Discussion: multiple tables in a SpatialData object

[kevinyamauchi]

This issue is to discuss the possibility of allowing multiple tables per SpatialData object. The implementation may be a list of tables, a dictionary of tables, or a MuData object containing multiple tables. Below are the notes from the Oct 28, 2022 hackathon in Munich. cc @giovp , @LucaMarconato , @ivirshup .

Single-table

Pro

• Intuitive compounds statements for queries on tables

idx = (sdata.tables['expression'].obs.cell_type == "Neuron") & (sdata.obs.n_counts > 100)
sdata = sdata[idx, :]

Cons

• We 100% need a SpatialDataContainer

sc = SpatialContainer('...')
sdata = sc.to_spatialdata(elements=['/images/...', '...'])

everyting = SpatialData('...')
sdata = everyhing.query.elements(elements=['/images/...', '...'])
subsdata = sdata.query.elements(elements=['/images/...'])

• Points can't be annotated with a table -> For Points (but not for Circles), you can store expression inside the Points.

Multi-table

Pro

• No need for SpatialDataContainer
• Points and Circles are unified, because both can be annotated by a Table

sdata0 = sdata.query.coordinate_system('coordinate_system_name', filter_rows=False)
sdata1 = sdata.query.bounding_box()
sdata1 = sdata.query.polygon('/polygons/annotations')
sdata1 = sdata.query.table(table_key=..., attr: ["obs", "var"], query="""""")

see https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.query.html

import squidpy as sq

adata = sq.datasets.mibitof()
df = adata.obs
out1 = df.query("(point == 23) & (cell_size > 200)")
out2 = df.loc[(df.point == 23) & (df.cell_size > 200)]
pd.testing.assert_frame_equal(out1, out2)

Cons

• Complex selection of multiple tables

idx = (sdata.table.obs.cell_type == "Neuron") & (sdata.table.obs.n_counts > 100)
sdata = sdata.query.table(key='expression', index=idx)

class SpatialData
    @property
    def table(self) -> AnnData:
        if len(self.tables) == 1:
            k = list(self.tables.keys())[0]
            return self.tables[k]
        else:
            raise ValueError("...")

('A > B' & 'cell_type == "Neuron"')