Spatial heterogeneity of infiltrating immune cells in the tumor microenvironment of non-small cell lung cancer

Highlights

• •Immune cells showed heterogeneous distribution in TME of NSCLC.
• •Efficientnet-B3-based model realized tumor and immune region segmentation.
• •TIME could be classified into three groups according to CD8+ T-cell density.
• •Single chromogenic IHC slices overlapping could realize mIHC simulation.

Abstract

Tumor-infiltrating lymphocytes (TILs) are essential components of the tumor microenvironment (TME) of non-small cell lung cancer (NSCLC). Still, it is difficult to describe due to their heterogeneity. In this study, five cell markers from NSCLC patients were analyzed. We segmented tumor cells (TCs) and TILs using Efficientnet-B3 and explored their quantitative information and spatial distribution. After that, we simulated multiplex immunohistochemistry (mIHC) by overlapping continuous single chromogenic IHCs slices. As a result, the proportion and the density of programmed cell death-ligand 1 (PD-L1)-positive TCs were the highest in the core. CD8+ T cells were the closest to the tumor (median distance: 41.71 μm), while PD-1+T cells were the most distant (median distance: 62.2μm), and our study found that most lymphocytes clustered together within the peritumoral range of 10-30 μm where cross-talk with TCs could be achieved. We also found that the classification of TME could be achieved using CD8+ T-cell density, which is correlated with the prognosis of patients. In addition, we achieved single chromogenic IHC slices overlap based on CD4-stained IHC slices. We explored the number and spatial distribution of cells in heterogeneous TME of NSCLC patients and achieved TME classification. We also found a way to show the co-expression of multiple molecules economically.

Introduction

Non-small cell lung cancer (NSCLC) remains one of the leading causes of cancer death worldwide [1,2], and immunotherapy represented by programmed cell death-1 (PD-1)/programmed cell death-ligand 1 (PD-L1) inhibitors have improved the prognosis of some patients [[3], [4], [5], [6], [7]]. PD-1/PD-L1 detection is now recommended by guidelines for selecting patients suitable for immune checkpoint inhibitors (ICIs) [8]. However, in various tumors, numerous patients with positive PD-L1 expression still do not respond to ICIs [9]. In addition, several studies have shown that patients with low PD-L1 expression on tumors show a durable response [[10], [11], [12]]. This situation may be partly because of the heterogeneous tumor microenvironment (TME) [13].

The tumor immune microenvironment (TIME) is a complex system containing different populations of T, B, and natural killer cells, and other components, which differs significantly between tumor types [14] and between individuals with the same kind of tumor [15,16]. The presence, variety, quantity, and spatial distribution of tumor-infiltrating lymphocytes (TILs) is increasingly considered essential biomarkers for therapeutic efficacy and prognostic prediction of various cancers [[17], [18], [19], [20]]. Among them, the prognostic value of CD8+ TILs has been widely recognized [21,22]. Understanding the composition and spatial distribution of TILs and determining their relationship with prognosis is critical for improving the efficacy of immunotherapy and implementing personalized anti-tumor therapy [23].

Based on this, several studies used immunohistochemistry (IHC) to establish a series of tumor prognostic scoring systems, which could identify patients who would benefit from immunotherapy. However, most of these studies are limited to quantitative analysis [[24], [25], [26], [27], [28]]. With the development of emerging technologies such as single-cell sequencing, heterogeneous TME has been more deeply recognized [29,30]. However, as these methods would compromise the structural integrity of tumors, the spatial distribution of TILs and their relationship to each other remains a critical issue to be resolved. Multiple immunohistochemistry (mIHC) and spatial transcriptomics have emerged as potential tools to explore cell interactions in a single section. However, due to the high cost and color interference caused by overlapping signals when multiple markers in mIHC are co-expressed [31,32], the widespread application of mIHC and spatial transcriptomics is challenged.

