Natural-language transcriptomics: from GEO data to a reproducible report.

1. Define the comparison before searching for data

The demonstration starts with a biological request: examine how LPS stimulation changes gene expression in human macrophages and produce an interpretable report.

Turn that request into explicit criteria—species, cell model, treatment, control, replication, data type, and desired outputs—before choosing a public study.

2. Use GEO discovery as a screening step, not a download button

Wisp Science searches candidate GEO series, then selects GSE248578 for the worked example. The visible report describes human GM-CSF-derived macrophages with matched control and LPS-stimulated samples from three donors.

Before analysis, verify the sample table, treatment duration, raw or processed file type, and whether the comparison is truly supported by the study design.

• Confirm accession, organism, cell model, and group labels on the GEO record.
• Check that biological replicates are not being treated as technical replicates.
• Read the source publication when a design detail is ambiguous.

3. Build an isolated environment when the machine is not analysis-ready

The session discovers that the required R environment is not already available. Instead of changing the whole computer, it bootstraps a project-scoped R/Bioconductor environment with pixi, conda-forge, and BiocManager, while preserving the lockfile, installation scripts, and logs.

The report shown in the walkthrough names DESeq2 and clusterProfiler alongside the R environment. Exact versions belong in the methods record because they affect reproducibility.

4. Put quality control before interpretation

A reliable workflow checks sample relationships and count behavior before discussing pathways. The generated artifact set includes quality-control and differential-expression views rather than only a final gene list.

Use PCA or related sample-level diagnostics to inspect grouping and outliers. If the design is paired by donor, preserve that structure in the statistical model instead of treating all samples as independent.

5. Connect differential expression to enrichment without overstating it

The workflow proceeds from differential-expression results to functional interpretation and generates multiple plots and tables in the project. Enrichment summarizes patterns in a ranked or selected gene set; it does not prove a pathway mechanism by itself.

Review thresholds, identifier mapping, background universe, direction of change, and whether the reported terms fit both the data and established LPS biology. The recorded report itself notes that donor effects were not modeled even though samples come from three donors, so the design matrix requires expert correction before scientific use.

6. Treat failure recovery as part of reproducibility

The recording does not hide the rough edges: package metadata is unavailable from one mirror, DESeq2 cannot be installed through the first Windows conda route, and a GSEA curve step fails. The workflow changes sources, uses BiocManager, and runs a supplementary script.

Keep the failed commands, environment changes, and recovery scripts beside the successful output. A clean final figure without that history is less reproducible than an imperfect run with a complete record.

7. Review the report as a bundle of evidence, code, and files

The final report remains open beside the session while generated code, environment files, tables, logs, and figures are visible in the project directory. The recorded session shows 11 figures spanning quality control, differential expression, enrichment, and GSEA; treat those as outputs of this run, not independently verified biological findings.

Before using the result, inspect every plot, check that each table can be regenerated, and compare the methods paragraph with the actual environment and code. The report also notes a single 16-hour time point and an unmodeled donor effect. The agent accelerates assembly; statistical and biological validity remain the researcher’s responsibility.