High inter‐follicular spatial co‐localization of CD8+FOXP3+ with CD4+CD8+ cells predicts favorable outcome in follicular lymphoma

Abstract The spatial architecture of the lymphoid tissue in follicular lymphoma (FL) presents unique challenges to studying its immune microenvironment. We investigated the spatial interplay of T cells, macrophages, myeloid cells and natural killer T cells using multispectral immunofluorescence images of diagnostic biopsies of 32 patients. A deep learning‐based image analysis pipeline was tailored to the needs of follicular lymphoma spatial histology research, enabling the identification of different immune cells within and outside neoplastic follicles. We analyzed the density and spatial co‐localization of immune cells in the inter‐follicular and intra‐follicular regions of follicular lymphoma. Low inter‐follicular density of CD8+FOXP3+ cells and co‐localization of CD8+FOXP3+ with CD4+CD8+ cells were significantly associated with relapse (p = 0.0057 and p = 0.0019, respectively) and shorter time to progression after first‐line treatment (Logrank p = 0.0097 and log‐rank p = 0.0093, respectively). A low inter‐follicular density of CD8+FOXP3+ cells is associated with increased risk of relapse independent of follicular lymphoma international prognostic index (FLIPI) (p = 0.038, Hazard ratio (HR) = 0.42 [0.19, 0.95], but not independent of co‐localization of CD8+FOXP3+ with CD4+CD8+ cells (p = 0.43). Co‐localization of CD8+FOXP3+ with CD4+CD8+ cells is predictors of time to relapse independent of the FLIPI score and density of CD8+FOXP3+ cells (p = 0.027, HR = 0.0019 [7.19 × 10−6, 0.49], This suggests a potential role of inter‐follicular CD8+FOXP3+ and CD4+CD8+ cells in the disease progression of FL, warranting further validation on larger patient cohorts.


| INTRODUCTION
In the western world, follicular lymphoma (FL) is the second most common subtype of non-Hodgkin lymphoma, accounting for between 20% and 25% of cases. 1,2 he disease tends to follow an indolent remitting and relapsing course, with great individual variability.4][5][6] Early identification of refractory/early relapsing cases and investigation of the biological basis is currently a major challenge. 7e tumor microenvironment (TME) plays a key role in the clinical course of FL.Two immune response gene expression signatures, IR1 and IR2, were identified to be predictive of long and short survival, respectively in FL. 8 The IR1 signature included genes encoding both T cells and macrophage molecules, whereas the IR2 comprised genes expressed in macrophages, dendritic cells or both.This and subsequent molecular studies, [8][9][10] suggested the potential importance of immune surveillance in FL raising the possibility of novel immune approaches.The role of immune T cells, [11][12][13][14][15][16][17][18] macrophages, 14,15,[19][20][21] NK/T cells 22,23 and myeloid cells 24,25 were investigated in FL generating inconsistent results.These studies were conducted on a different cohort of patients who might have different characteristics and analyzed using different computational pipelines that could potentially hamper consistency and comparison of the prognostic power of the different immune cells groups.Moreover, the composition of the intra-follicular areas, containing neoplastic cells, are distinct from the inter-follicular areas.However, most of the previous studies considered the FL TME as one homogenous ecosystem.The pattern of immune infiltration in these two sites is predictive of outcome. 15,16,26Thus, investigating the spatial interaction of immune cells in the two regions could provide new insight into the biology of FL.However, no computational image analysis software tailored to these cell compartments are available.
9][30] It has been shown that this technology also serves as a discovery tool to identify novel cell populations associated with tumor progression.Automated microscopy analysis is a more reliable approach to enumerate infiltrating cell populations but there has been limited use of deeplearning analysis to study the microenvironment in FL. 31,32 Thus, we decided to use multispectral immunofluorescence (M-IF) images containing 15 immune cell markers to (1) develop a deep learning-based method to identify cell phenotypes, (2) develop cell distribution and spatial analysis pipeline tailored to the tissue compartments of FL, and (3) identify novel immune phenotypes associated with time to progression (TTP).Such an integrated, high throughput approach enabled us to identify the clinically relevant spatial distribution of immune cell subsets in the inter-follicular area of FL TME. and performed in accordance with the ethical standards of the 1964 Helsinki declaration and its later amendments.All patients provided written informed consent.Time to progression, measured from diagnosis to relapse/progression time, was used as a clinical endpoint.

