INSIGHT: Spatially resolved survival modelling from routine histology crosslinked with molecular profiling reveals prognostic epithelial-immune axes in stage II/III colorectal cancer

TL;DR

INSIGHT addresses the challenge of prognosticating stage II/III colorectal cancer using routine histology by learning from spatial tissue organization with a graph neural network. It produces both patient-level survival risk and spatially resolved patch risk maps, and demonstrates superior prognostic performance relative to pTNM across internal and external cohorts. By anchoring histology-derived risk to multimodal molecular data, it reveals a cohesive epithelial-immune risk manifold characterized by fetal-like epithelial programs, myeloid-driven immunosuppression, and adaptive immune dysfunction, with particular relevance to MSI-High tumours. The framework enables mechanistic discovery and multimodal integration, offering actionable insights for targeted therapies and refined risk stratification in clinical practice.

Abstract

Routine histology contains rich prognostic information in stage II/III colorectal cancer, much of which is embedded in complex spatial tissue organisation. We present INSIGHT, a graph neural network that predicts survival directly from routine histology images. Trained and cross-validated on TCGA (n=342) and SURGEN (n=336), INSIGHT produces patient-level spatially resolved risk scores. Large independent validation showed superior prognostic performance compared with pTNM staging (C-index 0.68-0.69 vs 0.44-0.58). INSIGHT spatial risk maps recapitulated canonical prognostic histopathology and identified nuclear solidity and circularity as quantitative risk correlates. Integrating spatial risk with data-driven spatial transcriptomic signatures, spatial proteomics, bulk RNA-seq, and single-cell references revealed an epithelium-immune risk manifold capturing epithelial dedifferentiation and fetal programs, myeloid-driven stromal states including $\mathrm{SPP1}^{+}$ macrophages and $\mathrm{LAMP3}^{+}$ dendritic cells, and adaptive immune dysfunction. This analysis exposed patient-specific epithelial heterogeneity, stratification within MSI-High tumours, and high-risk routes of CDX2/HNF4A loss and CEACAM5/6-associated proliferative programs, highlighting coordinated therapeutic vulnerabilities.

Figures (13)

• Figure 1: Overview of the overall INSIGHT pipeline. Part i) shows the summary of different H&E WSI datasets used for model development and validation. Additional multi-modal datasets containing both H&E and an additional WSI modality were used for downstream discovery of prognostic factors associated with patient prognosis. The overview of the pipeline which utilizes H&E WSIs from Stage II/III Colorectal (CRC) specimens can be seen in ii). For each WSI we first perform tissue phenotyping to automatically restrict our analysis to tumour, stroma, and lymphocyte regions. These regions are then broken down into square patches for which we extract deep features. These are then used to construct a tissue graph, based on patch proximity. Such graphs can be used to train a Graph Neural Network (GNN) that directly predicts a patient-level personalized risk score. These risk scores can rank and stratify patients, as shown for the external test set PETACC-3 via the Kaplan-Meier Curve. Further, we can deconvolve the patient risk score into spatial patch-level risk. This can be seen in the WSI heatmap (red and blue signifying high and low risk respectively). These patch-level risk scores can be analysed by a pathologist as well as act as an automatic survival label. This allows us to discover patterns within different modalities that may not have a ground truth survival label. Here we analyse the relationship between spatial risk and various factors derived from mIHC, mIF and Spatial Transcriptomics from three external cohorts. This allows us to discover, in a time efficient manner, what biologically interpretable factors are most associated with prognosis. Further, patient-level risk scores allow performing subset discovery within clinical populations. These discoveries can be used to guide treatment decisions and future drug development that target these high-risk features. Further, we can combine all our risk factors into a single risk manifold that is biologically grounded and explains away INSIGHTS risk predictions
• Figure 1: PETACC3 Kaplan-Meier Curves. For different clinical subpopulations, we show the performance of INSIGHT via a Kaplan-Meier Curve. Patients are stratified into high and low risk groups based on an optimal risk score threshold learned during training.
• Figure 2: Morphological pattern analysis in high and low risk regions in external cohorts. In i) we see the representative high (red) and low (blue) risk patches taken from PETACC-3 where we see high risk regions showing a greater proportion of stroma and less well differentiated glands. In ii) we define the most important metrics in this study including nuclei solidity, aspect ratio and circularity that capture nuclei atypica. In iii) we systematically measure the relationship between patch-level risk and interpretable cell and gland features automatically generated using Cerberus in the Janssen cohort. This can be visualized via the bar chart showing the overall spearman correlation between a feature and localized risk.
• Figure 2: TCGA and SR386 Kaplan-Meier Curves. For each clinical subpopulation, we evaluate INSIGHT using Kaplan–Meier curves. Risk scores are obtained via 4-fold cross-validation: for each patient, we retain the risk score produced when that patient appeared in the test fold. This yields a combined dataset in which every patient is included exactly once and only as a test case. Patients are then stratified into high- and low-risk groups using an optimal risk score threshold learned during training.
• Figure 3: Association of spatially localized risk with multi-modal patch-level factors. In i)-iv) we focus on spatial transcriptomic profiles. In i) we begin with univariate analysis between patch-level gene expression derived from spatial transcriptomics and INSIGHT generated patch-level risk scores per patient. This is shown via the heatmap that shows the overall correlation between patient patches. To analyze finer grained detailed at a patient-level we show heatmap of patch-level risk scores and scaled patch-level gene expression within a patient, ii). This allows us to validate nuanced patient-level patterns, for example the opposite correlation of LGR5 to prognosis in IMU013 and IMU004. In iii) we perform multivariate gene analysis to regress risk with respect to data-driven patch-level Spatial Transcriptomic Signatures (STS) derived from 422 spatial transcriptomic genes. A given signature is represented via a word cloud, where red and blue genes represent if it is over or under expressed in a signature and the size of the gene its importance to the signature definition. We can then regress the binary signatures against localized risk via linear regression (OLS) allowing us to find the most important STSs, in this case STS- 8, STS -11 and STS -13. In iv) to further understand the STS expression patterns we show the proportion of the total patches in a patient that are activated for each signature, depicted via a heatmap, with blue and red depicting small and large proportion respectively. For each patient we also show several key point mutations and if they are mutated (purple) or not (green). In v) we show univariate analysis between localized predicted risk and concentration of mIHC stains and mIF proteins as additional validation of our discovers. This is visualized as a bar chart depicting the strength of correlations for different proteins. Finally in vi) we show our final derived risk manifold based on combining STS signatures, mIF protiens and micromorphological factors. With dimensionally reduction (see Methods) we can combine all these risk-associated features into a single risk manifold with three main axes. This is shown on the scatter plot where each point represents a single patch, coloured by its INSIGHT risk.