bayesage2 — `BayesAge2Clock`¶

BayesAge 2.0 implementation. Predicts transcriptomic age via Poisson log-likelihood maximization over a LOWESS-smoothed reference.

File: src/bayesage2.py
Test: unittests/test_bayesage2.py

Algorithm¶

Spearman correlation — rank-correlate every gene's frequency-normalized expression with age_days across Atlas training samples. Vectorized: all ~25K genes at once via rank-matrix operations.
LOWESS smoothing — fit a smoothed age–expression curve (LOWESS, frac=0.7) for the top lowess_top_n genes ranked by |Spearman r|. Produces a reference matrix of shape (genes × age_grid).
Prediction — for a query sample, select the top M genes by |Spearman r| and find the age that maximizes the Poisson log-PMF between the query's frequency-normalized counts and the reference curves.

The age grid spans 47–163 days at 1-day resolution (killifish-specific).

Constructor parameters¶

Parameter	Default	Description
`lowess_frac`	`0.7`	Fraction of points used per local regression
`lowess_top_n`	`500`	Gene pool size for LOWESS fitting

LOWESS fraction

The original BayesAge 2.0 notebook uses frac=0.7. An earlier version of this repo used frac=0.3, which was the single largest driver of differences from the original. Fixed in this implementation.

Methods¶

`build_reference(raw_counts, metadata)`¶

Builds the age-expression reference matrix on Atlas training data.

Stores internally: Spearman correlations, LOWESS-smoothed reference, age grid.

`predict(raw_counts, n_genes, reference=None)`¶

Predicts tAge for all samples in raw_counts using the top n_genes (M) genes. Optionally accepts a pre-loaded reference (from load_reference()).

Returns a pd.Series mapping sample ID → predicted tAge (days).

`loso_cv(raw_counts, metadata, m_values)`¶

Leave-one-sample-out cross-validation over multiple M values. Returns a DataFrame with columns [sample_id, age_days, tAge_M{m}, ...].

`load_reference(path)`¶

Loads a pre-saved reference .tsv instead of rebuilding from scratch.

`save_reference(path)`¶

Saves the reference matrix to a TSV file.

Example¶

from src.data_loader import DataLoader
from src.preprocessing import Preprocessor
from src.normalization import FrequencyNormalizer
from src.bayesage2 import BayesAge2Clock

loader = DataLoader()
counts, meta = loader.load_atlas()

pp = Preprocessor()
tissue_counts, tissue_meta = pp.stratify(counts, meta, tissue="Liver")
clean_counts, clean_meta, _ = pp.detect_outliers(tissue_counts)

freq = FrequencyNormalizer().normalize(clean_counts)

clock = BayesAge2Clock()
clock.build_reference(freq, clean_meta)
predictions = clock.predict(freq, n_genes=100)

Performance notes¶

Spearman correlation: ~0.85 s for 25K genes (vectorized)
LOWESS fitting: ~2–5 s for 500 genes
Prediction: ~0.1 s per sample

bayesage2 — BayesAge2Clock¶