Unlocking the future: mitochondrial genes and neural networks in predicting ovarian cancer prognosis and immunotherapy response

Introduction

Ovarian cancer (OC) is one of the most common gynecological cancers worldwide, with over 230,000 new cases and 150,000 deaths annually (1-3). According to studies from the United States and United Kingdom registries, one in six women dies within 90 days of diagnosis (1,2). Genetic factors, gene mutations (such as BRCA1/BRCA2), nulliparity, infertility, endometriosis, obesity, and age are associated with the incidence of OC, while pregnancy, oral contraceptives, and non-steroidal anti-inflammatory drugs are potential protective factors (1,4). High-grade serous OC is the most common subtype of OC, accounting for over 70% of cases (5). The treatment of OC still primarily involves surgery, chemotherapy, and targeted drugs. With the deepening of research, immunotherapy represented by immune checkpoint inhibitors (ICIs) is being explored for the diagnosis and treatment of OC (6).

The traditional view holds that anaerobic glycolysis in the Warburg effect is the main energy source for tumor growth (7). However, more and more evidence shows that the macromolecule synthesis of tumor cells depends on mitochondrial metabolism, and strategies targeting the mitochondrial oxidative respiratory chain can be developed for cancer treatment (8). For example, the MYC pathway, one of the most common pathways in tumor synthetic metabolism, which is associated with increased mitochondrial and oxygen consumption, has been found to be more highly expressed in OC (9,10). Changes in the MYC-CDK2/4-RB1 signaling pathway can be seen in 75% of ovarian clear cell carcinomas, and high expression of MYC is associated with platinum resistance and poorer prognosis. Moreover, MYC small interfering RNA has been proven to inhibit the growth of OC tumor cells (11-13).

This study collected data from The Cancer Genome Atlas Program (TCGA) and the Gene Expression Omnibus (GEO), sorted out mitochondria-located genes, and used neural network technology to establish models predicting the prognosis and immunotherapy response of OC patients. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-24-1233/rc).

Methods

Data sources

The TCGA data were downloaded from the UCSC-XENA website (https://xena.ucsc.edu/), selecting the OC dataset [Genomic Data Commons (GDC) TCGA Ovarian Cancer] and retaining samples with both prognostic data and next-generation sequencing (NGS) data for subsequent analysis. The OC immunotherapy dataset was downloaded from GEO (https://www.ncbi.nlm.nih.gov/geo/) (accession number GSE188249). Both TCGA and GEO data were analyzed using fragments per kilobase million (FPKM) sequencing results, and gene symbol conversion was performed using clusterprofiler. Mitochondrially localized genes were obtained from the MitoCarta3.0 database (https://www.broadinstitute.org/mitocarta).

This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). Given that all data were retrieved from the publicly available TCGA and GEO databases and patients’ private information has been concealed and is not traceable, this retrospective cohort was exempted from ethical considerations and written consent.

Study design

This retrospective study was designed as a diagnostic test. The main outcome was the overall survival (OS) time of OC patients, with secondary endpoint being the immunotherapy response. Sensitivity to immunotherapy was defined according to the data source authors’ definition, which meant sensitive patients with OC were those showed partial response (PR) or complete response (CR) after the use of ICIs (pembrolizumab), according to response evaluation criteria in solid tumors (RECIST) 1.1. TCGA data were used to construct an OC prognostic model, with patients randomly sampled into a discovery cohort (70% of total) or a validation cohort (remaining 30% of total), using the discovery cohort to build the prognostic model (based on neural network technology) and validating model performance in the validation cohort. Similarly, GEO data were used to construct an OC immunotherapy efficacy model, with 70% of the GEO data randomly selected as the discovery cohort (for model construction) and the remaining 30% as the validation cohort for model validation. There may be slight variations in gene expression detection across different assay kits, with some genes potentially undetectable. Therefore, in the prognostic model, we used the intersection of genes measured by TCGA and mitochondrially localized genes as predictors. Similarly, in the immunotherapy efficacy model, we used the intersection of genes measured by GEO and mitochondrially localized genes as predictors (Figure 1).

Figure 1 The flow chart of this study.

