StemCell Therapy

Clinico-pathological features of the GC patients

Eighty patients were enrolled in this study between July 2014 and December 2019 (Table1). The median age of the patients was 60 years (range, 5466 years), and most patients were men (76.3%). Among the 60 patients subjected to immunotherapy, 21 were treated with standard-of-care anti-PD-1/PD-L1 antibodies and 39 were treated as part of clinical trials (NCT03472365, NCT03713905). Archived pre-treatment samples were available from all patients. Ten (12.5%) patients were EBV(+) and 11 (13.75%) had confirmed deficient DNA mismatch repair (dMMR) GC.

To investigate the landscape of TIICs within the GC specimens, we quantified the density and spatial location of immune cells in 80 full-face formalin-fixed paraffin-embedded (FFPE) samples via m-IHC staining; the multiplex determination of the sub-cellular expression of 16 proteins was performed (Fig.1a). First, haematoxylin and eosin (H&E)-stained tissue sections were reviewed by two pathologists (S.Y. and H.Y.J.) to identify tumour core (TC), invasion margin (IM), and peri-tumoural normal (N) areas, which we refer to as regions of interest (ROIs) (Fig.1b). The m-IHC panels analysed are depicted in Fig.1cf. A total of 6488 high-power fields (TC: 4477, IM: 993, N: 1018) were imaged for all patients. A supervised image analysis system (inForm) was used to classify each image into tumour nests and stromal areas based on machine learning (Fig.1g). Cell segmentation showed nuclear, cytoplasmic and membranous outlines. Cell phenotyping data were obtained based on the positivity and relative intensity of all markers in one panel. The cell density, calculated for all regions (tumour + stroma), was measured separately in the tumour and stroma. Thereafter, TIICs were analysed at the single-cell level and 26 major populations were characterised (Supplementary Fig.1a).

a Schematic representation of the experimental design and analytical methods used in this study. b Selection of the regions of interest (ROIs) in representative images of haematoxylin and eosin (H&E)-stained formalin-fixed paraffin-embedded tissues. TC, tumour core; IM, invasion margin; N, normal tissue. Scale bar: 3mm. cf Representative composite and single-stained images of the multiplex immunohistochemistry panels used. Scale bar: 200m. g Overview of the automated image analysis pipeline.

To examine the distribution of TIICs within the tumour microenvironment, we analysed their spatial density in the TC, IM, and N areas. The enriched co-occurrence of immune populations defines a structured immune environment (Supplementary Fig.1a). A significant increase in the overall density of CD68+ cells was observed within the TC compared with that in the adjacent normal tissues; an opposite trend was observed for CD8+ and CD20+ cells (Fig.2a). Next, for a higher degree of detail, the distribution of each TIIC was explored. CD8+, CD8+PD-1LAG-3, CD20+ and CD68+CD163+HLA-DR cells accumulated at the IM and decreased toward the TC. In contrast, CD8+PD-1+TIM-3+, CD8+PD-1TIM3+, CD8+PD-1+LAG-3+TIM-3+, CD8+PD-1+LAG-3TIM-3+, CD4+FoxP3+CTLA-4+, CD4+FoxP3CTLA-4+, CD68+, CD68+HLA-DR+CD163 cells accumulated at the TC and decreased toward the IM. Interestingly, a higher density of CD4+FoxP3+ and CD4+FoxP3+PD-L1+ cells was found within the TC than in normal tissues (Fig.2b, Supplementary Fig.1b), highlighting the heterogeneous distribution of TIICs in GC.

a Constitution of the main tumour-infiltrating immune cell (TIIC) populations. KruskalWallis test with the Dunns multiple comparison test. b Density of TIICs across the regions of interest (n=80). TC, tumour core; IM, invasion margin; N, normal tissue. Immunofluorescence staining images refer to the co-expression of the corresponding markers and DAPI (nuclei). Scale bar: 20m. Box and whiskers represent mean1090 percentile. KruskalWallis test with Dunns multiple comparison test. c TIIC density grouped by subtypes. d Overall survival of 80 patients based on the density of TIICs. The individual TIICs were divided into high (>two-thirds of the patients; blue line) or low density (two-thirds of patients; red line). Log-rank (MantelCox) test was used. A two-sided P<0.05 was considered statistically significant.

