Differences in T-cell densities and neighborhood patterns in human colorectal adenomas and sessile serrated lesions

Abstract

Background:

T-cell responses influence recurrence and survival in colorectal cancer. T-cell subset distributions vary by molecular phenotype and anatomic location, shaping cytotoxic or immune-cold tumor immune microenvironments. However, the T-cell contexture and their spatial proximities within preinvasive lesions are not well characterized.

Methods:

We analyzed sessile serrated lesions (SSLs), tubulovillous/villous adenomas (TVs), and tubular adenomas (TAs) from three studies (N=120). Whole-slide multiplex immunofluorescence was used to quantify eight T-cell subsets (CD4⁺, CD8⁺, Th1 (CD4⁺TBX21⁺), Th17 (CD4⁺RORC⁺), Treg (CD4⁺FOXP3⁺), Tc1 (CD8⁺TBX21⁺), Tc17 (CD8⁺RORC⁺), TcTreg (CD8⁺FOXP3⁺). Densities were compared by histology using a generalized linear mixed model with a negative binomial distribution, including an offset for total cell area and adjusting for age, sex, anatomic location, and lesion size. Nearest-neighbor (NN) analyses assessed spatial proximities of T-cell pairs across lesion types.

Results:

TAs and SSLs had higher CD4⁺ and CD8⁺ T-cell densities than TVs (q<0.05). Compared with TVs, TAs also had higher Th17 cell densities, whereas SSLs showed a trend toward lower Treg densities (q=0.06). NN analysis showed greater Treg clustering in TVs than in SSLs and TAs. In contrast, TA versus SSL comparisons demonstrated predominant CD4⁺ clustering with Th17, Th1, and CD8⁺ subsets.

Conclusion:

TVs exhibited lower T-cell densities and greater regulatory T-cell clustering, consistent with an immune-cold environment. SSLs and TAs were more immune-infiltrated than TVs, but TAs had higher inflammation and CD4⁺ dominant clustering, suggesting stronger helper coordination.

Impact:

Preinvasive lesions demonstrate immune and spatial heterogeneity which may have implications for primary or secondary prevention.

Introduction

T-cell infiltration is a strong predictor of recurrence and survival in colorectal cancer (CRC). ^1–4 The distribution of helper T-cell subsets varies by tumor molecular phenotype and anatomic site. ^5–7,8 Microsatellite instability–high (MSI–H) CRCs, which arise predominantly in the proximal colon,^{9, 10} are enriched for CD8⁺ cytotoxic T-cells and Th1 cells, whereas microsatellite-stable (MSS) tumors, which are more common in the distal colorectum,^{7, 11} tend to have higher Th17 cell densities. Prior studies have shown better CRC outcomes for greater infiltration of Th1 and CD8⁺ cytotoxic T-cells.^1–4

Patterns of T-cell infiltration are characterized far less in preinvasive lesions than in CRCs. Sessile serrated lesions (SSLs), which are predominately located in the proximal colon and resemble MSI–H CRCs genomically and immunologically, are dense with intraepithelial lymphocytes including regulatory T-cells (Tregs) and CD4⁺ memory T-cells (CD4⁺CD45RA⁻CD45RO⁺);^{6, 9, 12, 13} tubular adenomas (TAs), part of the canonical MSS pathway, occur throughout the colorectum and often show Th1- and Th17-associated cytokine signatures;⁵ and tubulovillous/villous adenomas (TVs), frequently located in the distal colon and likely related to KRAS mutated MSS CMS3 CRCs, exhibit immune-cold profiles that intensify with increasing villous architecture.^{7, 10, 13}

Beyond differences in average densities, spatial proximity among T-cell subsets may shape immune function within the tumor microenvironment (TME).^{14, 15} T-cells cluster with other T-cells to form niches where close contact and reciprocal (paracrine) signaling help to sustain T-cell activation and function:^{16, 17}specific neighborhood relationships may reinforce cytotoxic-effector (Tc1/Th1) interactions versus immunoregulatory (Th17/Treg) milieus, whereas sparse or lack of T-cell neighbors may mark immune-cold regions or “immune deserts.”¹⁸ Although few studies have examined T-cell proximities in CRCs, Huang and colleagues¹⁴ reported strong Tc1-Th1 spatial co-localization (also called interactions) in MSI-H CRCs and conversely, higher Th17, Treg, and Th2 interactions in MSS tumors. In prior pilot work,¹⁹ we observed distinct spatial relationships in preinvasive lesions: SSLs contained abundant CD4⁺ T-cells but Tregs were sparse and spatially distant from other CD4⁺ T-cells, whereas conventional adenomas showed frequent CD4⁺/Treg commingling, consistent with a coordinated immune dampening response. However, to our knowledge, no study has compared T-cell neighbor relationships between different histological types of preinvasive lesions.