| Antibodies panels and multispectral immunofluorescence
Two to four micron thick formalin-fixed paraffin-embedded tissue sections of 32 FL and 4 normal tonsils were subjected to M-IF applying a series of antibodies (Supplementary Table

| Statistical analysis
All statistical analyses were carried out using the Python programming language.All correlation values were measured using the nonparametric Spearman test.The p-values were computed using the two-sided unpaired, non-parametric Wilcoxon method, considering p < 0.05 as significant.To correct for multiple testing, we applied Benjamini-Hochberg (BH) method.Time to progression was estimated using the Kaplan-Meier method and two-tailed log-rank test using Lifelines (v0.25.4)Python package. 33Multivariate Cox regression analysis was performed using the Lifelines (v0.25.4)Python package. 33

| Code availability
All methods and analyses were implemented in Python.For reproducibility and ease of sharing, the code and its dependencies are packed into a Docker container.The code runs on both local and high-performance clusters.The code is accessible upon request to the corresponding author or the first author.

| Patient clinical characteristics
The clinical characteristics of the 32 patients included in this study are summarized in Table 1.The histological grade for all patients who relapsed varied between grades 1 and 2, but only one patient showed a focal grade 3A pattern.After a median follow-up of 10.

| Deep learning for immune phenotyping in multispectral immunofluorescence images
The immune spatial phenotypes were investigated using four M-IF panels for T cells, macrophages, myeloid cells, and natural killer T (NK/T) cells.The M-IF images of diagnostic biopsies were acquired using the VECTRA 3 platform and the regions of interest defined as

Clinical characteristics
All patients (N; %)  "tiles" were selected from different areas to capture the heterogeneity in the tissue section (Figure 1B-D).
Though, M-IF is a high-throughput approach to characterize immune phenotypes landscape in tissue sections, the intermixing of colors deters accurate identification of cells and discernment of touching cells.In a computerized analysis of M-IF, color (intensity) is the main discriminating feature between different cell phenotypes.
Panels could have different colors (Figure 1C), and a supervised model trained on M-IF data from one panel might not be generalized to another.However, irrespective of the number of markers/colors used in the M-IF panels, the deconvoluted images in all panels have only brown (positive) and blue (DAPI, negative) colors (Figure 1D).
This suggests that a model trained on deconvoluted images from one panel data could be generalized to the other panels.Thus, our newly developed deep learning image analysis pipeline systematically detects and classifies cells in M-IF images from the deconvoluted images (Figure 2A and Methods in Supplementary Data).

| Image and spatial analysis tailored to follicular lymphoma cellular compartments
The composition and spatial organization of immune cells in FL was analyzed in the intra-follicular (within follicles) and inter-follicular (outside follicles) regions, by developing a tissue and follicles segmentation pipeline (Figure 2B and Methods in Supplementary Data).
This approach was designed to investigate whether distinct patterns of immune cell infiltrates in the two micro-ecosystems represent a robust tool to predict clinical outcome.
To quantify cells spatial co-localization and immune cell composition, we applied a Morisita-Horn index [34][35][36] to the regions within and outside the follicles separately (Figure 2C and Methods in Supplementary Data) and demonstrated differences between the two cellular compartments.

| Deep learning models accurately map single cells in multispectral immunofluorescence images
To enable the automated detection and classification of diverse cell types in M-IF images, we developed a deep learning pipeline (Figure 2A).The number of cells detected by the proposed deep learning method significantly correlated with the cells annotated by the expert pathologists (Spearman r = 0.94, p = 1.82 � 10 −12
For cell classification, we obtained an area under the curve (AUC) of 0.995 (Supplementary Figure 1a) on separately held test data.
Only 193 cells out of 10, 134 cells were wrongly classified (Supplementary Figure 1b).Visualization of the features learned by the convolutional neural network using uniform manifold approximation and projection (UMAP) dimensionality reduction demonstrated that cells of different classes are separated (Figure 3B and Supplementary Figure 1c).
The deep learning models were trained on the immune T cells panel data (Supplementary Table 2) and the models generalize to the other panels.The density of CD8+ cells across multiple panels is significantly correlated (Figure 3C-D and Supplementary Figure 1d) and cell classification AUC of 0.998 was obtained on single-cell collected from the macrophages and NK/T cell panels, panels unseen during model training (Supplementary Figure 1e-f).The deep learning models and co-expression analysis (Figure 2A and Methods in Supplementary Data) allowed us to identify different cell phenotypes in each panel (Figure 3E).

