Multivariate Bayesian variable selection regression

% Fine Mapping Benchmark % \url{https://github.com/gaow/mvarbvs/tree/master/dsc} % \today

Benchmark status

Data-set

Genotype

  • A GTEx sample region (FMO2) of size $N=698$, $P=7492$
  • A GUEVADIS sample region of size $N=343$, $P=1001$
  • Parameters to "trim" $P$, eg from 7492 also to, say, 1001

Phenotype

  • The original GTEx Throid and Lung expression for the GTEx sample region
    • should have around 3 eQTLs
  • A simple simulated GUEVADIS expression data from DAP-g paper

Simulation themes

Univariate

  • Simple point mass + rnorm() simulation, as in DAP-g paper
  • Point mass + mixture of normal, as in ASH paper
    • spiky, near-normal, flat-top, skew, big-normal, bimodal

Multivariate

  • Column-wise stacking of univariate simulations
  • Point mass + mixture of multivariate normal, as in MASH paper
    • All "canonical" prior covariances
    • Have to provide grid

About LD

There are mechansim to

  • Plot save LD heatmap for input data
  • Put signals to the most "LD-convoluted" blocks
  • Ensure signals are from independent LD blocks
  • Permute & break LD structure?

Fine-mapping methods

From Stephens Lab

  • varbvs
  • susie
  • M&M ASH

From the field

  • DAP-g
  • FINEMAP
  • CAVIAR

Each with multiple module "flavors" (parameters)

Single-replicate diagnosis

  • When avaiable, compare scattered plots of $\tilde{\beta}$ or $\hat{\beta}$ vs $\beta$ plots
  • Show PIP and log10BF
  • $\ldots$
  • Customized diagnostic plots for SSE methods: susie and M&M.

Still working on unifying output from methods and annotate with eg LD info.

Cross-replicate evaluation

Next to-do:

  • Power vs false positive: ROC
  • Point signal level: PIP / lfsr
  • Set signal level: cluster PIP

LD situation

GUEVADIS sample

GUEVADIS LD { width=80% }

Trimmed GTEx sample (to the size of GUEVADIS sample)

GTEx LD { width=80% }

Simple GUEVADIS simulation

Simple GUEVADIS simulation

Simple GUEVADIS simulation

varbvs

`varbvs`

susie

`susie`

FINEMAP

`FINEMAP` { height=80% }

DAP

DAP { height=80% }

CAVIAR

CAVIAR { height=80% }

GTEx with MASH "simple het" covariance

Trimmed GTEx, response 1

Simulated GTEx tissue 1

Trimmed GTEx, response 2

Simulated GTEx tissue 2

varbvs, response 1

`varbvs`

varbvs, response 2

`varbvs`

susie, response 1

`susie`

susie, response 2

`susie`

M&M, response 1

`M&M`

M&M, response 2

`M&M`

Copyright © 2016-2020 Gao Wang et al at Stephens Lab, University of Chicago