Therefore, in the present study, we used single-cell multiplex immunofluorescence (mIF) analysis to profile eight helper and cytotoxic T-cell subsets by histologic type (TA, TV, and SSL) adjusting for anatomic location, age, sex and lesion size using a convenience sample drawn from a larger multi-center retrospective study. We then examined pairwise spatial relationships among eight T-cell subsets by histologic type using a nearest neighbor (NN) approach.^{20, 21}

Materials and Methods

Our study is a cross-sectional analysis which compares T-cell densities across three lesion types. Samples came from participants in three completed studies: the Vitamin D/Calcium Polyp Prevention Study (ViCAPPS, 2004–2016),²² Diet and Health Studies III–V (DHS, 1998–2010),^{23, 24} and the Tennessee Colorectal Polyp Study (TCPS, 2003–2015).²⁵ In each study, participants had been diagnosed with one or more histologically confirmed colorectal adenomas or SSLs. Patient samples in ViCAPPS were obtained at baseline exam in most cases (2004–2013) and a small subset from the observational follow-up (2014–2016). Exclusion criteria included: familial CRC syndromes, inflammatory bowel disease, prior colon resection, and history of cancer other than non-melanoma skin cancer within at least 5 years prior to enrollment.

Ethics approval and consent to participate:

All subjects provided consent for participation in the original research studies. All research was performed in accordance with the Declaration of Helsinki. Data was deidentified prior to shipment to the data coordinating center at the Medical University of South Carolina. All research performed at MUSC was approved by the IRB-II -Medical University of South Carolina as exempt.

Each clinical study (ViCAPPS, DHS III-V, TCPS) provided four unstained 5 μm FFPE sections per subject. Prior to shipment, a site pathologist verified that the block contained ≥5 mm of tissue, was colonic in origin, and had adenomatous or serrated histology. Slides were barcoded, deidentified, and shipped in batches to the Data Coordinating Center at the Medical University of South Carolina (MUSC). Upon receipt all samples were verified against manifests by the pathology coordinator and stored in a desiccator in a cold room until multiplex immunofluorescence (mIF) analysis was started.

To ensure standardization of diagnosis across the study samples, a central pathology review was performed by an expert in gastrointestinal pathology (D.S.) blinded to original diagnosis and clinical information. For each digitized H&E, the pathologist identified the histologic type (e.g., TA, TV, SSL) and the diagnosis was recorded within a REDCap database. If there was discordance between the study pathologist (D.S.) and original diagnoses regarding histologic type of adenoma (TA, TV), the study pathologist’s diagnosis was accepted. If there was a disagreement between the study pathologist and the original diagnosis on whether the lesion was adenomatous, or serrated in origin, a second study gastrointestinal pathologist (D.L.) performed an independent review on the same digitized slide and was blinded to the first study pathologist diagnosis. To resolve differences between the first and second study pathologists, an adjudication protocol was used, and disagreements were resolved by reviewing images together and arriving at a consensus. Results presented here reflect the consensus diagnosis.

To reduce potential batch-effects across mIF runs, specimens were classified into four groups to ensure the groups were equally distributed across mIF batches: small (<10 mm) TAs, large (≥10 mm) TAs, TVs, and SSLs. Each 24-slide mIF batch contained 22 study samples balanced by clinical study (ViCAPPS, DHS, TCPS) and each of the four strata, plus two control tissues (normal colon mucosa and tonsil). Antibody panels and staining order were optimized using tonsil control, normal control and 20 colorectal neoplasms (10 adenomas (½ TAs, ½ TVs), 10 SSLs) balanced by anatomic location (proximal, distal, rectal). Control tissues for optimization were obtained from a commercial source or MUSC surgical pathology archives (normal control, adenomas, and SSLs) with aid from the MUSC study pathologist (D.L.).

Staining was performed on the Roche Ventana Discovery Ultra platform with Akoya Opal^™ tyramide signal amplification, following methods previously described.²⁶ Slides were stained for five immune markers (CD4, TBX21, FOXP3, RORC, CD8) and Pan Keratin (KRT), followed by DAPI nuclear counterstain for cell detection (see Supplemental Table 1 for additional details). Stained slides were mounted with ProLong^™ Gold Antifade reagent, coverslipped and imaged at 20× magnification on the Akoya Vectra^® Polaris^™ (PhenoImager^™ HT). Whole-slide images were reviewed and annotated in Phenochart^™ by staff blinded to the diagnosis.