| Decreased inter-follicular CD8+FOXP3+ cells is associated with relapse
To identify prognostic cell types outside the neoplastic follicles, we first computed cell density.A significantly lower density of CD8 +FOXP3+ cells outside the neoplastic follicles was found in diagnostic samples of patients who later relapsed, compared to those patients who did not relapse (BH corrected p = 0.0057, Figure 4A).
We also analyzed the association of the density of the remaining immune cells listed in Figure 3E in the inter-follicular and intrafollicular regions of FL with disease relapse and patient TTP (Supplementary Figures 2-3 and Supplementary Tables 3-4).However, none of these cells was significantly associated with disease relapse and patient TTP after applying multiple test corrections.

| Clinical relevance of immune cell co-localization
To understand the spatial interaction of the inter-follicular CD8 +FOXP3+ cells with the other T cell subsets in the TME, we first explored their spatial neighborhood using nearest neighbor (NN) analysis (Figure 5A-B).For each CD8+FOXP3+ cell, we identified the NN cell phenotype and computed the distance in the tissue space (Figure 5A).In the inter-follicular region, CD4+CD8+ and CD4 +FOXP3+ NN cells tend to localize closer to CD8+FOXP3+ cells than other T cell subsets including CD4-CD8+FOXP3-, CD4+CD8-FOXP3-, and CD4-CD8-FOXP3+ cells.
We then asked if the co-localization of these T cell subsets with CD8+FOXP3+ cells in the inter-follicular region is associated with relapse and TTP.To quantify spatial co-localization, we computed the Morisita-Horn index, which increases in value if there is a high degree of spatial colocalization between two variables (Methods in + with CD4+CD8+ cells was associated with longer TTP (Logrank p = 0.0093, Figure 5D).

| DISCUSSION
In Our study shows that in FL, the inter-follicular CD8+FOXP3+ T cells are prognostic and positively correlate with patients TTP.Even though these cells account for a small fraction of CD8+ immune T cells in the TME of FL, it has been shown that rare cell types such as antigenspecific T cells can play a crucial role in the development of cancer. 37,38 1970 Gershon and Kondo described a pool of CD8+ regulatory T cells which support tumorigenesis. 39This type of cells was subsequently described in prostate, 40,41 colon 42 and non-small cell lung cancer. 41In a mice model, Mayer et.al also showed that CD8 +FOXP3+ cells have a light suppressive function. 43However, other studies support our results and showed that CD8+FOXP3+ cells have anti-tumor cytotoxic activity.Using flow cytometry on mice treated with GMCSF secreting HER-2/neu vaccine, CD8+FOXP3+ T cells were abundantly found in regressing and immunogenic tumors. 44CD8 +FOXP3+ is a phenotype for anti-tumor T cells, and such cells have a similar expression profile of activated T cells. 43,44iggering an effective immune response promotes the expansion of CD8+FOXP3+ lymphocytes. 44In a mice model, Le et al.
demonstrated that CD4+ T cells promote the expansion of tumorspecific T cells such as CD8+FOXP3+ cells by secreting stimulatory cytokines like IL-2 and TGF-β. 44Moreover, K. Y et al. showed that CD8+FOXP3+ T cells are immunosuppressive, but, their inhibitor function could be altered using Toll-Like Receptor (TLR)-8 signaling 40,45 suggesting this could be utilized by immunotherapeutic strategies in cancer. 40,45Furthermore, it is reported that TLR signaling pathways interact with RCHOP immunochemotherapy that is used in FL. 46 Further functional studies are needed to understand whether the CD8+FOXP3+ T cells in the TME of FL have an "innate" anti-tumor function or this is modulated by exposure to the immunochemotherapy treatment.
To investigate the spatial interaction of inter-follicular CD8 +FOXP3+ cells with other cell types identified by our approach, we applied spatial co-localization analysis (Figure 2C and Methods in Supplementary Data).We found that higher co-localization of CD8 +FOXP3+ cells with CD4+CD8+ in the inter-follicular regions is associated with favorable TTP in FL.Previous studies described CD4 +CD8+ cells as effector anti-tumor T cells in a series of tumors for example, cutaneous T-cell lymphoma, 47,48 nodular lymphocytepredominant Hodgkin lymphoma, 49 and melanoma. 48The CD4 +CD8+ T cells have a high IL-2 cytokine secretion profile 47 and interestingly, a high level of IL-2 is reported to enhance the cytotoxic activity of CD8+ Tregs cells. 50These results suggest that CD4+CD8 + cells might have boosted the anti-tumoral activity of CD8+FOXP3 + cells through an IL-2 dependent pathway and thus resulting in a favorable patient outcome.
In this study, we showed that the combination of M-IF, deep learning and regional spatial analysis is a promising strategy to identify novel immune cell phenotypes in FL that could stratify relapsed versus not relapsed FL patients, and predict TTP.However, our study has some limitations.Firstly, the data used to train and evaluate the deep learning models were generated from the same lab and device, which could impact the generalizability of the models.
Secondly, the manual annotation of the neoplastic follicles is laborious, thus an automated deep learning methodology is a valuable development.Moreover, the small sample size and the highly variable indolent nature of FL 51 could introduce bias to the results.
In summary, our study showed that low density of CD8+FOXP3 + cells or low co-localization of CD8+FOXP3+ with CD4+CD8+ outside the neoplastic follicles (Figure 6) is associated with relapse and shorter TTP in FL patients treated with R-CHOP or R-CVP.The Patients diagnosed at Papa Giovanni XXIII Hospital (Bergamo, Italy) with grade I-IIIa FL between 01-Jan-2006 and 31-Dec-2015, treated with standard R-CHOP or R-CVP and with the availability of the diagnostic surgical biopsy were eligible for this study.Clinical information of 39 patients was gathered from the electronic charts.The diagnosis of FL was confirmed by three haematopathologists (Teresa Marafioti [TM], Alan Ramsay [AR], and Sabine Pomplun [SP]); who reviewed independently the morphology assessed by using H&E staining.The relevant immunostaining evaluated included CD20, CD3, BCL-2, BCL-6, CD10, CD21, MIB-1.All cases expressed BCL-2, CD10 and BCL-6 and no areas of diffuse growth pattern were present.The diagnosis of FL, followed the criteria of the revised fourth edition of the WHO classification of tumors of hematopoietic and lymphoid tissues.The exclusion criteria applied included: stage I disease, bendamustine therapy and rituximab maintenance.Seven cases were excluded of which six showed suboptimal tissue sections affecting staining, and the additional case had received bendamustine treatment.The final number of analyzed cases was 32.The study was approved by the Ethics Committee (approval number REG. 197/17) 542 -HAGOS ET AL.