Additionally, the localisation of TIICs with respect to the tumour nest and stroma areas (defined in Fig.1g) was further examined. CD8+, CD4+ and CD20+ cells were located primarily in the stroma and were less prevalent in the tumour nest. In contrast, CD66b+ cells were more prevalent in the tumour nest than in the stroma (Supplementary Fig.2a).

To evaluate the tumour immune microenvironment in GC, we compared the density of TIICs in the context of distinct clinico-pathological factors (Fig.2c, Supplementary Fig.3ae). Generally, there were few significant differences between Lauren classification, tumour differentiation and tumour location (oesophagogastric junction or not) with respect to densities of TIICs (Supplementary Tables15, Supplementary Fig.4a). Additionally, there were few differences in the density of TIICs between HER2-positive and -negative GC (Fig.2c). Overall, the density of total CD8+, CD4+ and CD68+ cells was associated with the disease stage. Additionally, advanced-stage GC (III-IV) samples showed a higher density of exhausted CD8+ T cells, CD4+FoxP3 cells and so on.

Furthermore, we analysed the density of TIICs in GC of different molecular subtypes (Supplementary Tables68). Interestingly, EBV-positive tumours showed higher densities of CD8+PD-1LAG-3 T cells than EBV-negative ones. EBV (+) GCs were characterised by abundant immune cell infiltration; however, not all EBV (+) patients responded to immunotherapy, indicating that specific immune cell infiltration is needed. Proficient MMR (pMMR) tumours showed a significantly higher abundance of total CD4+, CD68+, CD20+ and CD66b+ cells than dMMR tumours. Higher CD68+ and CD66b+ cells (neutrophils) are known to contribute to resistance to PD-1/PD-L1 treatment in several cancers13,19. We classified patients into four combined positive score (CPS) groups: CPS<1, 1 CPS<5, 5 CPS<10 and CPS10. Remarkably, the abundance of TIICs, including CD8+, CD4+, CD68+, CD20+ and CD66b+ cells, significantly increased with the increase in CPS, indicating a hotter tumour immune environment. However, the comparison between CPS 5-10 and CPS10 did not show a significant difference, providing evidence for the cut-off selection in clinical trials of anti-PD-1/PD-L1-based therapies. Altogether, as shown in Fig.2c, our results suggest that the infiltration pattern of immune cells depends on, but is not restricted to, GC molecular subtypes.

Next, we sought to understand whether the number of TIICs is correlated with patient survival. We found that higher levels of tumour-infiltrating T cell subsets, including CD8+PD-1+LAG-3+TIM-3+, CD4+FoxP3+CTLA-4+ T and CD68+STING+ cells, were associated with inferior overall survival (OS) in 80 patients (Fig.2d, Supplementary Fig.4b). CD8+PD-1+LAG-3+TIM-3+ cells [high vs. low, hazard ratio (HR) 1.98, 95% confidence interval (CI; 1.123.50)] and CD68+STING+ cells [high vs. low, HR 1.83, 95%CI (1.013.33)] were significantly associated with OS, as revealed by multivariate Cox analysis (Supplementary Table9). Collectively, these data highlight the clinical relevance of tumour-infiltrating T cells in the survival of GC patients.

Additionally, we analysed the prognostic value of the density of TIICs in the context of tumour and stromal cells. The data showed a similar trend for CD4+FoxP3CTLA-4+ T and CD4+FoxP3+CTLA-4+ T cells in both contexts. However, higher infiltration of CD8+PD-1+LAG-3+TIM-3+ T cells and CD68+ macrophages was associated with poorer OS with respect to tumour nests. In addition, higher infiltration of CD8+PD-1+TIM-3+ T cells, CD66b+ neutrophils and CD68+STING+ macrophages was related to a shorter OS with respect to the stroma (Supplementary Fig.2b). Therefore, these results highlight the value of studying immune cell density in defined tissue regions.

Given our ability to precisely define the positions of individual tumour cells and TIICs, we next sought to evaluate the clinical significance of the proximity between them. The observation that certain TIICs, including CD68+ cells, were enriched in the tumour region suggested that the proximity of TIICs to tumour cells might influence their phenotype. To further study these localisation patterns, a bioinformatics tool (pdist; see Methods) that determines the nucleus-to-nucleus distances between any two cell types was used. To incorporate both cell proximity and quantity, an effective score parameter was established: the proportion of TIICs near tumour cells (within the defined distance criteria introduced; Fig.3a). In other words, this score was calculated by the number of paired immune cells and tumour cells divided by the total number of immune cells across the whole slides to maintain the spatial variation to a large extent. Therefore, using this formula, a higher effective score indicates that within a certain distance, there is a higher density of tumour cells around the immune cells. Importantly, across the three distances considered (010/020/030m), CD8+PD-1+LAG-3+ T cells and CD66b+ neutrophils were the ones with higher effective scores (Fig.3b).

