#!/usr/bin/env bash
# DPYD baseline classifier — full pipeline (run EXPLICITLY; not part of scaffolding).
# No Vertex Workbench. Runs locally or in a short-lived Cloud Run job.
set -euo pipefail

BUCKET="${BUCKET:-gs://anukriti-ml-artifacts/dpyd-classifier}"
cd "$(dirname "$0")"

echo "[1/5] (re)pull gnomAD SAS/global AF for ALL ClinVar DPYD variants (full, uncapped)"
python -m src.fetch_gnomad --clinvar data/clinvar_dpyd.tsv --out data/gnomad_dpyd_sas.csv --cap 0

echo "[2/5] feature engineering -> training_data.csv + inference_set.csv"
python -m src.features

echo "[3/5] train RF/XGB/LightGBM + 5-fold CV (carries small-N caveat; use --force)"
python -m src.train --train data/training_data.csv --outdir . --n-splits 5 --force --upload "$BUCKET"

echo "[4/5] upload training data"
gcloud storage cp data/training_data.csv "$BUCKET/training_data.csv"

echo "[5/5] inference on candidate list (provide a REAL Scaria variant CSV)"
if [[ -f data/scaria_variants.csv ]]; then
  python -m src.infer --variants data/scaria_variants.csv --models models \
    --out results/scaria_variant_rankings.csv
  gcloud storage cp results/scaria_variant_rankings.csv "$BUCKET/results/scaria_variant_rankings.csv"
else
  echo "  SKIP: data/scaria_variants.csv not present (Scaria et al. list unresolved)."
fi
echo "done."