Objective analysis and clinical significance of the spatial tumor infiltrating lymphocyte patterns in non-small cell lung cancer.

Abstract

Purpose:

The spatial arrangement of lymphocytes in the tumor bed (e.g. immune infiltrated, immune excluded, immune desert) is expected to reflect distinct immune evasion mechanisms and to associate with immunotherapy outcomes. However, data supporting these associations is scant and limited by the lack of definition for lymphocyte infiltration patterns and the subjective nature of pathology-based approaches.

Methods:

We used multiplexed immunofluorescence to study major TIL subsets with single-cell resolution in baseline whole-tissue tumor samples from NSCLC patients treated with ICI. The spatial TIL patterns were analyzed using a qualitative pathologist-based approach, and an objective analysis of TIL density ratios in tumor/stromal tissues. The association of spatial patterns with outcomes was studied for different TIL markers.

Results:

The analysis of CD8-TILs patterns using qualitative assessment identified prominent limitations including the presence of a broad spectrum of phenotypes within most tumors and limited association with outcomes. The utilization of an objective method to classify NSCLCs showed the existence of at least three subgroups with partial overlap with those defined using visual patterns. Using this strategy, a subset of cases with “immune excluded-like” tumors showed prominently worse outcomes, suggesting reduced sensitivity to ICI, however these results need to be validated. The analysis for other TIL subsets showed distinct results, underscoring the relevance of the marker selected for spatial TIL pattern evaluation and opportunities for integrating markers.

Conclusions:

Our results identified major challenges associated with the qualitative spatial TIL pattern evaluation. We devised a novel objective strategy to overcome some of these limitations that has strong biomarker potential.

Introduction

Associations between the abundance of tumor infiltrating lymphocytes (TILs) and patient outcomes have been identified across multiple solid tumor types, including non-small cell lung cancer (NSCLC).(1–3) In addition, immunostimulatory anti-cancer therapies targeting T-cell co-inhibitory receptors such as PD-1 and CTLA-4 (e.g. immune checkpoint inhibitors [ICIs]) have shown prominent anti-tumor effect, particularly in a subset of patients with so-called “T-cell inflamed” tumor microenvironments (TMEs) displaying productive local adaptive immune responses.(4–7) These associations have supported a prominent biological role of TILs in sensitivity to ICIs and highlighted their biomarker potential for treatment/patient selection. In addition, changes in TIL levels or T-cell activation metrics are widely used as pharmacodynamic markers to assess the local impact of immunostimulatory therapies in clinical research and preclinical models.

Spatial analysis of TILs using pathology-based approaches has revealed marked differences in the spatial distribution of lymphocytes (and other immune cells) in the TME with expected associations with patient outcomes.(8,9) In an attempt to incorporate the spatial immune infiltration patterns into the evaluation of anti-tumor responses, tumors have been conceptually classified into three main groups based on the relative abundance of TILs in the tumor center/core and the invasive tumor margin into: ‘immune infiltrated’, ‘immune excluded’ and ‘immune desert’.(10,11) These spatial patterns are expected to reflect the underlying biology mediating tumor immune evasion and inform ICI sensitivity/resistance. However, the role and independence of spatial TIL patterns from other TIL-based metrics such as the density/abundance, functional profiles, or clonality remain uncertain.(12) In addition, the systematic study of TIL spatial patterns has been limited by the lack of a specific definition for such patterns, absence of strategies to unequivocally recognize them, and inconsistency across studies relative to the TIL marker(s) used, sample type/size analyzed, and impact on treatment-specific outcomes.(12–14) The development of new methods/criteria to unambiguously classify tumor samples based on distinct immune infiltration patterns could contribute to decipher the molecular mechanisms that regulate the recruitment, retention, and activation of TILs in the TME and eventually also support new therapeutic opportunities to treat these patients.

Here we developed an automated approach to classify full-face tumor biopsies based on the predominant location of TILs in the tumor-cell nest and in surrounding stromal regions. We studied the association of the different immune infiltration patterns with clinical benefit to ICI in a multi-institutional cohort of NSCLC patients. This approach allowed us to classify tumors in an unbiased and reproducible manner which could be used to study immune infiltration patterns and their significance in NSCLC and other tumor types.

