Ensemble biomarkers for guiding anti‐angiogenesis therapy for ovarian cancer using deep learning

Abstract

The central contribution of this research is to present highly effective ensemble biomarkers for guiding therapies to ovarian cancer, which has become the major cause of death from gynecologic cancer in industrialized countries1 whereas 90% of all ovarian cancer types is epithelial ovarian cancer (EOC). The typical therapy for EOC is a combination of cytoreductive surgery, platinum and taxane chemotherapy.2, 3 Bevacizumab (Avastin), which is the first FDA approved anti-angiogenic agent,4, 5 leads to normalization of tumor vasculature and improved effectiveness of standard therapy with promising result, but unfortunately it is associated with high degree of adverse effects, including arterial thromboembolism, high blood pressure, presence of excess proteins in the urine, haemorrhage, poor wound healing and ruptured bowel.6 The goal of this study is to examine whether the angiogenesis-related biomarkers, such as VEGF, Ang-2 and PKM2 could be used to build an effective therapeutic predictive system for patient selection, guiding ovarian cancer treatment. In this study, we investigate three angiogenesis-related potential biomarkers, including VEGF, Angiopoietin 2 (Ang-2) and Pyruvate kinase isoform M2 (PKM2), and develop an annotation-free instance boosting deep learning ensemble framework to forecast the treatment response to bevacizumab of patients with EOC or peritoneal serous papillary carcinoma (PSPC) using tissue microarrays. See Supporting Information (Section 1) for related works. To enable the development of therapeutic predictive AI systems, digitized whole slide images (WSIs) of tissue microarrays are collected from samples with EOC and PSPC, containing 720 female patient sample tissue cores in total. For the data inclusion and exclusion criteria, patients are eligible for the retrospective study if they are EOC or PSPC patients receiving 1) concurrent bevacizumab therapy, 2) second-line bevacizumab therapy post recurrence or 3) maintenance bevacizumab therapy. On the other hand, patients are ineligible if they are borderline malignant tumor or not primary ovarian or retroperitoneal carcinoma. The data include papillary serous carcinoma (444 samples), clear cell carcinoma (69 samples), endometrioid carcinoma (39 samples), mucinous carcinoma (10 samples), unclassified carcinoma (69 samples), and PSPC (89 samples) of 12 TMAs. This study has been approved by the hospital research ethics committee with the IRB approval number (TSGHIRB No.1-107-05-171 and No. B202005070), and we construct the database and trace patient data from the period of 2013/3 to 2021/1. Supplemental Table S0(a) presents the baseline characteristics of the database constructed in this study. The flowchart for training individual treatment effectiveness classifiers of the proposed method is illustrated in Figure 1, and the decision inference workflow of the proposed ensemble framework is illustrated in Figure 2. See Supporting Information (Section 2-1) for more data information and (Section 2-2) for the proposed method, respectively. For quantitative evaluation, the experiments were conducted in four parts. Firstly, the whole dataset (with 720 patient sample tissue cores) was randomly sampled and separated into two independent subsets, including 472 cores (66%) for training and 248 cores (34%) for testing as shown in Supplemental Table S0(b). Next, to further validate the effectiveness of individual models' predictions in how patients respond to the treatment, two generally adopted statistical analyses,7 including the univariate Kaplan-Meier survival (K-M) analysis and the multivariate Cox hazards analysis, were conducted. Fourthly, to examine the generalizability of individual models for unseen data, quantitative evaluation was performed using 5 -fold cross validation (CV). In the first experiment, the proposed biomarker system using PKM2 or Ang-2 achieves excellent performance in accuracy 98%, precision 100%, recall 97%, F1-score 99%, and AUC 99.7% (see Table 1a). According to the results using a single biomarker, the top two potential markers were then selected to build a hybrid marker-based system for further investigation of a possible even better prediction system. The proposed ensemble using both PKM2 and Ang-2 data achieves perfect prediction results, obtaining 100% in all measurements. Furthermore, Figure 3A compares the ROC curves of individual biomarker systems, and the proposed ensemble biomarker system provides the best treatment effectiveness classification accuracy, invariant to the threshold value. Secondly, Cox hazards analysis confirms that the proposed ensemble model is statistically significantly effective (p = 0.012), and (HR = 0.23) indicates that patients who are predicted with effective treatment responses are 0.23 times as likely to get cancer recurrence, compared to patients predicted with invalid responses (Table 1b). Thirdly, K-M analysis shows that there are very strong evidences that the proposed ensemble using both PKM2 and Ang-2 and the proposed models using PKM2 or Ang-2 are capable of identifying patients acquiring favorable treatment effects with low risk of recurrent cancer in the progression free survival time with high statistical significance ( p < 0.001 ) $(p < 0.001)$ (Figure 3B). See detailed evaluation and statistical analysis in Supporting Information (Section 3). Fourthly, to further examine the generalizability of individual models for unseen data, quantitative evaluation was performed based on 5-fold CV. As shown in Table 1c, even with 5-fold CV, the proposed ensemble model using both PKM2 and Ang-2 data consistently achieves near-to-perfect prediction in patient response to the treatment, obtaining accuracy 99 % ± 3 % $99{\rm{\% }} \pm 3{\rm{\% }}$ , precision 99 % ± 2 % $99{\rm{\% }} \pm 2{\rm{\% }}$ , recall 99 % ± 2 % $99{\rm{\% }} \pm 2{\rm{\% }}$ , f-measure 99 % ± 2 % $99{\rm{\% }} \pm 2{\rm{\% }}$ and AUC 100 % ± 0 % ${\rm{AUC\ }}100{\rm{\% }} \pm 0{\rm{\% }}$ . Moreover, the proposed biomarker system based on single PKM2 or Ang-2 data also obtains high accuracy and precision ( 0.91 ± 0.05 , 0.98 ± 0.03 $(0.91 \pm 0.05,0.98 \pm 0.03$ using PKM2); ( 0.96 ± 0.01 , 0.95 ± 0.03 $(0.96 \pm 0.01,0.95 \pm 0.03$ using Ang-2), respectively. In conclusion, these findings suggest that the proposed ensemble model could assist treatment planning for personalized medicines and may be used as a new potential biomarker system to predict the bevacizumab therapeutic effect. Further conclusion and discussion are in the supporting information (Section 4). This study is supported by the Ministry of Science and Technology of Taiwan, under a grant (MOST109-2221-E-011-018-MY3, MOST111-2321-B-016-003), Tri-Service General Hospital, Taipei, Taiwan (TSGHC108086, TSGH-D-109094, TSGH-D110036, TSGH-A-111010 and TSGH-A-112008) and National Taiwan University of Science and Technology—Tri-Service General Hospital (NTUST-TSGH-112-02) There is no competing interest, which influences the work reported in this study. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.

Topics & Concepts