Moreover, previous studies proposed that TME could be divided into three subgroups according to the association between immune cells and tumor cells: immune-inflamed tumors, immune-excluded tumors, and immune-desert tumors [33,34], among which ICIs had a good effect in immune-inflamed tumors [35]. However, due to the complex classification of these subgroups, there is currently no unified standard for the category of TME in NSCLC patients. Based on this, we plan to perform IHC staining of different immune checkpoints on consecutive slices, accomplish cell quantity and spatial analysis, and divide TMEs into the previously mentioned groups. However, manual labeling and cell counting on multiple slices require a lot of human resources and time and might exist intra- and inter-observer errors. With the development of artificial intelligence and digital pathology [[36], [37], [38]], it is possible to quantify and analyze cells automatically. Using this, we can break through the limitation of single chromogenic IHCs for cell recognition and analysis.

We used a cell staining- and morphological-based pattern recognition algorithm to explore intratumoral heterogeneity and cell interaction in the TME. With it, we identified four types of cells from consecutive single chromogenic IHC slices for quantitative and spatial analysis of immune checkpoint expression. In addition, we plan to explore the classification methods of TME, which may assist clinical treatment selection. At the same time, we are also attempting to apply deep learning methods on continuous single chromogenic IHC slices to realize mIHC simulation, which may enable low-cost, multi-immune signal analysis in the future.

Methods

Patients

Thirty-one primary tumor specimens from patients with postoperative pathological diagnosis of NSCLC were included. All patients received surgical treatment in Shanghai Pulmonary Hospital Affiliated to Tongji University in 2018 and were regularly followed up after surgery. No other treatment was received before surgery. The last follow-up date was October 2022. Excluded patients with an unclear diagnosis or lack of follow-up data. Patients were staged according to the 8th edition of the tumor, nodes, and metastasis (TNM) stage classification system. This study was approved by Shanghai Pulmonary Hospital Affiliated to Tongji University with written informed consent from all patients (ethical number K20-022).

Immunohistochemistry staining

Multi-point interval sampling (average interval: 1-1.5 cm) was performed on surgical specimens from 31 patients with NSCLC and recorded as A1-A6. Specific IHC staining of CD4 (MXB Biotechnologies, China), CD8 (MXB Biotechnologies, China), PD-1 (Golden bridge Zhongshan, Beijing ZM-0381), PD-L1 (22C3, Dako) and Foxp3(CST98377S, Cell Signaling Technology) was performed on serial slices obtained from each sampling point. Finally, we obtained 540 images. The specific information of the sample is shown in Fig. 1a.

Fig. 1.

The data information and an example of a training set image labeling and the segmentation process. a. The horizontal axis represents different patients, the vertical axis represents the sampling locations of the patients, and the points with different colors represent diverse immune checkpoints. b. Example of a training set image labeling (left to right: original image, circle annotation using LabelMe, circle annotation information map to the original image, mask of circle annotation). c. An region of interest image of 1500*1500 pixels shows the results of tumor region segmentation, immune region segmentation, and cell segmentation. The second column shows the predicted tumor and immune regions in purple and blue, the third column represents the segmented tumor and immune cells, and the fourth column circles the immune cells and the tumor cells in the original image.

Tumor cell segmentation

We used EfficientNet-B3 as the image segmentation model. Two types of semantic segmentation networks were constructed: tumor cell (TC) regions and tumor-associated immune cell regions. Two pathologists labeled the tumor areas and tumor-associated immune cell areas in randomly cropped 95 slices (1000*1000 pixels) of 40X magnification fields of view (FOV) via the LabelMe platform (http://labelme.csail.mit.edu/Release3.0/) (Fig. 1). The sketched images were randomly clipped to 384*384 size and were entered into EfficientNet-B3 as the training set of the segmentation model (batch size, 18; learning rate, 1.0^e-05). After 200 epochs of training, the dice value on the training set is 0.917, and the dice value on the test set is 0.908. After that, the obtained 540 original images (with an average of about 80,000 * 74,000 pixels) were divided into 4-6 sub-images from the upper left corner to the lower right corner and then put into the established segmentation model for the segmentation of tumor region and immune region (Fig. 1b).

Since the cell morphological structure in the IHC map was similar to that of the nucleus in the image stained by hematoxylin-eosin (HE) staining, we directly applied the trained multi-scale nuclear segmentation model with an attention mechanism to the cell segmentation of the IHC map using the method of transfer learning. Finally, we completed the segmentation of immune cells and TCs (Fig. 1c) [39].

