proDA

Note: This reading note was originally written in 2020 when I first encountered the preprint. I have expanded and updated it in June 2026 with proper citation details (the paper was published in Journal of Proteome Research after the initial bioRxiv version) and verified that the proDA package remains actively maintained on Bioconductor.

Why this one stuck

proDA does offer an option to do testing with missing values. Missing values in label-free data are not always random. You miss the low-intensity stuff systematically. Drop in a median or some small constant and you've quietly invented data points. Your matrix looks complete. Your downstream code runs without complaints. Your p-values are confident. They're also biased.

proDA refuses the shortcut. Instead of filling holes, it fits a sigmoidal dropout curve per sample. If a value is missing, the model says "this protein was probably below detection, here's the probability distribution for where it might have been." Then it carries that uncertainty through to the differential abundance test. The error bars grow. The significant hits shrink. The ones that survive are real.

proDA's probabilistic dropout model showing how missing values are modeled as censored observations below detection threshold

Figure

What you actually get

The practical difference is honest uncertainty. If you have a protein with 5 missing values and 1 weak observation, traditional imputation fills those 5 holes with guesses and pretends you measured 6 data points. proDA says "you measured one marginal signal, the rest fell below detection, here's how uncertain your mean estimate actually is."

Your significantly differentially abundant list gets shorter. That's the point. The proteins you lose were riding on invented data. The ones you keep passed a test that acknowledges how little you actually saw.

When it matters

This matters most when you're running differential abundance tests and your FDR depends on accurate variance estimates. If you're just making a heatmap for a figure or doing quick exploratory clustering, imputation is fine. The problem is when you impute for visualization and then forget you did it before running statistics.