Image Analysis

Image analysis of mIF-stained tissues was performed using inForm^® Tissue Analysis Software (v[2.6.0], Akoya Biosciences, Marlborough, MA). The workflow included Tissue Segmentation, Cell Segmentation, and Phenotyping. For training across these workflow features, at least 10 regions of interest were selected among the lesions for analysis. Lesions for training varied in histologic type to best capture the biological and morphological heterogeneity across lesion types. The regions of interest were used in the following steps:

Tissue Segmentation:

Tissue regions (epithelium, stroma) were defined using the Trainable Tissue Segmentation feature. For tissue segmentation training, specifications were required for Tissue Categories, Components for Training, Pattern Scale, and Segmentation Options. Three Tissue Categories were defined: Epithelium, Stroma, and Other (non-classifiable regions). DAPI (nuclear marker) and Keratin (KRT) (epithelial marker) were selected as Components for Training, that is, only DAPI and KRT and no other biomarkers were considered for defining tissue regions. A Large pattern scale was selected, as suggested by the inForm User Guide v.3.3.0. for tissue sections and larger scale structures. For smoothing edges along tissue regions, a Medium segmentation resolution was selected and Trim Edges was set to 5 pixels in Stroma regions within the Segmentation Options. Multiple iterations of training for tissue segmentation were implemented to ensure adequate performance of the segmentation algorithm across morphologically heterogeneous samples.

Cell Segmentation:

The Segment Cells feature was used for cell segmentation, including segmentation of the nuclei, cytoplasm, and cell membrane. Additional specifications were required for Nuclear Component Splitting, Assisted Component Splitting, Other Settings, and Membrane and Cytoplasm Settings. Under Components, DAPI was selected for nuclear detection and pan-keratin was selected to aid in membrane detection and nuclear splitting, given this marker is primarily expressed along the cell membrane. Specifications in Nuclear Component Splitting included a mixture of quality for Nuclear Staining, a Splitting Sensitivity of 0.33, and a Minimum Nuclear Size of 40 pixels. For Assisted Component Splitting, the Assisting Staining was “a mixture of quality,” a Splitting Sensitivity of 0.40, and a Minimum Nuclear Size of 35 pixels. Other Settings indicated to fill nuclear holes smaller than 2 pixels and to refine cells after segmentation. Lastly, the following specifications were selected for Membrane and Cytoplasm Settings: a Cytoplasm Thickness of 3 pixels, a Membrane Search Distance of 10 pixels, a mixture of quality for Membrane Staining, and a Membrane Signal Threshold of 0.50.

Cell Phenotyping:

Phenotyping was performed for the six markers of interest: CD4, CD8, pan-keratin, TBX21, RORC, and FOXP3. Together, these would indicate the following T-cell phenotypes: CD4 (CD4⁺), Th1 (CD4⁺TBX21⁺), Th17 (CD4⁺RORC⁺), Treg (CD4⁺FOXP3⁺), CD8 (CD8⁺), Tc1 (CD8⁺TBX21⁺), Tc17 (CD8⁺RORC⁺), and TcTreg (CD8⁺FOXP3⁺). In the Phenotyping Settings, a Phenotyping Schema was created for each biomarker and were set to a Standard Associated View. For each of the six biomarkers, cells were indicated as positive for the biomarker (e.g., FOXP3+ for cells positive for FOXP3) or “Other” for cells negative for the specified biomarker. To train this phenotyping classifier, at least 5 but no more than 40 cells were manually indicated as either positive or negative for each biomarker.

Downstream analysis was performed using the PhenoptrReports Open Source R Package (https://akoyabio.github.io/phenoptrReports/index.html, Akoya Biosciences, Marlborough, MA).

To avoid unnecessary noise due to inForm tiles containing a negligible amount of tissue, a minimum threshold of 50 total cells was implemented to filter out tiles with minimal tissue. DAPI stain, a nuclear marker, was used to define individual cells. KRT was selected to aid in epithelial membrane detection (KRT positive) to distinguish epithelial and stromal tissue regions.

Cells negative for all T-cell markers were classified as “Other” (i.e., negative for CD4/CD8/TBX21/RORC/FOXP3, positive or negative for KRT). Although these cells lack specific biological identity within the context of the current study, their presence, when occurring in contiguous regions, could indicate regions suspicious for T-cell deserts.⁷ That said, these T-cell sparse regions likely still contain other immune cell types or other markers we did not measure (e.g., macrophages, dendritic cells, collagens, mucins).

