Immune Escape Atlas as a practical tool for drug design and patient stratification in clinical trials
In recent years, a number of computational approaches provided different strategies to classify solid tumors based on the specific transcriptomic features of their tumor microenvironment (TME), predicting prognosis or elucidating responders to therapies. Still there is no unifying model that can account for different aspects of immune activity and tumor evasion within the same patient and guide development of novel immunotherapeutics. Tumor immune escape mechanisms, major causes of immunotherapy resistance, were also rarely included and intersected with TME states. In this study, we offered an approach to map a wide spectrum of known tumor IE mechanisms onto a large clinical cohort of cancer patients, harmonizing more than 37,000 patient samples, collected across 193 different datasets. Our findings discerned TME types that match existing approaches and are relevant to the developing TME-targeted therapies. We compared our TME classification framework to existing classifications and demonstrated a broadening method which allowed us to differentiate patients according to potential response to a number of targeted therapies. The proposed framework can be deployed by the medical and scientific community to interpret results of clinical trials, improve drug development and therapy decisions.
Features used in the study are a collection of mechanisms describing tumor, immune and their interaction. Immune Escape mechanisms are an important addition to an already described set of mechanisms covered by other TME classification tools, like Molecular Functional Portraits or Archetypes.
Immune Escape Mechanisms covered in the study
We defined 9 TME types based on full feature set of 45 mechanisms for major immune escape strategies, which can be used to guide decision-making for the patient.
Identified TME types
The identified Immune Escape TME types are as follows:| TME Type | Abbreviation | Description |
|---|---|---|
| Lymphoid-Cell-Enriched | IE/Lymph | Prevalence of lymphoid cells over myeloid cells in TME composition with relatively low stromal components and low hypoxia |
| B-cell-enriched, Angiogenic | IE/B | High level of infiltration with both lymphoid and myeloid cells, with a certain increase in B cells, high BCR diversity and prevalence of CXCL9+ macrophages over SPP1+ |
| Immune-Enriched, Hypoxic | IE/Hyp | High immune cell infiltration with prevalence of myeloid cells and high levels of hypoxia, glycolysis, EGFR and MAPK signaling |
| Highly Immune-Enriched, Inflamed | IE/Inf | The highest immune infiltration, especially with lymphoid cells, extremely high expression of both lymphoid- and myeloid-cell-associated checkpoint molecules and the highest TCR and BCR diversity. CXCL9+ macrophages were prevalent according to deconvolution, while stroma and hypoxia signals were low |
| Immune-Enriched, Fibrotic | IE/F | High immune cell infiltration with the highest macrophage-to-lymphocyte ratio and high TCR and BCR diversity, high stroma-associated features (high cancer-associated fibroblasts (CAFs) content with prevalence of myCAFs, high angiogenesis), and highly upregulated TGFβ pathway |
| Fibrotic, Angiogenic, Myeloid | F/Ang/Myel | Low immune infiltration, mostly by SPP1+ macrophages, and high content of CAFs, especially myCAFs. TGF-β signaling was highly upregulated, angiogenesis and endothelial cell levels were increased in this TME type |
| Fibrotic, Hypoxic | F/Hyp | Minimal immune cell infiltration and very high hypoxia level, with upregulated EGFR and MAPK signaling. CAF and endothelial cell levels were significantly lower than in other fibrotic types |
| Immune Desert | D | Cluster was marked by the lowest immune cell infiltration and lowest presence of stroma components, with occasional hypoxia. |
| Faintly Infiltrated, Angiogenic | D/Ang | Low immune cell content (mostly lymphocytes) with increased TCR and BCR diversity compared to other desert types, moderate angiogenesis and endothelial cell levels. |
Immune Escape class annotation for TCGA samples is in data/TCGA_IE_class_prediction.tsv ('IE_Class' column; rest — probabilities). Immune Escape features calculated for >37000 samples are in data/scaled_features.tsv
If software, data, and/or website are used in your publication, please cite [CITATION] and make a reference to this repository.
For more information visit BostonGene’s Scientific portal
See the SETUP.md to set up the environment. It is highly recommended. The following steps are performed using repo and the installed kernel; to clone:
git clone [email protected]:BostonGene/Immune_Escape.git
Sample is an Illumina RNA sequencing or microarray data of a human solid tumor biopsy in a format of dataframe in as tsv with genes as rows in HUGO format, samples as columns and values as logTPMs or log intensities.
We recommend to use the model for the samples with high purity (>80% of tumor cells) and only for diagnoses used in the study, i.e. solid cancers except sarcomas. The following TCGA cohorts were used as a discovery cohort: BRCA, HNSC, THCA, LUAD, LUSC, KIRC, SKCM, PRAD, BLCA, STAD, LIHC, OV, CESC, KIRP, COAD, ESCA, UCEC, PAAD, PCPG, READ, UVM, ACC, KICH, UCS, CHOL. See this page to decipher abbreviations. However, we tested the model to additionally classify unused diagnoses: Glioblastoma, Glioma, and Astrocytoma and showed that it was clinically relevant.
We described how to get the required data and perform quality control (QC) in the DATA_AND_QC.md.
Starting with identified batches of data, QC-checked, with removed outliers, and in a proper format, one can calculate and scale features. The are two ways:
- With sufficient sample size in a batch (10 as a minimal cutoff, however, the more is always the better, with low variety 10 also can be insuffiecient) — use Classify_samples.ipynb;
- With insufficient sample size in a batch — use Get_reference_cohort_and_classify.ipynb to find reference cohort out of metacohort, or get one manually and use Classify_samples.ipynb.
The workflow is as follows:
Diagram for sample classification
Notebooks description
-
Classify_samples.ipynb — notebook with example classification for the cohort with batch size > 10. Input: expression dataframe (HUGO genes in rows, samples in columns), output: series of Immune Escape classes predicted for the samples. See how to prepare expression dataframe in DATA_AND_QC.md.
-
Get_reference_cohort_and_classify.ipynb — notebook with additional step for finding the reference. Input: expression dataframe (HUGO genes in rows, samples in columns), output: series of Immune Escape classes predicted for the samples; predicted batch series for the samples. See how to prepare expression dataframe in DATA_AND_QC.md. See tests for the method in Test_mapper.ipynb; test show that for random 1000 samples for the cohort weighted F1 is F1=0.731 for k=5, not accounting for the unclassified samples.