Downstream analysis: recommended defaults, clearer labels and decision plots

[t0mdavid-m]

Requested by Tom · project thread

Before: the sidebar had "Results" and "Differential Protein Analysis", and it was unclear which table a plot showed. The heatmaps plotted the raw abundance table while PCA used the processed one. Filtering, Imputation, Normalization and Statistical Inference opened on arbitrary first options. With those options, "log2FC" was a difference of raw intensities, because the statistics function subtracts group means and assumes log2 input. Choices reset on every visit, and re-applying a step left the later steps' tables in place.

After: the sidebar reads Workflow Results, Downstream Analysis (1. Filtering, 2. Imputation, 3. Normalization, 4. Statistics) and Downstream Plots. Every plot page states which step's output it shows, and the heatmaps use the same processed table as PCA. Each step opens on a recommended setting and has a "Help me choose" plot. Every plot, including Volcano, PCA and the heatmaps, has a one- to two-sentence explanation. Choices persist per workspace, and re-applying a step clears the steps after it.

Step	Default	Decision plot
1. Filtering	Low Repeatability, ≤50% missing per group	Proteins kept across the threshold range
2. Imputation	MNAR, per-protein minimum	Mean intensity of proteins with vs without gaps; where imputed values land
3. Normalization	log2 + median, no scaling	Per-sample box plots before/after (live preview)
4. Statistics	limma-like + BH; warns if input is not log2	p-value histogram

Why per-protein minimum imputation. I ran a simulation with 500 proteins, 50 true 4-fold changes and 3 vs 3 samples. With 10% of values dropping out at random, the global minimum found 0 of 50 changes, because one imputed dropout inflates a protein's variance. The per-protein minimum found 36 of 50 with no false positives. When values were missing below a detection limit, it found 49 with 1 false positive.

How. Widgets are keyed on new postproc-* entries in default-parameters.json. postprocessing_param() seeds keys for older workspaces. clear_downstream_steps() drops stale step outputs. get_active_table() and show_pipeline_banner() give every plot page the same source and label. The plots live in src/common/postprocessing_plots.py, use Plotly and go through show_fig; OpenMS-Insight has no histogram or box plot component. Volcano, PCA and the heatmaps pass regenerate_cache=True, the same lines as #36. After #36, the only overlap is the sidebar entry in app.py.

Protein inference is unchanged: ProteomicsLFQ aggregation, picked protein FDR and unique peptides, and TMT ProteinInference with PEP/best/picked FDR. Both already match quantms' defaults.

Tested: I ran all four steps and the four plot pages in sequence with AppTest on a synthetic LFQ workspace, including an older params.json without the new keys, and checked each decision plot rendered from synthetic data. pytest tests passes (88 passed, 2 skipped). pylint --errors-only is clean apart from the pre-existing pyopenms import notices.

🤖 Generated with Claude Code

https://claude.ai/code/session_015XpSPmvsUELJs4GbbDXkrj