flowchart LR A[Reads or consensus FASTA] --> B[Nextflow orchestration] B --> C[Segment and reference selection] C --> D[Variant calling] D --> E[Normalization and MNV correction] E --> F[Reference-aware annotation] F --> G[Mutation watchlists] G --> H[Reports] G --> I[(PostgreSQL results)] D --> J[Independent validation] F --> J
Influenza surveillance
Validated mutation analysis across all eight genome segments
A case study in reference-aware influenza mutation surveillance, independent validation, and production workflow design.
$ case-study –system influenza-surveillance
The problem
Influenza A surveillance has to reconcile eight independently reassorting segments, multiple reference systems, biological numbering conventions, and continuously changing mutation definitions. A useful production system must do more than call variants: it must keep reference context attached to every result, detect invalid configurations before expensive computation, and produce evidence that downstream users can audit.
The system
The platform accepts either sequencing reads or consensus FASTAs. Nextflow coordinates reference selection and per-segment analysis; shared Python modules normalize variants, correct multi-nucleotide events, translate reference-aware coordinates, evaluate configurable mutation watchlists, and prepare structured results for reporting and persistence.
Key design choices include:
- a validated reference-bundle abstraction rather than scattered file paths;
- fail-fast checks that understand which workflow and input mode is active;
- cross-sample batching by segment to make efficient use of HPC resources;
- explicit handling of HA1/HA2 and arbitrary reference-coordinate systems;
- independently testable Python modules shared across workflow entry points.
Validation evidence
Validation used independent implementations and orthogonal comparisons rather than treating agreement with one tool as ground truth.
| Evaluation | Result |
|---|---|
| All SNVs, independent Biopython ground truth | F1 0.9915 across 192 samples × 8 segments |
| Missense variants | F1 0.9531 after resolving annotation failures |
| Independent calling-branch agreement | Macro Jaccard 0.9922 across the preregistered clean set |
Discrepancy analysis exposed two concrete failure modes: limited alignment reach in difficult cases and missing feature annotations. Fixing the annotation problem raised missense F1 from 0.6510 to 0.9531. This is the engineering value of the validation process: it produced changes to the system, not only a score.
Contribution boundary
The mutation-surveillance platform described above is solely architected and principally implemented by me. It operates within a wider ecosystem of shared public-health pipelines, databases, and infrastructure. My work in those shared repositories includes workflow features, validation, provenance, automation, and operational improvements; it is not presented as sole ownership of the upstream projects.
Public artifacts
BCCDC-PHL/fluviewer-nf— a shared public influenza workflow to which I contributed.CFIA-NCFAD/nf-flu— the maintained upstream pipeline used in the later laboratory transition.
The current production mutation-surveillance source is not public. This case study is intentionally limited to architecture and validation facts that can be described safely.