a Illustration of the distance analysis involving immune and tumour cells. Red dots: tumour cells; green dots: immune cells. The white translucent circle represents the radius. Effective score=number of paired immune cells and tumour cells/number of immune cells. Scale bar: 100m. b The distribution of the effective score of tumour-infiltrating immune cell (TIIC) populations in the tumour core in 10-, 20- and 30m increments (n=80). Error bars represent meanSEM. c Effective score of TIICs in patients grouped by gastric cancer subtypes. EBV, EpsteinBarr virus status; MMR, DNA mismatch repair; CPS, combined positive score. d Overall survival of the 80 patients based on the effective densities (010m and 020m) of TIICs. The individual immune infiltrate values were divided into high (> two-thirds of the patients in the cohort; blue line) or low density ( two-thirds of patients in the cohort; red line). Statistical relevance was defined using the log-rank (MantelCox) test. A two-sided P<0.05 was considered statistically significant.

We also calculated the distance between each TIIC and the closest tumour cell. Neutrophils, B cells and macrophages were located closer to tumour cells. We then analysed the distances between TIICs and tumour cells according to the PD-L1 CPS. In general, TIICs were located closer to tumour cells in patients with CPS10 (compared with the picture with respect to all other groups; Supplementary Fig.6a).

Interestingly, the effective scores also differed between different GC molecular subtypes, including those depending on the EBV, PD-L1 CPS, MMR and HER2 status (Supplementary Figs.5ae, 6b; Supplementary Tables1017). For instance, a significantly higher effective score of exhausted T cells (CD8+PD-1+LAG-3+TIM-3, CD8+PD-1TIM-3+), M1 (CD68+CD163+HLA-DR) and M2 (CD68+HLA-DR+CD163) macrophages within a 20m radius was observed in HER2-negative GC compared with that in HER2-positive GC (Fig.3c, Supplementary Fig.5b).

The combination of multiplexed imaging and machine learning implied that the density of TIICs within GC is linked to patient survival. For further detail, the effective density (the absolute number of TIICs near tumour cells within a 20m radius) was used as an additional measurement. This radius was pre-selected to identify immune cell populations most likely capable of effective, direct, cell-to-cell interactions with tumour cells, consistent with prior studies in multiple gastrointestinal tumour types11,20,21. Curiously, we found that patients with higher effective densities (radius of 020m) of CD68+STING+ macrophages, CD68+HLA-DR+CD163 STING+ macrophages and neutrophils showed significantly shorter OS than those with lower effective densities (Fig.3d). Importantly, the prognostic value was still significant after adjustment using the multivariate Cox model (Supplementary Table18). Other immune cell phenotypes were not associated with OS (Supplementary Figs.6c and 7c). These results indicate that the influence of TIICs on patient survival is dependent not only on the number of TIICs but also on their proximity to tumour cells. Overall, our data highlight that both the location and density of TIICs should be taken into consideration for prognosis predictions.

Human tumours contain exhausted T cells expressing multiple immune checkpoints; it has been proposed that these cells mediate resistance to PD-1 blockade. Thus, next, we investigated whether the density of TIICs and respective effective scores were associated with the clinical outcomes of anti-PD-1/PD-L1 immunotherapy. All 60 patients who received immunotherapy were assigned to the training (n=44, generated retrospectively from 15/11/2016 to 17/7/2019) and validation (n=16, generated prospectively from 29/7/2019 to 19/12/2019) cohorts. Importantly, we ensured that the clinical characteristics of the training and validation cohorts were balanced (Table2). We used logistic regression analysis to assess the association between TIICs and the objective response rate (ORR) in the training cohort. Importantly, we found that the density of CD4+FoxP3PD-L1+ T cells and the effective score of CD8+PD-1+LAG-3 T cells were closely associated with a positive response to anti-PD-1/PD-L1 therapy; conversely, CD8+PD-1LAG-3 T cells and CD68+STING+ macrophages were closely associated with a negative response to anti-PD-1/PD-L1 therapy (Supplementary Table19).