1
) panels detailed below to study specific immune-cell populations: (a) Immune T cells: CD4/CD8/FOXP3/PD-1; (b) Tumor-associated macrophages: CD68/CD163/CD206/PD-L1; (c) Myeloid cells: CD8/CD11b/CD14/ CD15 and (d) Natural killer T cells: CD8/Granzyme B (Granz B)/ Granulysin/CD16/CD56. Before data collection, experimental conditions (e.g., primary antibody dilution and time of incubation, order of antibodies immunostaining and fluorophores selection) of M-IF staining were optimized and validated against singleplex chromogenic immunostaining protocols on consecutive sections of reactive tonsil and classical FL specimens retrieved from the diagnostic files of TM, AR and SP's institution.Staining was carried out on a Leica BOND RX automated immunostainer (Leica Microsystems, Milton Keynes, UK) using tyramide signal amplification (TSA)-based Opal method (Opal 7-Color Automation immunohistochemistry Kit, Akoya Biosciences, Marlborough, MA, USA; Catalog No. NEL811001 KT).SOPs for each antibodies panels has been compiled and can be provided upon request to the corresponding authors.
4 years (range 0.25-15.2years), 23 patients remained alive.A total of nine deaths occurred and the causes were related to the progression of FL (3 cases); transformation to diffuse large B-cell lymphoma (1 case); secondary cancer (1 case), unknown (occurring >10 years posttreatment; 2 cases), complication of allogeneic stem-cell transplantation (1 case) and acute hepatitis (1 case).Fifteen patients relapsed/progressed after a median of 2.83 years (range 0.6-14.8years).The remaining 17 patients did not relapse after a median observation of 11.5 years (range 0.25-14.8years; Figure 1A) with 4 deaths not related with disease progression.

F I G U R E 1
Details of the study cohort and M-IF images.A Consolidated Standards of Reporting Trials (CONSORT) Diagram.B Illustrative image showing tile selection from a tissue section using VECTRA 3 platform.C Sample images from the four panels used in this study.Multispectral immunofluorescence images were acquired using VECTRA 3 with multiple markers in each panel.The scale bar is 10 μm.D Deconvoluted images for the immune T cell panel M-IF image in C. 544 -HAGOS ET AL.

