Bio-Babel pancreas case study

This is the compact reproducibility package for the GSE248369 case study. The workflow was:

1. manual preprocessing, quality control and cell annotation;
2. one biological-analysis prompt to Claude Code;
3. package discovery and API guidance through 19 read-only Bio-Babel MCP calls;
4. agent-written, executed and debugged downstream analysis; and
5. human review, completion of the genotype screen and biological interpretation.

Bio-Babel MCP supplied operating knowledge for Monocle2-python, DDRTree-python, tradeSeq-python and ggplot2-python. It did not execute the analysis itself.

Files

• preprocess.ipynb: manual preprocessing and annotation before the agent run.
• agent.ipynb: notebook generated and executed by Claude from the single prompt.
• followup.ipynb: human-reviewed follow-up that starts from frozen agent outputs; it does not repeat preprocessing, DDRTree reconstruction, BEAM testing or tradeSeq fitting.
• prompt.md, session.jsonl, session_meta.json, report.md: prompt, sanitized observable conversation, run metadata and Claude's preliminary report. The report is retained as an agent artifact, not as the final biological interpretation.
• tables/: full statistical results needed to audit BEAM, tradeSeq, the 1,116-gene genotype screen and the retained manuscript panels.
• code/: analysis helpers, the completed 1,116-gene screen and the supplementary-table builder.
• supplementary_tables.xlsx: publication-facing sample/QC, BEAM, genotype-screen and plotted-gene tables; it is regenerated by code/build_supplementary_tables.py (requires pandas and openpyxl).
• samples.tsv, cells.tsv.gz, provenance.json, software.tsv: sample and annotation metadata, input checksums, seeds and software commits.

followup.ipynb is distributed without cell outputs. To run it, place the deposited trajectory checkpoint at data/traj_ab.h5ad or set BIOBABEL_TRAJECTORY_INPUT to its location.

The raw Claude JSONL, biological matrices, H5AD checkpoints, environments, logs, temporary figures and superseded outputs are not included. Machine paths in the public notebook/transcript copies are redacted. The exact processed AnnData and trajectory checkpoint are identified by SHA-256 in provenance.json; they should be deposited separately rather than committed to Git.

Embryos are the biological replicates. Cell-level genotype-aware tradeSeq tests are candidate-screening statistics, not embryo-replicated inference. Harmony was used only during manual annotation and not as DDRTree input.