Work

Development of the model for predicting the efficacy of bevacizumab and prognosis in ovarian cancer using deep learning-based pathomics signatures

Year: 2025

Type: article

Abstract: To the Editor: Ovarian cancer (OC) is a highly aggressive malignancy that poses a significant threat to women’s health and is associated with a high mortality rate due to late diagnosis and chemotherapy resistance.[1] Bevacizumab, an angiogenesis inhibitor, is a Food and Drug Administration (FDA)-approved therapy for advanced OC, improving progression-free survival (PFS) and overall survival (OS) when combined with chemotherapy or used as maintenance therapy.[2] However, challenges persist. Identifying predictive biomarkers and optimizing bevacizumab treatment duration and timing are crucial. Resistance to bevacizumab, driven by reactivation of angiogenic pathways and changes in the tumor microenvironment, leads to treatment failure.[3] The development of new predictive methods is urgently needed due to their cost, potential toxicity, and limited patient benefit. Early diagnosis is key to improving OC prognosis. Histopathology is the gold standard for diagnosis and prognostic prediction, but it is complex and time-consuming. Recently, deep learning based on pathomics features has been used to construct predictive models, offering a faster alternative.[4,5] In this study, we used whole slide images (WSIs) from OC patients in The Cancer Genome Atlas Ovarian Cancer (TCGA-OV) dataset, the Cancer Imaging Archive (TCIA) cohort and Zhongda Hospital Southeast University (Zhongda cohort) to develop a predictive model via deep learning and the software Cellprofiler (Broad Institute, Cambridge, MA, England). The model accurately identified tumor regions and assessed the impact of bevacizumab on patient prognosis. We also attempted to cluster patients based on their treatment response to analyze genomic differences and guide patient prognosis. The studies involving human participants were reviewed and approved by the Ethics Committees and Institutional Review Boards of Zhongda Hospital Southeast University (No. ZDSYLL187-P04). Written informed consent for the publication of case details was obtained from the patients. The study flowchart is provided in Supplementary Figure 1, https://links.lww.com/CM9/C615, and detailed methods are available in the Supplementary text, https://links.lww.com/CM9/C615. Deep learning-based pathomics signatures for identifying tumor regions. The TCGA-OV and TCIA cohorts were included in the development of the deep learning-based model for tumor regions at the patch level. The TCIA cohort was used as the training set, while the TCGA-OV cohort was used as the testing set. The regions of interest (ROIs) of 332 WSIs were annotated by pathologists and tiled into 2,421,592 patches for feeding the tumor region classifier. The area under the curve (AUC) of the tumor region classifier in the training set was 0.999, and the heatmap of the confusion matrix revealed the distribution of each patch at the actual and predictive levels. Interestingly, the AUC was still 0.895 in the testing set indicating that the established classifier could effectively identify the tumor regions with high specificity [Supplementary Table 1, https://links.lww.com/CM9/C615]. Deep learning-based pathomics signatures for identifying bevacizumab efficacy. To construct a classifier model for calculating bevacizumab efficacy based on pathomics signatures, patches identified as tumors and their matched clinical signatures were fed to the bevacizumab response classifier to predict bevacizumab efficacy. The performance of this classifier was verified in the TCIA testing set at the patch level and in the TCIA hold-out testing set at the slide level. The AUC of the classifier at the patch level reached 0.937, and this model at the slide level achieved an AUC of 0.946 [Supplementary Table 2, https://links.lww.com/CM9/C615]. These results suggested that this classifier could predict bevacizumab efficacy based on pathomics information. Identification of bevacizumab efficacy in Zhongda cohort. This model could predict bevacizumab efficacy in the Zhongda cohort as an external validation set, as assessed by the area under the curve (AUC) in Figure 1A and the confusion matrix in Figure 1B. The performances of the models are listed in Supplementary Table 3, https://links.lww.com/CM9/C615. The AUC was 0.830, the accuracy was 0.855, and the sensitivity, specificity, F1-score, positive predictive value (PPV) and negative predictive value (NPV) were 0.792, 0.903, 0.826, 0.864, and 0.848, respectively. These data suggested that our model could predict bevacizumab-susceptible populations based on Hematoxylin–Eosin (HE)-stained pathological sections from primary tumors.Figure 1: Performance of pathomics signatures for identifying bevacizumab efficacy in the Zhongda cohort. ROC curves (A) and confusion matrix (B) of our model in predicting the efficacy of bevacizumab in the Zhongda cohort. In the confusion matrix, darker colors indicate greater quantities. AUC: Area under the curve; ROC: Receiver operating characteristic.Consensus clustering based on different treatment responses. To study the potential correlation between prognosis and hub pathomics characteristics in OC patients treated with bevacizumab, 41 significant deep learning and nuclear texture features associated with treatment response were selected via a least absolute shrinkage and selection operator (LASSO) regression model [Supplementary Figure 2A, https://links.lww.com/CM9/C615]. To explore the significance of the features extracted, the patients in the TCGA-OV cohort were divided into different clusters via the consensus clustering algorithm [Supplementary Figure 2B, https://links.lww.com/CM9/C615]. As shown in the figure, k = 2 was the most appropriate k value, indicating that the patients in the TCGA-OV cohort could be divided into two clusters. K–M curves revealed that individuals in the A cluster had worse OS than those in the B cluster did [Supplementary Figure 2C, https://links.lww.com/CM9/C615]. Furthermore, compared with the clinical characteristics, the