Phylogenetic Placement

The core statistical engine behind tools like pplacer and EPA-ng evaluates the likelihood of a query sequence residing at every possible branch of a reference tree. The algorithm computes the posterior probability distribution of placements, identifying the optimal insertion point that maximizes the evolutionary likelihood given a substitution model. This approach avoids the computational cost of full de novo tree inference while maintaining statistical rigor.

Placement accuracy is fundamentally constrained by the quality and comprehensiveness of the fixed reference phylogeny. The reference tree—typically built from full-length 16S rRNA genes, whole genomes, or multi-locus alignments—defines the topological space into which queries are inserted. Key considerations include:

• Taxon sampling density: Sparse reference clades reduce placement resolution
• Tree reconstruction method: Maximum likelihood or Bayesian trees preferred over neighbor-joining
• Alignment quality: Poorly aligned reference sequences propagate errors into placement

EPA-ng is the high-performance reimplementation of the original Evolutionary Placement Algorithm, optimized for large-scale metagenomic and metabarcoding studies. It processes thousands of query sequences by:

• Inserting each query independently onto every branch of the reference tree
• Computing likelihood scores using the RAxML phylogenetic likelihood kernel
• Reporting the maximum likelihood placement and its uncertainty via likelihood weight ratios EPA-ng scales linearly with the number of reference taxa and queries, making it suitable for datasets with millions of short reads.

Unlike hard taxonomic assignments, phylogenetic placement produces a likelihood weight ratio (LWR) distribution across multiple candidate branches. This probabilistic output quantifies placement uncertainty, revealing:

• Ambiguous placements: Queries that spread probability across sister clades
• Novel lineages: Sequences that place on deep branches with low confidence, suggesting divergent taxa absent from the reference
• Chimeric detection: Artifacts that distribute probability anomalously across distant clades Tools like guppy visualize these distributions as edge masses on the reference tree.

Phylogenetic placement differs fundamentally from taxonomic classification tools like Kraken2 or MetaPhlAn. While classifiers assign queries to named taxonomic bins using k-mer matching or marker genes, placement methods:

• Preserve evolutionary context: Queries are positioned relative to known organisms on a continuous tree
• Detect novelty: Sequences from undescribed lineages place on internal branches rather than being forced into existing taxa
• Enable phylogeny-aware downstream analysis: Placement distributions feed directly into edge principal components analysis and UniFrac distance calculations

Placement algorithms balance computational efficiency with statistical accuracy. Key performance factors include:

• Reference tree size: EPA-ng handles trees with tens of thousands of taxa; pplacer is optimized for smaller, curated references
• Query preprocessing: Aligning queries to the reference alignment using hmmalign or MAFFT is often the computational bottleneck
• Parallelization: EPA-ng supports multi-threaded execution, distributing queries across cores
• Memory footprint: The likelihood computation stores branch-length and substitution model parameters, scaling with reference tree edges For ultra-large metagenomic studies, pre-filtering queries by taxonomic markers before placement reduces runtime.

The statistical engine behind tools like pplacer and EPA-ng. For each query sequence, the algorithm evaluates the likelihood of it attaching to every possible branch of the reference tree. The placement that maximizes the likelihood score—given a substitution model like GTR+GAMMA—is selected as the most probable evolutionary origin. This approach avoids a full de novo tree reconstruction, making it computationally tractable for millions of short reads.

A fixed, pre-computed evolutionary tree built from full-length, high-quality sequences—typically 16S rRNA genes or whole genomes. The tree defines the backbone topology onto which query sequences are placed. Key requirements:

• Taxonomic breadth: Must span the expected diversity of the sample
• Branch length accuracy: Directly impacts placement precision
• Alignment integrity: The reference multiple sequence alignment must be curated to avoid introducing systematic bias during placement

The core algorithm implemented in EPA-ng and RAxML-EPA. It works by:

1. Aligning each short query read to the reference alignment using hmmalign or PaPaRa
2. Computing the likelihood of the query attaching to every branch of the reference tree
3. Assigning the query to the branch that yields the maximum likelihood score

EPA-ng accelerates this process using SIMD vectorization and parallel processing, enabling placement of millions of queries per hour.

A confidence metric for phylogenetic placements. Rather than assigning a query to a single branch, pplacer distributes the query's mass across multiple candidate branches, weighted by their relative likelihood. The LWR for a given branch is the proportion of total likelihood it accounts for. A placement with an LWR of 0.95+ indicates high confidence; lower values suggest ambiguity, often due to short read length or conserved regions with insufficient phylogenetic signal.

A probabilistic model of nucleotide or amino acid evolution that defines the rates at which one character state changes to another. Common models in phylogenetic placement:

• GTR (General Time Reversible): Allows six distinct substitution rates
• GAMMA rate heterogeneity: Models variation in evolutionary rates across sites using a discrete gamma distribution
• CAT model: A mixture model that assigns sites to rate categories, used in PhyloBayes for handling compositional heterogeneity

The choice of model directly impacts placement accuracy, especially for deeply divergent sequences.

Once queries are placed, the resulting jplace file enables rich ecological and evolutionary analyses:

• Phylogenetic diversity metrics: Faith's PD, weighted UniFrac
• Edge principal components analysis (Edge PCA): Ordination of samples based on placement distributions
• Squash clustering: Phylogenetically-aware clustering of samples
• Guppy: A visualization toolkit for exploring placement distributions on the reference tree

These methods provide statistically rigorous alternatives to OTU-based approaches.