From 6e89bd8902dc1b5dabb6cb34cf0f8ee0c28f815e Mon Sep 17 00:00:00 2001
From: Brian Clarke <9725212+bfclarke@users.noreply.github.com>
Date: Mon, 11 Nov 2024 10:57:25 +0100
Subject: [PATCH] drop scores lower than median and scale to fill [0, 1]

---
 deeprvat/deeprvat/associate.py | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/deeprvat/deeprvat/associate.py b/deeprvat/deeprvat/associate.py
index e9b073db..a147a055 100644
--- a/deeprvat/deeprvat/associate.py
+++ b/deeprvat/deeprvat/associate.py
@@ -442,6 +442,7 @@ def make_regenie_input_(
         pseudovar_pos = (this_gene_pos.End - this_gene_pos.Start).to_numpy().astype(int)
         ensgids = this_gene_pos.index.to_numpy()
 
+        median_score = np.median(burdens[:, 0]) # TODO: This is a hack!!!!
         logger.info(f"Writing pseudovariants to {bgen}")
         with BgenWriter(
             bgen,
@@ -468,8 +469,10 @@ def make_regenie_input_(
                 # 1. Warn if burdens are censored to remain > 0
                 # 2. Offset/scale more intelligently to fill out [0, 1] better
                 # 3. (maybe) Allow for setting offset/scale as parameter
-                offset = 0.251
+                offset = median_score
                 this_burdens = np.maximum(this_burdens - offset, 0)
+                max_burden = np.max(this_burdens)
+                this_burdens = this_burdens / max_burden
 
                 # REGENIE assumes by default genotypes are stored alt-first
                 genotypes = np.stack(