Statistical Methods

For each T-cell phenotype, we assessed differences in whole slide density across three histologic types (TA, TV, SSL). Because cell counts are often overdispersed (i.e., variance exceeds that expected by a Poisson distribution), we fit a generalized linear mixed model with a negative binomial (NB) distribution and log link²⁷. To account for variable tissue area/cellularity in different lesions we included the log of the total cell area (including T-cell and “Other” cells) as an offset in the model generating mean count densities. To avoid confounding, the model was adjusted for fixed effects of age, sex, anatomic location, and lesion size. To mitigate potential batch effects, clinical study and data-collection batch were adjusted for as random effects. In each model, we generated covariate-adjusted mean T-cell densities for each of the eight immune cell types and their 95% CIs for each histologic type (SSL, TA, TV) using estimated marginal means (emmeans) package in R. We also computed rate ratios (RRs) for each T-cell type across each lesion pair (SSL vs. TA, SSL vs. TV, TA vs. TV). For the 8 cell types evaluated, there were a total of 24 comparisons, and p-values were FDR-corrected for multiple comparisons, considering FDR q < 0.05 statistically significant.

To assess spatial co-localization between pairs of T-cells, we performed a neighborhood analysis by integrating established methods.^{20, 21, 28} For each lesion, a specific T-cell type A (the “anchor”) was selected, and for each of that type of T-cell, the nearest T-cell type B (its nearest neighbor (NN)) was recorded (Supplementary Figure 1). We then computed the proportion of these neighbors that were of type B (i.e., “NN”). For anchors and neighbors, we considered all previously defined cell types, i.e., CD4⁺, CD8⁺, CD4⁺TBX21⁺, CD4⁺RORC⁺, CD4⁺FOXP3⁺, CD8⁺TBX21⁺, CD8⁺RORC⁺, CD8⁺FOXP3⁺, for a total of 64 pairs. This observed proportion was adjusted by the expected proportion under the null hypothesis of complete spatial randomness, ^{29, 30} estimated via a bootstrap approach that permuted cell labels 50 times. (Supplementary Figure 1) It should be clarified that such an adjustment is crucial to ensure that the NN results are not driven simply by cell type abundance or underlying tissue deformities³⁰. The adjusted proportions were averaged within each subject, and group-level differences across histologic types were assessed using the Kruskal-Wallis test to detect differential directed interactions (A → B). We also completed a NN analysis using closest distance to the “Other” cell types. Here we also compared the three histologic types as above but did not permute an expected proportion under the null. Although exploratory, this analysis remained statistically stringent by limiting significance to comparisons with FDR q < 0.05.

Data availability:

The T-cell density code and data is available on Code Ocean https://codeocean.com/capsule/6311812/tree/v2

Results

We analyzed samples from 120 participants (Table 1). Lesion histologic types differed by sex, clinical study, lesion size, and anatomic location. Women had a higher proportion of TVs than men. TVs occurred more often in the distal colon and rectum and were larger, on average, than SSLs and TAs. Six traditional serrated adenomas (TSAs) were excluded from analysis due to small sample size.

Table 1.

Distribution of patient, lesion, and site variables across lesion histologic types (SSL, TA, TV).

	SSL	TA	TV
Variable	(N=27)	(N=77)	(N=16)	p
Sex, male (%)	17 (63.0)	47 (61.0)	4 (25.0)	0.023
Age, years (mean (SD))	60.41 (7.79)	57.14 (7.37)	56.50 (6.99)	0.114
Lesion size, mm (mean (SD))	7.89 (3.45)	7.52 (3.14)	14.81 (10.54)	<0.001
Location (%)				0.004
Proximal	18 (66.7)	32 (46.4)	3 (18.8)
Distal	9 (33.3)	27 (39.1)	7 (43.8)
Rectal	0 (0.0)	10 (14.5)	6 (37.5)
Clinical Study (%)				0.001
ViCAPPS	14 (51.9)	63 (81.8)	10 (62.5)
DHS III-V	7 (25.9)	3 (3.9)	0 (0.0)
TCPS	6 (22.2)	11 (14.3)	6 (37.5)

ViCAPPS: Vitamin D/Calcium Polyp Prevention Study; DHS III-V: Diet and Health Studies III-V; TCPS: Tennessee Colorectal Polyp Study

