Deep-learning–driven tumor microenvironment profiling improves immunotherapy response prediction

This dataset contains the test data for EcoNet (https://github.com/Gentles-lab/EcoNet), a deep-learning model that predicts immunotherapy response from bulk RNA-seq. test_data/: small example inputs for a quick end-to-end (ccRCC) smoke test of the four steps. Step 1: TME Representation step1_scRNA.h5ad: Subsampled ccRCC scRNA-seq (3,119 cells, 8 samples, 66 cell states), log-normalized. EcotyperModel/: ccRCC EcoTyper model (cell-state gene_info.txt + ecotypes.txt) defining ecotypes and their marker genes. NicheNet_DB/: NicheNet ligand-receptor and signaling database. step1_ecotype_abundance.txt: Ecotype abundances (11 ecotypes x 8 samples) linking ecotypes to the scRNA samples. Step 2: Graph Representation Training step2_expression.tsv: Bulk TPM expression (2,514 genes x 200 TCGA-KIRC samples) for GAT training. step2_abundance.txt: Matching ecotype abundances (11 ecotypes x 200 samples), the training target. Step 3: Response Prediction Training step3_gat_model.pth: Pretrained ccRCC GAT weights (input to transfer learning). step3_pretrain_expression.tsv: Immunotherapy-cohort bulk expression (2,487 genes x 150 samples). step3_pretrain_clinical.tsv: Matching clinical table (sample ID + R/NR response labels). Step 4: Prediction ccRCC_input_tpm.tsv: Example bulk TPM input (6 samples) for the ccRCC model. pancancer_input_tpm.tsv: Example bulk TPM input (6 samples) for the pan-cancer model. ccRCC_clinical.tsv: Example clinical table for the optional evaluation step.