Trimodal benchmarks

[alitinet]

Run

• [ ] stabMAP
• [ ] multiMAP

and

• [ ] plot UMAPs with results for both
• [ ] run scIB metrics for both

on Neurips 2021 data.

[alitinet]

Also see /lustre/groups/ml01/workspace/fabiola.curion/multigrade/ for stabmap example.

[alitinet]

data is in /lustre/groups/ml01/projects/2022_multigrate_anastasia.litinetskaya/trimodal/data/

[szalata]

As far as I see, multiMAP requires shared features between modalities that are used for creating shared latent manifolds. From the supplementary methods https://static-content.springer.com/esm/art%3A10.1186%2Fs13059-021-02565-y/MediaObjects/13059_2021_2565_MOESM2_ESM.pdf : "For each data point, MultiMAP finds a set of nearest data points in each modality. The distances to these nearest neighbors are converted to a geodesic distances on a shared latent manifold by normalizing with respect to a radius value"

Similarly, note in the documentation https://multimap.readthedocs.io/en/latest/ : "The .var spaces will be intersected across subsets of the objects to compute shared PCAs, so make sure that you have ample features in common between the objects. .X data will be used for computation."

[szalata]

StabMap is more robust because it doesn't require overlap in features between all of the datasets, but it still requires an overlap in features, which we don't have for surface proteins, afaik