Guided Sparse Factor Analysis Project

Introduction

Motivation

Genetic perturbation often regulates the expression of a network of genes via trans effect.

Current computational approaches to detect trans genetic effects include:

• Per-gene association analysis such as trans-eQTL analysis and differential expression analysis, but bears huge multiple testing burden;
  
• Sparse factor analysis which takes advantage of "gene modules", but subsequent analyses are necessary to interpret the biological meaning of factors.

Our approach to detect the effect of genetic perturbation:

• Identify genetically controlled factors that are correlated with the perturbation in a joint statistical framework.

We developed GSFA (Guided Sparse Factor Analysis), a factor analysis model that can infer unobserved intermediate factors given observed gene expression levels, with the advantage of inferred factors being sparse and their correlation with given sample-level conditions (e.g. genotype, CRISPR perturbation).

GSFA Model

Given a matrix \(Y \in \mathbb{R}^{N \times P}\) that holds the normalizd expression levels of \(P\) genes in \(N\) samples, and a guide matrix \(G \in \mathbb{R}^{N \times M}\) that holds \(M\) types of sample-level conditions:

\(Y = ZW^T+E\), where \(Z \in \mathbb{R}^{N \times K}\), \(W \in \mathbb{R}^{P \times K}\), \(E_{ij} \sim N(0,\psi_j)\),

\(Z = G \beta + \Phi\), where \(\beta \in \mathbb{R}^{M \times K}\), \(\Phi_{ik} \overset{i.i.d.}{\sim} N(0,1)\).

Both \(W\) and \(\beta\) have spike-and-slab priors.

Gibbs sampling is used to infer the model parameters from data.

Simulations

Single guide, multiple factors

• \(\beta\) = 0.1 to 0.6

Multiple guides, multiple factors

• Count data, \(\beta\) = 0.1 to 0.6, N=400, P=500;
• \(\beta\) = 0.1 to 0.6, N=400, P=500;
• \(\beta\) = 0.2 to 1, N=400, P=500;
• \(\beta\) = 0.1 to 0.6, N=4000, P=6000;

Applications

We applied GSFA to several published data sets of large-scale gene expression data with sample-level perturbations.

LUHMES CROP-seq Study

Source and Reference

High-throughput single-cell functional elucidation of neurodevelopmental disease-associated genes reveals convergent mechanisms altering neuronal differentiation, GEO accession: GSE142078.

Cells

Lund human mesencephalic (LUHMES) neural progenitor cells. (Cells were sequenced in 3 batches.)

Perturbations

CRISPR knock-down of 14 autism spectrum disorder (ASD)–associated genes (3 gRNAs per gene) + 5 non-targeting gRNAs.

Analyses

• Data overview;
• Pseudotime trajectory analysis;
• GSFA, normal-mixture prior;
• GSFA, normal-mixture prior, cells analyzed by early and late stages;
• (Archived) GSFA, spike-and-slab prior, 4 confounding factors corrected;
• (Archived) Enrichment analysis on GSFA result (spike-and-slab);
• (Archived) Gene discovery using LFSR, and comparison with DGE (spike-and-slab);
• (Archived) Permutation results (spike-and-slab);
• MUSIC topic modeling result

Primary Human T Cell CROP-seq Study

Source and Reference

Genome-wide CRISPR Screens in Primary Human T Cells Reveal Key Regulators of Immune Function, GEO accession: GSE119450.

Cells

Primary human CD8+ T cells from two healthy donors, with T cell receptor (TCR) stimulation.

Perturbations

CRISPR knock-out of 20 genes (2 gRNAs per gene) + 8 non-targeting gRNAs. Target genes were either found to regulate T cell responses in the genome-wide screens, or known checkpoint genes.

Analyses

• All T cells data overview, clustering analysis;
  GSFA, normal-mixture prior, donor not corrected, all cells analyzed by group;
  Permutation results for DGE methods;
  (Archived) GSFA, normal-mixture prior, donor corrected, all cells analyzed by group;
  
• Stimulated T cells data overview;
  Stimulated T cells pooled over 2 donors, batch effect corrected
  (Archived) GSFA, normal-mixture prior;
  (Archived) GSFA, spike-and-slab prior;
  (Archived) Enrichment analysis on GSFA result (spike-and-slab);
  
• Un-stimulated T cells data overview;
  Unstimulated T cells pooled over 2 donors, batch effect corrected
  (Archived) GSFA, normal-mixture prior;
  (Archived) GSFA, spike-and-slab prior;

MCF10A CROP-seq Study

Source and Reference

On the design of CRISPR-based single cell molecular screens, GEO accession: GSE108699.

Cells

MCF10A cells (normal human breast epithelial cells) with exposure to a DNA damaging agent, doxorubicin;

Perturbations

CRISPR knock-out of 29 tumor-suppressor genes (TP53, ...), 1 non-targeting control;
guide RNA readout measured on the single-cell level.

Analyses

• Data overview;
• GSFA, NTC condition regressed out, singleton cells only;
• GSFA, NTC condition regressed out;
• Permutation results for detection 10%, singleton;
• Gene discovery using LFSR, and comparison with DGE.

TCGA BRCA Somatic Mutation Study

References

Perspective on Oncogenic Processes at the End of the Beginning of Cancer Genomics.

Data sources

FireBrowse TCGA BRCA Archives.

• mRNA-seq file "illuminahiseq_rnaseqv2-RSEM_genes_normalized";
  
• Mutation annotation file "Mutation_Packager_Oncotated_Calls";
  
• Clinical file "Clinical_Pick_Tier1".

Samples

TCGA breast invasive carcinoma (BRCA) tumor samples (confined to female, Caucasian, and with somatic mutation annotation).

Perturbation

Somatic mutation status of selected frequently mutated driver genes (PIK3CA, TP53, TTN, GATA3, CDH1, MAP3K1, MAP2K4).

Analyses

• Data overview;
• Newest GSFA, female Caucasian only, missense + LOF mutations, TMM normalized;
• GSFA, female Caucasian only, missense + LOF mutations, TMM normalized;
• GSFA, female Caucasian only, missense + LOF mutations, RSEM normalized;
• GSFA, female Caucasian only, merged mutations;
• GSFA, female Caucasian only, merged mutations, corrected for subtype;