Materials and Methods

Samples and cohort

One hundred seventy-nine pre-treatment formalin-fixed paraffin-embedded (FPPE) whole-tissue sections from resections and core needle biopsies of NSCLC patients treated with PD-1 axis blockers were retrospectively collected from Yale University (n=75), Cleveland Clinic (n=22), UC Davis (n=21), Memorial Sloan Kettering Cancer Center (n=28) and Massachusetts General Hospital (n=33). Four of these samples were excluded due to limitations in the specimen quality including the presence of folded tissue areas, elevated autofluorescence or low cancer-cell content (<5% of malignant cells). Therefore, the total number of samples analyzed was one hundred and seventy-five. Information from the cases was locally collected from pathology reports and clinical records in each institution. This study was conducted in accordance with the principles of the Declaration of Helsinki and all tissue and clinical information were used in a de-identified fashion after approval from the Yale Internal Review Board (Yale Human Investigation Committee) protocols #9505008219 and #1608018220 or local institutional protocols, which approved the patient consent forms or waiver of consent. All patients provided written informed consent.

Multiplexed immunofluorescence staining

The following protocol was centrally performed at Yale to avoid batch effects between institutions. Whole-tissue section slides were deparaffinized and subjected to antigen retrieval using EDTA buffer (Sigma-Aldrich) pH= 8.0 and boiled for 1 hour at 96 °C in a pressure-boiling container (PT module, Lab Vision). Slides were then incubated with dual endogenous peroxidase block (#S2003: Dako) for 10 minutes at room temperature. Subsequent steps were carried out on the LabVision 360 Autostainer (Thermo-Scientific). Non-specific antigens were blocked by a 30-minute incubation in 0.3% BSA in TBST. The sequential multiplexed immunofluorescence protocol was performed using isotype-specific primary antibodies to detect epithelial tumor cells (Cytokeratin Alexa-488 conjugated, clone EA1/EA3, eBioscience), helper T-cells (CD4 IgG, 1:100, clone SP35, SpringBio), cytotoxic T-cells (CD8 IgG1k, 1:250, clone C8/144B, Dako), and B-cells (CD20 IgG2a, 1:150, clone L26, Dako). Secondary antibodies and reagents used were anti-rabbit Envision (Agilent Cat# K4003, RRID:AB_2630375) with biotinylated tyramide/Streptavidine-Alexa750 conjugate (PerkinElmer), anti-mouse IgG1k antibody (Thermo Fisher Scientific Cat# 18–4015-82, RRID:AB_11043414) with Cy3-tyramide (PerkinElmer), and anti-mouse IgG2a antibody (Abcam Cat# ab97245, RRID:AB_10680049) with Cy5-tyramide (PerkinElmer). Nuclei were stained using 4’,6-Diamidino-2-Phenylindole (DAPI). Residual horseradish peroxidase activity between incubations with secondary antibodies was eliminated by exposing the slides twice for seven minutes to a solution containing benzoic hydrazide (0.136mg) and hydrogen peroxide (50μl). Finally, slides were mounted with ProlongGold. A control TMA containing positive and negative controls was included in each batch to assess reproducibility.

Image acquisition and image analysis for TILs

The entire tumor and tumor surrounding stromal areas were scanned using a multispectral Vectra Polaris instrument (Akoya Biosciences). Each scan sample contained a range from 1–555 individual 20X magnification FOVs with a mean of 132 per case. Then, images were analyzed using InForm software (version 2.4.8, Akoya Biosciences) to obtain marker-based tissue compartments and single-cell populations. Briefly, a tissue segmentation algorithm was created and trained for each patient to define background, stroma and tumor areas. Then, cell segmentation parameters based on nuclear DAPI staining were established to obtain individual cell events, marker-based phenotype, and spatial location within the specimen. The resulting files include the compartment in which each cell was located (stroma, tumor or background) and the phenotype of each cell; helper T-cell, cytotoxic T-cells, or B-cells based on the positive expression of CD4, CD8, and CD20, respectively.