Mean T-cell densities and their RRs are shown in Table 2, Supplementary Figure 2. Compared with TVs, TAs had higher densities of both helper and cytotoxic T-cells: CD4⁺ (TA:TV ratio = 2.06, q<0.001) and CD8⁺ (TA:TV 1.52, q=0.04). CD4⁺RORC⁺ cell density was also higher in TAs than in TVs (TA:TV ratio = 1.77, q=0.04). SSLs showed a similar pattern versus TVs, with higher CD4⁺ (SSL:TV ratio = 1.85, q=0.004) and CD8⁺ (SSL:TV ratio 1.78, q=0.01). In contrast, CD4⁺FOXP3⁺ and CD8⁺FOXP3⁺ cell types were enriched in TVs, though did not retain statistical significance after correcting for multiple comparisons: the SSL:TV CD4⁺FOXP3⁺ and CD8⁺FOXP3⁺ ratios were 0.41 (q=0.06) and 0.37 (q=0.09), respectively. No statistically significant differences were observed between SSLs and TAs, although SSLs showed a nonsignificant tendency toward lower CD4⁺TBX21⁺ density relative to TAs (0.54, q=0.17)

Table 2.

Mean T-cell densities per mm² and 95% CIs (panel A) and Rate ratios (q-values) (panel B) in SSLs, TAs, and TVs, adjusted for age, sex, anatomic location and corrected at q < 0.05 (bold), q < 0.10 are italicized.

A. Mean T-cell densities
Cell Type	SSL Density (95% CI)	TA Density (95% CI)	TV Density (95% CI)
CD4	2,636 (2,185–3,182)	2,945 (2,627–3,302)	1,427 (1,086–1,875)
CD8	427 (349–523)	365 (322–414)	240 (177–324)
Th1	30.0 (15.4–58.4)	55.6 (31.7–97.6)	36.5 (16.7–79.7)
Th17	218 (162–294)	285 (239–340)	161 (107–242)
Treg	33.0 (21.4–50.9)	45.8 (35.6–58.8)	80.8 (44.4–146.8)
Tc1	21.9 (12.4–38.5)	28.6 (17.8–46.1)	25.6 (13.2–49.8)
Tc17	30.4 (21.9–42.2)	33.0 (27.4–39.9)	23.3 (15.1–35.9)
TcTreg	1.14 (0.66–1.97)	1.51 (1.10–2.07)	3.09 (1.46–6.55)
B. Rate ratios of mean T-cell densities
Cell Type	SSL vs TA (q)	SSL vs TV (q)	TA vs TV (q)
CD4	0.89 (0.43)	1.85 (0.004)	2.06 (<0.001)
CD8	1.17 (0.36)	1.78 (0.01)	1.52 (0.04)
Th1	0.54 (0.17)	0.82 (0.63)	1.52 (0.25)
Th17	0.76 (0.36)	1.36 (0.42)	1.77 (0.04)
Treg	0.72 (0.36)	0.41 (0.06)	0.57 (0.15)
Tc1	0.76 (0.36)	0.85 (0.63)	1.12 (0.70)
Tc17	0.92 (0.68)	1.31 (0.48)	1.42 (0.19)
TcTreg	0.76 (0.44)	0.37 (0.09)	0.49 (0.15)

In the neighborhood analyses comparing TAs vs TVs, we identified five statistically significant anchor-NN pairs (q<0.05).(Table 3; Figure 1). Four of five pairs involved a regulatory phenotype (CD4⁺FOXP3⁺ or CD8⁺FOXP3⁺) as the nearest neighbor. The most abundant anchor-NN pairs were CD4⁺/CD8⁺FOXP3⁺, CD8⁺/CD8⁺FOXP3⁺, CD4⁺/CD4⁺FOXP3⁺, CD8⁺FOXP3⁺/CD8⁺, and CD4⁺RORC⁺/CD4⁺FOXP3⁺. These results point to CD4⁺FOXP3⁺ and CD8⁺FOXP3⁺ as more common neighbors in TVs relative to TAs. Representative images (Supplementary Figure 3) illustrate tendency for TVs relative to TAs to exhibit closer proximities between CD4⁺FOXP3⁺ and other helper T-cell types.

Table 3.

Nearest neighbor analysis: Anchor and neighbor cell types, proportions (%), and q-values for TA vs. TV (panel A), SSL vs. TV (panel B), and SSL vs. TA (panel C) comparisons.

