Spatial characterization and stratification of colorectal adenomas by deep visual proteomics

Abstract

Colorectal adenomas (CRAs) are potential precursor lesions to adenocarcinomas, currently classified by morphological features. We aimed to establish a molecular feature-based risk allocation framework toward improved patient stratification. Deep visual proteomics (DVP) is an approach that combines image-based artificial intelligence with automated microdissection and ultra-high sensitive mass spectrometry. Here, we used DVP on formalin-fixed, paraffin-embedded (FFPE) CRA tissues from nine male patients, immunohistologically stained for caudal-type homeobox 2 (CDX2), a protein implicated in colorectal cancer, enabling the characterization of cellular heterogeneity within distinct tissue regions and across patients. DVP identified DMBT1, MARCKS, and CD99 as protein markers linked to recurrence, suggesting their potential for risk assessment. It also detected a metabolic shift to anaerobic glycolysis in cells with high CDX2 expression. Our findings underscore the potential of spatial proteomics to refine early stage detection and contribute to personalized patient management strategies and provided novel insights into metabolic reprogramming.

Keywords: Artificial intelligence; Cancer; Cancer system biology; Proteomics.

Graphical abstract

Figure 1

Study design and our multi-layered mass spectrometry-based proteomics approach (A) Schematic representation of the colorectal adenoma (CRA) cohort and the study design. Nine resected CRAs displayed high-grade (HG) dysplasia which led to the same pathological assessment and diagnosis but showed three different clinical outcomes. (B–E) Multi-layered and streamlined mass spectrometry (MS)-based proteomics approach applied to FFPE CRA tissues. (B) FFPE blocks of each polyp were cut and mounted onto PEN membrane slides. (C) For bulk proteomics analysis, each tissue was scraped, lysed, digested, and extensively fractionated. (D) Deep visual proteomics (DVP) workflow for the analysis of region and cell class-specific protein changes, including machine learning (ML)-based segmentation, RGB and morphology-dependent classification, followed by automated laser microdissection. (E) Bulk proteomics and low-input DVP samples were measured on the same EvoSep-timsTOF platform, either in data-dependent (ddaPASEF) or data-independent (diaPASEF) acquisition mode, followed by spectral identification and quantification with AlphaPept, MSFragger, or DIA-NN.

Figure 2

Bulk FFPE tissue proteomics of non-malignant colorectal adenomas (A) Number of proteins in each DDA-acquired fraction per CRA sample. (B) Normalized total intensity of all identified proteins in our deep CRA library created from 432 fraction samples. Highlighted in dark blue: colorectal cancer (CRC)-associated proteins, part of the TARGET (tumor alterations relevant of genomics-driven therapy) database. (C) Number of proteins per patient within a group. (D–F) (D) Pairwise proteomic comparison between C and NMN patient adenoma samples, acquired in DDA, or (F) DIA mode, and the respective (E) coefficient of variation (CV). DDA data originated from fractionated samples, DIA was measured as single run (50 ng). Significantly enriched proteins are colored and displayed above the black lines indicating statistical significance (two-sided t test, permutation-based FDR <0.05, s₀ = 0.1). (G) Number of significantly down- and upregulated proteins in the volcano plot analyses. (H) GO biological process enrichment (FDR <0.05) of significantly upregulated protein hits. (I and J) Gene Set Enrichment Analysis (GSEA) of diaPASEF acquired data, of (I) positively and (J) negatively enriched pathways.

Figure 3

DVP characterizes region-specific metabolic changes within strongly heterogenous CRA sample (A) IHC staining of patient tissue C3 with three annotated tumor areas. (B) Representative images of selected regions based on the degree of dysplasia, density of CDX2++ epithelial cells and lymphocyte infiltration (also see Figure S3D). The color code signifies: 1, high dysplasia with a high density of CDX2++ cells (purple); 2, low dysplasia and medium density of CDX2++ cells (yellow); 3, normal glandular architecture and lymphocyte infiltration (green). Scale bar, 100 μm. (C) Distribution of CDX2++, CDX2+, and CDX2- cells across annotated regions and the remaining whole tissue area (white). Note that percentages are rounded and may not add up to 100%. (D) Unique protein numbers identified in CDX2++, CDX2+, and CDX2- cells. 1000 contours collected in triplicates. (E) Principal-component analysis (PCA) of collected CDX2++ and CDX2- cells across respective regions. (F and G) Pairwise proteomic comparison of CDX2- and CDX2++ between the highly dysplastic area and the region with normal epithelium (two-sided t test, permutation-based FDR <0.05, s₀=0.1). (H) GO term enrichment (FDR <0.05) of positively (purple) and negatively (green) enriched proteins. (I) Cluster map of reactome-annotated pathways. Normalized and standardized intensity values were used as input.

Figure 4

Singly isolated CDX2- and CDX2++ cells from colorectal adenoma tissues from group C, HDA, and NMN reveal a potential biomarker set for patient stratification (A–D) Principal component analysis (PCA) of collected (A and B) CDX2-stromal and (C and D) CDX2++ epithelial cells across all nine colorectal adenoma tissues. (E and F) Pairwise proteomic comparison of (E) CDX2- and (F) CDX2++ cells comparing the cancer group to NMN (two-sided t test, FDR <0.01, s₀ = 0.1). (G) Ranked protein abundance of normalized mean intensities of all identified proteins within each CRA group. The potential markers for adenoma classification DMBT1, CD99, and MARCKS are highlighted and labeled. (H) Unsupervised hierarchical clustering of 244 ANOVA significant proteins (permutation-based FDR <0.01, s₀ = 0.1). (I) Line graphs of the top five proteins with the highest ANOVA q value per cluster. (J) GO term enrichment analysis of cluster 1 and 2 (FDR <0.05), highlighting biological process (BP) and reactome (R) pathways of proteins with a positive Z score.

Figure 5

Orthogonal assessment of DMBT1, MARCKS, and CD99 in a cell culture model and an extended colorectal adenoma cohort (A) Representative CRA images of IHC staining. Scale bar, 50 μm. (B) Signal quantification of DMBT1, MARCKS, and CD99 on IHC-stained images. (C) Schematic outline of the adenoma cell culture setup and the associated aggressiveness. (D) Log2 intensity values of these marker proteins across adenoma cell lines S/RG/C2 and PC/AA/C1, and colon carcinoma cell line HCT-15. (E and F) Study design of the extended CRA validation cohort and the tissue macrodissection of CDX2-positive HG dysplasia areas. (G) Number of precursors and proteins from macrodissected CRA tissue. (H) Coefficient of variation of each CRA group. (I) Overlap of proteins between each CRA group. (J) Log2 intensity values for CD99, MARCKS, DMBT1, and keratins 1, 2, and 10 from macrodissected, CDX2-positive HG dysplasia areas in the extended CRA validation cohort. (K) Correlation of fold change between C and NMN, to HDA and NMN, after a two-sided t test, FDR <0.01, s₀ = 0.1.