Pathologist-based evaluation of immune infiltrate patterns

All the samples were evaluated by a trained pathologist who was blinded to the clinicopathologic sample variables and patient outcomes.Due to the lack of a consensus definition to unequivocally identify these immune infiltration patterns, the predominant spatial distribution pattern of CD8 T-cells was used to classy the samples. Cases were considered immune infiltrated if there was a predominance (e.g. >2/3 of the tumor area) of immune cells within the tumor core and in close proximity with cytokeratin-positive malignant cells. Cases with a predominant location of T-cells in the tumor margin and distant from cancer-cells were categorized as immune excluded. Cases within the immune desert category were identified by the presence of <5% of TILs relative to the total cell population. The cases where two predominant immune patterns representing >1/3 of the tumor area recognized were classified as having a mixed pattern.

Objective analysis of TIL infiltration patterns

The data analyses from tumor and stromal TIL infiltration patterns were performed in R. The single-cell data were used to calculate the total number of cells in the stroma (n_S) and in the tumor (n_T) and cell counts for each lymphocyte type in both compartments (CD4/8/20_S, CD4/8/20_T) in each individual FOV. A small number of cells (0.46% of all cells) had missing phenotype data and were treated as “other” cells and excluded from the analysis. For each FOV, a ratio of the proportion of lymphocytes in the tumor $\left(n_{Lymph,T}\right)$ to the proportion of lymphocytes in the stroma $\left(n_{Lymph,S}\right)$ was calculated using the following formula: $Ratio=\frac{n_{Lymph,T}/n_T}{n_{Lymph,S}/n_S}$; where $n_T$ was the number of cells in tumor, and $n_S$ was the number of cells in stroma. This definition was used to control for the relative proportion of cells within each tissue compartment. Samples with only stroma or tumor tissue or with less than 1 mm² tumor area were excluded from the analysis (n=24).

K-means clustering of lymphocytes based on their compartment distribution.

K-means clustering is an unsupervised method used to classify data based on certain features. We used this technique to group cases independently for each TIL subset based on the distribution of their normalized ratios across all their FOVs. A distribution matrix (D) was defined in which each row represents a case, and each column corresponds to a range of ratios. The range of CD8, CD4 and CD20 was divided into 24, 21 and 22 unequally sized windows, respectively. For example, the first column in D contains the percentage of FOVs with ratio of 0 (no lymphocytes in tumor) for each patient. Di,j is the percentage of FOVs in patient i with the given lymphocyte ratio within window j. Additionally, FOVs that did not contain lymphocytes were removed. These FOVs represented 5.77%, 9.37% and 50.70% of all FOVs for CD8, CD4 and CD20, respectively. The high percentage of FOVs omitted from the CD20 analysis is due to expected focal/patchy distribution of CD20+ B-cells typically forming tertiary lymphoid structures and the absence of CD20+ lymphocytes in a large number of FOVs (29 patients did not have detectable CD20 cells). D was used to perform K-means clustering.. The determination of the optimal number of clusters was defined for each lymphocyte type based on consistency between hierarchical clustering, elbow curves and the silhouette method and was found to be K=3. The three groups were assigned to be stromal-predominant infiltrated, tumor-predominant infiltrated, or mixed infiltration according to the distributions of their ratios. To show stratification between the three groups, the clusters were visualized using heatmaps, in which increasing red intensity corresponds to larger D_i,j values.

Kappa agreement and statistical analysis

The Kappa agreement was calculated using the Kappa agreement online tool from GraphPad software (https://www.graphpad.com/quickcalcs/kappa1/). The three categories defined in the K-means clustering for each lymphocyte were used. A data table was created where the rows represented how each patient was classified by the first clustering (e.g. CD4) and the columns designate how the second or third clustering (e.g. CD8, CD20) classified the same patients. Generally, a Kappa agreement value inferior to 0.6 was considered to be moderate and <0.3 as low agreement.

