A common design has two grouping variables at once, for example cell type and tissue region, and the question is which interactions change per condition within a region. Two complementary strategies:
- Cross the labels: combine cell type and region into one label
(e.g.
Immune@Mucosa) so the sender/receiver axis carries region identity through the whole pipeline. - Subset: restrict to one region and run
compareLARISwithin it.
This tutorial shows strategy 1 on the MERFISH gut atlas ileum
(SPF vs germ-free, cell_class x anatomical_layer), using PIASO's
getCrossCategories to build the crossed labels. All outputs were
produced by exactly this code with LARIS v0.12.0.
Data: the same gut atlas as
tutorial 03
(Dryad 10.5061/dryad.p5hqbzm0z);
anatomical_layer holds Mucosa, Submucosa, Muscularis Propria,
Myenteric Plexus and GALT.
import pandas as pd
import laris as la
import piaso
for s in ileum_slices: # one slice per mouse here
sub = load_slice(s) # raw counts + X_spatial, as in tutorial 03
sub.obs["class_layer"] = piaso.pp.getCrossCategories(
sub.obs, "cell_class", "anatomical_layer")
sub.obs["class_layer"] = pd.Categorical(sub.obs["class_layer"])
lr_data = la.tl.prepareLRInteraction(sub, lr_df,
use_rep_spatial="X_spatial")
_, res = la.tl.runLARIS(lr_data, data=sub,
use_rep="X_spatial",
use_rep_spatial="X_spatial",
groupby="class_layer",
calculate_pvalues=False,
specificity_reference="all")
results[s] = res20211027_WT_ile_slice_2: 30,307 cells, 30 crossed groups
...
Crossing multiplies the group count (here about 30 = 7 cell classes x 5
layers, minus empty combinations), so the triple table grows
accordingly. Keep calculate_pvalues=False; the comparison below does
the inference.
lr_cmp, triple_cmp = la.tl.compareLARIS(
results, conditionMap=cond, referenceCondition="WT",
sampleToSubject=mouse_of)L1 tested 129; triples tested 5983
The triple table now answers region-resolved questions directly:
triple_cmp[triple_cmp.pvalue.notna()].nsmallest(8, "pvalue") sender receiver interaction_name log_diff pvalue pvalue_fdr
ENS@Muscularis Propria Immune@Muscularis Propria Tgfb2::Tgfbr2 8.6358 0.0001 0.0167
Endothelial@Muscularis Propria Smooth Muscle@Myenteric Plexus Jag1::Notch2 6.5534 0.0003 0.0326
Endothelial@Muscularis Propria Smooth Muscle@Myenteric Plexus Igf2::Igf1r 6.4763 0.0003 0.0252
Fibroblast@Myenteric Plexus Smooth Muscle@Submucosa Wnt2b::Fzd7 -7.5059 0.0004 0.0514
Endothelial@Muscularis Propria Endothelial@Muscularis Propria Jag2::Notch1 6.8688 0.0011 0.0264
ENS@Muscularis Propria Interstitial@Muscularis Propria Tgfb2::Tgfbr2 7.4664 0.0013 0.1046
Fibroblast@Muscularis Propria Fibroblast@Myenteric Plexus Wnt9a::Lrp6 5.7122 0.0014 0.1378
Interstitial@Submucosa Smooth Muscle@Mucosa Tgfb3::Tgfbr2 7.3985 0.0017 0.2393
The leading calls localize the germ-free changes anatomically: enteric nervous system to immune Tgfb2 signalling specifically in the muscularis, and endothelial Notch/Igf2 signalling at the myenteric plexus. A plain cell-class analysis reports Tgfb up somewhere; the crossed labels say where.
la.pl.plotCompareLARIS(triple_cmp, effect_col="log_diff",
label_col="label", # "sender -> receiver\nLR pair"
condition_labels=("SPF", "germ-free"),
title="Crossed cell type x layer triples")
Because every point is now a (cell type @ layer) pair, the volcano reads anatomically. Aggregating the calls by layer shows where the germ-free changes concentrate; the rate is shown rather than raw counts because layer pairs contribute very different numbers of testable triples:
Muscularis Propria to itself (2.2%) and to the Myenteric Plexus (2.3%) stand out against Mucosa to Mucosa (0%). With one slice per mouse and roughly 6,000 triples only 4 clear FDR, so this panel uses nominal p < 0.01 as a descriptive rate; treat it as a map of where to look, and take calibrated per-LR calls from Level 1 or from tutorial 03.
The labels are worth plotting on the tissue too, next to the interaction they localize:
Reading one interaction across layers works the same way:
triple_cmp.query("interaction_name == 'H2-Eb1::Cd4'").nsmallest(6, "pvalue") sender receiver log_diff pvalue
Endothelial@Mucosa Immune@Mucosa -4.4212 0.0551
Epithelial@Mucosa Immune@Submucosa -2.8628 0.1000
Epithelial@Mucosa Immune@Mucosa -8.2260 0.1000
Fibroblast@Mucosa Immune@Submucosa 2.8365 0.1163
Fibroblast@Submucosa Immune@Submucosa -2.3928 0.1287
Epithelial@GALT Immune@GALT -4.8382 0.1997
The H2-Eb1::Cd4 loss concentrates in the mucosa and GALT, exactly the compartments where microbiota-driven antigen presentation lives. (With one slice per mouse and 30 groups these per-triple tests are underpowered; the trend is the point of this readout, the calibrated per-LR calls come from Level 1 or from tutorial 03's matched estimator.)
- PIASO's spatial tutorials cover the upstream steps (QC, clustering, annotation, regulon analysis) that produce the labels crossed here.
getCrossCategories(obs, col1, col2)is PIASO's helper; without PIASO the same labels are one line of pandas:obs["class_layer"] = obs["cell_class"].astype(str) + "@" + obs["anatomical_layer"].astype(str).- Crossing shrinks the cells per group. Groups below
n_cells_expressed_thresholdcontribute little; consider merging rare combinations first. - Strategy 2 (subset to one region, then
compareLARISwithin it) is preferable when only one region is of interest, since it avoids the group explosion entirely. - Deconvolved data: LARIS uses discrete labels, so with per-spot proportions use the dominant label per spot; where one cell type dominates everywhere, region or cluster labels usually carry more contrast.