Consistency with the PEAgent browser

The contribution scores served by the PEAgent genome browser were produced by the analysis pipeline that accompanies the manuscript, not by this package. If peagent_tool is to be useful for designing new sequences, the numbers it computes from sequence must be the same numbers.

They are.

What was compared

Five crownRoot cell-type-specific ACRs of the maize atlas were recomputed from sequence alone with

scorer = ContributionScorer("maize")
scorer.compute_many({acr_id: sequence},
                    targets=scorer.celltypes_in_tissue("crownRoot"),
                    normalize="tissue")

and each of the resulting 23 cell-type tracks was compared, base by base, with the stored pipeline array that the corresponding browser BigWig was built from. The raw (normalize="none") global track was checked the same way against the pipeline's constitutive-global output for three constitutive ACRs.

Result

Normalization	Tracks compared	Median Pearson r	Minimum r	Max absolute difference	Max difference relative to track amplitude
"tissue"	115	0.999995	0.999975	3.2 × 10⁻⁴	0.42 %
"none"	3	0.999999	0.999999	2.0 × 10⁻⁴	0.09 %

The residual differences are float32 accumulation noise — the two computations ran on different hardware, in different batch orders, years apart. There is no systematic offset.

Why the conventions match

The pipeline computes a cell-type track as:

1. mutate every position of the 1,344 bp window to all four bases and take the change in the shared 64-d bottleneck embedding;
2. project onto the mean final-layer weight vector of the cell group;
3. centre across A/C/G/T at each position;
4. multiply by the reference one-hot, so only the reference base carries a score;
5. subtract the mean of step 4 over all valid tissue_celltype groups of the same tissue.

Steps 1–4 are normalize="none"; step 5 is what normalize="tissue" adds. The set of groups averaged in step 5 is taken from the cell-group metadata, which is the same set the pipeline used — for crownRoot, all 23 cell types.

AttributionResult.tissue_reference lists exactly which groups went into the subtraction, so the reference set is never implicit:

result = scorer.compute(sequence, target="crownRoot_cortex", normalize="tissue")
result.tissue_reference
# ('crownRoot_LRC', 'crownRoot_QC', 'crownRoot_atrichoblast', ...)

Reproducing this check

The comparison script is script/56-peagent-tool-docs/01_validate_tissue_norm.py in the analysis repository. It needs the stored pipeline arrays, which are not distributed with the package; the tutorial notebooks are the runnable equivalent.