Detecting antibody reactivities in Phage ImmunoPrecipitation Sequencing data

PhIP-Seq (Phage ImmunoPrecipitation Sequencing) lets researchers measure which of hundreds of thousands of candidate peptides a person’s antibodies bind to — producing a read-count matrix that, on the surface, looks much like RNA-Seq output. Because of that resemblance, many labs have repurposed RNA-Seq analysis tools (edgeR, DESeq2, voom) for PhIP-Seq. But the underlying biology and experimental design are different: PhIP-Seq compares one serum sample against a small panel of antibody-free “mock IP” (beads-only) wells, and low-count peptides may carry real biological signal rather than being filterable noise.

The authors first stress-test edgeR on PhIP-Seq data and find it works surprisingly well, capturing the dominant variance structure even though it was designed for two-group RNA-Seq comparisons. They then introduce BEER (Bayesian Enrichment Estimation in R), a framework explicitly built around PhIP-Seq’s data-generating mechanism: a Bayesian model that uses edgeR’s parameter estimates as informative priors, then computes per-peptide posterior probabilities of enrichment. Across simulations and two real datasets — HIV elite controllers and SARS-CoV-2 convalescent samples — BEER delivers higher sensitivity for moderately enriched peptides while keeping false positives tightly controlled.

Translation for working immunologists: if your VirScan or AllerScan analysis pipeline pivots on edgeR, the results are likely defensible — but you are leaving discoveries on the table at intermediate fold-changes, exactly the regime where weak but biologically meaningful antibody reactivities live. BEER recovers them at the cost of additional CPU time for MCMC sampling, and is distributed via Bioconductor with the full pipeline at github.com/athchen/beer_manuscript.