Andrew D Kern
University of Oregon, OR·Ecology & Evolution
Genome sequencing now produces DNA data from huge numbers of individuals, but the tools to interpret it have not kept up. Classic population genetics models often ignore where organisms live and how genomes are actually shaped, so they miss information in modern data.
- Fields
Ecology & Evolution
Genetics
Computer Science
- Topics
population genetics
population genomics
deep learning
- Grants
- 2 known grants, through Jun 2029
What they work on
- Develop methods to infer dispersal, how far individuals move, directly from population genomic data, and assess how spatial structure affects GWAS, genome-wide studies linking variants to traits, with Anopheles gambiae data in mind.
- Build deep learning methods that detect beneficial alleles by relating a gene variant's geographic spread to surrounding genomic variation, and selection-detection methods built on inferred population-scale genealogies.
- Expand stdpopsim, an open-source genome simulation library, to model selection realistically using published selection estimates, and use it to benchmark common methods for inferring demography and selection.
- Design deep neural networks that compute directly on ancestral recombination graphs (ARGs), maps of shared ancestry along the genome, using graph learning to scale to whole-chromosome data, and release them as software.
- Evaluate methods for estimating uncertainty in deep learning predictions for population genetics, and apply self-supervised learning techniques from generative AI to genomic data.
More about the research
Making sense of this flood of data requires new statistical and machine learning methods. The group has pioneered deep learning, a type of machine learning built on layered neural networks, for tasks such as detecting natural selection, finding DNA inherited from other populations, mapping recombination and predicting geographic origin. Its two projects extend this work in complementary directions: one focuses on geography, adaptation and simulation, and the other on scaling deep learning to whole-genome data.
Why it matters
These methods could help researchers extract more reliable conclusions from large genome datasets, including tracking how malaria mosquitoes adapt to control efforts in Africa and avoiding false signals in genetic association studies. Open-source tools such as stdpopsim and scalable deep learning software would give geneticists reproducible, benchmarked ways to study history, movement and adaptation.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Ryan D. Hernandez
University of California-San Francisco, CA·Genetics
Measure how race, ethnicity and genetic ancestry each contribute to molecular traits and to heart, lung, blood and sleep disorders, using TOPMed whole genome sequencing, RNA-seq gene expression and metabolomic data from over 180,000 people.
Volodymyr Kuleshov
Cornell University, NY·Computer Science·Early-career
Develop a deep generative neural network model of genetic sequences that replaces classical hidden Markov models and handles high dimensionality and long-range dependencies along the genome.
David E Reich
Harvard University, MA·Genetics
Assemble an ancient DNA time series of 34,000 West Eurasians, enough to measure natural selection coefficients as small as about 0.25%, and release it as a public resource before publication.
The school
University of Oregon: 14 labs on PI Pages, $175M a year on research, 219 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.