Tommy Carstensen

Seven years at the Wellcome Sanger Institute on the African Genome Variation Project and the Uganda Genome Resource: whole-genome sequence from thousands of individuals, and a haplotype reference panel for a continent that reference panels had largely skipped. Then deep learning on UK Biobank electronic health records at Queen Mary University of London, two years of data science at Novo Nordisk, and now clinical machine learning with the Benros group in Copenhagen, on national registry and EHR data.

Selected publications

• 2024 · Nature Medicine
  High-resolution African HLA resource uncovers HLA-DRB1 expression effects underlying vaccine response Mentzer AJ, Dilthey AT, Pollard M, et al. · 6th of 46 authors
• 2023 · Nature
  Africa-specific human genetic variation near CHD1L associates with HIV-1 load McLaren PJ, Porreca I, Iaconis G, et al. · 14th of 71 authors
• 2020 · Nature Communications
  Distinct genetic architectures and environmental factors associate with host response to γ2-herpesvirus infections Sallah N, et al. · 4th of 20 authors
• 2019 · Cell
  Uganda Genome Resource enables insights into population history and genomic discovery in Africa Gurdasani D, Carstensen T, Fatumo S, et al. · 2nd of 57 authors
• 2019 · Diabetologia
  Genome-wide association study of type 2 diabetes in Africa Chen J, Sun M, Adeyemo A, et al. · 5th of 17 authors
• 2015 · Nature
  The African Genome Variation Project shapes medical genetics in Africa Gurdasani D, Carstensen T, Tekola-Ayele F, et al. · 2nd of 42 authors

full list on Google Scholar

Registry-scale data breaks the standard statistical packages long before the statistics stop being valid. These implement the same estimators without holding the table in memory.

coxstream

Python · R · Cython · survival analysis

Exact out-of-core Cox proportional hazards regression. Each Newton–Raphson step becomes a sum over parquet row groups, so peak memory is one row group plus O(p²) state rather than the whole table. The estimates stay exact: coefficients agree with R's survival::coxph to 2e-14 at two million rows, and with lifelines to 3e-14 on data rounded to 120 distinct event times, where ties necessarily straddle chunk boundaries. Standard errors agree to 2e-13.

pip install coxstream / CRAN / R source / doi:10.5281/zenodo.20693783

renew-glm

Python · NumPy · SciPy · generalized linear models

Streaming generalized linear models with bounded memory, implementing the renewable estimation of Luo and Song (2020). One pass per iteration accumulates the sufficient statistics, so the streaming path holds one chunk plus O(p²) state, independent of the number of rows. The chunk-buffered path trades that for standard errors. Gaussian, binomial and Poisson, checked against statsmodels in the test suite.

pip install renew-glm / doi:10.5281/zenodo.21950314

PlotPick

Python · Claude Vision · Streamlit · PyMuPDF

Recovers the numbers behind published figures. Upload images, PDFs or ZIPs and it reads boxplots, bar charts and line plots back into structured tables. Benchmarked against ChartX and PlotQA with Claude, GPT, Gemini and DePlot baselines.

hosted app (Streamlit sign-in)

RAM- and CPU-efficient data processing

23 slides · Quarto / reveal.js · Python and R · synthetic registry-shaped data

open the deck / all slides

Reports and fixes accepted into the tools this work depends on.

• polars #27860 Reported with a reproducer at 200 million rows: parquet min/max statistics exclude NaN while polars orders NaN as the largest value, so float >= filters could never skip row groups. Labelled accepted upstream; the nan_count-gated fix is verified locally and not yet merged.
• samtools/bcftools Six merged pull requests: error messages that name the accepted types and options, and documentation corrections for the two-VCF-input statistics.
• biopython Merged: modified residue 4BF now maps to phenylalanine rather than tyrosine.