0.5 -w vs2/ all
VIBRANT_run.py -i assembly/contigs.fa -folder vibrant/
genomad end-to-end assembly/contigs.fa genomad/ genomad_db/

# Consensus
python3 consensus_viral_prediction.py \
    --vs2 vs2/final-viral-combined.fa \
    --vibrant vibrant/phages.fa \
    --genomad genomad/viral.fna \
    --min-tools 2 \
    --output consensus_viral.fa

# 4. Quality check
checkv end_to_end consensus_viral.fa checkv/ -t 8

# 5. Annotation
pharokka.py -i checkv/viruses.fna -o pharokka/ -t 8
DRAM-v.py annotate -i checkv/viruses.fna -o dramv/ --threads 8

# 6. Taxonomy
blastn -query checkv/viruses.fna -db nt -outfmt 6 > blast_tax.txt
vcontact2 --raw-proteins pharokka/proteins.faa --output vcontact2/

# 7. Abundance
coverm contig --coupled clean_R*.fq --reference checkv/viruses.fna \
    --methods trimmed_mean --output-file abundance.tsv

# 8. Host prediction
iphop predict --fa_file checkv/viruses.fna --db_dir iphop_db/ --out_dir iphop/

Tool Compatibility Matrix

Upstream Tool	Compatible Downstream Tools	Notes
metaviralSPAdes	VirSorter2, VIBRANT, geNomad	All viral ID tools
VirSorter2	CheckV, Pharokka, DRAMv	Standard workflow
VIBRANT	CheckV, DRAMv	Built-in annotation
CheckV	Pharokka, DRAMv, vConTACT2	Use `viruses.fna` output
Pharokka	vConTACT2, DRAMv	Use protein FAA
DRAMv	Custom scripts	Metabolic analysis

Common Pitfalls in Tool Combinations

❌ Anti-Patterns (Avoid These)

Using only one viral ID tool
False positive rate too high (10-30%)
Solution: Always use ≥2 tools
Not running CheckV after viral ID
Can't assess quality or remove contamination
Solution: Always run CheckV
Over-relying on machine learning tools
Need validation with sequence-based methods
Solution: Combine ML with BLAST/CRISPR
Ignoring tool version differences
Databases and algorithms change
Solution: Record versions, use same version within study
Combining incompatible tools
E.g., using DNA assembler for RNA viruses
Solution: Check tool documentation

✅ Best Practices

Consensus predictions (≥2 tools agree)
CheckV filtering (remove low quality)
Multiple evidence types (CRISPR + homology + ML)
Version control (document all tool versions)
Appropriate thresholds (adjust for your goals)

Computational Resource Considerations

Tool Combination	RAM	CPUs	Runtime (1000 contigs)
VirSorter2 + geNomad	32GB	16	2-3 hours
VirSorter2 + VIBRANT + geNomad	64GB	16	4-6 hours
Full stack (assembly + ID + annotation)	128GB	32	12-24 hours

Optimization tips: - Run tools in parallel when possible - Use high-memory nodes for assembly - Cache databases (don't re-download)