Tumor immune dysfunction and exclusion evaluation and chemoimmunotherapy response prediction in lung adenocarcinoma using pathomic-based approach

To the Editor: Lung cancer, specifically lung adenocarcinoma (LUAD), is one of the primary cause of cancer-related mortality globally.[1,2] Nevertheless, only a small subset of individuals with LUAD have derived clinical benefits from chemoimmunotherapy in either first-line or subsequent treatment settings. Both programmed death-ligand 1 (PD-L1) expression and tumor mutational burden (TMB) have proven inadequate in accurately predicting treatment outcomes in these scenarios.[3] Consequently, there exists a pressing necessity to identify a reliable biomarker to inform treatment decisions. The tumor immune dysfunction and exclusion (TIDE) transcriptome biomarker platform has the capability to produce signatures indicative of T-cell dysfunction, T-cell exclusion, and the ability to forecast responses to immune checkpoint inhibitors (ICIs).[4,5] A high TIDE score typically correlates with heightened resistance to ICIs. Therefore, a biomarker to assess TIDE status and predict the efficacy of chemoimmunotherapy in LUAD would be valuable. Pathomics, also known as digital pathology, involves the utilization of image analysis to extract subvisual and quantitative features from whole slide images (WSIs) in a high-throughput manner.[6,7] To our knowledge, few studies have utilized a pathomic-based approach to assess TIDE status and predict the response to chemoimmunotherapy in patients with LUAD. The objective of this study was to establish a pathomic model of the TIDE classifier and subsequently validate its efficacy in predicting response to chemoimmunotherapy in patients with LUAD. This was a retrospective cohort study, and the process of patient selection is shown in Supplementary Figure 1, https://links.lww.com/CM9/C215. A total of 522 LUAD patients from The Cancer Genome Atlas (TCGA) were screened. After excluding (1) no histopathological images (n = 44), (2) image quality not qualified (n = 102), and (3) lack of clinical data or RNA-seq data (n = 49), 327 LUAD patients were included for developing and validating the model. A total of 256 advanced LUAD patients without driver mutations receiving chemoimmunotherapy from Shanghai Chest Hospital were screened. The patients were excluded for the following reasons: (1) not first-line treatment (n = 72); (2) lack of histopathological image (n = 63); (3) biopsy from extrapulmonary lesion (n = 67); and (4) image quality not qualified (n = 9). Finally, 45 patients with pathological slides were enrolled in the analysis set [Supplementary Figure 1, https://links.lww.com/CM9/C215]. This study was approved by the Ethics Committee of Shanghai Chest Hospital (No. KS24007). Written informed consent was obtained from patients or their guardians. The processes of preprocessing of WSIs, histopathological features extraction and selection, pathomic model development, and statistical analysis can be found in the methods section of the Supplementary Materials, https://links.lww.com/CM9/C215. The overall study design is shown in Supplementary Figure 2, https://links.lww.com/CM9/C215. A total of 327 LUAD patients with TIDE scores were included. X-tile was used to determine the optimal cutoff value of TIDE scores and the cutoff value was 0.03 [Supplementary Figure 3A–C, https://links.lww.com/CM9/C215]. The clinical characteristics of LUAD patients with low and high TIDE scores are shown in Supplementary Table 1, https://links.lww.com/CM9/C215. Patients with low TIDE scores had significantly longer overall survival (OS) than the ones with high scores [Supplementary Figure 3D, https://links.lww.com/CM9/C215]. A total of 327 LUAD patients with WSIs and TIDE scores were included to conduct pathomic model. These patients were randomly divided into training and validation groups in a 7:3 ratio, with 229 patients in the training group and 98 patients in the validation group. The baseline demographic and clinical characteristics were found to be similar between the two groups [Supplementary Table 2, https://links.lww.com/CM9/C215]. Eleven features were identified by using the maximal relevance and minimal redundancy (mRMR) method and recursive feature elimination (RFE) method [Supplementary Figure 4A, https://links.lww.com/CM9/C215]. The gradient boosting machine (GBM) was utilized to construct pathomic model, and the importance of the features on model prediction of TIDE status is visually represented in Supplementary Figure 4B, https://links.lww.com/CM9/C215. The