Out of five screenings that have already been analysed, the agent has confidently identified a single gene in more cells than the researchers did, and turns every error into a proposed fix for the open-source project.
Grouped optical screening studies link genetic disruptions to changes in cell shape or behaviour, but converting microscopy images into actionable results requires several days of analysis. Even with an established pipeline such as Brieflow, an end-to-end open-source tool, researchers must select parameters, inspect outputs and troubleshoot failures for each experiment.
A ‘wrapper’ around an existing pipeline
A team has published Brieflow-auto on arXiv, a plugin for Claude Code that enables a language model specialised in code to drive Brieflow. The researcher provides the experimental data and approves the run; the agent adjusts the parameters in an interactive notebook, whilst the harness verifies the configuration, monitors execution, recovers from common failures and checks the quality of the results. When it detects a pipeline fault, the agent opens a ticket on the project, proposing a backwards-compatible fix.
Out of five screenings that had previously been analysed manually, the automated runs confidently assigned a single gene to more cells than the original analyses. In one screen for which the pipeline had not been designed, every failure resulted in a reviewed fix. The authors see this as a general model: wrapping an existing scientific tool within a harness to speed up its use and enhance its reliability.
Read the full article on the original source
Go to the source →