2 F I G U R E 3
Computational deep learning and image processing pipelines.A Multi-stage deep learning cell detection and classification pipeline for multispectral immunofluorescence (M-IF).Cell detection and classification were applied to the N deconvoluted images.For cell detection, we applied CONCORDe-Net (Methods in Supplementary Data).CONCORDe-Net generates the x and y locations of the center of the nucleus of the cells.A patch centered on these cells' location was extracted and fed to a VGG style convolutional neural network based cell classifier (conv = convolution, maxp = maxpooling layer).The classifier generates the probability of the input patch being positive (p +ve ) and negative (p -ve ) for a marker.The class of the cell was assigned to positive (p +ve ≥ p -ve ) or negative (p +ve < p -ve ).This repeats for the N deconvoluted images.The positive cells from the N deconvoluted images were mapped onto one plane for co-expression analysis.Then, cells were spatially mapped on M-IF images.The bar indicates a 10 μm resolution.B A diagram of tissue and follicles segmentation.Follicles are manually segmented by an expert pathologist.Morphological ops = morphological operations.C Spatial Voronoi tessellation of within and outside follicles tissue compartments for Morisita-horn index spatial analysis.Following the follicle segmentation in (B), the area within the follicle and outside follicle were divided into smaller polygons called "Voronoi".Using location cells data in (A) cells can be mapped onto these polygons to apply spatial analysis 546 -HAGOS ET AL.Evaluation and validation of deep learning models.A Correlation between the number of cells annotated by an expert and the number of cells detected by deep learning (DL).A dot represents a human-annotated rectangular region, which contains around 450 cells.B Two-dimensional representation of deep learned features after Uniform manifold approximation and projection (UMAP) dimensionality reduction along with their marginal distributions.Negative and positive classes represent cells negative and positive for a marker, respectively.C, D Deep learning model validation.The deep learning models were trained on immune T cell panel data.The trained model was then applied to all panels.The density of CD8+ cells (cells per 1000 μm 2 ) across different panels was significantly correlated.A dot represents a sample or patient.All correlation values were computed using a non-parametric Spearman correlation.E After detecting cells on the deconvoluted images, using the proposed co-expression analysis (Methods in Supplementary Data), we were able to spatially map cells expressing single or multiple markers in all panels, which allows us to visually validate the deep learning models and co-expression analysis on m-IF images.The scale bar is 10 μm F I G U R E 5 Prognostic spatial co-localization of cell populations outside the neoplastic follicles.A Graphical representation of CD8 +FOXP3+ cells nearest neighbor (NN) cells outside the neoplastic follicles.B The distribution of the distance of NN cells of different phenotypes.C Boxplot showing the difference in co-localization of CD8+FOXP3+ with CD4+CD8+ cells outside follicles between relapsed (n = 15) and not relapsed (n = 17).D Kaplan-Meier curves illustrating time to progression (TTP) of patients dichotomized using median colocalization of CD8+FOXP3+ with CD4+CD8+ cells outside follicles.E Forest plots showing multivariate Cox regression analyses.Continuous values were used for the density and spatial localization parameters.Follicular lymphoma international prognostic index (FLIPI).For statistical comparisons among groups, a two-sided, nonparametric, unpaired, Wilcoxon signed-rank test was used, unless stated otherwise.To correct for multiple testing, we applied Benjamini-Hochberg (BH) this study, we developed a deep learning-based image processing pipeline for M-IF images to decipher the immune microenvironment in FL.To the best of our knowledge, this is the first study to analyze the distribution and spatial interaction of immune cells in the interand intra-follicular compartments of FL using high throughput M-IF images and deep learning image analysis.In FL, the abundance and distribution of immune cells within and outside neoplastic follicles are distinct and heterogeneous, 26 and thus the spatial interaction of the cells.The combination of M-IF and deep learning-based image analysis focused on FL compartments enabled us to identify novel prognostic cell populations and spatial patterns in FL.
density of CD8+FOXP3+ and co-localization of CD8 +FOXP3+ with CD4+CD8+ appear to be predictive of TTP independent of FLIPI score, and combining these features with FLIPI scores could improve prognostication.These findings require validation on a large cohort of FL patients treated with the same or different treatment regimens.