For the statistical analysis, comparisons were performed using Mann-Whitney test for two groups and Chi-square test for contingency tables using GraphPad V9.5.1 (RRID:SCR_000306). The Kaplan-Meier analysis was performed using RStudio (V2023.03.1+446). Patients were divided according to the visual classification (Immune infiltrated, excluded, desert or mixed) or according to the clustering classification for CD8, CD4 and CD20 (Stroma infiltrated-like, Tumor infiltrated-like, Mix-like). Exploratory analysis combining two or more phenotype groups were performed based on the similarity with survival patterns. The survival analysis included 5-year overall survival and the p-values and Hazard Ratios (HR) reported are based on the Log-rank (Mantel-Cox) test. All p values were based on two-sided test, and all values <0.05 were considered statistically significant.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author.

Results

To investigate the clinical significance of TIL spatial patterns in NSCLC patients treated with ICI, we evaluated the predominant spatial pattern of CD8+ effector TILs using qualitative, pathology-based visual analysis of pre-treatment whole-tissue section samples from 175 NSCLC patients treated with PD-1 axis blockers stained with a multiplexed immunofluorescence panel including markers for cancer-cells (cytokeratin), T- (CD4 and CD8) and B-cells (CD20) ((6) and methods section). The analysis of the tumor specimens captured the entire tissue area, allowing for the evaluation of TIL patterns across cancer-cell nests and in the surrounding (non-tumor) stromal areas. The visual inspection of the samples aimed to classify the tumors based on the three previously described patterns⁹ (Figure 1).

Figure 1. Graphical representation of previously described tumor-immune infiltration patterns.

Three immune infiltration patters have been described based on the predominant spatial distribution of CD8 T-cells in the tumor core and/or invasive margin. The immune infiltrated pattern is characterized by a predominance of CD8 T-cells in the tumor core (left panels), the immune excluded phenotype is denoted by predominance of T-cells in the invasive margin (middle panels), whereas the immune desert pattern can be identified by the lack of CD8 T-cells (right panels). The panels show a graphical representation (upper panel) and a representative fluorescent image from NSCLCs (bottom panel) showing such patterns. Cytokeratin (CK) was used for epithelial cancer-cells (green), CD8 for effector T-cells (red) and DAPI for all cellular nuclei (blue).

The initial evaluation demonstrated that nearly all the samples showed multiple CD8+ T-cell infiltration spatial patterns with variable proportions across a continuous spectrum (Figure 2A). The visual analysis of samples also revealed the existence of previously unrecognized/undesignated predominant spatial TIL patterns with overlapping features with the three previously described groups. For example, we identified a subset of cases showing abundant TILs in the tumor center/core area, but with clear immune-cell exclusion features characterized by CD8+ T-cells located almost exclusively in the (non-tumor cell) stromal area surrounding the cancer-cell nests and limited contact with malignant cells (Figure 2B). Using the qualitative classification based on T-cell abundance in the tumor core and the perilesional/invasive margin, these samples would be classified as “immune infiltrated” despite the clear microscopic immune exclusion.

Figure 2. Pathologist-based classification of tumor-immune patterns based on CD8 T-cell location.

(A) Representative immunofluorescence caption from a NSCLC with a mixed TIL infiltration pattern, where areas with immune desert-(A1) and immune-infiltrated patterns (A2) can be simultaneously recognized. (B) Immunofluorescence image showing a representative NSCLC with tumor immune infiltration in the tumor center but absence of T-cell penetration into the cancer-cell nests, consistent with microscopic exclusion. (C) Pie chart showing the percentage of cases within each spatial pattern based on pathology-based visual analysis (n=175). (D) Comparison of the total tissue area for the different immune infiltration patterns in NSCLC. (E and F) Kaplan-Meier analysis of the 5-year overall survival of NSCLC patients stratified by pathology-based immune infiltration patterns (E); and for the combination of immune infiltrated and mixed patterns versus excluded pattern (F).