Defining positive and negative cells and Regions of interest (ROIs)

In the color interval of the IHC staining, the region where the segmentation mask overlapped with the brown region was regarded as the positive cell region. In contrast, the non-overlapping region was regarded as the negative cell region. To further study the tumor and lymphocyte landscape in NSCLC, we used a simple U-Net to distinguish the tumor area from the stroma area for all IHC images and calculated the tumor and stroma area size. The ROI was divided into three parts: core, inner, and outer, with the maximum difference of tumor stroma ratio between groups as the cut-off value.

Calculation of the cell density and distance

Cell density (cell/mm²), as a parameter with quantitative and spatial information, was calculated as the ratio of the number of positive cells to the area of the tissue section. To further explore the spatial data between cells, we defined the median value of the closest distance between the two types of cells as the most relative distance between the two types of cells. Fast Library for Approximate Nearest Neighbors (FLANN) is a collection of algorithms for searching large data sets and high-dimensional features. Based on the previous cell segmentation results, we can obtain the coordinate information of any cells. We used the FLANN algorithm to calculate the closest distance between different cells, which was implemented in Python using the pyflann package. This study used this method to define the distance between tumor cells and positive T cells in each slice (CD4/CD8/Foxp3/PD-L1/PD-1).

Virtual mIHCs

We used the image matching algorithm to match the reduced and rotated graphs of continuous slices marked with different immune molecules of the same sample and calculate the rotation angle. In the feature matching of the image, we adopted accelerated-KAZE (AKAZE) feature algorithm, an improved version of the scale-invariant feature transform (SIFT) feature algorithm. The algorithm assumed nonlinear diffusion filtering to construct the scale space of the image and retained more large-image edge features. The descriptors obtained by AKAZE feature algorithm have rotation invariance, scale invariance, illumination invariance, and space invariance. We used the AKAZE feature extraction algorithm integrated with OpenCV to extract the features of the images. Brute Force (BF) match-matching was used to match the feature information of the two pictures and obtain the position information of key matching points. Moreover, we received the transformation matrix through the perspective transformation of the images. After matching through the transmission transformation, finally, the new photos were obtained.

The next step is to calculate the overlapping information of the positive immune regions of the matched successive slices. Therefore, we need to match the original IHC map, carry out image transmission transformation according to the immune region and cell segmentation results, and calculate the corresponding overlapping information. Due to computational limitations, 16x-32x down-sampling results are used for transmission transformation.

Statistical Analysis

Kaplan-Meier methods were used to assess patients' OS and DFS. Spearman and Pearson correlation was used to evaluate cell density correlation. The Mann–Whitney U and Kruskal-Wallis test was applied to analyze the differences for continuous variables, and the Chi-square analysis was applied for categorical variables. P<0.05 was considered to indicate statistical significance. All analyses were accomplished by Python (version 3.6.5) and R software (version 3.6.3).

Results

Characteristics of the patient cohort

Our cohort included 540 IHC images of 31 NSCLC patients. In this cohort, most patients (n=22, 71.0%) were male and 10 (32.3%) were smokers. Twenty-two patients (71.0%) were younger than age 70 years. Three patients (9.7%) had stage IA NSCLC, 11 patients (35.5%) had stage IB NSCLC, four patients (12.9%) had stage IIA NSCLC, six patients (19.4%) had stage IIB NSCLC, and six patients (19.4%) had stage IIIA NSCLC. The remaining patients had stage IIIB NSCLC (n=1, 3.1%). Twenty patients (64.5%) had lung adenocarcinoma (LUAD), and the rest were squamous cell carcinoma (LUSC). The surgical method of 26 patients was lobe resection (83.9%) (Table 1).

Table 1.

Clinical characteristics of non-small cell lung cancer patients.