pathomic model demonstrated an area under the curve (AUC) of 0.811 (95% confidence interval [CI]: 0.756–0.867) for predicting the TIDE classifier in the training cohort, as depicted in Supplementary Figure 5A, https://links.lww.com/CM9/C215. Similarly, the model exhibited a moderate level of discriminability with an AUC of 0.730 in the validation cohort [Supplementary Figure 5B, https://links.lww.com/CM9/C215]. Additional performance metrics, such as overall accuracy, sensitivity, and specificity, can be found in Supplementary Table 3, https://links.lww.com/CM9/C215. The calibration curves indicated good agreement between prediction and observation in both the training and validation cohorts [Supplementary Figure 5C, D, https://links.lww.com/CM9/C215]. The decision curve analysis indicated that the pathomic model exhibited strong performance in both the training and validation cohorts [Supplementary Figure 5E, F, https://links.lww.com/CM9/C215]. Consistent with expectations, the pathomic score (PS) generated by the model was notably higher in the high TIDE group compared to the low TIDE group in both the training and validation cohorts [Supplementary Figure 5G, H, https://links.lww.com/CM9/C215]. Based on the optimal cutoff determined by the R package (version 4.2.2, https://www.r-project.org/) “survminer” in both the training and validation cohorts, patients were stratified into a PS-low group (PS <0.450) and a PS-high group (PS ≥0.450). The detailed demographic and clinicopathological characteristics of the patients in the entire cohort (n = 327), PS-low subgroup (n = 193), and PS-high subgroup (n = 134) are presented in Supplementary Table 4, https://links.lww.com/CM9/C215. The Kaplan–Meier survival curve demonstrated a significant disparity in OS (hazard ratio [HR] = 1.53; 95% CI: 1.05–2.21, P = 0.025) between the high and low PS groups [Supplementary Figure 6A, https://links.lww.com/CM9/C215]. Patient stratification based on PS was effective in predicting OS across various subgroups [Supplementary Figure 6B, https://links.lww.com/CM9/C215]. A multivariate Cox regression analysis was conducted to adjust for clinicopathological variables, revealing that PS remained a significant independent prognostic factor for OS prediction [Supplementary Figure 6C, https://links.lww.com/CM9/C215]. Furthermore, a pathomic nomogram incorporating PS and clinicopathological factors was developed using both training and validation cohorts [Supplementary Figure 6D, https://links.lww.com/CM9/C215]. The AUCs at different follow-up intervals (1 year, 2 years, and 3 years) demonstrated the strong prognostic accuracy of this model [Supplementary Figure 6E, https://links.lww.com/CM9/C215]. The calibration curves demonstrated strong concordance at 1 year, 2 years, and 3 years between predicted and observed outcomes [Supplementary Figure 7A, https://links.lww.com/CM9/C215]. Decision curve analysis indicated a favorable net benefit for the model at various threshold probabilities [Supplementary Figure 7B–D, https://links.lww.com/CM9/C215]. These findings suggested that the PS may serve as an independent predictor of outcomes in patients with LUAD. We examined the correlation between PS and objective response rate (ORR) to chemoimmunotherapy in a cohort of 45 patients with advanced LUAD. The PS-based ORR classifier demonstrated an AUC of 0.688 in the ROC curve, indicating an optimal PS cut-off value of 0.175 for predicting ORR [Supplementary Figure 8, https://links.lww.com/CM9/C215]. Clinical characteristics of patients in the low and high PS groups in the analysis set are presented in Supplementary Table 5, https://links.lww.com/CM9/C215. Patients in the low PS group exhibited a significantly higher ORR compared to those in the high PS group (76.9% vs. 47.4%, P = 0.041, Supplementary Figure 9A, https://links.lww.com/CM9/C215). The median PFS in the low PS group (14.8 months) was significantly longer than that in the high PS group (6.0 months, Supplementary Figure 9B, https://links.lww.com/CM9/C215), and the adjusted HR was 0.372 (95% CI: 0.180–0.771, P = 0.008). The median OS of patients with low PS was not reached and 25.5 months for those with high PS (Supplementary Figure 9C, https://links.lww.com/CM9/C215), and the adjusted HR was 0.267 (95% CI: 0.084–0.844, P = 0.024). In this cohort, the study found that PD-L1 expression did not show a significant association with ORR, as indicated by an AUC of 0.556 (95% CI: 0.375–0.737) [Supplementary Figure 9D, E, https://links.lww.com/CM9/C215]. Multivariate regression analysis revealed that PS had a more