Sign in

Jeffrey Pullin

@jeffreypullin.bsky.social
196 followers 318 following 51 posts

PhD Student, MRC Biostatistics Unit University of Cambridge Gates Cambridge Scholar Bioinformatics, genetics, single-cell, statistics Australian 🇦🇺

PostsRepliesMedia
Jeffrey Pullin @jeffreypullin.bsky.social · 22/07/2025
When comparing methods we found that mixed model methods did not have better performance, but that, as previously reported, count distribution methods increased power. Overall we recommend the negative binomial GLM model, using the APL, as the method with the best overall performance.
Statistical power of negative binomial and linear model methods across the
OneK1K dataset a) Number of eQTLs detected by the quasar linear model and negative binomial GLM with adjusted profile likelihood dispersion estimation methods across all cell types
in the OneK1K dataset. b) Number of eGenes detected by the quasar linear model and negative
binomial GLM with adjusted profile likelihood dispersion estimation methods across all cell types
in the OneK1K dataset.
100
Jeffrey Pullin @jeffreypullin.bsky.social · 22/07/2025
When run on CPUs quasar is quite a bit faster (up to ~40x) than exisiting methods, while producing concordant output when the statistical model aligns.
 Histograms of Pearson correlation of − log10 transformed variant-level p-values for each gene, correlating the output of output
of quasar against that uses the same statistical model (LM: tensorQTL, NB-GLM : jaxQTL, LMM:
apex. All results are computed for the B IN cluster. b) Speed of methods across the three representative cell types. All methods were run on CPUs. Methods are labelled by the options used
to run them: for tensorQTL and jaxQTL ‘cis’ computes significance at the level of genes while ‘cis
nominal’ computes significance at the level of variants.
100
Jeffrey Pullin @jeffreypullin.bsky.social · 22/07/2025
2. We also show that negative binomial models can fail to appropriately control the Type 1 error, which we fix in quasar by implementing the Cox-Reid adjusted profile likelihood (APL), a core part of edgeR and DESeq2.
Bar charts of number of discoveries across different tools and thresholds in a paper about eQTL mapping
100