Use geometric mean to calculate peptide score

CHANGELOG.md

@@ -9,6 +9,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 ### Changed
 
 - Instead of having to specify `train_from_scratch` in the config file, training will proceed from an existing model weights file if this is given as an argument to `casanovo train`.
+- Use geometric mean instead of arithmetic mean to calculate peptide score.
 
 ## [4.0.0] - 2023-12-22
 

casanovo/denovo/model.py

@@ -6,6 +6,7 @@
 
 import depthcharge.masses
 import einops
+import scipy.stats.mstats
 import torch
 import numpy as np
 import lightning.pytorch as pl
@@ -1032,8 +1033,11 @@ def _aa_pep_score(
     Calculate amino acid and peptide-level confidence score from the raw amino
     acid scores.
 
-    The peptide score is the mean of the raw amino acid scores. The amino acid
-    scores are the mean of the raw amino acid scores and the peptide score.
+    The peptide score is the geometric mean of the raw amino acid scores.
+    Peptide scores that don't fit the precursor m/z tolerance are penalized by
+    subtracting 1.
+    The amino acid scores are the arithmetic mean of the raw amino acid scores
+    and the peptide score.
 
     Parameters
     ----------
@@ -1049,7 +1053,7 @@ def _aa_pep_score(
     peptide_score : float
         The peptide score.
     """
-    peptide_score = np.mean(aa_scores)
+    peptide_score = scipy.stats.mstats.gmean(aa_scores)
     aa_scores = (aa_scores + peptide_score) / 2
     if not fits_precursor_mz:
         peptide_score -= 1

pyproject.toml

@@ -32,6 +32,7 @@ dependencies = [
     "requests",
     "rich-click>=1.6.1",
     "scikit-learn",
+    "scipy",
     "spectrum_utils",
     "tensorboard",
     "torch>=2.1",