Model training

The discovery cohort served as the basis for model training, while the validation cohort was used for further evaluation of the model’s performance. The model was constructed based on neural network theory, and to enhance its performance, batch normalization layers (these layers automatically standardized the data by subtracting the mean and dividing by the standard deviation), and batch training functions (these functions can automatically extract subsets of data for training) were adopted. To prevent overfitting, dropout layers (they were configured to silence 20% of neurons during each training iteration to prevent certain neurons from becoming overly dominant), and early stopping functions were employed (which automatically terminate training when there is no significant improvement in model performance after several rounds). Adam was chosen as the optimizer, with the learning rate set between 0.01 and 0.05. In predicting patients’ survival, considering this is a time-to-event classification task rather than a traditional classification task, we built the neural network based on the DeepSurv theory by Katzman et al. (14). The model was built in python 3.9 and using packages pytorch, torchtuples, pandas, matplotlib and numpy.

Model evaluation and compression

We evaluated the model from two aspects: discrimination and calibration. The primary metric for evaluating discrimination was the area under the receiver operating characteristic curve (AUC). Generally, an AUC closer to 1 indicates better model performance, while an AUC closer to 0.5 suggests that the model’s predictions are akin to random guessing. The model has good performance when the AUC is greater than 0.7. Other indicators include sensitivity, specificity, accuracy, ‎negative predictive value (NPV) and positive predictive value (PPV).

Besides, we used the Hosmer-Lemeshow goodness of fit test to judge the calibration ability of the models. If the P value is greater than 0.05, it indicates that the model has good calibration ability.

We truncated the survival data at 1, 2, and 3 years respectively, and evaluated the predictive performance of the model in detail.

Ultimately, we used the PySide6 software package to compress the entire model into a Windows executable program for ease of use by clinicians.

Statistical analysis

Statistical analysis was performed using R (version 4.2.0). Numeric data were compared using the Wilcoxon test, while categorical data were compared using the Chi-squared test or Fisher’s exact test. A two-sided P value of less than 0.05 was considered statistically significant.

Results

Clinical characteristics of patients

In the prognostic model, a total of 375 TCGA OC patients were included for analysis, with 262 patients randomly drawn into the discovery cohort and 113 into the validation cohort. The median survival time in the discovery cohort was 1,012 days [interquartile range (IQR), 547–1,712 days], and in the validation cohort, it was 1,032 days (IQR, 394–1,562 days), with no significant difference between the two groups (P=0.54). In the discovery cohort, 104 patients (39.69%) survived, and in the validation cohort, 41 patients (36.28%) survived, with no significant statistical difference (P=0.53). In the immunotherapy efficacy model, 18 patients were assigned to the discovery cohort, of which 50.00% (9 patients) were sensitive to treatment. In the validation cohort, 25.00% (2 out of 8 patients) were sensitive to immunotherapy. There was no significant statistical difference between the two groups (P=0.39) (Table 1).

OC prognostic model

We intersected TCGA sequenced genes with mitochondrially localized genes, and 1,113 genes were detected simultaneously (Figure 2A). Then, using these 1,113 genes as predictive variables and the patient’s OS as the outcome variable, we built a neural network in Python. After 68 rounds of training, the early termination function automatically ended the training (Figure S1A). The final OC prognostic model included 10 hidden layers, including an input layer of 1,113×8, a rectified linear unit (ReLU) activation layer, a normalization layer, a 20% drop out layer, an 8×4 linear layer, another ReLU activation layer, a normalization layer, a 20% drop out layer, a 4×1 linear layer, and a Sigmoid activation layer (Figure 2B). The average AUC of the prognostic model was 0.7268 [95% confidence interval (CI): 0.7258–0.7278] for the discovery cohort and 0.6475 (95% CI: 0.6466–0.6484) for the validation cohort (Table 2).

Figure 2 The process of constructing prognostic and immunotherapy efficacy model of ovarian cancer. The prognostic model’s predictive genes (A), neural network structure (B), and receiver operating characteristic curve (C). The immunotherapy efficacy model’s predictive genes (D), neural network structure (E), and receiver operating characteristic curve (F). TCGA, The Cancer Genome Atlas Program; GEO, Gene Expression Omnibus; ReLU, rectified linear unit.

