Skip to content

Module : DGE with SCDE

Jump to bottom
GBrelurut edited this page Jan 5, 2018 · 4 revisions

Module : DGE with SCDE

This module calculate for each gene the probability of differential expression.

  • Internal name : scde

  • Avalaible : local mode

  • Input Ports :

    • matrix : filtered count matrix (tsv)
    • cells : filtered cells metadata (tsv)
    • genes : genes metadata (tsv)

  • Output Ports :

    • dgeoutput : differential expression result (tsv)

  • Optional parameters :

    • Main Parameters
Parameter Type Description Default Value
n.cores int Number of cores to use 1
model.group.col string Name of the column if cells must be grouped for model fitting (NULL if no grouping is necessary) NULL
prior.length int Number of points for prior calculation 400
batch.col string Name of the column indicating batches, if batch correction is required (else NULL) NULL
n.randomizations int Number of randomization for testing 150
  • Parameters for model fitting
Parameter Type Description Default Value
min.observation int Minimal number of observations for a gene to be used for model fitting 3
min.genes int Minimum number of genes for model fitting 2000
threshold.segmentation boolean Use or not threshold segmentation to accelerate failure estimation TRUE
failure.threshold int Number of reads indicating a gene failed amplification 4
max.pairs int Maximum number of comparisons that should be performed per group for estimation of dropout rate 5000
min.pairs int Minimum number of comparisons that should be performed per group for estimation of dropout rate 10
poisson.param float Parameter of the Poisson distribution used to model failures 0.1
linear.fit boolean Weither to use linear fit for model fitting (highly recommanded) TRUE
min.theta float Minimum for the dispersion parameter of the negative binomial 0.01
max.theta float Maximum for the dispersion parameter of the negative binomial 100
  • Parameters for prior calculation
Parameter Type Description Default Value
save.prior.plot boolean Weither to save or not prior plot TRUE
pseudocount int Pseudocount to add to observation before log transforming them 1
quantile float Quantile used to set maximum expression value 0,999
max.value float Alternative to quantile, maximum expression value NULL
  • Parameters for test
Parameter Type Description Default Value
return.posteriors boolean Weither to return or not the posteriors TRUE
  • Configuration example
<step id="DGE" skip="false">
	<module>scde</module>
	<parameters>
		<parameter>
			<name>prior.length</name>
			<value>400</value>	
		</parameter>
		<parameter>
			<name>n.randomizations</name>
			<value>200</value>	
		</parameter>
		<parameter>
			<name>n.cores</name>
			<value>12</value>	
		</parameter>
	  </parameters>
</step>

Interpreting output files

Introduction to scde

Evaluation of model fitting and misfit removal

In order to evaluate goodness of fit of the model, the module calculate the amount of variance from the measured values explained by the model (i.e. r-squared of a linear model where predictive value is the measured value and predicted value is the model value). The model is plotted on a goodness of fit plot :

FittingPlot

Considering that model would be fitted for the majority of cells, we expect the distribution of this values to be "nearly Gaussian" :

densityFiltered

Misfits are expected to be outliers showing lesser values (see distribution plot below), thus cells showing lesser value are removed one by one, until Shapiro's test returns sufficient probability under Gaussian assumption.

densityAll

Shapiro's P-values show an increase of several order of magnitudes after some removals (see p-value as a function of number of removals below). This increase indicates "nearly Gaussian" distribution.

probaPlot

The module also plot goodness of fit as a function of Michaelis Menten model maximum, before and after outliers depletion. This allow for visual inspection of the process.

qualityAll qualityFiltered

Scatter Plot

After cleaning data, the module produces two scatter plot, showing all cells in term of number of feature (y-axis) and number of reads (x-axis).

Raw_Cellplot

The first one, show all cells, the ones in red are those being eliminated.

Filtered_cellplot

The second one shows cells remaining after filtering. At the end of the filtering, cells should behave like a mixture of gaussian, i.e. you can wrap them in a given number of ellipses.

Clone this wiki locally