Goal: explain why specific genes are selected (mutated) in a given tissue by integrating multi-omics and identifying features that separate gene-mut vs gene-WT samples.

• Download & stage TCGA-like data (Expression, CNV, Mutation).
• Preprocess
  • Expression: log2(TPM + 1)
  • CNV: thresholded/continuous as configured
  • Mutation: per-gene event counts (e.g., MISSENSE / TRUNC)
  • Align sample IDs and take the intersection across views.
• Integrate omics with MOFA (per cohort)
  • Train an initial (raw) model and inspect variance explained.
  • Retrain a final model with selected factors and export variance plots, factors and weights.
• Per-gene LASSO (per cohort × gene)
  • Build labels for a target gene (e.g., TP53_mut vs TP53_wt).
  • Use a no-leak mutation design (drop target-gene mutation columns) and run L1-logistic regression to rank driver features.
• Predefined genes
  • --predef forces genes into MOFA inputs / top-weights (expanded per view; e.g., _MIS, _TRUNC for Mutation).

Tested on HPC (SLURM); also runnable locally for small tests.

nextflow run wf_main.nf -profile slurm  \
  --projects TCGA-BRCA  \
  --genes TP53,PIK3CA  \
  --predef "BRCA1,BRCA2,PIK3CA,KEAP1,EGFR,KRAS,TP53"  \
  --results_base /scratch/yazbecka/mola-fsel-nf/results_BRCA_release44 \
  --data_base    /scratch/yazbecka/mola-fsel-nf/data/raw/  \
  --nfolds 5 --seed 1 --mofa_topn 100  \
  --topn_expr 20 --topn_mut 20 --topn_cnv 20  \
  -with-dag dagRelease44.svg -with-report reportRelease44.html

data/<PROJECT>/ # from --data_base (cache/prepared inputs)
  └─ raw/
      ├─ cnv_thresholded_by_genes.rds
      ├─ cnv_thresholded_by_genes.tsv
      ├─ expression_se.rds
      ├─ mutations_df.rds;
      └─ mutations_maf.rds
results/<PROJECT>/
  ├─ mofa/
  │   ├─ model_raw.rds / model_final.rds / model_final.hdf5
  │   ├─ Model_raw_variance_explained.pdf
  │   ├─ Model_final_variance_explained.pdf
  │   ├─ factors.tsv
  │   └─ weights_list.rds
  ├─ weights/<GENE>/                # top features per view (for LASSO)
  └─ lasso/<GENE>/                  # coefficients, CV metrics, (plots will be added later)

• R libraries: scripts/00_globals.R may set .libPaths("/scratch/.../R/x86_64.../4.x"). Edit to your path or remove and rely on the conda/container env.
• MOFA Python env: scripts/legacy/PrepareMOFA.R uses reticulate::use_condaenv("/scratch/.../CondaMofa/conda_envs/mofa", required=TRUE). Point this to your env or switch to use_basilisk = TRUE.
• MOFA options: In the same file adjust
  model_opts$num_factors,
  model_opts$likelihoods = c(Expression="gaussian", Mutation="poisson", CNV="gaussian"),
  train_opts$drop_factor_threshold,
  data_opts$scale_views.
• Top-N feature selection: Controlled by CLI flags --mofa_topn, --topn_expr, --topn_mut, --topn_cnv (Nextflow) or --topn in legacy R scripts.
• Mutation events: mut_mat_mis_trunc() in scripts/00_globals.R defines how MIS/TRUNC counts are built—tweak here if needed.
• CNV thresholds: Thresholding/continuous handling lives in scripts/00_globals.R (CNV helpers) — edit cutoffs there.
• Reproducibility knobs: --seed and --nfolds (LASSO CV); set explicitly for stable outputs.
• Predefined genes: Use --predef "KRAS,TP53,..." to force inclusion (view-normalized names are auto-expanded, e.g., _MIS/_TRUNC for Mutation).

• Reveals tissue-specific selective pressures behind why a gene is mutated in one tissue but not another.
• Produces interpretable features across Expression / CNV / Mutation (not a black box).
• Uses no-leak modeling around the target gene to keep results valid.
• Reproducible & HPC-ready (Nextflow/SLURM), easy to extend (e.g., add other modules like xCell/PROGENy; in process and coming soon).

• Nextflow ≥ 21.04.3
• R ≥ 4.1: optparse, fs, dplyr, readr, SummarizedExperiment, MOFA2, reticulate, ggplot2, maftools
• Python (for MOFA via reticulate): mofapy2==0.7.0
  (Alternatively, MOFA2 can use use_basilisk = TRUE.)

Scripts resolve paths relative to the repo (workflow.projectDir); no cluster-specific hardcoding is required.

• Labels can be auto-generated per target gene (*_mut vs *_wt).
• LASSO uses a no-leak mutation design (drops target-gene mutation features).
• --predef ensures specific genes are included in the MOFA/weights stage.

• Free to use, copy, change, and share.
• No guarantees. It may have bugs. Use at your own risk.
• I’m not liable for any problems or losses.
• Please keep my name and this license when you share it.
• Please cite MOFA2 and your data source (e.g., TCGA).
• Full legal text: see the LICENSE file (MIT).