A. TA vs. TV
Anchor Cell Type	NN Cell Type	TA (%)	TV (%)	q-value
CD4+	TcTreg	0.004	0.04	0.003
CD4+	Treg	0.51	1.31	0.02
Th17	Treg	0.64	1.46	0.02
TcTreg	CD8+	0.98	2.29	0.007
CD8+	TcTreg	0.04	0.15	0.05
B. SSL vs. TV
Anchor Cell Type	NN Cell Type	SSL (%)	TV (%)	q-value
CD4+	Treg	0.32	1.31	0.004
Treg	Treg	0.84	2.71	0.005
TcTreg	CD8+	0.86	2.29	0.01
Tc17	Tc17	0.75	1.66	0.03
Th17	Treg	0.37	1.46	0.005
Tc17	Treg	0.12	0.88	0.004
CD8+	Treg	0.06	0.69	0.004
CD4+	TcTreg	0.01	0.04	0.004
Treg	Tc17	0.04	0.35	0.04
TcTreg	Treg	0.04	1.09	0.007
C. SSL vs. TA
Anchor Cell Type	NN Cell Type	SSL (%)	TA (%)	q-value
CD8+	CD4+	8.94	15.68	0.002
CD4+	CD4+	21.65	28.99	0.004
Th17	CD4+	16.99	23.68	0.004
Tc17	CD4+	6.43	12.16	0.008
Tc1	CD4+	4.23	7.76	0.036

Figure 1.

Comparison of T-cell nearest neighbor interactions for TV vs. TA, TV vs. SSL, and TA vs. SSL. Directed arrows represent higher anchor-to-nearest neighbor T-cell type interaction in the first listed histologic type relative to the second listed histologic type. Panel A (left) depicts higher anchor-to-nearest neighbor T-cell type interactions in TVs relative to TAs. Panel B (middle) depicts higher anchor-to-nearest neighbor T-cell type interaction in TVs relative to SSLs, and Panel C (right) indicates higher anchor-to-nearest neighbor T-cell type interaction in TAs relative to SSLs, Anchor/neighbor pairs shown are statistically significant at q < 0.05

Comparing SSLs to TVs, we observed ten significant neighbor pairs (q<0.05; Figure 1). Seven of the ten anchor-NN pairs involved CD4⁺FOXP3⁺/CD8⁺FOXP3⁺ as the nearest neighbor. The most common pairs were CD8⁺FOXP3⁺/CD4⁺FOXP3⁺, followed by CD8⁺/CD4⁺FOXP3⁺, CD4⁺FOXP3⁺/CD8⁺RORC⁺, CD8⁺RORC⁺/CD4⁺FOXP3⁺, and CD4⁺/CD8⁺FOXP3⁺. (Table 3) TVs vs. SSLs also show a tendency to form CD8⁺RORC⁺-CD4⁺FOXP3⁺ reciprocal pairs, which was not observed in comparison to TAs. Representative images (Supplementary Figure 3) illustrate tendency for TVs relative to SSLs to exhibit closer proximities between CD4⁺FOXP3⁺ and other T-cells, particularly the RORC⁺ T-cell subtypes (CD4⁺RORC⁺, CD8⁺RORC⁺).

In contrast to the density analysis where no significant differences were observed between SSLs and TAs, the neighborhood analysis identified five significant anchor-NN pairs enriched in TAs relative to SSLs (q<0.05; Table 3, Figure 1). These pairs all featured CD4⁺ as the nearest neighbor in TAs relative to SSLs, with CD8⁺RORC⁺/CD4⁺ as the most prominent pair, followed by CD8⁺TBX21⁺/CD4⁺, CD8⁺/CD4⁺, CD4⁺RORC⁺/CD4⁺, and CD4⁺/CD4⁺.This pattern positions CD4⁺ as the primary neighbor in TAs relative to SSLs, as shown in the representative TA image (Supplementary Figure 3).

Next, we considered a neighborhood analysis contrasting the abundance of “Other” cells (i.e., non-immune or other immune cells) by histologic type. These analyses were not adjusted for random labeling. “Other” cells appeared as NNs more often in TVs than in TAs and, to a lesser extent, in TVs than in SSLs (Supplementary Figure 4). Six of eight T-cell phenotypes as anchor cells had higher “Other” NNs in TVs vs TAs. In SSLs, only CD4⁺FOXP3⁺, CD8⁺RORC⁺, and CD8⁺FOXP3⁺ as anchor cells showed lower “Other” neighbors relative to TVs. These patterns suggest a relative sparseness of T-cells in TVs relative to both TAs and SSLs.

