TME state characterization

library(CellTFusion)
#> 

Once latent factors have been extracted (see Cell groups and latent factors), each factor can be linked to known biological programs:

1. Hallmark GSEA — associate each factor with MSigDB Hallmark gene sets.
2. Meta-program mapping — match each factor to cancer type-specific reference programs derived from TCGA.
3. TME subtype annotation — relate factors to established TME subtypes.

These steps use:

• counts.norm — normalized expression matrix (genes x samples).
• latent_spaces — output of compute.latent_factors(); latent_spaces$Z is the samples x factors matrix.

Hallmark GSEA per latent factor

compute_factor_gsea() fits a multivariate limma model with the latent factor scores as covariates, ranks genes by their moderated t-statistic for each factor, and runs a pre-ranked GSEA with fgsea (Korotkevich et al. 2021) against the MSigDB Hallmark collection (Liberzon et al. 2015). The Hallmark collection summarizes about 50 well-defined biological processes (e.g. EMT, interferon response, hypoxia), a natural vocabulary to interpret data-driven factors. The samples of features_df must be in the same order as the columns of RNA.tpm.

gsea_results <- compute_factor_gsea(
  RNA.tpm     = counts.norm,
  features_df = latent_spaces$Z,
  plot_dot    = TRUE,
  top_n       = 10,
  file_name   = "Tutorial"
)

Element	Description
$DE_results	Named list of limma results, one per factor
$GSEA_results	Named list of fgsea results, one per factor

With plot_dot = TRUE, a dot plot of the top Hallmarks is saved per factor (dot size = significance, color = normalized enrichment score, NES):

Mapping factors to TCGA meta-programs

What are meta-programs?

A meta-program (MP) is a recurrent transcriptional program reflecting a cell state (e.g. cell cycle, hypoxia, epithelial-mesenchymal transition, interferon response) that recurs across tumors and cancer types. Gavish et al. (Gavish et al. 2023) characterized such programs from single-cell RNA-seq across many cancers. CellTFusion follows the same logic with bulk TCGA data: it derives cancer type-specific meta-programs from TCGA, so that factors found in a new study can be described with a shared vocabulary of TME states.

How map_factors_to_metaprograms() works

1. Reference (pre-built, per cancer type): CellTFusion was run on a TCGA cohort, and the Hallmarks were clustered by their NES profiles across the TCGA factors (derive_meta_programs()). Each resulting meta-program is a set of Hallmarks, annotated with a TME subtype (see below). References are shipped for "blca" (bladder cancer), "luad" (lung adenocarcinoma) and "skcm" (melanoma).
2. Study factors: the Hallmark NES of each study factor come from compute_factor_gsea().
3. Matching: for each study factor and each meta-program, the mean NES of the meta-program’s Hallmarks is computed. The meta-program with the highest positive mean NES is the factor’s best_MP.

In other words: which known program does this factor’s Hallmark signature resemble most?

mp_mapping <- map_factors_to_metaprograms(
  gsea_study  = gsea_results,
  cancer_type = "skcm",
  plot        = TRUE,
  file_name   = "Tutorial"
)

Element	Description
$factor_mapping	One row per study factor: factor, best_MP, best_score (its mean NES), all_scores (mean NES for every meta-program) and TME_subtype of the best meta-program
$reference	The meta-program reference: meta_program, its hallmarks and TME_subtype

A custom reference can be used with mp_file, either a data frame like $reference or the path to an .RData file containing an object named meta_programs.

With plot = TRUE, each panel shows a study factor: grey bars are the mean NES for every meta-program, and the best-matching one is colored by its TME subtype:

TME subtypes

map_factors_to_TME() uses the TCGA annotations of Bagaev et al. (Bagaev et al. 2021), who defined four pan-cancer TME subtypes (Molecular Functional Portraits, MFP):

• IE (immune-enriched, non-fibrotic) — high immune infiltration, low stromal signal.
• IE/F (immune-enriched, fibrotic) — high immune infiltration with a fibrotic/stromal signature.
• F (fibrotic) — stroma-dominated, low immune infiltration.
• D (depleted) — low immune infiltration and low fibrosis.

These subtypes are associated with response to immune checkpoint blockade. For a TCGA cancer type, map_factors_to_TME() matches the TCGA patients in Z to their MFP subtype and tests each factor across the four subtypes with a Kruskal-Wallis test. Factors with p < 0.05 are labeled with the subtype that has the highest median score; the others are labeled "uncharacterized". This is how the TME subtypes of the TCGA meta-program references were assigned (annotate_metaprograms_TME()).

tme_annotation <- map_factors_to_TME(
  cancer_name = "skcm",
  Z           = latent_spaces$Z,   # TCGA samples x factors
  plot        = TRUE,
  file_name   = "Tutorial"
)

Deriving meta-programs

To build your own reference, for example from several cohorts, derive_meta_programs() clusters the Hallmarks by their NES profiles across factors. The number of meta-programs k is chosen at the elbow of the within-cluster sum of squares if NULL.

meta_programs <- derive_meta_programs(
  gsea_results = gsea_results,
  k            = NULL,
  file_name    = "Tutorial",
  plot         = TRUE
)

Running everything with CellTFusion()

The GSEA and meta-program mapping are run by CellTFusion() when cancer_type is given:

res <- CellTFusion(raw.counts = raw.counts, cancer_type = "skcm")

res$TME_states              # factor-to-meta-program mapping
res$Metaprograms_reference  # meta-program reference used

References

Bagaev, Alexander, Nikita Kotlov, Krystle Nomie, et al. 2021. “Conserved Pan-Cancer Microenvironment Subtypes Predict Response to Immunotherapy.” Cancer Cell 39 (6): 845–865.e7. https://doi.org/10.1016/j.ccell.2021.04.014.

Gavish, Avishai, Michael Tyler, Alissa C. Greenwald, et al. 2023. “Hallmarks of Transcriptional Intratumour Heterogeneity Across a Thousand Tumours.” Nature 618 (7965): 598–606. https://doi.org/10.1038/s41586-023-06130-4.

Korotkevich, Gennady, Vladimir Sukhov, Nikolay Budin, Boris Shpak, Maxim N. Artyomov, and Alexey Sergushichev. 2021. “Fast Gene Set Enrichment Analysis.” bioRxiv, ahead of print. https://doi.org/10.1101/060012.

Liberzon, Arthur, Chet Birger, Helga Thorvaldsdóttir, Mahmoud Ghandi, Jill P. Mesirov, and Pablo Tamayo. 2015. “The Molecular Signatures Database (MSigDB) Hallmark Gene Set Collection.” Cell Systems 1 (6): 417–25. https://doi.org/10.1016/j.cels.2015.12.004.