,
9 tweets,
4 min read
Read on Twitter
Our preprint on gene regulatory network inference from single-cell RNA-seq is up doi.org/10.1101/534834
The SCINGE method uses ensembles of Granger Causality models to estimate regulator-gene interactions from single-cell data with pseudotimes. @atulpdeshpande @ron_stewart2 1/9
The SCINGE method uses ensembles of Granger Causality models to estimate regulator-gene interactions from single-cell data with pseudotimes. @atulpdeshpande @ron_stewart2 1/9
Granger Causality works well for long time series, which we rarely have with bulk gene expression data. Single-cell data is appealing because we have many samples making a long (pseudo) time series. However, the samples are irregularly-spaced in pseudotime. 2/9
We adopt the Generalized Lasso Granger test to smooth the irregular pseudo time series data with a kernel function. This can also smooth missing values in the single-cell data. 3/9 
To our surprise, in some cases using the pseudotime values actually hurt the performance of our method and others that consider pseudotimes. It was better to use the cell ordering from the pseudotimes but ignore the pseudotime values. 4/9
This finding relates to studies from @coletrapnell (doi.org/10.1101/426981) and Allon Klein (doi.org/10.1073/pnas.1…) about limitations of pseudotemporal expression data for regulatory network inference. 5/9
We used precision-recall curves to evaluate the network inference methods - ROC curves are inappropriate for imbalanced datasets - but caution that even PR curves can obscure important aspects of an evaluation. 6/9
Sometimes the PR curve looked pretty good (better than random), but the average precision for most individual regulators was close to random. Ranking the regulators with more edges in the gold standard above those with fewer edges gave a fairly good PR curve overall. 7/9
A huge thanks to @CHTC_UW who helped us access the @opensciencegrid to develop and test SCINGE, providing millions of CPU hours. 8/9
Our SCINGE software is on GitHub github.com/gitter-lab/SCI… and archived on Zenodo doi.org/10.5281/zenodo…. Yes, this version is in MATLAB. We're happy to explore porting it to your favorite language if there is strong demand for the method. 9/9
