ReCAST: Single-Cell-Informed Tumour Microenvironment Profiling for Translational Prognostic Assessment in Esophageal Cancer

Abstract

Background/Objectives: Single-cell transcriptomics provides detailed maps of the tumour microenvironment, but translating these maps to large clinical cohorts requires methods that can recover biologically meaningful cell states from routinely available bulk transcriptomes and establish whether these states provide information beyond standard clinical factors. The objective of this study was to develop ReCAST, a donor-aware method that estimates tumour microenvironment cell states from bulk tumour transcriptomes, and to test whether these estimates improved survival prediction in esophageal cancer beyond routine clinical variables. Methods: ReCAST constructed outcome-independent cell-state anchors from single-cell transcriptomes by first aggregating expression within donors and then across donors, followed by robust non-negative projection of bulk tumour profiles. The framework was evaluated using synthetic mixtures, 720 genuine-cell pseudo-bulk mixtures from 60 held-out donors, cross-study mixtures from an independent single-cell cohort, an independent adenocarcinoma single-cell cohort, and independently measured tumour compositions in 182 TCGA esophageal cancers. Biological validity was examined against Human Protein Atlas cell-type annotations. ReCAST-derived tumour microenvironment features were subsequently tested for incremental prognostic value beyond age, sex, histology, and stage using repeated cross-validation, and the clinical prediction models were evaluated by locked external validation in an independent cohort of 60 patients with squamous-cell carcinoma. Results: ReCAST recovered 13 biologically coherent cell states whose markers showed strong enrichment for corresponding independent cell-type annotations (fold enrichment 5.1–29.8; FDR < 10−15). Cell-state profiles were reproducible across held-out donors and independent studies and donor-aware reference construction improved cross-study transfer. ReCAST outperformed BisqueRNA and unbalanced optimal transport on held-out pseudo-bulk mixtures and, because it worked on within-sample ranks, extended cell-state profiling to data without raw counts, such as variance-stabilised, microarray or summarised reference data; where raw counts were available, count-based methods (MuSiC, DWLS, BayesPrism and a CIBERSORT-type ν-support-vector regression) achieved lower errors and were the preferred complement. In clinical tumours, inferred immune composition tracked independently measured lymphocyte infiltration and leukocyte fraction. Routine clinical variables remained the strongest predictors of survival and were transferred to the independent squamous-cell carcinoma cohort (Uno C-index 0.61); adding ReCAST-derived features produced no incremental improvement (ΔUno C-index −0.033, 95% CI −0.089 to 0.023), a result that was consistent within each histology and with an adenocarcinoma-matched reference. Conclusions: ReCAST provides reproducible cell-state profiles across donors, studies and histologies and extends tumour microenvironment profiling to expression data without raw counts. In esophageal cancer, routine clinical variables already capture the prognostic information carried by cell-state composition, which positions cell-state profiling as a tool for biological characterisation rather than risk prediction. Establishing measurement validity and clinical utility separately provides a robust route for evaluating molecular biomarkers.