Afanc: taxonomic deconvolution and probabilistic lineage assignment for pathogen genomic surveillance

Technical Summary of "Afanc: taxonomic deconvolution and probabilistic lineage assignment for pathogen genomic surveillance"

Problem Statement
Reliable pathogen genomic surveillance requires tools that can simultaneously achieve broad taxonomic sensitivity and high-resolution lineage or clade assignment. Current methodologies typically address only one of these tasks. Genomic distance or sequence content-based methods (e.g., Kraken 2, Bracken) offer rapid, broad-spectrum screening but struggle with specificity when related taxa occupy overlapping sequence spaces, often resulting in ambiguous or false-positive low-level assignments. Conversely, marker-based approaches (e.g., Mykrobe, TB-Profiler) provide high-resolution classification within well-characterized systems but rely on curated marker sets, assume sufficient separation between classes, and perform poorly when evidence is incomplete, overlapping, or hierarchically ambiguous. Furthermore, existing workflows often fail to explicitly represent uncertainty or handle mixed taxonomic signals, forcing overconfident assignments in complex scenarios.

Methodology
The authors present Afanc, an extensible framework designed to integrate database construction, taxonomic deconvolution, and probabilistic lineage assignment. The system consists of four discrete sub-tools:

1. get_dataset: Downloads genome assemblies from GenBank based on user-defined taxonomic lists.
2. autodatabase: Constructs a reference database by reconciling input assemblies with NCBI taxonomy. It employs a Mash-based distance index to filter assemblies, retaining only representative, high-quality members of each taxon panel to remove outliers and redundant sequences before database construction.
3. screen: Profiles input read sets against the constructed database. This module executes a five-stage workflow:
   • Kraken 2 Reporting: Generates an initial taxonomic report.
   • Report Deconvolution: A novel algorithm parses the hierarchical Kraken 2 tree to identify supported taxa. It redistributes ambiguous reads from unsupported nodes to supported neighboring taxa using evidence-weighted read redistribution. This process utilizes elastic thresholds calculated from parent-child similarity values (derived from the Mash index) rather than fixed global cutoffs, allowing for sensitive detection of low-abundance taxa while minimizing false positives.
   • Competitive Mapping: Reads are mapped to reference assemblies associated with detected taxa.
   • Variant Calling: SNPs are called from the mapped reads.
   • Lineage Classification: If a lineage model exists for the detected taxon, the SNP profile is passed to a Bayesian classifier.
4. classify: Allows the application of lineage models to existing VCF files without re-running the full screening workflow.

The lineage classification engine utilizes a Bayesian SNP-based classifier operating on two model families: Canonical-Epsilon-Bayesian (CEB) models, which use pseudo-empirical emissions based on canonical markers, and Semi-Bayesian (SB) empirical models, which estimate emission probabilities from empirical allele frequencies. The classifier explicitly distinguishes between "callable absence" (negative evidence) and "missingness" (uncalled sites), excluding the latter from likelihood calculations to avoid penalizing low-coverage samples.

Key Contributions

• Hybrid Framework: Afanc bridges the gap between broad taxonomic screening and fine-scale lineage assignment, integrating database curation, report deconvolution, and probabilistic classification into a single workflow.
• Elastic Thresholding and Read Redistribution: The development of a novel algorithm for Kraken 2 report deconvolution that uses elastic, taxon-specific thresholds and evidence-weighted read redistribution to resolve complex, compound datasets without relying on brittle global cutoffs.
• Explicit Uncertainty Handling: The framework explicitly represents ambiguity and incomplete evidence, reporting unclassified samples or closest-lineage signals rather than forcing incorrect discrete assignments.
• Quality-Controlled Database Construction: The autodatabase module ensures reference quality by filtering out atypical or low-quality assemblies using Mash distance metrics before classification.

Results
The authors evaluated Afanc across three distinct pathogen systems:

1. Mycobacteriaceae (Taxonomic Profiling): Afanc correctly assigned species and subspecies across all simulated coverage cohorts, including those with only 1% genome breadth. It outperformed Kraken 2, Bracken, KrakenUniq, and Mykrobe in both sensitivity and precision, particularly in low-coverage and mixed-sample scenarios. In mixed datasets, Afanc recovered all expected taxa, whereas comparator tools produced spurious assignments or failed to detect constituents at lower coverage.
2. Mycobacterium tuberculosis (Lineage Classification): In a dataset of 958 samples, Afanc assigned 955 samples (99.7%) at harmonized resolution with zero wrong-branch calls. It outperformed TB-Profiler (99.6% correct, 2 wrong-branch) and Mykrobe (98.7% correct, 1 wrong-branch). At full resolution, Afanc achieved 99.3% exact or hierarchy-compatible assignments.
3. Plasmodium falciparum (Geolineage Prediction): Afanc assigned 18,830/20,712 samples (90.9%) to regional lineages, outperforming Malaria-Profiler (75.1%) by 15.8 percentage points. The improvement was consistent across regions, though Central Africa remained a challenging class for both tools due to biological overlap.
4. Mpox Virus (Clade/Lineage Assignment): Afanc produced no wrong-branch assignments. When predictions were harmonized to the resolution of available metadata, it achieved 99.4% compatibility, successfully distinguishing shallow, rapidly evolving outbreak lineages.

Significance and Claims
The paper claims that Afanc provides a reusable framework for converting noisy, broad taxonomic evidence into interpretable lineage-level inferences across diverse pathogen systems (bacteria, parasites, and viruses). Its primary contribution is not a new marker catalogue for a specific organism, but a unified workflow that links database curation, taxonomic deconvolution, and probabilistic classification.

The authors position Afanc as a classification layer for genomic surveillance that addresses critical failure points in current workflows: poor reference database quality, the misinterpretation of complex taxonomic reports, and the inability to handle mixed or low-coverage signals. By preserving uncertainty and explicitly handling incomplete evidence, Afanc supports epidemiological and evolutionary interpretations where competing classes are weakly separated. The study concludes that high-resolution pathogen classification can be embedded within a single extensible framework, making it suitable for surveillance environments where pathogen targets and reference data evolve over time.