Validation-of-CAF-Subtypes

Team Leader: Jenny Chien
Tech Leader: Canping Chen
Writer: Alex Xu
Team: Alice Browne, Brian Mott, Filippo Pederzoli, Jing Zhu, Nathan Wong

What are CAFs

Cancer-associated fibroblasts are a diverse set of stromal cells populating the tumor microenvironment (TME), easily recognized by a spindle-like nuclear and membrane structure.

They are widely responsible for shaping the extracellular matrix and the physico-chemical properties of the tumor, and executing pro- and anti-tumor functions by secreting proteases, cytokines, and growth factors that promote tumor growth, suppress immunity, and more. CAFs receive a fraction of the attention given to immune cells, where heterogeneity is resolved by lymphoid/myeloid lineage and further by immune cell subtypes and exhaustion state. In contrast, CAFs have been defined as a monolith, expressing markers such as alpha-SMA or vimentin. More recently, CAF heterogeneity of clinical significance has been discovered, via signaling such as podoplanin linked to poor outcomes. In 2023, deep single cell sequencing of fibroblast populations defined 9 subtypes, including matrix CAFs, inflammatory CAFs, vascular CAFs, and others.

We hypothesize that mapping CAF heterogeneity onto single cell and spatial omics data will better resolve patient and sample heterogeneity stemming from CAFs. Further, by defining CAF subtypes, we can propose distinctive markers of CAF subtypes and propagate CAF subtypes to other forms of spatial data that lack the molecular resolution of scRNAseq, allowing CAF subtypes to be defined in datasets such as spatial protein (e.g. IMC, CODEX, etc) and probe-based spatial RNA (e.g. MERFISH, CosMx) assays.

Objective

We will develop computational methods to annotate CAF subtypes across a range of spatial omic assays available in HTAN. If successful, users will be able to:

1. Train and validate a prediction model to identify CAF subtypes
2. Identify CAF subtypes in scRNAseq datasets
3. Apply CAF subtyping to spatial transcriptomic data
4. Identify CAF subtype signatures in spatial protein and imaging data
5. Apply CAF subtyping in other spatial assays
6. Confirm the location of CAF subtypes by comparing to imaging data

Requirements

💥INSERT PACKAGES AND VERSION NUMBERS HERE

• Python
  • scanpy - 1.10.1
  • anndata - 0.10.7
• R
  • Seurat - 5.0.2
  • anndata - 0.7.5.6
  • reticulate - 1.39.0
  • ggplot2 - 3.5.1
  • data.table - 1.16.2

Data

CAF Subtype
HTAN Multi-omic
HTAN Multi-omic CellxGene

Pipeline

💥INCLUDING PACKAGE NAMES AND SPECIFIC PIPELINES/TECHNIQUES USED

1. CellTypist: create model using CAF subtype reference data
   • Use CAF reference data
2. sklearn: train model via 10-fold validation
3. Preprocess data
   • Use multi-omic data
   • RCTD: attribute spatial data to CAF type
   • 💥WHAT ARE ALL THE PACKAGES WE USED AND WHEN
   • Identify CAF signatures in custom spatial probe data
4. Classify CAF subtypes using scripts
5. Xena viewer: validate CAF subtyping via multi-omics

Strategy

Preparation

To achieve our goal of generalized CAF subphenotyping of spatial omics data, we first generate a subtyping model with CellTypist using 10-fold validation of the CAF scRNAseq reference data set from Cords 2023, using SCRIPT.
Model performance plateaued at ~1000 genes.
The model differentiates CAFs, with example gene signatures below.
Application of the classifier to scRNAseq data showed no tissue-specific CAF subtype preference and an increase in expected antigen-presenting CAFs.
Next we apply sample preprocessing to spatial omic modalities to prepare for subtyping.
Input formats: scRNAseq, snRNAseq, SlideSeq, MERFISH, CODEX, MIBI
Data input/output can be performed packages like XYZ.
The primary constraint is the information depth and resolution of the spatial omic input data. We apply our strategy first to the data of Klughammer et al, a recent multi-omic study of breast cancer.

Oversampled Input Data

Sequencing-based spatial omics generates low density, high coverage molecular profiles that often lack true single cell resolution. We used RCTD to approximate single CAF data points and applied the CAF subtyping model.

