The spatial organization of multicellular ecosystems underpins tissue homeostasis and disease progression. Single-cell atlases have extensively characterized cellular compositions and their pathological remodeling. However, single-cell omics technology separates molecular profiles from their native spatial context. This separation makes it difficult to identify condition-specific microenvironments associated with pathological phenotypes. Here, we present DREAM (Dual-stream Representation & Explicit Attribution Modeling), a computational framework for interpretable attribution via concept-driven modeling. DREAM leverages context-aware semantic transfer to construct robust niche semantic representations, synergistically encoding intrinsic biological semantics and extrinsic spatial topology through a dual-stream architecture. By incorporating a concept bottleneck mechanism, the framework maintains a balance between clinical predictive accuracy and biological interpretability. DREAM was benchmarked across five spatial proteomics and transcriptomics datasets. It consistently outperformed existing methods in identifying reproducible tissue domains. Applied to colorectal and liver cancer cohorts, condition-specific microenvironments were characterized, and slice-level pathological phenotypes were accurately predicted. These results were supported by downstream computational analyses. Additionally, the identified microenvironments were shown to be significant prognostic indicators, and key immune-active regions, such as tertiary lymphoid structures (TLS), were accurately localized. Ultimately, DREAM demonstrates how concept-driven modeling can highlight potential pathological associations from complex spatial omics data, providing a novel computational perspective for understanding spatial heterogeneity in complex diseases.
Figure.1 Overall architecture of DREAM. a The workflow begins with the Input dataset module. b The Context-aware niche semantic transfer paradigm (CAST) leverages neighborhood aggregation to transform context-aware niche feature matrices into neighborhood-scale niche semantic representations. c The Dual-stream driven semantic-topological synergistic representation learning framework reconciles feature representation trade-offs. It synergistically encodes biological semantics through a Semantic Stream and spatial topology through a Topology Stream. d The Hybrid concept bottleneck interpretability architecture functions by first mapping fused features to intermediate “concepts” (spatial domains), subsequently constructing a predictor for clinical phenotypes, and allowing for the explicit quantification of domain contributions via gradient back-propagation. e Downstream Analysis utilizes the interpretable model to quantitatively dissect pathological mechanisms, identifying which condition-specific microenvironments drive disease progression and distilling clinically actionable insights.
All analyzed datasets are publicly available and can be accessed via the following links: (1) the mouse spleen CODEX dataset [https://data.mendeley.com/ datasets/zjnpwh8m5b/1]; (2) the human UTUC IMC dataset [https://doi.org/10.5281/zenodo.6376766]; (3) the mouse V1 neocortex STARmap dataset [https://zenodo.org/record/ 7830764#.ZDpObi-1HUI]; (4) the MERFISH Frontal cortex dataset [https://cellxgene.cziscience.com/ collections/31937775-0602-4e52-a799-b6acdd2bac2e]; (5) the human DLPFC 10X Visium dataset [http://spatial.libd.org/spatialLIBD/] [https://www.ncbi.nlm.nih.gov/geo/]; (6) the CRCLM dataset can be accessed via the link [https://drive.google.com/file/d/1QsQIT0iwcWBFzUBcLUPKYuSnSBxfaME/view?usp=drive_link] [https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE132465]; (7) the liver datasets [https://ngdc.cncb.ac.cn/gsa-human/browse/HRA000437] [https://db.cngb.org/search/project/CNP0000650].
We provided codes for reproducing the experiments of the paper "Interpretable attribution of clinical phenotypes to condition-specific microenvironment via concept-driven modeling", and comprehensive tutorials for using DREAM. Please check the tutorial website for more details.
Upon the usage the users are requested to use the following citation:
· D. Zhang, R. Qi, and B. Liu, "Interpretable attribution of clinical phenotypes to condition-specific microenvironment via concept-driven modeling,"