config.yaml

# ============================================
# MLOmics Project Configuration
# ============================================
# Central configuration file — ALL hyperparameters, paths, and seeds live here.
# Loaded via: from src.utils import load_config; config = load_config()
# ============================================

project:
  name: "MLOmics"
  version: "1.0"
  seed: 42
  cancer_types: ["GS-BRCA", "GS-COAD"]

paths:
  raw_data: "data/raw"
  preprocessed: "data/preprocessed"
  toy_data: "data/toy"
  models: "models"
  results: "results"
  cv_folds: "data/cv_folds.json"
  sample_map: "data/sample_map.csv"
  experiment_log: "experiment_log.csv"

modalities:
  mrna:
    name: "mRNA"
    feature_count: 5000
    file_pattern: "{cancer}_mRNA_top.csv"
  mirna:
    name: "miRNA"
    feature_count: null  # varies by cancer: BRCA=366, COAD=200 — read dynamically
    file_pattern: "{cancer}_miRNA_top.csv"
  methy:
    name: "Methylation"
    feature_count: 5000
    file_pattern: "{cancer}_Methy_top.csv"
  cnv:
    name: "CNV"
    feature_count: 5000
    file_pattern: "{cancer}_CNV_top.csv"

preprocessing:
  drop_threshold: 0.5 # Drop samples with >50% missing modalities
  scaling: "zscore" # z-score normalization

baselines:
  xgboost:
    n_estimators: 300
    max_depth: 6
    learning_rate: 0.1
    random_state: 42
  random_forest:
    n_estimators: 300
    max_depth: null # None = unlimited depth
    random_state: 42

fusion:
  encoder_hidden: 256
  latent_dims: [32, 64, 128] # Candidates to try
  default_latent_dim: 64
  classifier_hidden: 128
  dropout: 0.3
  batch_size: 32
  epochs: 100
  patience: 10 # Early stopping
  learning_rate: 0.001
  weight_decay: 0.0001

evaluation:
  n_folds: 5
  metrics: ["precision", "recall", "f1", "nmi", "ari"]