Built a small Nextflow DSL2 pipeline for bulk RNA-seq differential expression (AML vs. healthy) with a deliberately dual-mode design: a fast simulated-counts mode with planted log-fold-changes on canonical AML genes for CI and self-checking, and a --real_data mode that pulls TCGA-LAML and GTEx whole-blood gene-level counts from the recount3 project (uniform Monorail / STAR / GENCODE v26 alignment) so the same pipeline operates on real public RNA-seq cohorts.

End-to-end on 50 AML vs. 50 healthy whole-blood samples × ~20 k expressed genes: pipeline runs in ~11 s, recovers all 20 canonical AML signature genes (FLT3, KIT, MEIS1, HOXA9, MPO, CD34, RUNX1, GATA2, …) as significant at FDR<0.05, with 14/20 in the published-literature direction. The 6 “direction mismatches” are progenitor-associated TFs (RUNX1, GATA2, DNMT3A, ASXL1, TET2) — a known confound when comparing AML bone-marrow blasts against mature peripheral blood. The signature-recovery stage surfaces it explicitly rather than hiding it.

Five Nextflow processes — LOAD_REAL_COUNTS (or SIMULATE_COUNTS) → NORMALIZE_COUNTS → RUN_DE (Welch t-test on log2-CPM + hand-rolled BH FDR) → MAKE_VOLCANO (interactive Plotly) → SIGNATURE_RECOVERY (recall / precision / median rank vs. a curated 20-gene AML signature). Pinned conda env, project-relative paths, pytest smoke test on the simulated mode.

Platforms & Tools: Nextflow DSL2, Python (numpy / pandas / scipy / plotly / kaleido), recount3, GENCODE v26, conda, pytest

Pipeline source and the full real-data run report live in bioinformatics-public/aml_rnaseq_nf — see docs/REPORT.md for the narrative walkthrough (methods, dataset provenance, embedded volcano, signature recovery table, runtime profile).

Volcano plot from the real-data run: TCGA-LAML AML vs. GTEx whole-blood healthy, with the curated AML signature genes labeled. All 20 markers sit above the FDR line.