Step 01
Define the comparison before running the graph
The important design choice is not the tool. It is what will be compared: linear-reference outputs against graph-aware outputs, with the same samples and a fixed set of downstream statistics.
Precision medicine Pangenome
Case study · Precision medicine · CSIRO, Mayo Clinic, Paris-Saclay
The reference genome is biased toward one ancestry. Adapt it to your cohort with a pangenome, and the answer changes completely.
Variant calling leans on a single linear reference genome built largely from European-ancestry samples. For underrepresented populations, that reference introduces measurable bias — making precision medicine least accurate for exactly the people it already underserves. I led the multi-institution project that quantified it, using Middle Eastern and Horn of Africa populations as the test case.
For the specialist Pangenome graph pipelines in Python and Bash on HPC, run on Illumina short-read WGS data. SNP/INDEL calling accuracy, PCA, admixture modelling and F-statistics, each measured against the graph-based approach. Built on doctoral work in reference bias, ancient-DNA mapping and reproducibility standards.
4
Team led: 2 researchers + 2 postgraduate students
3
Institutions · 3 time zones
Pangenome
Graph workflow on HPC
Q1 paper
BioMolecules, 2025
The result
A reference genome can become an invisible assumption.
For precision medicine in underrepresented populations, the question is not only which model performs best. It is whether the reference itself is shaping the signal before any model sees the data — and for Middle Eastern and Horn of Africa populations specifically, it is.
Linear reference versus pangenome graph
Reference sequenceAlternative sequence
Linear reference
Alternative sequence forced onto one path.
Pangenome graph
Alternative sequence kept as graph paths.
- Same cohort, same samples
- Two variant sets, one comparison
- 4 population statistics, explored below
SNP & INDEL calling accuracy
Variant calls made against the pangenome graph are compared, sample by sample, to calls made against the linear reference — the first and most direct place reference choice can silently change a result.
Principal component analysis
PCA run on both variant sets shows whether individuals cluster by true ancestry, or partly by which reference happened to align their reads more easily.
Admixture modelling
Ancestry-proportion estimates per sample are compared across the two variant sets, to see whether reference choice shifts inferred ancestry and not just raw genotype calls.
D- and F-statistics
These test for reference-driven bias directly: if a population is systematically pulled toward the reference's own lineage, D- and F-statistics move even when nothing biological has changed.
Why it matters
A biased variant call isn't a rounding error. It's a missed or wrong result.
Reference bias doesn't stay abstract once it reaches a report. A variant the reference makes harder to call correctly can be flagged pathogenic when it isn't, or missed when it matters; a polygenic risk score built on those calls is systematically off for that ancestry group; a drug-response genotype (see the pharmacogenomics case study) is only as good as the calls underneath it. Precision medicine that works for one ancestry and silently degrades for others isn't precision medicine.
Not just a methods paper: coronary heart disease, Somali population
The same reference-bias-aware approach was applied to a specific disease in a specific underrepresented group — coronary heart disease genetics in the Somali population, presented as a poster at GA4GH Planetary 2024. The method generalises to any under-represented population and any condition with a genetic component, not just the Horn of Africa / Middle Eastern cohort in the published paper.
In the published paper
What changed when the reference changed.
With no Somali genomes to build from, the study used proxy populations. The result shows that the reference changes the answer. It is not a Somali reference panel.
The comparison design
Build the graphMozabite Illumina whole genomes, Human Genome Diversity Project
Test populationBedouins
Compare with hg38Effective population size, allele frequencies, GWAS markers
Coalescent effective population size, same population, two references
Log scale. Only the graph figure fell inside the 95% interval from simulations. Graph variants also had significantly lower allele frequencies (p < 2.2 × 10-16), and the GWAS markers found on the standard reference were not the ones found on the pangenome.
The method
Six steps from representation to interpretation.
Step 02
Make ancestry and cohort structure explicit
Pangenome benefits are not uniform. The workflow keeps population labels, relatedness checks and cohort metadata close to the analysis so shifts in PCA or D/F-statistics can be interpreted, not just observed.
Step 03
Run on HPC as a reproducible workflow
Graph workflows are too heavy for ad hoc notebooks. Jobs need pinned references, resource requests, resumable stages and logs that survive peer review across three institutions and three time zones.
Workflow shape
prepare-reference --linear GRCh38.fa --graph cohort.pg map-cohort --samples samples.tsv --backend hpc --out mappings/ call-or-project --mappings mappings/ --out variants/ stats-panel --variants variants/ --pca --dstat --fstat
Step 04
Quantify movement, not just accuracy
The useful result is often a change in structure: PCA separation, allele-frequency estimates, D-statistics or F-statistics. Those shifts show where the reference was influencing interpretation.
Step 05
Separate biological signal from tooling artefact
Every observed shift is checked against coverage, mappability, filtering and batch effects. Otherwise the graph becomes a new black box in place of the old one.
Step 06
Package the figures for scientific scrutiny
The final output needs reproducible figures, a methods trail, parameter values and enough narrative for collaborators from medicine, bioinformatics and biostatistics — across CSIRO, Mayo Clinic and Paris-Saclay — to review the same result.
Limits
Pangenomes do not magically fix population genomics.
The graph can introduce its own bias
If the graph is built from an uneven cohort, it can encode new representation problems. Graph composition is part of the method, not a neutral detail.
Compute and operational complexity are real
The pipeline needs HPC/cloud discipline: storage, indexing, resumability and versioned references. This is production engineering, not just method selection.
Clinical claims require separate validation
The project supports precision medicine research and method development. Clinical deployment still needs assay-specific validation and governance.
Track record
Four years of reference-bias work, not one paper.
This project builds directly on doctoral research at the University of Adelaide (2018–2022) on reducing reference bias in genomic data analysis using pangenome graphs and benchmarking.
Related peer-reviewed output
Q1 journals2025
Oliva et al., “A Pangenomic Approach to Improve Population Genetics Analysis and Reference Bias in Underrepresented Middle Eastern and Horn of Africa Populations”
BioMolecules (Q1) · DOI 10.3390/biom15040582
2021
Oliva et al., “Systematic Benchmark of Ancient DNA Read Mapping”
Briefings in Bioinformatics (Q1, IF 9.5) · DOI 10.1093/bib/bbab076
2021
Oliva et al., “Additional evaluations show that specific BWA-aln settings still outperform BWA-mem for ancient DNA data alignment”
Ecology and Evolution (Q1) · DOI 10.1002/ece3.8297
2025
Souilmi, Oliva et al., “Lessons learned: recommendations for reproducible paleogenomic data analyses”
The American Journal of Human Genetics 112(12), 2830–2841 · co-first author
Also presented as a poster at GA4GH Planetary 2024, on coronary heart disease genetics in the Somali population. The multi-institution work sits inside an NIH-funded programme (∼US$850K/yr, held by Mayo Clinic) on pangenome methods for precision medicine. Related CSIRO work: PrecisionGo, with Sydney University and Westmead Hospital, applied similar time-series methods to genetic patterns in liver-transplant donors and recipients — I was the main contact and project manager.
Need a genomics workflow that can survive peer review?
Start with a diagnostic, then decide whether graph methods, linear workflows or a hybrid approach are justified for your cohort.
Keep reading
Other case studies
Case studyGeneGuardian
Dynamic consent for genomic data, live at a pilot hospital
Read it →Case studyMicrobiomeNanopore 16S: sample to pathogen report in 2 days
Read it →Case studyPharmacogenomicsStar-allele calling in 2,503 people
Read it →Case studyNGS & WGSRaw reads to a decision, across platforms
Read it →