config.yaml

# EPInformer batch pipeline configuration
samples_table: samples.tsv

# ---- Shared reference files ----
reference:
  fasta: ./data/reference/hg38/hg38.fa
  gene_bed: ./data/reference/hg38/XpressoGeneBounds.hg38.bed
  chrom_sizes: ./data/reference/hg38/GRCh38_EBV.chrom.sizes.tsv
  expression_csv: ./data/GM12878_K562_18377_gene_expr_fromXpresso.csv
  blacklist: null
  # Split ENCODE/rE2G average Hi-C (ENCFF134PUN, 5 kb). This is selected
  # automatically for a sample with no usable cell-specific .hic file.
  average_hic_dir: ./data/reference/abc_avg_hic/by_chrom

# ---- Element-gene link parameters (Stage 1: BAM -> predictions) ----
# Activity = sqrt(mean(H3K27ac rep RPMs) * mean(DNase rep RPMs)); replicates are listed
# per-assay in samples.tsv (accessibility_bams / h3k27ac_bams, mean-pooled). The SAME
# replicate BAMs feed BOTH the encoder pre-training activity target AND the ABC neighborhoods
# activity feature, so pre-training and EPInformer training quantify activity from identical
# BAMs. GM12878 uses DNase ENCFF729UYK+ENCFF020WZB and H3K27ac ENCFF269GKF+ENCFF201OHW
# (reaches 0.617). K562 SHIPS single-rep by design (accessibility_bams/h3k27ac_bams left
# empty): its 3 H3K27ac reps differ hugely in depth (rep3 ENCFF232RQF 70.7M vs 10.7M/4.27M),
# so equal-weight mean-pool DILUTED the deep rep and dropped the encoder 0.745 -> 0.729.
# GM12878 pools 2 reps/assay because its reps are similar depth and pooling reaches 0.617.
# (Candidate MACS2 peak-calling still uses the single accessibility_bam; only read-counting
# for activity is pooled — the candidate region SET only defines where activity is measured.)
abc_params:
  n_top_peaks: 150000
  peak_extend: 250
  max_distance: 2500000
  gamma: 0.87
  tss_slop: 500
  hic_resolution: 5000
  neg_fraction: 0.05
  max_encoder_peaks: 100000
  include_self_promoter: false
  include_promoter_region: false
  n_threads: 12  # keep in sync with slurm/run_pipeline_cpu.slurm #SBATCH --cpus-per-task
  # Hi-C processing (Broad ABC defaults)
  hic_gamma: 1.024238616787792
  hic_scale: 5.9594510043736655
  hic_gamma_reference: 0.87
  hic_pseudocount_distance: 5000
  scale_hic_using_powerlaw: true
  tss_hic_contribution: 100.0

# ---- Encoding parameters (Stage 2: predictions -> HDF5) ----
preprocessing_params:
  min_distance: 0
  max_distance: 100000
  n_enhancer: 60
  max_seq_len: 2000
  include_self_promoter: true
  tss_column: TSS_xpresso
  no_bigwig: true  # sequence-only; no BigWig signal tracks in HDF5

# ---- Output directory layout ----
output:
  base_dir: ./batch_output
  # Per-sample dirs: {base_dir}/{cell_type}/links/ and {base_dir}/{cell_type}/encoding/