Analyze What Papers Actually Reported

Step A — Data ingestion & harmonization (inputs & outputs)

• Harmonize gene IDs to Ensembl/UniProt, liftover coordinates if needed, reformat somatic calls to MAF/VCF standard for downstream tools.
• Record provenance/MD5 checksums; store data in parquet/AnnData for scalable IO.

Step B — QC, sample & variant filtering

• Sample QC: tumor purity (ABSOLUTE/ASCAT), read depth, contamination, RNA integrity — drop samples failing thresholds; document thresholds for reproducibility.
• Variant QC: filter strand-bias, mapping artifacts, panel-derived hotspots; use matched-normal when available to remove germline/mosaic calls.
• For FFPE or low-input datasets, apply HRM-informed sensitivity expectations for hotspot detection (note HRM sensitivity ranges) (HRM assay sensitivity for FFPE hotspots).

Step C — Variant calling, CNAs, structural calls

• Use ensemble somatic-calling (e.g. 2+ callers) for SNV/indel; annotate with VEP/Oncotator; call CNAs with ASCAT/Sequenza; call SVs with Manta/GRIDSS. Keep raw BAMs for reanalysis.
• Produce per-sample mutation catalogs (trinucleotide contexts) for signature inference.

Step D — Mutational-signature & topography analysis

Step E — Driver prioritization: molecular & age-aware evidence

1. Molecular mechanism scoring: use interpretable models (MutaPheno-style) that integrate structure, ΔΔG, PPI/PLI sites, domain context, and stability/binding disruption for missense variants; this improves mechanistic interpretability and generalization (MutaPheno molecular driver predictor).
2. Age-aware causation: compute per-gene carcinogenic effect estimates using the age-distribution framework to separate selection in normal tissue from true carcinogenic drivers; prioritize genes with young-age enrichment and high w (carcinogenic effect) as causal candidates (Age-based carcinogenic effect framework).
3. Cross-validate driver lists with IntOGen and tumor-type recurrence; flag discordant genes (selected in normal tissue but low carcinogenicity) for cautious interpretation (IntOGen driver reference).

Step F — Multi-omics integration & interpretable ML

• Feature construction: combine molecular driver scores, mutational signatures, CNAs, expression (tumor vs normal), immune infiltration scores, and clinical covariates (age, stage).
• Modeling approach: (i) transparent models for biomarker discovery (XGBoost + SHAP or L1-regularized GLMs) with nested CV; (ii) built-in interpretability (SHAP/feature importance) and stability selection for reproducibility (XGBoost biomarker framework).
• Benchmarking: use MLOmics standardized splits for training/holdout and IntOGen/ICGA for external validation (MLOmics benchmarking resource).

Step G — Experimental/clinical validation prioritization

• Prioritize variants/genes with convergent evidence: molecular disruption (MutaPheno-like high score), young-age enrichment (high w), recurrence across cohorts, and biochemical plausibility for follow-up assays (in vitro functional assays, patient-derived organoids, or targeted sequencing panels).
• Use cancerSimCraft synthetic datasets to benchmark method sensitivity/specificity under controlled clonal architectures before wet-lab validation (cancerSimCraft for method benchmarking).

Step H — Reproducibility, code & data release

• Package pipelines using nf-core/Snakemake containers; publish exact environment (Conda/Docker) and parameter files; release driver-scoring code and data splits on GitHub with DOI (Zenodo).
• Report limitations explicitly: cohort biases, missing matched normals for many TCGA/COSMIC entries, variation in sequencing depth and capture, and the inability of structure-predicted features to capture all conformations (MutaPheno caveats on structure reliance).