Home

Summary and overview

• Data description
• Seminar talk 2018-04-16
• Some code for generating plots for the paper
  • code/paper.plot.images.R

Model training and prediction results

• Early analysis using trendfilter and evaluating model performance

• Investigate the property and quality of the cell time labels derived from GFP and RFP: PVE of intensities by cell times, comparing PVEs by cell times from DAPI and FUCCI vs from FUCCI; comparing prediction before and after removing PC outliers (after is slightly better), and in different sets of genes.

• Investigate the property of Seurat classes and cell time

• Noisy label analysis
  • PVE in DAPI, GFP, and RFP before vs after removing noisy labels
  • Comparing results excluding/including noisy training labels on top 10 and top 101 cyclical genes, when excluding noisy labels from each fold validation, hence unequal sample numbers per fold as opposed to excluding noisy labels before making fold valdiation, hence equal nubmer per fold
• Analyze the training datasets
  • Comparing trendfilter order 2 and order 3 with npcirc.nw on the odd-numbered genes in the top 101 cyclical genes
  • Comparing resuls of the various classifers (supervised, PC, unsupervised) on top 101 cyclical genes, and on top 10 cyclial genes.
• Analyze the withheld samples
  • Evaluate the withheld sample including noisy labels on the top 10 and top 101 cyclical genes] and comparing peco methods with seurat predictions.
  • Evaluate the withheld samples removing noisy labels on the top 10 and top 101 cyclical genes and comparing peco methods with seurat predictions.
• Finalizing training for predicting cell times for cells from individuals
  • Compute and compare cell time estimates derived under two different assumptions: assuming equal variance in PC1 and PC2 (circle) vs. unequal variances in PC1 and PC2.
  • Results assuming equal variance, considering when
    • Cell times derived from fucci only: cyclical genes selected, and gene annotations
    • Cell times derived from fucci and dapi: cyclical genes selected
    • Comparisons: Prediction error smaller when based on fucci time only than based on fucci and dapi time, though only a very small difference. However, the genes selected in these two are similar, both found to have 37 genes enriched for the Cell Cycle GO (0007049).
• Finalizing training for predicting cell times for one individual
  • Results assuming equal variance, considering when
    • Cell times derived from fucci only: cyclical genes selected, and gene annotations

Approaches to fitting cyclical trend in gene expression data

• Apply spherically projected multivariate linear model
  • First trying out spml package
  • To predictd cell times using traing/test samples set-up

• Learning about correlation for circular response variable: some simulations to check its property and also relations with Fisher’s z transformation

• CellcycleR 0.1.6
  • Model convergence assessment
  • Fitting on intensities across plates and individuals
  • Fitting on Leng data)
  • Fitting on fucci-seq RNA-seq data)

Cell cycle signal in gene expression data

1. We investigated cell cycle signals in the sequencing data alone.
   • Consider transgene count in sequencing data
   • Compute linear correlation between gene expression levels with DAPI and FUCCI intensities
2. We then assign categorical labels of cell cycle and explored the expresson profiles of these categories.
   • Cluster samples by Partition around medoids(PAM)
   • Cluster samples by Guassian mixture modeling
   • Select a subset of samples that are closet to the cluster centers (cluster representives) using silhouette index
   • Examine gene pression scores defined by the Macosko paper in the selected cluster representatives before confounding correction and after confounding correction
   • Examine gene expression scores defined by the Macosk paer in the sorted cells of the Leng et al. 2015 paper
3. We ordered cells on a circle using FUCCI intensities alone.
   • First, I used GFP and RFP intensities to estimate a least-square fit of unit circle which approximate the relative ordering of cells on cell cycle
   • I computed the PCs of GFP and RFP and used these to infer the relative ordering on a circle (i.e., polar coordinates), and to evaluate the circle fit, I computed circular-circular and circular-linear correlation with DAPI and gene expression
   • Here I put together lists of genes identified as significantly correlatd with the PC-based fit by linear correlation and circular-linear correlation, also cyclical genes by smash
4. I used nonparametric methods to identify genes that may be cyclical along cell cycle phases.
   • Fit smash and kernel regression on circular variables on a subset of genes with detection rate > .8.
   • Fit trendfilter on a subset of genes (5) that are observed (visually) to have cyclical pattern. trendfilter is robust to small proportion of undetected cells, approx 2 or 3%. In cases of simulation when increasing proportion of undetected cells to 20%, we observed a flat line in gene expression for genes previously identified to tend to a cyclical pattern.
   • Next, we fit trendfilter on all genes after transforming the data to follow standard normal distribution, permutation-based p-values for PVE are used to select 101 significant cyclical genes.

RNA-seq data preprcessing

1. The first step in preprocessing RNA-seq data consists of QC and filtering.
   • Sample QC and filtering
     • Sample QC criteria
       
     • Sequencing depth
       
     • Reads versus molecules
   • Gene QC and filtering
     • gene filtering
     • PCA with technical fators
2. We then analyzed and corrected for batch effect due to C1 plate in the sequencing data
   • Estimate variance explained in IBD and correct for batch effects

Microscopy image analysis

• Processing images - from images to intensities

We evaluated and pre-processed the results of image analysis as follows:

1. We visually inspect images deteced to have none or more than one nucleus. For cases that are inconsistent with visual inspection, we correct the number of nuclei detected.
   • Inspect images with multiple nuclei
   • Inspect images with no nucleus
2. We applied background correction to the intensity measurements of GFP, RFP and DAPI based on the following analyses.
   • CONFESS results
   • QC analysis including no. nuclei detected, DAPI, and intensity variation
   • Explore using log10 sum pixel intensity for signal metrics
   • Compare correction approaches using median versus mean background
   • Explore associations between nucleus shape metrics vs intensities
3. We analyzed intensity variation across individuals and batches and considers approaches for removing batch effects in the data.
   • Visualize signal variation by plate and individual identity
   • Visualize the structure of signal variation by individual identity
   • Quantile normalization for GFP, RFP and DAPI
   • Estimate variance explained in IBD and correct for batch effects in intensities
4. We investigated the cell time estimates based on FUCCI intensities.
   • Consider the mean-adjusted FUCCI intensities, what’s the relationship between cell time estimates and DAPI and sample molecule count?
   • Consider the quantile-normalized FUCCI intensities versus the mean-adjusted intensities (for adjusting batch effect). What’s the differences between cell time estimates from the two normalization approaches?

One-time investigations

• Why some gene symbols (genes) correspond to multiple Ensembl IDs?

• I selected a set of cell cycle genes that belong to GO term Cell Cycle and looked at the overlap with the detected genes in our data.

• Replicate Seurat example of scoring cell cycle phases