Therefore, the predominant pattern present in an estimated >2/3 of the tumor area was used to classify the samples. As shown in Figure 2C, the majority of the cases were classified as “immune infiltrated” (76%), followed by smaller fractions of “immune excluded” (12%) and “immune desert” (6%). In addition, another 6% of cases showed different CD8+ TILs spatial patterns present in >1/3 of the tumor area and were therefore classified as displaying a “mixed” pattern. The examination of NSCLC cases also revealed a possible role of the T-cell density and sample size variation across cases as factors influencing the predominant qualitative TIL spatial pattern, where cases with higher TIL density and larger specimen area were more commonly classified using the predefined patterns than cases with low immune infiltration and/or smaller tumor tissue area. In this regard, cases in the “immune desert” group included cases with prominently smaller tissue sample size/area than the other groups with a mean tissue area of 10.35 mm² as compared to 77.25 mm², 100.48 mm² and 121.31 mm² for “immune infiltrated”, “excluded” and “mixed”, respectively (Figure 2D).

The analysis of the overall survival of patients with tumors showing the four spatial TIL infiltration patterns showed overlapping results with no clear survival advantage across the groups (Figure 2E). A numerically higher overall survival hazard ratio was seen in cases with an immune excluded phenotype relative to the other spatial patterns suggesting worse outcomes in this patient population, but the difference was mild and did not meet statistical significance (HR:1.35 [CI:0.70–2.58]; Figure 2F). Together, these results show the limitations of classifying NSCLCs based on pre-defined qualitative TIL spatial patterns using pathologist-based estimations, particularly in light of the heterogeneity in biopsy size, sites and approach, as is common seen in NSCLC.

To objectively represent and analyze the spatial TIL infiltration patterns, we used single-cell segmentation and fluorescence co-localization strategies to establish the number and density of CD8+ T-cells in the cytokeratin-expressing cancer-cell nests and within cytokeratin negative non-tumor stromal areas across cases⁶. To assess the specific spatial patterns, we calculated the ratio between the number of CD8+ T-cells in the tumor and the stromal areas using digital sample tiles (e.g. fields of view or FOVs) acquired using 20X microscopic magnification. The entire specimens were captured allowing the analysis of the T-cell compartment location predominance and assessment of regional differences within and across cases. To classify the samples into subgroups, we performed K-means clustering of the CD8+ intratumor to stromal ratios resulting in three stable groups with overlapping features as compared to previously reported qualitative spatial patterns (Figure 3A). The first group, designated as “immune infiltrated-like”, comprised 19% of the cohort and showed a predominance of CD8 T-cell infiltration in cancer-cell nests. The second group, designated as “immune excluded-like”, constituted 19% of the cohort and included cases in which the majority of the specimen showed low intratumor T-cells and higher peritumoral stromal infiltration. The third group designated as, “immune mixed-like” represented 62% of the population and was characterized by the presence of both tumor-predominant infiltrated areas and stromal-predominant infiltrated areas. Cases with zero values in the numerator (no CD8 T-cells in tumor) and denominator (no CD8 T-cells in stroma) were considered “immune-desert like” but none of the samples met these criteria supporting that all cases had detectable CD8+ TILs (Figure 3B).

Figure 3. Objective analysis of spatial TIL patterns in NSCLC based on predominant CD8 T-cell compartment location.

(A) Heatmap showing the K-means clustering based on the distribution of the normalized ratios for CD8 T-cells across all fields of view (FOVs), each column represents the ratio distribution of an individual NSCLC patient. The bar color represents the percentage of FOVs within a given ratio. (B) Pie chart representing the percentage of cases within each pattern based on the K-means clustering of CD8-T cells. (C and D) Kaplan-Meier analysis for 5-year overall survival of NSCLC patients stratified based on CD8 T-cell K- means clustering groups (C) and for the combination of immune infiltrated and mixed-like patterns versus excluded-like pattern (D).

