Compute UMAP from pre-existing distance matrix

[ivirshup]

It should be easier to compute a UMAP layout from a pre-existing distance matrix. Currently, we only really support this for pre-existing connectivity matrices.

The missing step here is computing the umap weighted graph from a distance matrix.

The functionality for this largely already exists in the sc.neighbors module in private functions (example: #2139 (comment)). We would just need an API.

Previously discussed in: #1561

[ivirshup]

@flying-sheep did this get implemented by your recent work with the neighbors function?

[flying-sheep]

Not conveniently, no:

@dataclass
class DummyTranformer:
    distances: sparse.csr_matrix

    def get_params(self):
        return dict(n_neighbors=self.distances.getnnz(1)[0])

    def fit_transform(self, _):
        return self.distances

pbmc3k = sc.datasets.pbmc3k()
sc.pp.neighbors(
    pbmc3k,
    use_rep="X",  # prevent it from doing a PCA
    transformer=DummyTranformer(random_dist_mat(pbmc3k.n_obs)),
)

and this still computes connectivites.

We’re much closer though!

[Reovirus]

@flying-sheep
Hello. Should we implement separate method for connectivity matrix? Now connectivity matrix is strongly computed in umap AND in gauss methods. Compute it in gauss only will be inconsistent (I can be assignee)

[flying-sheep]

Hm, maybe the best approach would be to rename the method parameter to connectivities, and allow specifying an array/matrix instead of a string for it.

Then if an array/matrix is passed, only its dimensions are validated, but no connectivities are computed.