Comparison with LineageOT about moscot HOT 6 CLOSED

The notebook can be found here: https://github.com/theislab/moscot/blob/feature/lineageot-comparison/notebooks/MK_2021-09-17_lineageot.ipynb

from moscot.

Have checked also bifurcating, same story. Will convert this notebook to a more easy-to-run comparison and try FGW as well.

from moscot.

Hi Mike, thanks! I wouldn't expect GW to perform well here, the idea would really be to try and use FGW. There may be a few parameters to tune here, e.g.

• weight alpha between GW (barcodes) and W (gene expression distance) terms
• distance metric used for the W term, e.g. euclidean distance in PCA space, shortest path distance based on KNN graph (see NovoSpaRc), euclidian distance in scVI latent space
• distance metric used for the GW term, e.g. hamming distance, modified hamming distance, check out whatever Cassiopeia uses. Note that Cassiopeia was developed for a different computational task, lineage tree building based on barcodes. However, for that task, you also need to define a barcode distance. Alternatively, rather than working with barcoding distances directly, we could use Cassiopeia to build a lineage tree first and use distances along tree edges for the GW term. This could potentially get rid of some noise in the barcodes, at the cost of working with processed data without uncertainty quantification.
• regularization strength epsilon

So this looks like a grid search to me. However, we can automate this since we can quantify our performance.

from moscot.

Even better for tree fitting: http://www.microbesonline.org/fasttree/

from moscot.

Problems/notes

• Scaling for LinegeOT: their implementation is really slow. We can start by keeping the setup as it is, don't need to re-run lineageOT. If that doens't work (low number of genes, meaningless distance), have to re-run.
• Normalizing distance matrices so that GW and W terms are comparable

from moscot.

For now, I consider this done.

from moscot.