We first validated the DSP technology by exploring gene-wise and ROI-wise association across all samples/genes. We selected tumor and stroma ROIs, given their expected distinct expression characteristics. The representative images of full scan and ROIs from either tumor or stroma are shown in Figure 1A. Representative ROIs from patients who received either chemotherapy only or combined with immunotherapy are shown in Figure 1B. Additional full scan images and ROIs are shown in Supplementary Figure S1. As expected, based on the expression profiles of 84 genes at the transcriptional level and 40 at the protein level in all ROIs, tumor and stromal regions showed strong region-specific expression patterns, regardless of individual tissue characters and different treatments (Figures 1C,D). The mRNA data from tumor ROIs also showed a therapy-differentiated expression pattern. A clear separation of patient groups with different neoadjuvant therapies by mRNA expression in tumor but not stroma ROIs suggests a potential immune signature in the tumor-enriched TME compartments (Figure 1C). Protein expression did not separate patient groups potentially due to the limited ROI numbers and protein targets in this study (Figure 1D). Also, the RNA or protein data for the 17 genes in common between 84-plex RNA and 40-plex protein panels did not separate patients according to different treatments (Supplementary Figures S2A,B).

For between-gene association analysis, common epithelial cell markers such as AKT1, Cytokeratin, EPCAM, Ki-67, and S6 showed a strong association at both the RNA and protein levels (Supplementary Figure S2C,D). At the protein level, the data showed a strong co-expression pattern between the T-cell markers (CD3, CD4, and CD8) and the myeloid-derived cell markers (CD14 and CD163). They are also significantly associated with the total immune cell markers (CD45/CD45RO/HLA-DR) and to a lesser extent with the B-cell markers (CD20), indicating an overall consistency of those markers being co-expressed within particular microenvironments (Supplementary Figure S2D).

Based on the ROI selection strategy of the TME regions, we compared the tumor-surrounding stromal and tumor epithelial areas (ROIs) within or between the treatment groups. Chemotherapy patients (patients 1 and 2) showed a generally minimal immune marker expression at tumor-enriched regions at the RNA and protein levels, indicating low tumor-infiltrating lymphocytes (TILs) (Figures 2A,B). All immune-related cells resided in the surrounding stroma, with an apparent exclusion from tumors demonstrated by the overall low expression of the immune markers (Figures 2A,B). The RNA expression within the tumor-enriched regions revealed distinct expression profiling between these two patients, whereas those same sets of markers showed minimal differences within the stromal compartments (Figure 2C and Supplementary Tables S4, S5). Compared with patient 2, patient 1 showed higher levels of GZMB, TNF (tumor necrosis factor), IL12B (interleukin 12B), IL6 (interleukin 6), CD8A, CD3E, CD4, and CXCL9/10, suggesting a more active status of immune infiltration. Patient 1 also showed higher levels of PTEN (phosphatase and tensin homolog), AKT1, and Ki67. Patient 2 showed an increased STAT3 (signaling transducer and activator of transcription 3) (Figure 2C), likely due to the primary tumor difference or adaptation to the treatment. The differential expression above could only be discovered when the tumor-enriched regions and stroma regions were explored separately (Figure 2C, and Supplementary Tables S4, S5 for the entire gene lists). When comparing the 17 common genes, mRNA difference was not captured at the protein level in the same ROI profiling, emphasizing the importance of proteomics data in complementing transcriptomic data to stratify patients.

As expected, combinational therapy (chemo/immuno) induced various immune infiltration methods into the tumor regions manifested by an elevated expression of many immune markers, including CD4, CD68, GZMB, CD40LG (CD40 ligand), CD86, CD276, and LAG3 (lymphocyte-activation gene 3) at the mRNA level in multiple tumor ROIs (Figure 2D). Chemokines and cytokines, such as CXCL10 and IL15, IL12B, and IL6, and the receptor CXCR6 (C-X-C chemokine receptor type 6), also increased at RNA levels. Concomitant increases of PD-L1, IDO1 (indoleamine-pyrrole 2,3-dioxygenase), OX40 (tumor necrosis factor receptor superfamily member 4), CD66b, and CD68 at the protein level were also observed (Figure 2E compared to Figure 2B).

