The problem
The challenge and the data
Built from the bench: start from a biological question and use a sequence model to answer it with something discrete and reproducible.
The hackathon
Built with Claude: Life Sciences is a research hackathon run with the Gladstone Institutes. More than 6,000 people applied within 24 hours, and 500 were selected to take part. The Research track asks participants to begin from a biological question and use Claude Science and Claude Code to produce a finding, a trained model, or a reproducible analysis. One of the suggested directions, from the Pollard lab, is to train a model that reads DNA to predict regulatory activity, then ask what a single-base change does to it. That is the direction this project takes.
The biological question
Non-coding variants drive a large share of disease risk, yet the effect of a single base change in a regulatory element is hard to read from sequence. A massively parallel reporter assay measures that effect directly. Saturation mutagenesis takes it to the limit and measures the effect of every possible single-base substitution in an element. The question I ask is simple. Can a pretrained DNA sequence model, shown no measured value, redraw that map base by base, well enough to point at a causal variant.
The dataset
The ground truth is the saturation-mutagenesis atlas of Kircher et al., Nature Communications 2019. It measures a log2 activity effect for every single-base substitution across about twenty disease-associated regulatory elements, released as 29 element tracks because some loci were assayed in several cell types.
The elements are the CAGI5 Regulation Saturation community benchmark, so published supervised models set a reference point. The model is used with no parameters fitted to these data at all.
Why this test set can be trusted
The benchmark is exhaustive. Every single-base substitution was measured, not a curated subset, so a model cannot look good just because the answer was already known for the famous variants. There is no easy subset to land on and no popularity bias in what was tested. A strong result here reflects the model, not the choice of test cases.