henhaobio
Sign in

bioRxivGeneticsPreprint

DetectGxT: detecting gene-by-treatment interactions on molecular count phenotypes accounting for allelic additivity

Yuriko HarigayaiD, Michael I. LoveiD, William S. J. ValdariD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

MotivationIdentifying the mechanisms by which genetic variants affect the molecular response to an applied treatment is important across multiple biological fields, and an effective approach to this end is interaction molecular QTL mapping. However, the statistical models commonly used to detect such gene-by-treatment interactions (GxT) are non-trivially misspecified, and this can lead to decreased power. ResultsWe developed an R software package, DetectGxT, that uses nonlinear regression to more accurately model the relationship between the genotype and the transformed molecular count phenotypes. It also optionally models donor or polygenic random effects. Simulations show that nonlinear regression can increase the power to detect interactions. In existing interaction expression QTL mapping data from primary human neural progenitor cells, nonlinear and linear regression approaches identified overlapping but distinct sets of gene-SNP pairs with significant GxT interactions. Overall, our results suggest an advantage of nonlinear regression over linear regression in detecting GxT interactions on molecular phenotypes. AvailabilityThe DetectGxT software is available at https://github.com/yharigaya/detectgxt. Contactmilove@email.unc.edu, william.valdar@unc.edu

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (5 Aug 2026) · CC BY 4.0

R package DetectGxT fits allelic-additivity nonlinear models with optional donor or polygenic random effects, raising simulated power to detect gene-by-treatment interaction QTLs over linear regression.

Interaction molecular QTL studies usually log-transform counts and assume the genotype acts linearly on that scale. This is a misspecification when allelic additivity holds on the raw count scale, and it can cost power; existing tools cannot readily fit nonlinear mixed-effects interaction models, so this package supplies one for the second step of two-step mapping.

Key results

  1. SuggestedIn paired simulations generated under allelic additivity, nonlinear regression with a donor random effect had higher power to detect interactions and a better TPR/FDR trade-off than linear or RINT-based linear models. Both of those models showed substantial bias in their effect estimates.

    Paired design with 78 or 156 donors (total n 156 or 312). With a donor random-effect term in the model, nonlinear beat linear and RINT on power and on empirical TPR vs FDR (nominal FDR 0.01, 0.05, 0.1; Fig. 1 shows the n = 156 case with no simulated random effect). Its estimates largely matched true values, while RINT P values were anti-conservative at the larger n. · Results; Figure legends

  2. SuggestedIn unpaired simulations (one sample per donor) with polygenic random effects, nonlinear regression kept P values calibrated while linear and RINT models were anti-conservative. Nonlinear also gave the highest power and the most accurate effect estimates.

    Total n 156 and 312, with GRMs from each of four 1000 Genomes ancestry populations. Nonlinear P values were calibrated with or without the polygenic term; linear and RINT P values skewed toward zero. Nonlinear had the highest power across G×T effect sizes and the best TPR/FDR and estimation, consistently across ancestries. · Results

  3. SuggestedIn paired simulations where the data contained donor random effects, leaving the random-effect term out of the model degraded performance for every functional form. This suggests that donor random effects should be accounted for.

    Data were simulated with zero, medium or high random-effect SD and fitted with or without a random-effect term (six scenarios). When a donor effect was present but not modeled, calibration, power, overall performance and estimation all worsened for nonlinear, linear and RINT alike (Fig. S17). · Results

  4. ShownApplied to hNPC RNA-seq with and without growth stimulation, the package's linear mixed-model implementation reproduced nearly all previously published interaction eQTLs. This supports the fidelity of the implementation.

    Recovered 79 out of 83 previously identified interaction eQTLs, using linear mixed-effects models with a polygenic random effect on fully paired data from 78 genetically diverse fetal donors. · Results

  5. SuggestedOn the same hNPC data, the nonlinear and linear approaches agreed on most interaction eQTLs, but each detected a few gene-SNP pairs that the other did not.

    A large portion of interaction eQTLs was shared between the nonlinear and linear approaches, with a handful of gene-SNP pairs unique to each (Fig. S44). The main text gives no counts and reports no follow-up of the approach-specific pairs. · Abstract; Results

Methods, data and reproducibility

Two-step interaction QTL mapping. Condition-stratified QTL scans with standard linear tools define candidate feature-SNP pairs. DetectGxT's map_qtl then tests these pairs jointly across conditions with a nonlinear model encoding allelic additivity on the count scale, optionally adding donor or polygenic random effects. It was compared with linear and RINT models in simulations.

  • Simulated paired count data (one sample per donor per condition) generated under allelic additivity with donor random effects · n = 78 or 156 donors (total 156 or 312 samples)
  • Simulated unpaired count data (one sample per donor) with polygenic random effects from GRMs of four 1000 Genomes ancestry populations · human (GRMs from 1000 Genomes populations) · n = total sample sizes 156 and 312
  • Published RNA-seq of primary human neural progenitor cells from genetically diverse fetal donors, with and without growth stimulation (Matoba et al. 2024) · human · n = 78 donors · dbGaP phs003642.v1.p1; doi:10.1038/s41593-024-01773-6
Methods & tools
DetectGxT (R package; map_qtl, preprocess_pheno), MatrixEQTL, TensorQTL, limix_qtl, ClassifyGxT, Nonlinear mixed-effects regression (allelic additivity), Linear mixed-effects interaction model (baseline), Rank-based inverse normal transformation (RINT) plus linear model (baseline), Genomic relationship matrices from 1000 Genomes ancestry populations
Assay conditions & compute
Computational efficiency is described only qualitatively, as adequate across all simulation scenarios and replicates (Figs. S15, S16). No hardware, runtime values or memory are stated in the main text.
Reproducibility gaps
The hNPC data must be retrieved through dbGaP (phs003642.v1.p1). The package and paper code are also archived on Zenodo (10.5281/zenodo.21460613, 10.5281/zenodo.21446264). No R or package versions are given.
Limitations (5)
  • Nonlinear fitting costs more computation, so the authors limit it to the second step (candidate pairs only) rather than genome-wide marginal QTL scans. Those scans still rely on linear models of transformed counts. · stated by the authors
  • SNPs where one homozygote is missing in at least one condition (AA-AB SNPs) can make genotype and interaction effects practically non-identifiable and misestimated in the nonlinear model. The authors recommend excluding them. · stated by the authors
  • In unpaired simulations the polygenic random-effect term did not affect calibration. The authors attribute this to randomly drawn genotypes and expect mixed models to calibrate better with real genotypes, which was not simulated. · stated by the authors
  • Simulated counts came from the same nonlinear allelic-additivity model that DetectGxT fits, so part of its advantage may come from matching the data-generating process. The main text reports no robustness test against other generative models. · noted in this summary
  • In the hNPC reanalysis, the counts of shared and approach-specific interaction eQTLs are left to Fig. S44. Nothing independent, such as allele-specific expression, tests whether the nonlinear-only hits are real interactions. · noted in this summary

Read this if you map interaction eQTLs or other molecular QTLs across treatment conditions and want a nonlinear, allelic-additivity alternative to log-scale linear mixed models.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISHarigaya '25Wang '25Zhang '24Aschard '15Zhang '24Yamamoto '21
76% semantic matchGenetics
Yuriko Harigaya, Nana Matoba, Brandon D. Le+4 more·PLoS Genetics·Apr 9, 2025· 2 citations