Finally, we examined stroma-specific, epithelium-specific, and combined stromal-epithelial (overall minus the “Other” region) mean T-cell densities and rate ratios for histologic comparisons (Supplementary Tables 2–8). Epithelial findings were largely consistent with the overall results. Within stromal regions, SSLs had lower CD4⁺ and CD4⁺TBX21⁺ densities than TAs; otherwise, the SSL versus TV and TA versus TV contrasts generally paralleled the overall findings.

Discussion

In prior work,⁵ increasing villous histology in colorectal adenomas has been associated with lower immune marker densities of common T-cell cytokines (e.g., IL17a, IL6) and NK cell ligands, consistent with an immune-colder³¹ environment relative to TAs. Here, we extend those observations by showing that TVs have significantly lower adjusted CD4⁺ and CD8⁺ mean densities compared with both TAs and SSLs. Low T-cell densities, particularly low CD4⁺ effector and CD8⁺ cytotoxic cells, are associated with reduced immune-surveillance activity, pointing to a higher likelihood for progression to CRC and consistent with the aggressive nature of the villous lesions.^{13, 32, 33} Moreover, TVs also showed reduced RRs for Th17 relative to TAs. Th17-mediated homeostatic responses play a central role in epithelial protection, barrier maintenance, and elimination of microbial products,^{34, 35} pointing to potential barrier compromise in TVs,³⁵ also linked with tumor progression.³⁴

The neighborhood analysis comparing TVs to TAs and SSLs highlights the emergence of an immune suppressive contexture centered around Treg and TcTreg cell types. In our prior spatial analysis¹⁹ comparing adenomas (TAs/TVs) to SSLs, we observed close clustering of CD4⁺ cells with Tregs, consistent with a coordinated dampening response. In the present study, TVs showed increased proximity of T-cells to Tregs and TcTregs cell types compared to SSLs and TAs. Tregs can suppress anti-tumor immunity by promoting apoptosis of cytotoxic T-cells and diminishing the effector responses.^36–38 High abundance of Tregs are often linked with progression or poorer outcomes but in some instances may quell the proinflammatory environment, diminishing tissue damage and minimizing the chances of progression to CRC. The T-cell neighbors of Tregs seem to matter: Tregs with nearby T-cells have poorer prognosis whereas Tregs with greater proportional proximities from T-cells (like in SSLs) appear to lead to better outcomes.¹⁵ Recent evidence in nasopharyngeal cancer³⁹ and CRC²¹ identified regions with high densities of cytotoxic T-cells clustering with Tregs as poorly prognostic. In our data, we observed that TVs had a higher proportion of reciprocal CD8⁺ to TcTreg interactions than in TAs (Figure 1). TcTregs, while a rare population, are even more suppressive than Tregs.⁴⁰ TVs also exhibited greater reciprocal Tc17-to-Treg proximity in comparison to SSLs. Tc17 cells are weakly cytolytic and, in cancer, represent a dysfunctional phenotype.^{41, 42} Together our results suggest that the greater proximities to CD8⁺, TcTregs, and Tc17 cell types in TVs may foster immune suppressive and dysfunctional neighborhoods.

TVs also demonstrated frequent proximity to “Other” (non–T-cell) cell types relative to both TAs and SSLs. T-cell cold regions are reminiscent of the immune-excluded niches of CMS3 CRCs^{7, 15} or immune deserts described for CMS2 CRCs, highlighting potential importance of underlying genetic drivers of T-cell sparseness in TVs. TVs also exhibit enhanced ECM remodeling and collagen cross-linking (e.g., elevated PLOD3 ⁴³) which may physically restrict T-cell adhesion and trafficking across the TiME. TVs also show higher proliferation ⁴⁴ (e.g., Ki-67, c-Myc, cyclin D1), suggesting high turnover and metabolic stress which may predispose T-cell exhaustion and apoptosis,⁴⁵ in turn influencing immune neighborhoods.

Contrary to expectations, TAs were CD4⁺ and CD8⁺ rich lesions overall, and more like SSLs, relative to TVs. However, neighborhood analyses showed that CD8⁺ , CD4⁺, Th17, Tc1, and Tc17 cells in TAs had closer proportional proximities to CD4⁺ cells relative to SSLs, suggesting a stronger effector response as opposed to cytotoxic. These findings are in line with prior work which consistently reports that SSLs, relative to adenomas, exhibit stronger cytotoxic signaling, whereas adenomas have higher T-helper populations.^46–48 More recent studies further demonstrate that MSS CRCs, for which TAs and TVs are the parent lesions, have higher Th17, Treg, and Th2 cell populations than MSI-H CRCs which have higher Tc1 and Th1 responses.¹⁴ Moreover, spatial analysis in their work found that the interactions between Tregs, Th17, and Th2 cells in MSS CRCs clustered closer together with farther distances to Th1 and Tc1. Clustering patterns in TAs observed herein suggest that these lesions are more immunogenic than their MSS CRC counterparts. Future analysis will need to examine how TAs change in immunogenicity as they grow.