Characteristics	All patients (n = 31)	%
Smoking Status
Former/Current	10	32.3
Never	21	67.7
Sex
Male	22	71.0
Female	9	29.0
Age, Years
<70	22	71.0
≥70	9	29.0
TNM-Stage
IA	3	9.7
IB	11	35.5
IIA	4	12.9
IIB	6	19.4
IIIA	6	19.4
IIIB	1	3.1
Histology
Squamous carcinoma	11	35.5
Adenocarcinoma	20	64.5
Surgery procedures
Lobectomy	26	83.9
Non-lobectomy	5	16.1

Quantitative analysis of TILs and TCs

The quantitative information of positive lymphocytes at A1-A6 is shown in Fig. 2a. According to the most considerable difference in tumor stroma ratio of ROI, they were divided into core, inner, and outer types. We analyzed the density and proportion of positive cell populations in these categories to explore the distribution of positive lymphocytes and positive TCs at different ROIs. The percentage of positive TCs and TILs is calculated as the ratio of positive TCs (or TILs) on the slice to the number of all TCs (or TILs). The proportion (core: 12.38%; inner: 7.67%; outer: 10.22%) and number (core: 886.22 cells/mm²; inner: 557.87 cells/mm²; outer: 583.60 cells/mm²) of PD-L1-positive tumor cells in all ROIs were the highest, the number of cells was higher in outer, and core, but lower in inner, and the density was the highest in the core. The proportion and density of PD-L1-positive tumor cells showed a similar tendency, reaching the maximum value at the core. The frequency and proportion of Foxp3-positive tumor cells were the lowest in the outer (45.51 cells/mm², 0.68%) and the core (40.76 cells/mm², 0.69%) among all immune molecules, while the density and proportion of PD-1-positive tumor cells (40.44 cells/mm², 0.40%) in the inner region were the lowest. Foxp3-positive tumor cells were evenly distributed in the three ROIs. The proportion and number of PD-1-positive tumor cells in the outer were significantly higher than those in the inner and core (core: 0.76%, 41.46 cells/mm²; inner: 0.40%, 40.44 cells/mm²; outer: 0.91%, 62.30 cells/mm², P<0.05)(Fig. 2b-c).

Fig. 2.

Quantitative analysis. a. Proportion of positive tumor infiltrating lymphocytes (TILs) at A1-A6; b-c. Density distribution of different tumor cells (TCs) and positive TILs at core/inner/outer; d-e. Proportion distribution of different positive TILs and TCs at core/inner/outer.

CD4-, CD8-, and PD-L1-positive lymphocytes were abundant in all ROIs. The number and proportion of PD-L1- and CD8-positive T cells in the core were basically the same (PD-L1: 2418.33 cells/mm², 36.28%; CD8: 2480.92 cells/mm², 35.35%). CD4+ T cells, compared with the above two kinds of T cells, the density was lower but accounted for more in the core (2038.89 cells/mm², 39.14%). The number and proportion of CD8+ and CD4+ T cells in the inner and the outer regions (inner: CD8: 2098.64 cells/mm², 28.70%; CD4: 2224.33 cells/mm², 40.94%; outer: CD8: 2269.17 cells/mm², 32.56%; CD4: 2240.30 cells/mm², 41.54%) was basically the same, and significantly higher than PD-L1+ T cells (inner: 1797.49 cells/mm², 25.73%; outer: 1487.74 cells/mm², 22.99%). CD4+ T cells decreased gradually from the outer to the core, but CD8+ and PD-L1+ T cells gradually increased in density and proportion, reaching the maximum number and proportion in the core (P<0.05). The density and proportion of PD-1+ T cells (core: 475.88 cells/mm², 7.02%; inner: 443.13 cells/mm², 6.53%; outer: 606.60 cells/mm², 8.91%) and Foxp3+ T cells (core: 651.65 cells/mm², 10.00%; inner: 575.93 cells/mm², 8.56%; outer: 608.92 cells/mm², 9.76%) were relatively low in TILs. PD-1+ T cells showed a gradual increase trend from core to outer and reached the maximum at the outer (P<0.05), while the number and proportion of Foxp3+ T cells did not change significantly in the three ROIs (P>0.05)(Fig. 2d-e).

Spatial analysis of TILs and TCs

The median distance of TCs to positive lymphocytes at A1-A6 are shown in Fig. 3a. To explore the spatial information between TCs and T cells in-depth, we divided ROI into three groups as the method mentioned earlier and analyzed the distance between cells at the core. As shown in Fig. 3b, CD8+ T cells were the closest cell group to the tumor (median distance: 41.71 μm), while PD1+ T cells were the farthest from the tumor (median distance: 62.2 μm).

