normalization — FrequencyNormalizer, DESeq2Normalizer

Two normalization strategies, each required by a different subset of clocks.

File: src/normalization.py
Test: unittests/test_normalization.py

FrequencyNormalizer

Divides each gene's count by the total read count for that sample, producing relative frequencies in the range [0, 1].

freq_ij = count_ij / sum_j(count_ij)

Used by: BayesAge 2.0

Why raw counts: BayesAge 2.0 models gene expression as a Poisson process. Relative frequencies are the natural input; applying DESeq2 first would double-normalize.

Usage

from src.normalization import FrequencyNormalizer

norm = FrequencyNormalizer()
freq_counts = norm.normalize(raw_counts)   # DataFrame, same shape

---

DESeq2Normalizer

Implements median-of-ratios size-factor normalization via pydeseq2.

Default design factors: ["age_days"]

Used by: Elastic Net clock, PCR clock

Usage

from src.normalization import DESeq2Normalizer

norm = DESeq2Normalizer(design_factors=["age_days"])
norm_counts = norm.normalize(raw_counts, metadata)

# Optionally save to disk at the same time
norm_counts = norm.normalize(raw_counts, metadata, save_path=Path("outputs/normalized/Atlas_DESeq2_normalized.csv"))

Caching

The pre-normalized Atlas matrix is cached by src/normalize_reference.py:

python src/normalize_reference.py
# → outputs/normalized/Atlas_DESeq2_normalized.csv

All run scripts load this cached file rather than re-normalizing each time.

---

Normalization choice per clock

Clock	Input	Normalizer
BayesAge 2.0	Raw counts	FrequencyNormalizer
PCR	DESeq2-normalized	DESeq2Normalizer
Elastic Net	DESeq2-normalized	DESeq2Normalizer