We then compared inter-treatment group expression profiles and observed an overall high expression of immune markers at both the RNA and protein levels in the combinational treatment group, suggesting induction of immune infiltration (Figures 3A,B). As expected, we observed a significant (p < 0.001) tumor-specific increase of PD-L1 expression at both the mRNA and protein levels for patients receiving combinational therapy (Figures 3C,D). We validated this finding by conventional immunohistochemistry (IHC), where we observed an increased expression of PD-L1 in combination therapy–treated patients reconfirming the DSP data (Figure 3E). Another biomarker, B7-H3 (CD276), also increased at the mRNA and protein levels in the tumor regions of combination therapy–treated patients (Figures 4A,B). The IDO1 expression was also up-regulated markedly at both the mRNA and protein levels within the tumor regions of combination therapy patients (Figures 4C,D). Of note, CD45 only increased at the protein level but not at the mRNA level (Figure 4E), suggesting the need to evaluate the TME at both the omic levels. In addition, stromal ROI profiling identified an increased protein expression of STING (stimulator of interferon response CGAMP interactor 1) and CD14 in the chemotherapy group, a sign of immune exclusion in the presence of the chemo agent alone (Figures 4F,G).

Of further importance, a holistic view of the tumor region–based clustering (both unsupervised hierarchical clustering and principal component analysis (PCA)) resulted in concordant findings that the RNA expression profiling correlates strongly with therapeutic options (Figures 5A–C). Further to these findings, both unsupervised hierarchical clustering and PCA analyses also yielded two distinct therapy-guided groups at the mRNA level in the stromal regions (Figures 6A–C). These ROI region–defined grouping could not be observed when tumor and stroma are mixed (Figures 6D,E), emphasizing the importance of spatially oriented information for patient stratification and biomarker discovery.

We compared the DSP profiling of RNAs and proteins in common within the same regions at the subhistological resolution of a few hundred square micrometers in size. The purpose was to gain insights into the dynamic changes between RNA and protein. Seventeen common targets [AGR1, CD4, CD8 (CD8A), and CD20 (MS4A1), CD68, CTLA4 (cytotoxic T-lymphocyte–associated protein 4), GZMB, 4-1BB (TNFRSF9), IDO1, LAG3, PD-1 (PDCD4), CD45 (PTPRC), PDL1 (CD274), FOXP3 (forkhead box P3), VISTA (V-domain immunoglobulin suppressor of T cell activation), B7-H3 (CD276), and Tim-3 (T-cell immunoglobulin and mucin domain 3)] were filtered out and plotted based on their expression correlation for individual ROIs (Figure 7A). An overall correlation between RNAs and proteins was moderate (R ² = 0.58, p < 0.05). Genes including PD-L1, CD4, and VISTA exhibited strong correlations at the RNA and protein levels. CTLA4, FOXP3, and GZMB exhibited weak correlations (Figure 7B). The transcriptomic and proteomic correlation differed between therapy groups, with an overall higher correlation in the chemotherapy group than in the combinational therapy group (p < 0.05) (Figures 7C–E). The most pronounced alteration of GZMB and CD276 showed high concordance in the chemotherapy patients but minimal or negative correlations in combination therapy patients (Figures 7C,D). Despite the likelihood of low levels of transcripts in specific ROIs (in the case of GZMB and CD276), low correlations of FOXP3 and CTLA4 due to the low level of protein below the baseline suggest that transcriptional information may not be fully interpreted at protein levels to execute their biological function. Therefore, posttranslational modifications likely play pivotal roles in a particular cellular context (Shiratori et al., 1997; Chikuma et al., 2000).

Tumor specimens from four patients were obtained during surgery from Beijing Cancer Hospital. Samples were immediately fixed and paraffin-embedded. All patients were TNM classified as stage III colorectal cancer (T3) with no metastasis at diagnosis. Before surgery, two patients received neoadjuvant chemotherapy with capecitabine and oxaliplatin (CAPEOX). The other two patients received the CAPEOX chemotherapy combined with sintilimab (a monoclonal antibody that binds to PD-1). All four patients were pathologically evaluated pre-and posttreatment. All patients were microsatellite stable (MSS) based on IHC data of MLH1, MSH2, MSH6, and PMS2. Anti-MLH1 antibody (GM002), anti-MSH2 antibody (RED2), anti-MSH6 antibody (EP49), and anti-PMS2 antibody (EP51) were obtained from Gene Tech Biotechnology (Shanghai China). All four patients had histological tumor regression with AJCC/NCCN grade (TRG) of 2/3 based on posttreatment pathological evaluation. Three serial sections of tissue samples (5 μm thickness) from individuals were freshly prepared for the DSP protein, RNA profiling, and IHC, respectively, to allow parallel comparison. Detailed patient information is provided in supplementary (Supplementary Table S1). All tissue samples had informed consent from the patients as documents in Beijing Cancer Hospital for research purposes.