Fig. 3.

Spatial analysis. a. The median distribution box diagrams show the distance between tumor cells (TCs) and the nearest positive tumor infiltrating lymphocytes (TILs) at A1-A6; b. Distribution of the median distance between TCs and the nearest positive TILs in the core; c-d. Quantity and proportion of positive TILs in different distance ranges of TCs at the core.

To further integrate the information on quantity and spatial distance between cells, we counted the number of positive lymphocytes in the core within a radius of 50 μm around the tumor. The results showed that CD8+ T cells were the most, followed by CD4+ and PD-L1+ T cells. Within 10 μm, the number of CD8+, CD4+, and PD-L1+ T cells was significantly higher than that of other T cells (P<0.001), and the number of CD8+ T cells was the highest (density: 2467 cell/mm²). When the radius was expanded to 10-20 μm, CD8+, CD4+, and PD-L1+ T cells were still the main groups most associated with TCs, and the number of these three types of cells increased significantly and reached a peak. When the radius expanded to 20-30 μm, 30-40 μm, and 40-50 μm, CD8+, CD4+, and PD-L1+ T cells were still the main cell groups, but their numbers decreased gradually (Fig. 3c).

We also counted the proportion of positive T cells within the radius of 50 μm around the tumor in the core. The results showed that within this range, CD8+, CD4+, and PD-L1+ T cells accounted for a relatively high proportion, and there was no significant difference within these three groups (median percentage of 4.0% vs. 4.1% vs. 3.8%). The proportions of Foxp3+ T cells and PD1+ T cells were low, and there was no significant difference within the two groups. Within 10 μm, the ratio of CD8+, CD4+, and PD-L1+ T cells was significantly more substantial than other lymphocytes (P<0.001). When the radius was expanded to 10-20 μm, CD8+, CD4+, and PD-L1+ T cells were still the main groups most associated with tumor cells, and their proportion reached the maximum. This observation is also maintained when the radius is enlarged to 20-30 μm. Compared with cells within a radius of 10 μm, the number and proportion of all T cells within 10-30 μm were significantly increased. These results indicated that a large number of lymphocytes were located at the core of tumor and most of which were located 10-30 μm away from the tumor, a distance that allowed cells to interact directly with each other (Fig. 3d).

Colocalization of different TILs in the TME

The coexistence of different types of TILs in the TME has been demonstrated, but their interactions need to be further explored. To understand the relationship between various TILs in the core, Spearman coefficient was used to analyze the correlation between different types of positive T cells [40]. Rank correlation coefficients of different positive lymphocyte densities showed that PD1+ T cells was positively correlated with CD4+ T cells (r=0.41), CD8+ T cells (r=0.44), and Foxp3+ T cells (r=0.67). CD8+ T cells were positively correlated with Foxp3+ T cells (r=0.48). However, there was no significant correlation between PD-L1+T cells and the other four types of lymphocytes regarding colocalization (Fig. 4a).

Fig. 4.

The correlation of tumor infiltrating lymphocytes (TILs) and the classification of tumor-associated immune microenvironments (TIMEs). a. Spearman coefficient of positive TILs in the core; b. K-means clustering, TIME was divided into immune-inflamed, immune-excluded, and immune-desert; c. The mean density of CD8+T cells in three types of TIMEs; d. Average distance between CD8+T cells and TCs in three kinds of TIMEs; e. Overall survival (OS) of the three categories of patients; f. Disease-free survival (DFS) of the three categories of patients.

Classification of TIMEs and associated prognosis

Previous studies have shown that CD8+ T-cell density could distinguish three major cancer immunophenotypes: “immune-desert,” “immune-excluded,” and “immune-inflamed” [34]. Based on this, we took each individual patient as a unit, conducted K-means clustering based on the average CD8+ T-cell density of A1 to A6. Subsequently, we found the cutoff values of immune-inflamed, immune-excluded, and immune-desert, and then we classified patients into three types of cancer immunophenotypes based on this where cluster1 corresponds to "immune-inflamed," cluster2 corresponds to "immune-excluded," and cluster3 corresponds to "immune-desert" (Fig. 4b). The CD8+ T-cell density of patients in the three categories is shown in Fig. 4c, and the average distance between CD8+ T cells and TCs in the three categories is shown in Fig. 4d. To verify the predictive effect of TIME typing on patient prognosis, we further explored the overall survival (OS) and disease-free survival (DFS) of the three categories of patients. The results showed that the TIME typing based on average CD8+T cell density had a tendency to distinguish patients' prognosis (P < 0.05) (Fig. 4e-f).

