# Attach the DeLorean data frame to access members

Fixed hyperparameters

The standard deviation of cell pseudotimes around the cell capture times has been set as $$ \sigma_\tau = r opts$sigma.tau $$ The length scale of the gene expression profiles has been set as $$ l = r hyper$l $$

Variance decomposition into temporal variation and noise

We examine the variance both between- and within-time points. We expect some of the variation within a time point to be associated with noise in the temporal dimension, that is estimating the pseudotime correctly should reduce the within-time variance. The rest of the variation within a time point will be due to noise. Conversely we expect most of the variation between time points to be due to temporal variation although some will be due to noise.

Let $$ k_c \in {\kappa_1, \dots, \kappa_T} $$ be the time point at which cell $c$ was captured. We partition the cells by their captured time points indexed by $1 \le t \le T$: $$ \mathcal{K}t = {c: k_c = \kappa_t} $$ We group the expression measurements by gene and observed capture time to calculate means and variances: $$ \begin{align} \mathbb{M}{g,t} &= \text{Mean}{c \in \mathcal{K}_t}{x{g,c}} \\ \mathbb{V}{g,t} &= \text{Var}{c \in \mathcal{K}t}{x{g,c}} \end{align} $$

Noise estimates

We estimate the gene-specific noise levels by assuming that all the within-time variation in the data is due to noise. This should be a slight overestimate as some variation will be due to noise in the pseudotemporal dimension. $$ \hat{\omega}g = \textrm{Mean}_t{\mathbb{V}{g,t}} $$ Giving us this fit for our empirical Bayes prior on the $\log \hat{\omega}_g$

(ggplot(gene.var, aes(x=log(omega.hat))) + geom_density() + geom_rug()
    + stat_function(fun=function(x) dnorm(x,
                    colour="blue", alpha=.7, linetype="dashed")

Temporal variation estimates

We estimate the temporal variation by calculating the expected variance of samples at the capture times from the expression profile of a gene $g$. These depend on our fixed length scale and unknown temporal variance $\psi_g$. The covariance of these samples will be $\psi_g \hat{\Sigma}$ where $$ \hat{\Sigma}{t_1,t_2} = \Sigma\tau(\kappa_{t_1}, \kappa_{t_2}) $$ The expected variance of our samples is $\psi_g V_{\hat{\Sigma}}$ where $$ V_{\hat{\Sigma}} = \textrm{Mean}{\textrm{Diag}(\hat{\Sigma})} - \textrm{Mean}{\hat{\Sigma}} $$ and we can slightly overestimate the temporal variances $\psi_g$ by ignoring the noise in our data $$ \hat{\psi}g = \frac{\textrm{Var}_t{\mathbb{M}{g,t}}}{V_{\hat{\Sigma}}} $$ Giving us this fit for our empirical Bayes prior on the temporal variation $\log \hat{\psi}_g$

(ggplot(gene.var, aes(x=log(psi.hat))) + geom_density() + geom_rug()
    + stat_function(fun=function(x) dnorm(x,
                    colour="blue", alpha=.7, linetype="dashed")

Correlation between temporal variation and noise

The correlation between the temporal variation and noise estimates is r with(gene.var, round(cor(psi.hat, omega.hat), digits=2)). Plotting the estimated noise levels against temporal variation gives:

    ggplot(gene.var, aes(x=log10(psi.hat), y=log10(omega.hat)))
    + geom_point()
    + geom_abline(intercept=0, slope=1, linetype="dashed", alpha=.7)

Cell sizes relation to gene variation

How big are the cell size estimates relative to the average variation in the gene expression? The line represents the square root of the mean variance of the gene expressions.

gene.vars <- apply(expr, 1, var)
mean.gene.variation <- sqrt(mean(gene.vars))
ggplot(cell.sizes, aes(x=S.hat)) +
    geom_histogram() +
# Detach the previously attached DeLorean data frame

Try the DeLorean package in your browser

Any scripts or data that you put into this service are public.

DeLorean documentation built on May 2, 2019, 9:24 a.m.