pronounced impact on ORR compared to PD-L1 expression [Supplementary Figure 9F, https://links.lww.com/CM9/C215]. Additionally, PS was found to be correlated with PFS and OS in various PD-L1 expression subgroups [Supplementary Figure 10, https://links.lww.com/CM9/C215]. Consequently, the researchers developed an integrated model combining PD-L1 expression and PS [Supplementary Figure 9D, https://links.lww.com/CM9/C215], which significantly enhanced the accuracy of response prediction (AUC 0.702, 95% CI: 0.551–0.852). Pathway enrichment analysis was conducted utilizing the gene set enrichment analysis (GSEA) method to investigate the potential associations between the model and the outcomes. The results indicated that tumors within the high PS group exhibited significant enrichment for cancer hallmark-related pathways, including myelocytomatosis oncogene (MYC) signaling, oxidative phosphorylation, purine metabolism, and pyrimidine metabolism pathways, which are known to be associated with aggressive tumor phenotypes [Supplementary Figure 11A, B, https://links.lww.com/CM9/C215]. This result might explain why LUAD patients with high PS showed shorter OS. Analysis of transcriptome data revealed that low PS tumors demonstrated heightened infiltration of cluster of differentiation 8 positive (CD8+) T cells and CD8+ T native cells [Supplementary Figure 11C, https://links.lww.com/CM9/C215], suggesting that this pathomic model might involve in cancer immunity and affect the sensitivity to ICI treatment. Low PS tumors exhibited a lower level of tumor mutational burden, yet demonstrated similar levels of neoantigens compared to high PS tumors [Supplementary Figure 11D, E, https://links.lww.com/CM9/C215]. Notably, low PS tumors displayed a significant correlation with elevated cytolytic activity and immunophenoscore [Supplementary Figure 11F, G, https://links.lww.com/CM9/C215]. These findings suggested a potential relationship between the pathomic model and T cell-mediated anti-tumor immune response. The tumor immune microenvironment has been identified as a crucial determinant of treatment response and outcomes in various cancers. For instance, the TIDE status has shown promise in predicting patient outcomes following initial ICI therapy.[4] Nevertheless, significant obstacles persist in the assessment of TIDE, including limited availability of tumor tissue, intratumor heterogeneity, and the substantial costs associated with testing. Consequently, the establishment of a pathomic model for evaluating TIDE holds significant importance for enhancing clinical decision-making. Furthermore, the correlation between this model and the infiltration of CD8+ T cells, cytolytic activity, and immunophenoscore elucidated its ability to forecast clinical outcomes of chemoimmunotherapy in patients with LUAD. This study has some limitations. First, this study was retrospective. Thus, the possibility of residual confounding cannot be ruled out. Second, the sample size of validation set was moderate, which limited the power to draw a definite conclusion. Further multicenter studies with larger sample sizes are needed to validate the results of this study. Third, the pathomic model was validated using a cohort of patients with advanced LUAD. Its value in early stage setting needs further investigation. Fourth, the correlation between pathomic model and treatment efficacy in patients with advanced LUAD receiving chemotherapy should also be assessed. Last but not least, it would be better to use advanced LUAD specimens for TIDE calculation and construct a pathomic model. In summary, a pathomic model was created to evaluate TIDE status through the analysis of histopathological images. This model shows potential as a biomarker for predicting the efficacy of chemoimmunotherapy in LUAD. Additional prospective studies are necessary to confirm these findings. Funding This study was supported by grants from the National Natural Science Foundation of China (Nos. 82373425 and 82372722), the Medical Innovation Research Special Project of the Science and Technology Commission of Shanghai Municipality (No. 23Y11904200), the Shanghai Innovative Medical Device Application Demonstration Project 2023 (No. 23SHS02600), the “Science and Technology Innovation Action Plan” Medical Innovation Research Special Project of Shanghai (No. 21Y11913500), the key project of the Medical and Health Technology Development Research Center of the National Health Commission (No. WKZX2023CX030003), and the Shanghai Key Laboratory Open Project (No. STCSM 22DZ2229005). Conflicts of interest None.