Virtual mIHC

We took the CD4 IHC staining slices of each sample for reference and rotated and matched other IHC staining slice maps, as shown in Fig. 5. Because the transmission transformation of feature matching algorithm in computer vision was adopted in this study, IHC maps of some markers could not be matched. Ultimately, 57 matching results were obtained after calculation (57 tissue samples simultaneously stained for CD4, CD8, PD1, and Foxp3).

Fig. 5.

Matching and overlapping process of continuous slices (1808888-A3 as an example)

After matching, we calculated the proportion of overlapping immune regions occupying the total area of immune regions. In this study, we calculated CD4+ CD8- Foxp3- PD1+, CD4+ CD8- Foxp3- PD-1-, CD8+ CD4- Foxp3- PD-1+, CD8+ CD4- Foxp3- PD-1-, CD4+ CD8+ Foxp3- PD-1+, CD4+ CD8+ Foxp3- PD-1-, CD4+ CD8- Foxp3+ PD-1+ and CD4+ CD8- Foxp3+ PD-1- overlapping area of T cells in the total immune area. At the same time we draw a box plot for matching proportions. The results showed that the overlaps of CD4+ CD8- Foxp3- PD-1-, CD8+ CD4- Foxp3- PD-1- and CD4+ CD8+ Foxp3- PD-1- were large, among which CD4+ CD8+ Foxp3- PD-1- accounted for the largest proportion (Fig. 6). This indicated that two types of T cells, CD4+ and CD8+, often accumulated in the same area.

Fig. 6.

Boxplot based on the proportional values of the eight categories of overlapping regions.

Discussion

Our study established tumor and immune region segmentation models based on Efficientnet-B3. Based on this, we performed the quantitative and spatial analysis of five immune molecules on TCs and TILs, revealed the spatial heterogeneity distribution of immune cells in the TIME. Then, the TIME was classified according to CD8+ T-cell density and the prognosis was explored. In addition, we economically and efficiently demonstrated the co-expression of multiple antibody markers could be explored by overlapping multiple staining of successive mono-stained IHC slices.

The pathological assessment of TILs is emerging as a promising biomarker for solid tumors. Multiple studies have concluded that high CD8+ T-cell density is strongly associated with better overall survival (OS) in patients [41,42]. Therefore, we explored the quantity and spatial distribution of TILs in surgical patients with NSCLC. The results showed that CD8 and CD4 positive lymphocytes were mainly distributed in the core, and most of them were located within 10-30 μm of the tumor, where they could directly interact with tumor cells [43]. Previous studies have shown that effective anti-tumor response was mainly maintained through closed cross-talk between effector immune cells and tumor cells, and CD8+ TILs near tumor cells are closely associated with better DFS and OS [43]. Based on this, we classified the TIME as "immune-desert," "immune-excluded," and "immune-inflamed" based on the density of CD8+ T cells in the core. The results showed that the density of CD8+ T cells in the core area was correlated with OS and DFS of patients, which further proved the potential clinical application value of TILs evaluation in tumor core area.

PD-L1 is expressed on various cell surfaces and can bind to the PD-1 receptor on T cells to inhibit T-cell activation [42]. TCs achieve immune escape through the high expression of PD-L1. In contrast, immune checkpoint inhibitors block the PD-1/PD-L1 signaling pathway to restore the normal function of T cells to kill TCs, thus achieving clinical benefits for patients [44,45]. Food and Drug Administration has approved the expression of PD-L1 as a biomarker to guide the choice of ICIs in advanced NSCLC [46]. Still, the expression of PD-L1 only refers explicitly to the expression of TCs, and the expression of PD-L1 expressed on the surface of other non-tumor cells is not counted. However, more and more clinical evidence shows that PD-L1 expressed on the non-tumor surface also inhibits T-cell activation [47,48]. Some patients with negative PD-L1 expression responded to ICIs [49], which might be related to it [50]. Our results showed that T cells could also expressed large amounts of PD-L1, which further confirmed the conclusions of previous studies [51].

