ProbGen24 Day1

Shamil Sunyaev

Effect sizes are small. Number of loci vs median effect size is linear with a negative slope. More than 12000 associated variants for height GWAS.

• What is the balance of evolutionary forces stably maintaining polygenic variation?

Connally et al. eLife 2022 showed that many genes under GWAS peaks and with independent evidence of the biological roles have eQTLs in the relevant tissue. But this eQTLs do not colocalize with GWAS variants.

Joint distribution of allele frequencies and effect sizes

U-shaped plot with low MAF and high MAF have larger effect sizes, smaller effect sizes with common variants. This is not news. The strange property is the symmetry.

Mental model – neutrality

• Many human diseases are late onset and the underlying variation is truly neutral – Strong version.
• The traits are under selection, but variants identify by GWAS have so small effects that they are insensitive to selection (below the drift barrier) – Weak version.

The weak version is consistent with pleiotropy while the strong version is not.

See Gazal et. al. 2018.

Is selection acting on human common diseases?

• Monogenic diabetes and familial hypercholesterolemia are very rare in the population.
• Polygenic form of T2D is common. Cholesterol levels vary in the population, even though the trait is heritable.

Rare variants for some complex traits are under directional selection. Ganna et. al. 2018, Huang et. al. 2023.

Influx of mutations increases disease liability and reduces fitness.

Bulmer, 1972. The genetic variability of polygenic characters under optimizing selection, mutation and drift.

Approach: \[\psi(x \mid \beta)\] –> allele frequency \(x\) and effect size \(\beta\) distribution of allele frequency conditional on effect size.

GWAS ascertainment, LD, Sampling noise.

Simulations: directional selection has assymetry in the distribution. 1-D stabilizing and High-D stabilizing shows symmetric distribution.

LD Score Regression \(\righarrow\) LDSPEC.

High proportion of zero effects kills correlations. Clustering of functional effects rescues correlations.

One-sided mutation: Hill-Robertson effect

Take-Home: Stabilizing selection is consistent with distribution of allele frequencies and effect sizes, also consistent with effect size correlations.

Bridging biology and genetics

• Increase power by integrating network of biological assays.

Rare variant analysis at the level of pathways and gene networks. - incorporate network relationship of genes.

NERINE. Network-based rare variant enrichment. \(\alpha \sim MVN(0, \theta \cdot \Sigma)\) pathway effect \(\theta\).

Khurana lab. Khurana, Peng et al. Cell Systems 2017. Gene network in Parkinson’s disease.

Independent neuronal survival CRIPRi screen in \(\alpha\)-Sunucleinopathy model. Several genes from this network are associated with “PD” neurons.

Molly Przeworski

Sources of germline mutations in humans

Germline mutation \(\rightarrow\) Point mutation

Source: Replication error, DNA damage + repair.

Paternal bias in germline mutation. Many more germ-cell divisions over age. Gao et al. 2019.

Maternal age also affects number of phased DNMs per proband.

Wu et al. 2020. Narrow range of paternal bias in mutation across mammals. Degree of paternal bias is uncorrelated with paternal age. Stable with age in humans and other mammals.

Hypothesis: Mutations does not arise with replication, but from DNA damage + repair. Repair – inefficient vs efficient but error-prone?

https://cancer.sanger.ac.uk/signatures/

SBS1 and SBS5 germline mutations. SBS5: Etiology unknown, SBS1: Spontaneous deamination of methylated CpGs.

SBS1 accrues with age in dividing cells, SBS5 accrues with age in dividing cells and post-mitotic ones.

• Mutuations can accumulate with time due to distinct mechanisms: constant rates of cell divisions or constant rates of damage.
• Most germline mutations are assigned to SBS5.

Spisak et al. 2023 bioRxiv.

Laurent Duret

Meiotic recombination –> segregation of homologous chromosomes, maintain genetic diversity.

Recombination hotspots.

Fine-scale recombination maps available for a few species. Hotspots in promoter-like regions – evolutionarily stable. Some organisms do not have hotspots, e.g. Drosophila, C elegans. In other organisms, hotspots are directed by PRDM9, e.g. Humans, Mice, Salmonids – hotspots are located away from promoters – very rapid evolution of hotspot location (self-destruction caused by PRDM9 activity).

PRDM9 dependent hotspots - away from promoter-like regions (CpG islands). PRDM9 independent hotspots: near promoter-like regions (CpG islands).

Is the location of default hotspots conserved across PRDM-9 deficient mammals?

Mutant mice - no PRDM-9. Dog LD-based recombination rate: Laplace with peak near 0, plotted against distance to MDH loci of mutant mice.

93% of mouse default hotspots are hypomethaled.

What is the recombination activity at default hotspots in mammals with a functional PRDM9 gene?

GC-biased gene conversion (gBGC) signatures can be used as a measure of recombination activity.

Take-Home: PRDM-9 independent hotspots are also in promoter-like regions.

See julian Joseph, Djivan Prentout, Alexandre Laverre, Theo Tricou and Larent Duret, BioRxiv. High prevalence of PRDM9-independent recombination hotspots in placental mammals.

Machiki Katori, Tsukuba Tokyo

Can mutation rates be biased by experiences? Do epigenetic patterns predict mutation rates?

Accumulation line of Arabidopsis thaliana. Plants rely more extensively on epigenetic mechanisms compared to mammals.

epiNMF

Epigenetic features - NMF - clusterings - find epigenetic patterns - GLM with exons.

Take-home: Epigenetic patterns partially predict mutation rates.

Impacts of the epigenome-induced mutations on genetic adaptation.

Is a purely random mutation process sufficient? Cisplatin induced various types of DNA damages. Resistant cells appered after 7 days from treatment.

Marta Pelizzola

Integration of mutational opportunities to estimate robust mutational signatures

COSMIC repository of mutational signatures. Tate et a;.

Input data: rows \(\rightarrow\) patients. columns \(\rightarrow\) mutation types. Each cell contains count of the mutation type.

Perform NMF.

With zero flanking nucleotides, there are 6 mutation types. With one flanking nucleotide, there are 96 types – trinucleotide context. One can also consider pentanucleotide or heptanucleotide context.

Negative Binomial NMF with opportunites.

Metric: Reconstruction error on the BRCA dataset, GKL divergence plotted against the number of signatures.

Model with opportunities (NB-NMF) improves prediction.

Take-Home: Extended sequence context facilitate the investigation into the reasons for the excess of certain types of mutations. Including mutation opportunities improve prediction.

Mutational opportunites: number of sites for a mutation type to occur, e.g. TTT has a larger opportunity than other types.

Pier Palamara

Threads algorithm for ARG. - Add (“thread”) a new sample to an existing genealogy.

Zhang et al, Nat Genet, 2023. And inferred ARG allows inferring the presence of unobserved variants. Complex traits can be studied using ARGs. Linear mixed model, using GRM (genetic relationship matrix) derived from ARG.

h2 using GCTA - \(O(n^3)\). ARG-RHE reduces time complexity to \(O(n)\).

Hrushikesh Loya

GhostBuster.

Large-scale genealogies.

An admixture is defined by its inverse coalescence rates with reference panel. Assume there a re 3 types of trees in the generalogy. Comp1 ICRs, 1 coalescece faster than 2 to give target. Comp2 ICRs, 2 coalescece faster than 1 to give target. Comp3 ICRs, both have similar rates.

The Eurasian ancestry explains the Neanderthal ancestry in Africa and is 5000-20000 years old.

Jonathan Terhorst

Bayesian algorithm for PSMC.

• Parallelization using exponential forgetting.

phlash. https://github.com/jthlab/phlash