To determine the clinical significance of the objective spatial CD8+ TIL patterns, we analyzed the association with survival after ICI. As shown in Figure 3C the overall survival curves in Kaplan-Meier analysis showed overlaps between the “Immune infiltrated-like” and “Mixed-like” groups with a numerically longer survival as compared to the “Immune excluded-like” cluster. An exploratory subgroup analysis comparing the survival of patients with immune-excluded like tumors to all the other spatial subsets combined, showed a significantly shorter survival of this group with 2.16 higher mortality risk (HR: 2.16 [1.15–4.03]) (Figure 3D).

To determine the possible impact of using different TIL subsets to spatially classify NSCLCs, we conducted tumor/stromal ratio clustering analysis of the cohort using CD4+ T-cells or CD20+ B-cells. For CD4+ T-cells, the same three groups were obtained after the K-means clustering analysis (Figure 4A). However, the proportion of cases in each group was very different with the “immune infiltrated-like” being the largest cluster (41%), followed by ‘mixed-like” (33%) and “Immune excluded-like” (26%). Similarly, to CD8 T-cell infiltration, none of the cases were classified as “Immune desert-like” (Figure 4B). The survival analysis for CD4+ T-cells clusters didn’t show differences in survival (Figure 4C); however, a numerically higher survival ratio was observed for patients in the “Immune infiltrated-like” subgroup as compared to the rest (HR: 1.40 (0.93–2.12), Figure 4D).

Figure 4. Classification of the spatial TIL patterns in NSCLC based on CD4 T-cells.

(A) K-means clustering based on the distribution of the normalized ratios for CD4 T-cells across all fields of view (FOVs) and patients. (B) Pie chart representing the percentage of cases within each pattern based on the K-means clustering of CD4-T cells. (C and D) Kaplan-Meier analysis for 5-year overall survival of NSCLC patients stratified based on CD4 T-cell K-means clustering groups (C) and for the comparison of infiltrated-like patterns versus the rest of patterns (D).

The analysis for CD20+ B-cells typically forming tertiary lymphoid structures in the peritumoral area showed a majority of cases with “Immune excluded-like” cases (45%), followed by “Mixed like” (31%) and “Immune infiltrated-like” (12%). Interestingly, B-cells were the only subset of TILs in which we could identify an “Immune desert-like” phenotype (12%), where most of the FOVs were absent of B-cells in the tumor-cell nests and stromal compartment and was associated with poor survival (Figures 5A-C). The survival analysis for CD20 B-cell clusters showed that cases with an “Immune infiltrated-like” pattern had a significant longer overall survival as compared to the rest of the groups (HR: 1.76 (1.00–3.09), Figures 5D and E). The concordance between the pattern classification of NSCLCs using the same strategy but different TIL subpopulations was moderate to low with kappa coefficients ranging from (0.05 to 0.42), indicating that the TIL analyzed prominently impacts the spatial classification.

Figure 5. Classification of the spatial TIL patterns in NSCLC based on CD20 B-cells.

(A) K-means clustering based on the distribution of the normalized ratios for CD20 B-cells across all FOVs and patients. (B) Pie chart representing the percentage of cases within each pattern based on the K-means clustering of CD20 B-cells. (C) Bar plot depicting the kappa agreement between immune infiltration patterns for different TILs subsets. (D and E) Kaplan-Meier analysis for 5-year overall survival of NSCLC patients stratified based on CD20 B-cell K-means clustering groups (D) and for the comparison of infiltrated-like patterns versus the rest of patterns (E).

We finally analyzed the association between the objective analysis of spatial patterns and major clinicopathological variables. We found a significantly higher proportion of CD8+ infiltrated-like cases in PD-L1 positive tumors than in PD-L1 negative (27% vs 6%) and a higher proportion of CD8+ Mixed-like cases in PD-L1 negative samples (75% vs 55%). Similar results were observed for CD20 B-cells, where patients with Infiltrated-like pattern were more likely to be PD-L1 positive (17% vs 6%), whereas there was a higher proportion of Mixed-like cases in PD-L1 negative samples (45% vs 20%). No other significant associations were found, indicating that the TIL infiltration patterns do not segregate with specific clinicopathologic patient or tumor subsets (Table 1).

