The question
How does WRN gene effect vary with continuous MSI score in colorectal cancer cell models? Does the association change when related models count once per source donor?
What I built
A browser workbench for five genes in a pinned DepMap cohort. Readers can inspect observations, switch between individual models and donor medians, see missing measurements and download a reproducible analysis. A separate human-tumor view keeps expression data distinct from gene dependency.
How it works
Python checks source files and joins. TypeScript validates the resulting datasets before display, including donor membership, coverage and calculated summaries.
The donor analysis uses complete measurement pairs before aggregation. Otherwise, two incomplete records could accidentally produce a pair that was never measured. Exported dossiers preserve the data and configuration needed to replay the analysis. WRN uncertainty is computed separately in Python.
The chart leads the interface. Selecting a crowded point reveals nearby observations without moving their coordinates. Missingness and the observation unit stay visible beside the result.
Result
Of 96 models, 63 have both WRN and MSI measurements, representing 61 source donors. Spearman correlation changes from −0.603 for individual models to −0.604 for donor medians. The donor estimate's nominal 95% bootstrap interval is −0.762 to −0.384.
Aggregation changes this association little. It does not resolve selection, missingness or confounding, and it is not independent replication or evidence of treatment benefit.
Twenty-three Python tests and interface checks cover joins, validation, export replay and failure states. The release check reproduced three interface artifacts byte-for-byte using verified cached inputs. Independent researcher feedback remains open.
Repository
Public repository pending. The local release package includes the analysis pipeline, interface, setup instructions and separate data notices. Observation dossiers can be checked with npm run replay -- dossier.json.
What I'd do next
Observe whether a researcher can inspect and reproduce an observation without assistance, then use that feedback to refine the interface.