GenSep Overview

GenSep implements a liability-threshold framework for case–case subtype discrimination — how well genetics can tell two disease subtypes apart. The central quantity is the genetic separation variance

V_S = λ1² h1² + λ2² h2² − 2 λ1 λ2 rg h1 h2,

built from the two subtypes’ liability-scale SNP heritabilities (h1², h2²), their genetic correlation (rg), and the prevalence-driven selection intensities (λ_i). V_S is large when the subtypes have strong subtype-specific genetic components and small when their genetic effects are largely shared. From V_S, GenSep derives the oracle case–case AUC (the maximum achievable AUC if the true genetic values were known) and its leading-order approximation, and the balanced observed-scale case–case heritability h²_cc = V_S / (V_S + 4) — a bounded, AUC-linked summary of genetic separation. Every quantity comes with a standard error.

GenSep computes these three interchangeable ways:

From GWAS summary statistics — a tagging file and the two subtypes’ summaries; GenSep estimates the heritabilities and genetic correlation (SumHer) on one common SNP set and derives everything with a fused block-jackknife SE.
From point estimates — observed-scale heritabilities and genetic correlation with their SEs, from LDAK / SumHer, LDSC, or any other method; the SE is propagated by Monte-Carlo or the delta method.
Plus finite-PRS AUC — given each subtype’s PRS case/control AUC, GenSep also reports the AUC achievable with those polygenic scores and the PRS recovery V_PRS / V_S.

GenSep is an open-source, statically linked C++17 program with no runtime dependencies (Eigen is header-only and vendored). GenSep is developed and tested on 64-bit Linux, for which a prebuilt executable is provided; on Windows it runs under WSL, and on macOS a Linux container or VM is the supported route.

User’s Guide: Installation · Tutorial

Citation

Chao Ning, Jasper Hof and Doug Speed. Quantifying the genetic separability of disease subtypes. medRxiv (2026). doi:10.64898/2026.09.18.26363395

The heritability and genetic-correlation solvers are a C++ re-implementation of SumHer (Speed & Balding, Nature Genetics 2019), originally implemented in LDAK.

Contact

For questions, open an issue on GitHub or email me at chao.ning@qgg.au.dk

For other tools, see chaoning.github.io/software.html.