2023, June

2023-06-30

  • Knowles-Raj meeting, presented by Aline. I did not take meeting notes.
  • Continue simulation setup. Study about correlated variables, and validation of clustering: internal and external.
  • How to benchmark different methods using real data? In real data, we know the true labels (broad phenotypes). Each method provides a reduced rank approximation, \(Y_{\mathrm{opt}}\). One idea is to classify based on \(Y_{\mathrm{opt}}\) and compare the clusters with the true labels. We assume that the best methods will give the best clusters.

2023-06-29

  • Journal club presentation, Chapter 10 of Graham Coop’s book. Presented by Anjali and Karin.
  • Simulation of correlated feature variables.

2023-06-28

  • Simulation setup for z-scores. Generative model using correlated loadings and orthogonal factors.
  • Group meeting by Claire. Meeting notes.
  • NYGC photoshoot.

2023-06-27

  • Literature review of nuclear norm minimization.

  • Think how to evaluate the different methods. Write the tasks to evaluate in latex.

  • Notes:

2023-06-23

  • GradVI. Rerun simulation for comparing GradVI with CAVI in three different settings: independent variables, correlated variables with block-diagonal covariance matrix, trendfiltering.

2023-06-22

  • Read Jaggi 2013 paper on Frank-Wolfe algorithm.
  • Write algorithm and theory for weighted nuclear norm matrix dimensionality reduction.
  • Journal club presentation, Chapter 10 of Graham Coop’s book. Presented by Anjali and Karin.

2023-06-21

  • Group meeting by Chirag Lakhani. Meeting notes.

  • Meeting with David. We discussed about 3 main topics:

    • NPD Sumstats. We discussed about the Frank-Wolfe algorithm. David told that the rank of the optimized matrix was not always at the boundary of the \(\ell_1\) ball. He fixed some errors in the R implementation of weighted PCA – step size was incorrect. Variant 1 of step size ($ = 2 / (2 + t)$) is slower than Variant 2 optimum \(\gamma\) obtained by minimizing the function \(f(\mathbf{X} - \gamma \mathbf{D})\), where \(\mathbf{D} = \mathbf{X} - \mathbf{S}\). Removing duplicate GWAS from IEU helps delineating the clusters in lower dimension. Can we setup simulations to benchmark the methods? Can we use weights to improve results? How to harmonize the weights? Will Niamh be able to help us in harmonizing the GWAS summary statistics?
    • Second project. We discussed whether to work on splicing project in collaboration with Ivan Iossifov or applying Tejaas on the BigBrain project. We decided to do some pilot work on both before deciding on any one of them.
    • Finemapping using annotations. Gilead is trying to rerun the simulations by increasing PVE.

2023-06-20

Rerun the exploratory analyses on NPD summary statistics using the newly preprocessed data. I had to include several phenotypes in the meta file, trait_meta.csv.

2023-06-16

Clean the NPD summary statistics data. In particular, remove duplicate PGC GWAS from IEU OpenGWAS project, include the GWAS summary statistics imputed on GTEx variants. Earlier preprocessing ignored many GWAS imputed on GTEx.

2023-06-15

2023-06-14

  • Group meeting presentation at NYGC.
  • Look at data preprocessing. I found that many PGC GWAS were included twice (as they were also included in the IEU OpenGWAS project).

2023-06-13

Mostly preparation for group meeting.

2023-06-12

  • Start preparing presentation for group meeting.
  • Read about heterogeneity in neuropsychiatric disorders.
  • Create a custom NYGC theme for Apple Keynote.
  • Update Linkedin profile.

2023-06-09

Work from Home

Read, write and submit review for an article in Scientific Reports.

2023-06-08

Work from Home

I read about using Log Bayes Factor as genetic distance between populations to transform the genotype data matrix. Delrieu and Bowman, 2005 discuss the mathematical principles. Delrieu and Bowman, 2006 discuss how it can be applied and visualized. From the Bayes Factor, they use the scene-anchored argument of Evett and Weir, 1998, “Interpreting DNA Evidence, Statistical Genetics for Forensic Scientists” to arrive at a statistic which resembles the likelihood ratio. For details of derivation, see Evett and Weir, 1991, “Flawed reasoning in court”.

2023-06-07

2023-06-06

Work from Home

  • Read about trend filtering.
  • Update personal website.

2023-06-05

  • Compare RobustPCA and nuclear norm matrix factorization using Frank-Wolfe algorithm. See Notebook

2023-06-02

Work from Home

  • Meeting with Gilead Turok. He is working on finemapping using annotations.
  • Attend the Raj-Knowles group meeting.
  • Read about PCA.
  • Update workbench. Install micromamba and Quarto in MacOS. See details.

2023-06-01