Foxp3, as one of the key transcription factors controlling the development and immunosuppressive function of regulatory T cells (Tregs), is the signature molecule of Tregs. Treg inhibits T cells through direct intercellular action and cytokine inhibition [52]. Studies have shown that high Foxp3 expression was associated with poor prognosis in multiple tumors [53]. In addition, studies have shown that breast cancer tumor cells could also express a small amount of Foxp3 [54], and similar results were found in our research: a small number of Foxp3-positive TCs were observed. Because Foxp3 has an immunosuppressive effect, TCs might suppress the immune response through Foxp3 expression to achieve immune escape. The specific cellular and molecular pathways still need further study. At the same time, our study found that Foxp3 mainly accumulated at a distance from the tumor to play an immunosuppressive role. Previous studies have shown that baseline tumor size was associated with prognosis in patients receiving immunotherapy [55,56]. Together with our findings, it might be related to the distribution of tumor-associated immune cells in the tumor microenvironment.

Multi-immunohistochemical (mIHC) staining is a reliable high-throughput method that simultaneously identifies multiple immune checkpoints [57,58]. However, there are problems with the cross-expression of antibodies and tissue damage during antibody stripping [59]. High costs also limit its widespread use [32]. Meanwhile, as a clinical transformation platform, the mIHC test platform also has high standards for pathology laboratories regarding computing power, storage space, analysis speed, and server [60]. At the same time, traditional IHC is a practical and cost-effective test method. To compensate for the deficiency of a single target in precision medicine, we proposed matching and overlapping continuous single chromogenic IHC slices to generate virtual mIHC images to quantify antibody expression in target cell populations. This provides an economical means of detecting co-expression of multiple immune checkpoints.

Our study still has some limitations: First, it was retrospective and included a relatively small cohort, which made it challenging to adjust for variables such as TNM stage and to assess differences between groups accurately. Therefore, this study can be regarded as a preliminary exploration of the heterogeneous TME of NSCLC, laying a foundation for larger scale investigation and verification in the future. We will expand the sample size in future studies to further explore the heterogeneity of TME. Second, we took multi-point interval sampling of the tumor area, which may cause the information of the middle part of the sample to be ignored. However, previous studies have shown that random sampling of the tumor center could better represent the whole tumor [41]. Of course, we will continue to conduct in-depth research in the future and apply existing techniques to whole-slide images to verify our conclusions further. Third, our study classified the TIME into immune-inflamed, immune-excluded, and immune-desert. Because it is a biomarker related to immunotherapy efficacy, subsequent immunotherapy cohorts are needed to validate the value of our TIME classification in prognostic prediction.

In conclusion, our findings provided an overview of the spatial heterogeneity of TILs in NSCLC and offered clues to subsequent immunotherapy-related studies. At the same time, we also implemented a virtual m-IHC technique with overlapping continuous single IHC-stained slices, providing a convenient and economical method for TME analysis.

Data availability

Source data for figures are provided with the paper. Other data that support the findings of this study are available from the corresponding author upon reasonable request.

Novelty and impact

Based on quantitative and spatial exploration of heterogeneous tumor microenvironment, we could classify tumor microenvironment of non-small cell lung cancer patients into different types that were associated with prognosis. Furthermore, virtual multiplex immunohistochemistry (IHC) by overlapping consecutive single chromogenic IHCs slices could show the co-expression of multiple molecules economically.

CRediT authorship contribution statement

Xinyue Liu: Conceptualization, Writing – original draft, Investigation, Methodology, Formal analysis, Project administration. Yan Kong: Methodology, Software, Formal analysis. Youwen Qian: Validation, Visualization. Haoyue Guo: Conceptualization. Lishu Zhao: Data curation. Hao Wang: Data curation. Kandi Xu: Data curation. Li Ye: Validation. Yujin Liu: Validation. Hui Lu: Software. Yayi He: Funding acquisition, Writing – review & editing.

Declaration of competing interest

There were no conflicts and interests.