Research
Rare variation, non-additivity, and the statistics of very large biobanks.
My work sits at the point where statistical genetics stops being a clean estimation problem and starts being an engineering one. The questions are simple to state and awkward to answer: which genes carry rare variation that changes disease risk, how much of the genome acts in ways that are not simply additive, and how do we combine evidence across biobanks that were never designed to be combined?
The Biobank Rare-Variant Consortium
BRaVa is a global federated effort to analyse rare coding variation at scale. Sixteen biobanks and cohorts to date contribute association results under a shared protocol: harmonised phenotype definitions, a common variant annotation pipeline, and gene-based tests that can be meta-analysed without moving individual-level data across borders.
The first flagship analysis spans over 1.2 million sequenced individuals across ten biobanks and diverse ancestries, covering 33 clinical endpoints and 11 quantitative traits. Around a third of the gene-level associations it recovers are invisible in any single contributing biobank, which is the entire argument for doing this the hard way.
brava-genetics.github.io →Life beyond the additive model
Almost all of human complex trait genetics assumes additivity because it is convenient. I have spent a good deal of time asking what that assumption costs. That includes a genome-wide scan for dominance effects in the UK Biobank, exome-wide evidence for compound heterozygous effects, recessive meta-analysis across six global biobanks, and rare-variant-driven deviations from additivity at biobank scale.
A related thread: individuals whose measured phenotype sits far from their polygenic expectation turn out to be enriched for rare damaging variants in genes that cause rare disease. Residuals are not noise, they are a screen.
Psychiatric, reproductive and metabolic traits
Exome sequencing across nearly 14,000 people with bipolar disorder implicated AKAP11, a gene shared with schizophrenia whose protein interacts with GSK3B, the presumed target of lithium. Alongside that, work on schizophrenia and autism gene discovery, a genome-wide study of infertility and reproductive hormones, and the genetic architecture of how adiposity changes across adulthood in electronic health records.
Phenotypes are the bottleneck
Genotypes are cheap and standardised. Phenotypes are neither. I work on principled distillation of high-dimensional biobank phenotype data into interpretable structure, on curating phenotype definitions that mean the same thing in Helsinki and Osaka and Boston, and on the unglamorous machinery that makes cross-biobank comparison honest.
Pathogen evolution
My DPhil work built Bayesian models of HIV-1 evolution within and between hosts, using genealogical and dynamical modelling to infer rates of immune escape and reversion, and scaling those inferences to very large sequence databases while controlling for population structure and recombination. The statistical instincts carried over intact.
Nature Communications, 2019 →Oxford
I am a SMARTbiomed Senior Research Fellow, jointly based at the Big Data Institute and the Department of Statistics, where I also co-lead the (gen)omics theme. SMARTbiomed is a Novo Nordisk Foundation funded centre for biomedical data science.
Always happy to hear from prospective students, postdocs and collaborators.
duncan.palmer@stats.ox.ac.uk →