We performed cutoffs at 1, 2, and 3 years to measure the performance of the OC prognostic model in detail. The receiver operating characteristic (ROC) curves of OC prognostic model at 1, 2, and 3 years are shown in Figure 2C. In the discovery cohort, at 1 year, the AUC was 0.7597 (95% CI: 0.6252–0.8942), specificity 0.7869, sensitivity 0.6667, accuracy 0.7786, NPV 0.9697, PPV 0.1875; at 2 years, AUC 0.7734 (95% CI: 0.6908–0.8560), specificity 0.9481, sensitivity 0.5800, accuracy 0.8779, NPV 0.9054, PPV 0.7250; at 3 years, AUC 0.7461 (95% CI: 0.6789–0.8134), specificity 0.9022, sensitivity 0.5897, accuracy 0.8092, NPV 0.8384, PPV 0.7188. In the validation cohort, at 1 year, the AUC was 0.6827 (95% CI: 0.5114–0.8541), specificity 0.7451, sensitivity 0.6364, accuracy 0.7345, NPV 0.9500, PPV0.2121; at 2 years, AUC 0.7250 (95% CI: 0.6128–0.8373), specificity 0.7931, sensitivity 0.6538, accuracy 0.7611, NPV 0.8846, PPV 0.4857; at 3 years, AUC 0.7087 (95% CI: 0.6126–0.8049), specificity 0.7895, sensitivity 0.6216, accuracy 0.7345, NPV 0.8108, PPV 0.5897.

The goodness-of-fit test results showed P values of 0.67, 0.63, and 0.43 at 1, 2, and 3 years, respectively, in the discovery cohort. In the validation cohort, the P values were 0.08, 0.10, and 0.11 at 1, 2, and 3 years, respectively (Table 3).

OC immunotherapy efficacy model

We intersected GEO sequenced genes with mitochondrially localized genes, and 1,114 genes were detected simultaneously. We then used these 1,114 genes as predictive variables and the patient’s sensitivity to immunotherapy (CR + PR) as the outcome variable to build a neural network in Python (Figure 2D). After 114 rounds of training, the early termination function automatically ended the training (Figure S1B). The final OC immunotherapy efficacy model included 10 hidden layers, including an input layer of 1,114×4, a ReLU activation layer, a normalization layer, a 20% drop out layer, a 4×2 linear layer, another ReLU activation layer, a normalization layer, a 20% drop out layer, a 2×1 linear layer, and a Sigmoid activation layer (Figure 2E).

The ROC of the OC immunotherapy efficacy model is shown in Figure 2F. In the discovery cohort, the model AUC was 0.9444 (95% CI: 0.8333–1.0000), specificity 1.0000, sensitivity 0.8889, accuracy 0.9444, NPV 0.9000, PPV 1.0000. In the validation cohort, the model AUC was 0.9167 (95% CI: 0.6667–1.0000), specificity 0.8333, sensitivity 1.0000, accuracy 0.8750, NPV 1.0000, PPV 0.6667.

The results of the goodness-of-fit test showed that in the discovery cohort, the P value of the model was 0.37, while in the validation cohort the P value was 0.12 (Table 4). The model’s parameters are presented in the supplement file (available at https://cdn.amegroups.cn/static/public/TCR-24-1233-Supplementary.pdf).

Model packaging

Considering the black box characteristics of neural networks, we packaged the OC immunotherapy efficacy model into a Windows 64-bit executable program for ease of use by clinicians (Figure 3). After launching the tool, users can prepare the patient’s NGS FPKM file following the format of the built-in example file. By clicking “Choose NGS file and predict”, users can select the sequencing file to activate the built-in pre-trained neural network. After the calculation is completed, the software will display the probability of the patient being sensitive to immunotherapy in the “Result” part.

Figure 3 The use of the tool for predicting immunotherapy response of ovarian cancer. NGS, next-generation sequencing; FPKM, fragments per kilobase million.