two clusters both presented several significant clinical features that might affect OS, including tumor grade, International Federation of Gynecology and Obstetrics (FIGO) stage and age [Supplementary Figure 2D, https://links.lww.com/CM9/C615]. These results demonstrated the essential role of hub pathomics characteristics, such as nuclear texture features and tumor microenvironment (TME) features, in influencing patient prognosis. Differences in the TME and response to chemotherapy among different consensus clusters. To investigate the role of immune cells in the TME and the response to chemotherapy, immune cell infiltration and chemotherapy sensitivity were analyzed via the single-sample gene set enrichment analysis (ssGSEA) algorithm based on transcriptional profiles of the TCGA cohort. Compared with those in the A cluster, the number of activated CD4+ T cells was significantly greater, whereas the number of CD56bright natural killer (NK) cells was lower in the B cluster [Supplementary Figure 3A, https://links.lww.com/CM9/C615]. The sensitivity of commonly used chemotherapeutic agents was evaluated via the “pRRophetic” package. The B cluster was more sensitive to paclitaxel and gemcitabine than the A cluster was but had a lower response to cisplatin [Supplementary Figure 3B, https://links.lww.com/CM9/C615]. These results suggest that there is an association between chemotherapy and NK cell suppression. CD56bright NK cells produce proinflammatory cytokines and have immunoregulatory properties, suggesting that patients in the B cluster may benefit from cytokine management rather than direct CD4 cytotoxicity. Genomic landscape in different consensus clusters. To explore the potential signaling pathways enriched in the two clusters, we performed gene set variation analysis (GSVA). Patients with worse OS in the A cluster were enriched in the NOTCH and hedgehog signaling pathways, whereas those in the B cluster were enriched in DNA repair, oxidative phosphorylation and several classical signaling pathways, such as the E2F, G2M, MYC, and MTORC1 signaling pathways [Supplementary Figure 4A, https://links.lww.com/CM9/C615]. Differentially expressed genes (DEGs) between the two clusters were also analyzed in the TCGA cohort, and four significant DEGs were selected, including the upregulated genes CYB561D2 and PCC1A and the downregulated genes ZNF12 and RIN2 [Supplementary Figure 4B, https://links.lww.com/CM9/C615]. Finally, we screened the genes with the most differential mutations in the two clusters. Interestingly, in the top 10 mutated genes in the total cohort, the A cluster had a smaller sample size of TTN mutations than did the B cluster [Supplementary Figure 4C, https://links.lww.com/CM9/C615]. These results indicated that various enriched genes, mutations and signaling pathways existed between the two clusters with different hub pathomics characteristics. This study is dedicated to developing a deep learning-based predictive model aimed at accurately identifying patients who are sensitive to bevacizumab. The model is specifically designed for the postoperative evaluation phase. Clinicians can upload postoperative pathological images of the tumor tissue to the system, and the model will provide predictive results based on image feature analysis. This information will assist clinicians in assessing the potential benefits of bevacizumab treatment for patients. However, given the complexity of individual patient differences, the model’s predictive results should only serve as one of the references for clinical decision-making. The final decision on whether to use bevacizumab treatment should still be made by a professional medical team, considering the specific clinical context of the patient. Our study indeed has certain limitations. First, the WSIs utilized in this study were sourced from public databases, specifically the TCGA-OV and TCIA cohorts, in addition to the medium–large cohort. It is acknowledged that there may be some variation between these different cohorts. Moreover, owing to the suboptimal quality of some WSIs, despite the normalization of images, the inherent defects or artifacts of the WSIs themselves may still affect the model’s output results, leading to data contamination. In the future, generative adversarial networks (GANs) could be employed to increase the standardized performance of the model. Second, our external independent validation set included only 33 patients, necessitating larger clinical trials to further validate our results. Additionally, during the initial delineation of tumor areas, no secondary delineation was performed, introducing a certain degree of subjective error. In future research, we will continue to employ techniques such as cross-validation, regularization, and early stopping to further enhance the robustness and generalizability of our models. This work will allow us to assess the performance of the model across a broader spectrum of patient demographics and clinical practices. In summary, pathomics information extracted from WSIs can be used to identify tumor regions and predict bevacizumab efficacy in OC. Tumor nuclear features, TME features and DEGs are closely associated with patient OS and chemotherapy effects, suggesting the potential role of the TME in benefiting bevacizumab treatment. The genes screened in the two clusters may be promising markers for predicting bevacizumab efficacy. Our study may support further development of screening strategies for bevacizumab-sensitive patients and may provide new ideas for individualized OC-targeted therapy. Funding This study was supported by grants from Zhongda Hospital Affiliated to Southeast University, Jiangsu Province High-Level Hospital (Nos. 2023GSPKY11, GSP-LCYJFH01), National Clinical Key Discipline Construction Fund (No. gynecology czxm-zk-40), National Natural Science Foundation of China (Nos. 82372126 and 82072078), Beijing Xisike Clinical Oncology Research Foundation (No. Y-zai2022/ms-0126), and Nanjng Science and Technology Major Project (No. 202305039). Conflicts of interest None.

Source: Chinese Medical Journal

DOI: 10.1097/cm9.0000000000003823