JuliaCon 2026

Mass spectrometry-based interactomics experiments produce lists of hundreds to thousands of candidate protein-protein interactions, but a large fraction of these are non-specific contaminants or experimental artifacts. Existing tools typically apply a single statistical test and threshold -- a t-test on fold changes, or a simple scoring scheme -- discarding the rich multi-dimensional structure of the data. A protein might be modestly enriched but detected with striking consistency, or show a clear dose-response trend that a fold-change filter would miss entirely.

BayesInteractomics.jl takes a different approach. For each candidate interaction, the package computes Bayes factors from three independent statistical models that each capture a distinct aspect of the data:

1. A Beta-Bernoulli model that evaluates whether a protein is detected more consistently in bait samples than in controls.
2. A hierarchical Bayesian model (via RxInfer.jl) that estimates quantitative enrichment (log2 fold change) while sharing information across experimental protocols.
3. A Bayesian linear regression that tests for dose-response correlation between prey and bait abundance.

Rather than multiplying these Bayes factors under a naive independence assumption, the package uses Copulas.jl to model the dependency structure between evidence sources. This matters because enrichment and detection evidence are positively correlated under both hypotheses -- a genuinely enriched protein is also more likely to be consistently detected -- so treating them as independent inflates the combined evidence and drives up false discovery rates. An EM algorithm fits a two-component copula mixture (H0 vs. H1) that accounts for this correlation, selecting among Clayton, Frank, Gumbel, Gaussian, and Joe copula families via model comparison. The result is a joint Bayes factor and calibrated posterior probability for every protein that properly reflects the shared information content across evidence types.

This talk will cover the statistical design choices, how the Julia ecosystem made them practical, and lessons learned building a research-grade Bayesian analysis package. Specific topics include:

• How RxInfer.jl's reactive message-passing enables fast variational inference for the hierarchical enrichment model, and why this matters when you need to fit the same model thousands of times.
• Using Copulas.jl for method-of-moments fitting of Archimedean copulas, and how multiple dispatch made it straightforward to support six copula families through a single interface.
• Automated data curation via the STRING database API, including protein group resolution and synonym mapping
• Package extensions for optional functionality: a network analysis extension (Graphs.jl, GraphPlot.jl) for building and visualizing interaction networks, and an experimental structural docking extension (BioStructures.jl) -- both loaded only when the user imports the relevant packages.
• Self-contained HTML reports (client-side Plotly.js and DataTables.js, no additional Julia dependencies) so that collaborators who do not have Julia installed can explore results interactively in a browser.

The talk is aimed at Julia users interested in Bayesian statistics, scientific computing, or computational biology. No proteomics background is assumed and I will demonstrate a complete analysis from raw data to interactive report on a real dataset.