Inferring tissue element identities from sample-level compositional data in cancer

The relative abundance of cellular and structural components in cancer tissues reflects disease biology and outcomes. High-throughput technologies, e.g., digital pathology and single-cell transcriptomics (scRNA-seq), profile tissue regions or cells, revealing the heterogeneous building blocks (elements) that comprise bulk phenotypes. Yet, translational applications increasingly demand that models not only infer sample-level distributions but also provide element-level characterizations (e.g., of specific histology regions, cells) that explain tumor macroscopic behavior. Achieving such granularity from sample-level compositional data without laborious annotations remains a challenge, requiring models capable of connecting local properties to global phenotypes. Existing methods, using attention to weight element contributions to sample-level predictions, lack probabilistic grounding and biological interpretability; thus, have limited utility in clinical settings. Here, we explicitly model sample-level compositional constraints with element-level assignments using Optimal Transport (OT). Particularly, we introduce Composer , a domain-agnostic dual-task machine learning model trained without element-level annotations to 1) estimate sample-level compositions, and, 2) importantly infer element labels as interpretable compositional allocations. The model's performance was evaluated across data modalities, cancer types, and tasks. We highlight 3 concepts: a) Tissue type classification on whole-slide images (WSIs): On 60 hematoxylin & eosin (H&E) WSIs from high-grade serous ovarian cancer (HGSOC, Boehm et al. 2022), Composer predicted the overall WSI tissue composition (Jensen-Shannon divergence (JSD): 0.26; mean absolute error: 0.11), classifying correctly (AUROC 0.98) the tissue type (tumor, stroma, necrosis, adipose, other) of individual regions within WSIs without training on regional annotations. b) Tumor segmentation on WSIs: On 185 H&E WSIs spanning HGSOC, breast, and colorectal cancers (MSKCC), Composer estimated the overall tumor fraction per WSI (JSD 0.12; MAE 0.15), distinguishing efficiently tumor from non-tumor regions (AUROC 0.91) within WSIs. c) Single-cell type annotation in scRNA-seq: Using 156 scRNA-seq HGSOC samples (Vazquez-Garcia et al. 2022), Composer not only inferred bulk cell type distributions (JSD: 0.15; MAE: 0.06), but also accurately classified (AUROC: 0.97) individual cells to their designated type (T cells, monocytes, fibroblasts, cancer cells, other). In summary, the proposed OT-based weakly supervised framework provides an effective approach for linking element-level representations to sample-level compositional profiles in cancer. Its applicability across data analyses empowers spatial and molecular characterization of tumor ecosystems, biomarker quantification, and potential applications to precision oncology.