Built with Claude: Life Sciences hackathon
Research track · 2026
A DNA sequence model reads a disease regulatory element, base by base.
I asked AlphaGenome, a pretrained sequence model, to predict the effect of every single-base change in a panel of disease regulatory elements, from DNA alone. It reproduces the wet-lab saturation-mutagenesis maps, and, shown no clinical information, places a canonical cancer driver at the very top of its ranking. As a check I also trained my own model on AWS with the CLI, orchestrated by Claude Code, and the pretrained model holds its own with no training at all.
The result. Blind to every annotation, the model's single highest-impact predicted variant is chr5:1295135 G>A, the pathogenic TERT C250T promoter mutation, a recurrent driver in melanoma and glioblastoma. The neighbouring C228T hotspot appears in the same ranking. Both are recovered from DNA alone.
A trained comparator. To calibrate the zero-shot result, I also trained a supervised gradient-boosted model on the measured effects, run as a managed AWS SageMaker job from the command line. Head to head, the pretrained model matches it across the atlas by the median, pulls ahead on the TERT promoter, and transfers to a held-out locus without ever fitting a measured value, where the trained model collapses.
What is here
Challenge & dataset
The hackathon, the Kircher saturation-mutagenesis atlas, the biological question, and why an exhaustively measured test set makes the result trustworthy.
The methodHow I used Claude
Claude Code orchestrated and built the pipeline and ran a five-reviewer peer review. Claude Science verified the result independently.
The scienceThe report
A full scientific paper: the zero-shot recovery, the 29-element atlas, the pre-registered nomination, controls, and honest limitations.
The toolExplore the maps
An interactive explorer: pick an element, compare predicted and measured effects base by base, and hover to find the clinically significant variants.
The assistantAsk the data
The explorer plus a grounded Claude assistant: ask about any base, variant, element, or disease, and it answers only from the study's measured and predicted effects.
