Slide 1
Slide 1 text
Flexible GLMM for Testing Cell Colocalization
in Spatial Immunofluorescence Data
PSB Workshop
Siyuan Ma
2023/01/03
Slide 2
Slide 2 text
Multiplexed immunofluorescence (MxIF) data workflow
2
* Amitay et al., bioRxiv, 2022
Slide 3
Slide 3 text
• Do B-cells interact
(i.e., colocalize) with
CD8- T-cells, more
than expected by
chance?
• Do B-cells interact
with CD8- T-cells
differently in
patients that
survived vs. didn’t
survive at five years?
* VectraPolarisData, Wrobel, Ghosh, Bioconductor, 2022
* Steinhart et al., Mol Cancer Res, 2021 3
Slide 4
Slide 4 text
Significance of colocalization depends on the
neighborhood topology of the image
4
Cell A Cell B Others
Slide 5
Slide 5 text
Approach 1: permutation
5
* histoCAT, Schapiro et al., Nat. Methods, 2017
• Issues:
• Null distribution difficult to generalize across images
• How to compare between e.g. disease conditions?
Slide 6
Slide 6 text
6
B-cell/CD8- T-cell
interactions are
enriched but also
heterogeneous
across images
Slide 7
Slide 7 text
Approach 2: per-cell modeling
7
• 7 out of 16 neighbors are
CD8- T-cells.
• In the entire image,
~1.31% are CD8- T-cells.
• The difference quantifies
the enrichment of B-
cell/CD8- T-cell
interactions.
• Number of CD8- T-cell
neighbors can be
modelled via binomial
distribution.
Slide 8
Slide 8 text
Generalized linear mixed effects model for immune interactions
8
For sample 𝑖, B-cell 𝑗:
log
𝜋!"
1 − 𝜋!"
= log
*
𝜋!
1 − *
𝜋!
+ 𝛽 + 𝑒!"
• 𝜋!": proportion of CD8- T-cells interacting with the 𝑗-th B-cell in sample 𝑖
• *
𝜋!: overall proportion of CD8- T-cells in sample 𝑖 (included as offset)
• 𝛽: difference between the two, i.e., enrichment of B-cell/CD8- T-cell interactions
compared random chance.
• 𝑒!": spatial random effect. Cells close together maybe more closely correlated!
* Kondo, Ma, et al., Gastro, 2021
Slide 9
Slide 9 text
Model flexibly incorporates covariates
9
log
𝜋𝑖𝑗
1 − 𝜋𝑖𝑗
= log
*
𝜋𝑖
1 − *
𝜋𝑖
+ 𝛽 + 𝛽%&'()%*%+,
∗ vysurvival + 𝑒𝑖𝑗
Slide 10
Slide 10 text
Summary
10
• Rigorous parameter and test statistics (effect sizes, standard errors, p-
values) for immune interaction enrichment.
• Cell-specific modelling generalizes across images. Random effects
address spatial correlations.
• Flexibly incorporate covariates.
• Analytical p-values (no permutations).
• Available with the R package spaMM.