Digital spatial profiling (DSP) was performed on slide-mounted FFPE samples as described previously (Wang et al., 2021a). For protein profiling, slides were preprocessed with deparaffinization and rehydration before incubation with a tricolor fluorescence morphology marker panel targeting Pan-CK (epithelial and tumoral regions), CD45 (immne cells), and SYTO13 (nuclear) together with target-specific oligo-conjugated primary antibodies cocktails for 40 proteins. For RNA profiling, a panel of probes for 84 mRNAs were hybridized at 37°C overnight in a hybridization oven and then incubated with the aforementioned morphological antibody panel. Detailed protein and RNA target information can be found in supplementary (Supplementary Tables S2, S3). For both the protein and RNA profiling, multicolored morphology markers allow compartment visualization to guide the selection of regions of interest (ROIs). This one-step overnight incubation was then followed by 20x high-precision scanning on a GeoMx DSP system and circular ROI selection. To ensure reliable quantification and inter-ROI data comparison, surface areas of ROIs were drawn between 2 × 10⁴ and 1.2 × 10⁵ μm² for protein profiling or between 3.8 × 10⁴ and 3.8 × 10⁵ μm² for RNA profiling. Conjugated target-specific oligos were released upon UV light illumination and collected in 96-well plates. Heat-dried oligos were then hybridized to unique NanoString barcodes, purified on the nCounter Prep station, and counted on the nCounter analysis system using standard procedures (Merritt et al., 2020; Wang et al., 2021b).

Immunohistochemistry (IHC) was performed on (FFPE) patient sections using prevalidated PD-L1 antibody (13684S, Cell Signaling Technology Danvers, MA). After deparaffinization and rehydration in xylene and ethanol, antigen retrieval was performed in 1x EDTA retrieval solution (pH 9.0) with heating. Slides were then blocked with goat serum (x0907, Dako) for 1 h and followed by primary antibody incubation at 1:200 dilution. After overnight incubation, slides were incubated with goat anti-rabbit antibody (E046201, Dako), then developed with DAB substrate (k3468, Dako), and counterstained with hematoxylin (CTS-1090, Biotechnologies). Slides were scanned with a brightfield microscope (Aperio CS2, Leica) and processed by ImageScope software (Leica).

To adjust system and experimental bias and to counteract ROI size variation effects, raw digital count files (RCC) for individual ROIs were normalized by ERCC RNA spike-in controls before downstream processing. This quality control step generated normalization positive factors from individual ROIs. The ROI inclusion criteria were limited on a minimum surface area of 1.6 × 10³ μm² for protein and 1.6 × 10⁴ μm² for RNA, and minimum nuclei counts of 20 for protein and 200 for RNA generally. Any ROIs resulting in a normalization positive factor higher than 3 or lower than 0.3 were excluded from the downstream analysis. QC-qualified ROI count files were then normalized by the geometric mean of housekeeping genes (Histone H6, GAPDH, and S6 were used for protein. RAB7A, OAZ1, UBB, POLR2A, and SDHA were used for RNA). The normalized data were log-transformed with or without being median-centered before comparison and plotting. For proteogenomic comparison, gene and protein ID were matched by Entrez ID for the following genes: AGR1, CD4, CD8 (CD8A), CD20 (MS4A1), CD68, CTLA4, GZMB, 4-1BB (TNFRSF9), IDO1, LAG3, PD-1 (PDCD4), CD45 (PTPRC), PDL1 (CD274), FOXP3, VISTA (VSIR), B7-H3 (CD276), and Tim-3 (HAVCR2). All data were processed and analyzed in DSP analysis software and R version 3.6.0 with relevant packages. The correlation analysis was computed using the “Pearson” method. Hierarchical clustering and correlation matrix were done with “pheatmap” package. The principal component analysis (PCA) was conducted by “FactoMineR” and “factoextra” packages. Volcano plots were created with log2FC set at 1 and adjusted p-value at 0.05 for cut-off (dashed lines). Venn plots were created with the “VennDiagram” package. For the differential expression analysis, a non-parametric Mann–Whitney U-test was used, and p-value was set to 0.05 at a significant cut-off. Due to the limited number of probes and samples, in some cases, p-value was presented without adjustment. Other relevant plots were generated by “ggplot2” package. For function and pathway annotation and enrichment analysis, differentially expressed genes (gene symbol) were submitted and processed in Metascape online interface (https://metascape.org).