Table 1.

Association between TILs immune infiltration patterns and clinicopathological variables

Parameter	CD4 Infiltrated-like No. (%)	CD4 Mixed pattern No. (%)	CD4 Excluded-like No. (%)	P value	CD8 Infiltrated-like No. (%)	CD8 Mixed pattern No. (%)	CD8 Excluded-like No. (%)	P value	CD20 Infiltrated-like No. (%)	CD20 Mixed pattern No. (%)	CD20 Excluded-like No. (%)	CD20 Desert-like No. (%)	P value
Age, y
<65	29 (45%)	21 (33%)	14 (22%)	0.43	17 (27%)	37 (58%)	10 (16%)	0.29	13 (20%)	18 (28%)	27 (42%)	6 (9%)	0.24
≥65	22 (37%)	19 (32%)	19 (32%)		9 (15%)	40 (67%)	11 (18%)		5 (8%)	23 (38%)	25 (42%)	7 (12%)
Sex
Male	30 (44%)	21 (31%)	17 (25%)	0.74	14 (21%)	43 (63%)	11 (16%)	0.98	9 (13%)	27 (40%)	24 (35%)	8 (12%)	0.24
Female	26 (38%)	24 (35%)	19 (28%)		14 (20%)	43 (62%)	12 (17%)		10 (14%)	18 (26%)	35 (51%)	6 (9%)
Smoker
No	11 (46%)	6 (25%)	7 (29%)	0.70	4 (17%)	15 (63%)	5 (21%)	0.79	4 (17%)	7 (29%)	8 (33%)	5 (21%)	0.26
Yes	45 (40%)	38 (34%)	29 (26%)		24 (21%)	70 (63%)	18 (16%)		15 (13%)	38 (34%)	50 (45%)	9 (8%)
Histology
ADC	39 (43%)	30 (33%)	21 (23%)	0.65	18 (20%)	56 (62%)	16 (18%)	0.96	14 (16%)	27 (30%)	41 (46%)	8 (9%)	0.86
SCC	13 (36%)	12 (33%)	11 (31%)		8 (22%)	22 (61%)	6 (17%)		5 (14%)	11 (31%)	15 (42%)	5 (14%)
Stage
I-II	8 (28%)	13 (45%)	8 (28%)	0.20	5 (17%)	19(66%)	5 (17%)	0.87	3 (10%)	11 (38%)	14 (48%)	1 (3%)	0.45
III-IV	48 (45%)	32 (30%)	26 (25%)		23 (22%)	66 (62%)	17 (16%)		16 (15%)	33 (31%)	44 (42%)	13 (12%)
PD-L1 Status
PD-L1<1%	20 (39%)	16 (31%)	15 (29%)	0.52	3 (6%)	38 (75%)	10 (20%)	0.01	3 (6%)	23 (45%)	21 (41%)	4 (8%)	0.02
PD-L1≥1%	29 (46%)	20 (32%)	14 (22%)		17 (27%)	35 (55%)	12 (19%)		11 (17%)	13 (20%)	30 (47%)	10 (16%)

Discussion

Previous studies have endorsed the implementation of pathology-based assessment of T-cell infiltration patterns in the TME as a biomarker to predict response to immunotherapy in solid tumors.(8,10,15) As example, the Immunoscore is a digital pathology-based method to classify tumors based on the predominant spatial location of CD3 and CD8 T-cells in the core and invasive margin of tumors and has been associated with prognosis in colorectal cancer.(16,17) Similar studies in breast cancer have also identified distinct immune infiltration patterns with different clinical course.(8,11,18) However, the data supporting this association in other tumor types, and its possible role with immunotherapy outcomes are very limited, which is further complicated by the lack of agreement in the definition of these patterns and the subjective nature of pathology-based approaches.(12)

