attribute

Train a single MLP and compute per-SNP Integrated Gradients attribution, writing a PLINK-.glm-style TSV. See Concepts and methodology for what the model and attribution step are doing, and Output formats for the output column layout.

posthoc attribute

Usage

posthoc attribute [OPTIONS]

Options

--pfile <pfile>: Required Prefix of PLINK2 .pgen/.pvar/.psam fileset.

--pheno <pheno>: Required Phenotype file.

--pheno-name <pheno_name>: Phenotype column name (default: first value column).

--covar <covar>: Covariate file.

--logistic: Binary phenotype (BCE loss).

--linear: Continuous phenotype (MSE loss).

--model <model_name>

Options:: mlp

--attribution <attribution_name>

Options:: integrated_gradients

--maf <min_maf>: Minimum minor allele frequency.

--geno <max_missing>: Max per-variant missingness.

--indep-pairwise <indep_pairwise>: LD pruning: WINDOW_SIZE STEP R2_THRESHOLD (requires plink2 on PATH).

--hidden-dims <hidden_dims>: Comma-separated MLP hidden layer sizes.

--dropout <dropout>: Dropout parameter of neural network

--epochs <max_epochs>

--patience <patience>

--lr <lr>

--val-fraction <val_fraction>

--n-steps <n_steps>: Integrated Gradients interpolation steps.

--ig-baseline <ig_baseline>

Baseline reference input for Integrated Gradients.

Options:: zero | mean

--seed <seed>

--device <device>

--out <out_path>: Required Output .glm path.

Behavior notes

Exactly one of --logistic/--linear must be given.
--model currently only accepts mlp; --attribution currently only accepts integrated_gradients. Both are exposed as explicit choices to keep the CLI stable if/when additional models or attribution methods are added.
--hidden-dims is a comma-separated list of integers, e.g. 256,64 for a two-hidden-layer MLP with 256 then 64 units.
Attribution is computed only on the held-out validation split used during training (train_result.val_idx), not the full dataset — so N in the output reflects the validation split size, and --val-fraction indirectly controls how many samples the attribution statistics are based on.
--ig-baseline mean uses the per-feature mean of the validation design matrix as the Integrated Gradients reference input, instead of an all-zeros baseline. This is generally more appropriate for genotype data, where 0 is a meaningful allele count rather than a neutral reference.

Example

posthoc attribute \
  --pfile datasets/data/processed/chr22_subset \
  --pheno outputs/simulated.pheno \
  --pheno-name PHENO1 \
  --logistic \
  --hidden-dims 256,64 \
  --epochs 200 \
  --patience 15 \
  --n-steps 50 \
  --ig-baseline mean \
  --out outputs/ig_results.glm