SlideSeq Test

CAF subtyping of SlideSeq data 💥Show results of CAF subtyping (table, bar plot) 💥SHOW RESULTS PROJECtED ONTO SPATIAL COORDINATES
Using VisualizeCAFSubtypes.R, the spatial location of CAF subtypes can be visualized.
Validation is performed by projecting CAF subtyping results against corresponding H&E, visualized in Xena Browser.

Existing Issues

• 💥 WHERE DID WE GET STUCK? SMALL THINGS GO HERE, BIG PICTURE STUFF GOES IN THE FUTURE DIRECTIONS SECTION BELOW
• Rare cell types are lost during harmonization steps, and low cell numbers can result in math and processing errors

Understampled Input Data

CAF subtypes must be mapped onto the available probes Probe-based spatial omics use sparse probe sets that rarely coincide with markers of CAF subtypes. Preprocessing results in relatively well-defined cells via segmentation, but projecting CAF phenotypes onto the dataset may require a bespoke process depending on format.
We attempted to preprocess MERFISH, CODEX, and H&E images to correlate with transcriptomically-defined CAF subtypes.
MERFISH: ~300 genes, focused on tumor and immune genes of interest (10 gene final overlap with CAF subtype DEG XXXXX)
CODEX: 50 proteins, focused on tumor and immune markers (3 gene final overlap with CAF subtype DEG XXXXX)
H&E: 2 color, baso/acidophilic staining

Considerations for undersampled data preprocessing

1. Reference data depth

• Full scRNA CAF subtype
• CAF subtype DEGs
• CAF subtype protein

2. Target data format

• Spatial or single cell
• RNA or protein

3. Target data depth

• Full transcriptome
• Probe set
• Targeted proteins Construct a mapping between CAF signatures and available markers.

Tools for bridging spatial and single cell: RCTD, STdeconvolve
Tools for bridging full transcriptome to targeted probes: 1. Exact match 2. Shared pathways 3. Functional annotated relationships

MERFISH Test

We used the DEG table (Supplementary Table 3, CAF reference) to detect overlapping genes between the MERFISH panel and define MERFISH-specific CAF subtype signatures. Subtype signature scores were calculated as the mean expression of all genes in the signature for each cell scanpy gene_scores().

We projected MERFISH CAF subtype scores onto H&E images usimg htan_jamboree_merfish_binned.ipynb or merfish_not_binned.ipynb, finding concordance with SlideSeq spatial projections. * - denotes genes used in the MERFISH gene signature.

CODEX Test

Only 3 overlapping gene/protein pairs were found with CAF subtypes and the provided leiden clusters were poor, CAFs could not be easily identified.
CAF subtyping was unsuccessful. Authors provided 10-20 clusters on 50 proteins across all subtypes, some clusters were defined functionally (i.e. Ki67+), CAFs and immune cells were co-clustered. Remaining markers would not be able to resolve CAF subtypes either (only 3 exact matches, shared pathway analysis not performed).

Note: VIM+ clusters overlap with PTPRC(CD45)+ and other immune markers, or PECAM1(CD31)+ endothelial marker. KRT19 tumor marker is uniformly expressed. Additional preprocessing of CODEX phenotyping would be needed.

Existing Issues

• For understampled data, confidence in assigned CAF subtypes is lower, we can score CAF subtypes using MERFISH but we have not classified
• CAF subtypes are not morphologically distinct and cannot be distinguished by most protein panels, so assignments are difficult to validate with non-RNA data

Final results

• CAF subtype classifier
• CAF classifier validation
• SlideSeq preprocessing
• SlideSeq: Classifier applied
• SlideSeq: Images validated
• MERFISH preprocessing
• MERFISH: Score established
• MERFISH: CAFs classified
• CODEX preprocessing
• Image overlay protocol

Next Steps

• Convert MERFISH CAF subtype gene signature scores to a classifier
• Re-analyze CODEX to obtain CAF populations or request from authors

Future Directions

Generalizability of CAF reference data: Are 9 CAF subtypes specific to breast cancer or are they identifiable in other tissues?
True multi-omic vs. adjacent slides: Can we expect alignment between slides, to validate CAF subtyping between modalities in the "same" tissue?
Scope of CAF subtyping: Can we subtype at broad and coarse scales simultaneously, or do we need to identify fibroblasts first before subtyping separately?

References

CAF Subtypes Training Data

EXPLAIN RESULTS OF OUTPUTS, WHERE IT CAME FROM, BE EXPLICIT LIST PLANNED FEATURES AND FUTURE IDEAS DEFINE YOUR JARGON