The analysis of cases using qualitative assessment of pre-defined visual patterns identified prominent limitations of this approach including the presence of a broad spectrum of phenotypes within most tumors, where the existence of a single spatial pattern was exceptional. In addition, the overwhelming majority of cases were assigned to a single spatial subtype (e.g. immune infiltrated). The limitations of pathology-based TILs assessments due to low to moderate inter- and intra- observer agreement have been extensively reported.(19,20) In addition, new evidence has challenged the idea of uniform immune spatial patterns due to the demonstration of differential levels of intra-tumor spatial immune heterogeneity.(6) Additionally, we found prominent associations between the sample size and the presence of specific spatial TIL patterns, where cases with smaller tissue samples were more likely to be classified into the immune desert phenotype, indicating the possibility of sampling/scoring bias. Future studies should consider special handling or even exclusion of small samples showing an immune desert pattern. Careful evaluation of the cases also identified spatial patterns not previously described, such as those with T-cell infiltration within the tumor center core, but marked T-cell exclusion from cancer-cell nests. These results underscore the difficulties of employing pathology-based approaches to characterize the spatial distribution of TILs.

The utilization of an objective method to classify NSCLCs based on their predominant spatial TIL distribution showed the existence of at least three predominant subgroups with more balanced representation within the cohort and partial overlap with those defined using visual inspection patterns. Interestingly, using the method, we described a new spatial pattern, reported as mixed pattern, with characteristics of the immune infiltrated and immune excluded phenotypes, again highlighting the relevance of regional differences within areas of a same tumor sample. Using this strategy, a subset of cases with “immune excluded-like” tumors corresponding to 19% of the cohort showed significantly worse outcomes suggesting reduced sensitivity to ICI.

Recent studies have reported a positive clinical association between higher levels of CD4+ T-cells and B-cells and response to immunotherapy in multiple tumor types.(21–23) Despite this clinically meaningful associations, the role of the spatial distribution of these cells has not been previously explored. The analysis of cases using CD4 as a marker for helper T-cells showed a dissimilar distribution of spatial patterns as compared to CD8+ T-cells and a trend towards better prognosis in those patients with CD4+ T-cell infiltrated-like tumors. The analysis for CD20+ B-cells was the only cell subtype where we could identify an immune desert-phenotype, which was also associated with the worse survival, supporting a prominent role of B-cell infiltration pattern. In addition, previous studies have reported a positive correlation between the levels of different TILs subsets;(3,6) however, the low agreement in the spatial distribution of these subsets suggests differential mechanism of recruitment and retention of immune cells into the TME, underscoring the relevance of the marker selected for spatial TIL pattern evaluation and the opportunities for marker integration.

Our results are limited by the use of a single retrospective multi-institutional NSCLC patient cohort receiving ICI treatment in an uncontrolled fashion. In addition, some samples included archive resection specimens preceding the immunotherapy treatment initiation by variable time periods, which could affect the results in unpredictable ways. However, the use of such a cohort is more likely to represent a real-world scenario and reflect the conditions in which NSCLC patients are evaluated and managed clinically. Future studies will be required to validate these findings in a prospective setting.

In conclusion, we have conducted a detailed analysis of the spatial TIL patterns in NSCLC. Our results identified major challenges associated with qualitative pattern evaluation and devised a novel objective strategy to overcome some of these limitations that has strong biomarker potential.

Translational Relevance.

The spatial arrangement of lymphocytes in the tumor bed (e.g., immune infiltrated, immune excluded and immune desert) has been considered as a prominent tumor immune rejection feature and is expected to predict response to immunotherapy in solid tumors. However, data supporting these associations is limited due to the lack of consensus definitions and the subjective nature of pathology-based approaches to identify such spatial TIL infiltration patterns. We developed an objective and automated approach to classify tumors based on the predominant TIL infiltration pattern in the tumor bed. Using unsupervised analysis, we identified three major TIL infiltration patterns with distinct clinical outcomes to PD-1 axis inhibitors in patients with advanced NSCLC. The strategy established in our study reveals the impact of spatial TIL infiltration patterns in immunotherapy outcomes of NSCLC patients and establishes an objective framework to use this feature in clinical research and potentially as a predictive biomarker.