Saturation-mutagenesis explorer

Reading a disease regulatory element, base by base

A DNA sequence model predicts the effect of every single-base substitution in a regulatory element, from sequence alone. Compare the zero-shot AlphaGenome map, and a supervised model trained on the assay, against the wet-lab measurement, and find the bases that carry a known clinical variant. Measurements are from the Kircher et al. 2019 reporter assay.

—
Spearman, predicted vs measured

Measured  effect on activity (Kircher MPRA)

ACGT

Predicted  DNASE from DNA alone, zero-shot AlphaGenome

ACGT

Trained model  supervised on this element,

ACGT
position in element
silencing activating clinically significant variant hover the maps; click to pin a base

Clinically significant variants in this element

    Selected base

    Hover or click a cell in the maps to inspect a single-base substitution.

    How to read it. Each column is a position in the element; each row is an alternate base (A, C, G, T). Colour is the effect on regulatory activity, on the same zero-centred scale in all three maps: vermillion increases it, blue reduces it. The measured map is the assay effect. The predicted map is AlphaGenome, zero-shot. The trained-model map is a supervised baseline scored on positions held out within the same element, the reference the report compares against; it collapses on unseen loci, so read it as a within-element result only. A ring marks a base that Ensembl annotates as clinically significant.
    Data. Measured effects: Kircher et al., Nature Communications 2019 (GEO GSE126550). Predictions: AlphaGenome, run zero-shot; supervised baseline trained on AWS SageMaker. Clinical annotation: Ensembl variation (GRCh38). Built with Claude Code for Built with Claude: Life Sciences, 2026.