Step 01
Ancient DNA and reference bias: Illumina WGS, benchmarked (PhD, Adelaide, 2018–2022)
Almost every genome study maps reads to a single linear reference, and that reference is not neutral: a read carrying the reference allele maps more easily than one that does not, so results drift towards the reference. Ancient DNA makes it worse, because the molecules are short and damaged, with cytosines reading as thymines near fragment ends.
I worked with Illumina data from published ancient genomes and from our own lab at the Australian Centre for Ancient DNA. Real data cannot tell you where a read truly belongs, so I also simulated reads from human chromosome 22 with Mitty, carrying variants from African, East Asian and European individuals in the 1000 Genomes Project, and added realistic post-mortem damage with gargammel: about 16% C-to-T misincorporation, modelled on a 7,000-year-old sample, in untreated, half-UDG and full-UDG versions.
I then benchmarked 30 mapping strategies across BWA-aln, BWA-mem, Bowtie2 and NovoAlign (including NovoAlign against a reference augmented with IUPAC codes at known variant sites), scoring precision, reference bias at SNPs, downstream effects on D-statistics and PCA, and CPU hours. Published in Briefings in Bioinformatics (2021). Well-tuned BWA-aln stayed the most reliable choice, and mapping-quality filtering cut the bias for every mapper. When another group proposed a new BWA-mem setting as a replacement, the same benchmark still put BWA-aln ahead (Ecology and Evolution, 2021).
How the benchmark was built
Result: well-tuned BWA-aln stayed the most reliable, and mapping-quality filtering cut the bias for every mapper.
Toolchain around it
AdapterRemoval · samtools · Picard read processing mapDamage damage profiles ANGSD genotype likelihoods, low coverage PLINK · ADMIXTURE · smartpca · ADMIXTOOLS structure, admixture, f-statistics vg (construct, map) · minigraph pangenome variation graphs