Application Note: Production Variant Calling and Long-Read Assembly

Reference architecture

Use AxiKeystone for method development and memory-heavy single-node work; AxiLattice for the production Slurm queue; and AxiForge for validated GPU-accelerated or imaging/protein stages. Protected NAS holds raw reads, references and approved outputs. Per-job NVMe holds sorting, temporary BAM/CRAM and assembler scratch.

Workflow stages

1. Validate sample sheet, checksums and metadata.
2. Run read QC and contamination/identity checks.
3. Align, sort and mark/process reads in immutable containers.
4. Call per-sample variants, then joint genotype where required.
5. Apply approved filters and annotation.
6. For long reads, reserve a high-memory node for assembly/polishing.
7. Publish results with a run manifest and QC report.

Scheduler profile

Create resource labels for standard CPU, high-memory and GPU stages. Let Nextflow request the correct Slurm partition and memory, rather than embedding hostnames. Cap concurrency for metadata-heavy stages and pre-stage reference bundles to node-local storage. Use retry only for clearly transient failures; scientific tool errors need review.

Acceptance evidence

Run recognised truth or internal reference samples and compare sensitivity/precision, genotype concordance, coverage, contamination and structural metrics. Record wall time, samples/day, peak RAM and scratch. Validate restore, interrupted-run resume and reference/version rollback.

Operational controls

Encrypt governed storage, audit access and restrict exports. Do not include identifiers in generic monitoring labels. Separate research findings from clinically reportable outputs. Revalidate when changing caller, reference, container, major driver/runtime or workflow logic.