Bioanalytical data science · Python

ELISA Analysis Pipeline

A tested, end-to-end pipeline that turns raw 96-well plate-reader data from sandwich ELISAs into quality-controlled analyte concentrations, statistical comparisons between treatment groups, and an audit-ready report.

▶ Launch the live app View source on GitHub Example report

The app runs entirely in your browser (Python compiled to WebAssembly). The first load takes about 30 seconds while the scientific Python stack downloads.

135
automated tests · 97% coverage
< 8%
median error recovering simulated ground truth
2
public real-world datasets re-analysed
0
servers needed for the live app
The live app showing plate QC results for the demo study

What it does

Each stage is a documented module with the scientific reasoning in its docstrings.

Parse & validate platesBlank subtraction Weighted 4PL / 5PL fit (AICc)Back-calculation Outliers · %CV · recoveryLOD / LLOQ / ULOQ Controls & plate pass/failDilution linearity Inter-plate normalizationWelch / Mann-Whitney / ANOVA · BH-FDR HTML report

Outputs

Standard curve with quantifiable range and residuals
Weighted 4PL fit, validated quantification range, LOD, and weighted residuals.
IL-6 by treatment group
Group comparison with Holm-adjusted significance. Hollow points are censored values.
Bridge controls before and after normalization
Bridge controls before and after inter-plate normalization (IL-6 control CV ~25% → ~5%).
Plate map heatmap
Raw-OD plate map. Edge effects and pipetting gradients show up here.

Real data, honest conclusions

Notebook · biology Serum IL-6 after sleep fragmentation A re-analysis of a published mouse study (Dryad, CC0).
  • The original straight-line calibrations recover standards at 28–142%. The weighted 4PL recovers 94–102%.
  • 100 / 110 samples fall below the LLOQ.
  • A labelled sensitivity analysis still reproduces the IL-6 rise at 6, 12 and 24 h.
 Notebook · QC SoftMax Pro plates across two kit lots Five real plates (ELISAtools, MIT).
  • Rule-based outlier QC matches the analyst on 4 / 4 wells, and exposes 13 discretionary masks.
  • A shared control reveals a ~30% kit-lot shift, which normalization corrects.

Reports

Self-contained HTML, regenerated from the code on every deploy.

SimulatedDemo study Two cytokines, three treatment groups, four plates with calibrator bias, edge effects and outlier wells. Real dataMouse IL-6 Four plates and 110 samples, with plate acceptance, censoring and flags. Real dataSoftMax plates Triplicate QC and kit-lot normalization.

Built with

Python 3.11–3.13NumPypandas SciPy (curve_fit)statsmodelsmatplotlib Streamlitstlite / Pyodide (WebAssembly) pytestruffGitHub ActionsGitHub Pages