The density of CD4+FoxP3PD-L1+ T cells, CD8+PD-1LAG3 T cells and CD68+STING+ macrophages, and the effective score of CD8+PD-1+LAG3 T cells were used to define a TIIC signature (Fig.4a), with the potential to improve the ability of identifying responders to anti-PD-1/PD-L1 immunotherapy. We used four types of machine learning models and calculated the area under the curve (AUC) of the training and validation cohorts, including extra tree classifier (ETC), AdaBoost classifier (ABC), gradient boosting classifier (GBC) and multi-layer perceptron (MLP) models. In the validation cohort, the average AUCs of the four algorithms were 0.80, 0.85, 0.77 and 0.75, respectively (Fig.4b, c, Supplementary Table20). The corresponding 95% CIs were narrow, suggesting that the TIIC signature can indeed be used to predict the response to immunotherapy (Supplementary Table20). Importantly, the four algorithms showed a similar performance before and after adjusting for the hyper-parameters, indicating the strength of the predictive value of the TIIC signature itself (Supplementary Fig.7a). Furthermore, we explored the predictive power of the TIIC score combined with CPS, EBV status and MMR status. The combined TIIC signature had a better AUC in the ETC, GBC and ABC models, but not in the MLP model (Supplementary Table21).

a Definition of the tumour-infiltrating immune cell (TIIC) signature. Red arrows highlight specific immune cells. b Average area under the curve (AUC) of TIIC signature and combined TIIC signature (TIIC+ EpsteinBarr virus status+mismatch repair status+PD-L1 combined positive score) in the four machine learning models in the validation cohort. c Representative receiver operating characteristic (ROC) curves for the performance of the identified TIIC signature and combined TIIC signature in gastric cancer patients subjected to immunotherapy in the validation cohort. ETC extra tree classifier, GBC gradient boosting classifier, ABC AdaBoost classifier, MLP multi-layer perceptron.

We quantified the contribution of each marker in the prediction models through feature importance using the scikit-learn package (Supplementary Tables22, 23). We outputted the feature importance and the average value of each parameter to present its contribution. As shown in Fig.5a, the effective score of CD8+PD-1+LAG-3 cells had higher feature importance than the density of CD68+STING+, CD4+FoxP3PD-L1+, or CD8+PD-1LAG-3 cells in ETC, GBC and ABC machine learning models. As presented in Fig.5b, the effective score of CD8+PD-1+LAG-3 cells had higher feature importance than that of the other three immune cell types, EBV, MMR and PD-L1 CPS. Thus, the dominant predictive marker is the spatial organisation for response to immunotherapy.

a, b The feature importance of each marker in the prediction model. c, d KaplanMeier curves of the (c) immune-related progression-free survival (irPFS) and (d) immune-related overall survival (irOS) of anti-PD-1/PD-L1-treated patients stratified by the tumour-infiltrating immune cell (TIIC) signature in the validation cohort. Log-rank (MantelCox) test was used for analysis.

We also evaluated the predictive values of other candidate biomarkers. AUCs of 0.58 and 0.76 in the training and validation cohorts, respectively, were defined for PD-L1 CPS (Supplementary Fig.7b). We analysed the treatment response based on EBV status, MMR status and HER2 expression in univariate and multivariable logistic regression models (Supplementary Table24). EBV-positive status and dMMR tended to be associated with a better response. The association of HER2 expression with treatment response was not consistent between univariate and multivariable models. Therefore, taken together, our data suggest that the TIIC signature has a greater power for patient stratification (Supplementary Fig.7b).

Next, we investigated the prognostic use of the TIIC signature; the univariate Cox proportional hazard regression model was used to calculate the HR of each indicator. Then, we used the HR of each indicator as the weight to multiply the value of the indicator itself and then calculated the weighted sum of the four indicators. In this analysis, we categorised patients into high- and low-score groups based on the TIIC signature. The difference in the survival probability over time between the groups was calculated using the KaplanMeier method. As expected, we observed a significant difference in both immune-related progression-free survival (irPFS) and immune-related overall survival (irOS) in the validation cohort (Fig.5c, d, Supplementary Table25). Therefore, the TIIC signature might be useful to identify patients that will show active anti-tumour immune responses a priori.

See the article here:
Predicting response to immunotherapy in gastric cancer via multi-dimensional analyses of the tumour immune microenvironment - Nature.com

This entry was posted on August 19, 2022 at 2:08 am and is filed under Immune System. You can follow any responses to this entry through the RSS 2.0 feed. Both comments and pings are currently closed. |