SSLs, the precursors of MSI-H CRCs, demonstrated lower Th17 cells and lower regulatory T-cell types compared to TVs. Our prior small study found that SSLs and TSAs had higher NK ligand density than conventional adenomas,⁵ and others have found that SSLs compared to conventional adenomas showed enhanced expression of cytotoxic cells (e.g., CD8⁺, NK, γδT).⁴⁶ Intraepithelial lymphocytes and PD-1/PD-L1 densities increase with progression in the serrated pathway,⁹ consistent with epithelial dominance and a tendency toward T-cell exhaustion rather than depletion of T-cells as observed in TVs. In our previous spatial analysis study, contrasting SSLs and adenomas, we observed SSLs contained abundant CD4⁺ cells near the tumor center but Treg cells were scanty and spatially distant,¹⁹ consistent with present findings. A greater spatial distance between CD8⁺ T-cells and Tregs is linked to better clinical outcomes.¹⁵ Huang et al.¹⁴ reported that Th1 and Tc1 cells closely interacted within MSI-H CRCs but were spatially distant from Treg, Th2, and Th17 cells. Our NN results confirm that SSLs relative to TAs showed reduced clustering near Treg or Th17 populations. At the same time, we have no evidence of strong Tc1 and Th1 spatial interactions in SSLs, as is observed in MSI-H CRCs, which could reflect the unexpectedly high immunogenicity of TAs and earlier phase of development of SSLs. It will be important to consider the spatial relationships in advanced SSLs (i.e., dysplasia) in future work.

Strengths of this study include single-cell resolution with simultaneous phenotyping and permutation-based neighborhood testing, which complemented density summaries. Several limitations should also be considered. First, the numbers of TVs and SSLs were smaller than the number of TAs, and the study was based on a convenience sample drawn from a larger cross-sectional study. Moreover, we were unable to determine whether the observed immune environments characterized lesions predisposed to progression or lesions in which protective, potentially cytotoxic immune infiltrates had limited further growth. Longitudinal studies are needed to define how immune contexture changes within lesion types over the course of early colorectal carcinogenesis. In addition, differences across contributing studies, including tissue storage practices before study initiation, geographic location, sample type, and sample age, may have introduced selection-related bias and may limit the generalizability of the findings, despite adjustment for clinical site and batch and randomization of samples across staining runs. A further limitation is that a single section may not fully represent the entire lesion or capture its full heterogeneity. However, whole-slide sections were used to reduce sampling bias relative to tissue microarrays, and in most cases the diagnostic pathologist (DS) reviewed the section most proximate to the immunofluorescence slide. Finally, the biological significance of low cell proportions observed in the nearest-neighbor analysis for TVs relative to other lesion types is difficult to interpret given the small numbers of positive cells. Future work should 1) investigate whether increasing villous component spatially correlates with reduced CD4⁺ and CD8⁺ infiltration and greater Treg clustering, as our findings suggest that spatial organization may shift with advancing villous architecture, 2) incorporate lesion size, degree of dysplasia (for adenomas), and potential markers of inflammation as proxies for progression as the TiME likely changes as the lesion progresses towards malignancy, 3) further evaluate tissue-specific (stroma, epithelium) neighborhood relationships among T-cells and “Other” cell types, 4) incorporate intensity-based methods to evaluate FOXP3-high and FOXP3-low subsets as recently described¹³ in the colorectal literature. 5). Future investigations will be needed to assess associations between immune contexture and underlying genomic drivers, including somatic mutations (e.g., KRAS, BRAF, APC, TP53), epigenetic alterations (e.g., CpG island methylation and MLH1 promoter methylation), as well as single-nucleotide variant analyses. In summary, regulatory proximity, especially Treg-centered clustering, and Th17-Treg neighborhoods distinguished TVs from SSLs and TAs even when bulk T-cell densities were similar. TAs showed helper-dominant, Th17-skewed architectures, whereas SSLs appeared less effector-skewed. These spatial features refine our understanding of preinvasive colorectal immunobiology and, together with prior work, point to hypotheses that warrant testing in larger, longitudinal studies.