Discussion

As one of the most common lethal tumors in women, OC has a 5-year survival rate of approximately 45% (15). The standard treatment for OC encompasses methods such as surgery and chemotherapy. In recent years, poly-ADP-ribose polymerase inhibitors (PARPi) and immunotherapy have emerged as new therapies, attracting increasing attention (16-18). Most immunotherapeutic drugs for OC are still undergoing clinical trials, with the objective response rate for ICIs remaining limited to about 6–15% (18).

Moreover, the efficacy of immunotherapy is still lacking effective predictive factors. The predictive value of classical indicators, such as programmed death-ligand 1 (PD-L1) and tumor mutational burden (TMB), remains limited. A study by Song et al. found that the AUC value for PD-L1 in predicting immunotherapy efficacy was only 0.569; similarly, the ability of TMB to predict the prognosis of immunotherapy patients and its reproducibility across different samples have been questioned (19-22). Therefore, there is an urgent need for a new indicator or algorithm to identify the population sensitive to immunotherapy, in order to avoid unnecessary drug treatments and adverse events.

Traditional prognostic predictions for patients often employ the Cox proportional hazards model. This model, based on linear assumptions, fits various risks and calculates the probability of positive events. However, real-world situations are often more complex and non-linear, which limits the application of Cox regression. Machine learning algorithms, particularly deep learning neural networks, have gained increasing recognition from clinicians due to their superior predictive performance (23-25). Our preliminary studies also found that neural networks show excellent predictive value for both prognostic prediction and immunotherapy efficacy (26-28).

Increasing evidence suggests that mitochondria play a role in supplying energy to key metabolites in tumors. And inhibiting mitochondrial oxidative phosphorylation and corresponding nucleotide metabolism, the tricarboxylic acid cycle, etc., are potential methods for tumor treatment (8,29). It is revealed that mitochondria-related signaling pathways and genes are associated with patient prognosis and drug resistance (11-13). Therefore, the combination of mitochondrial-related genes and neural network algorithms to predict patient prognosis and drug response holds both theoretical and clinical value.

In this study, we integrated data from TCGA and GEO to establish neural network models. These models use mitochondrial localization genes as predictive variables to predict the prognosis and immunotherapy response of OC patients. The average AUC for the OC prognostic model was 0.7268 for the discovery cohort and 0.6475 for the validation cohort. For the OC immunotherapy efficacy model, the average AUC was 0.9444 for the discovery cohort and 0.9167 for the validation cohort. This study shows that the use of mitochondrial localization genes and neural networks can predict the prognosis and immunotherapy response of OC patients. Therefore, we packaged it as a Windows executable tool for the convenience of clinicians, considering there is limited way to predict OC patients’ immunotherapy response.

Yang et al. constructed OC prognostic model based on the investigation of ferroptosis-related long non-coding RNA, and it scored 0.793 AUC in discovery data and 0.681 AUC in validation data (30). Li et al. developed prognostic models for OC based on ferroptosis and necroptosis, and the model had about 0.584–0.728 AUC (31). Jiang et al. built OC prognostic model by combination of transcriptomic and proteomic data, which showed around 0.596–0.749 AUC (32). Except for Shaoyi Yang’s model (30), which had a similar AUC to ours, the rest of the models performed worse. This indicates both the feasibility of predicting OC prognosis at the genetic level and the complexity of OC patient prognosis. A multi-dimensional approach may be required to further enhance the model’s performance. Chen et al. found that autophagy-related genes could be used to predict OC survival and were related to their immunotherapy response (33). Wang et al. also observed a correlation between TMB-based genes and immune infiltration in OC (34). Unfortunately, they did not further build upon these findings to create predictive models for assessing OC immunotherapy efficacy.

There are limitations in this study. As a retrospective study, it may suffer from selection bias and information bias. More prospective studies and larger sample data could better improve and validate the developed models.

Conclusions

The application of mitochondrial genes and neural networks shows potential in predicting the prognosis and immunotherapy response in OC patients. And this approach could provide valuable insights for personalized treatment strategies.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-24-1233/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-24-1233/prf

Funding: This work was supported by the National Key Technology Research and Developmental Program of China (Program Nos. 2022YFC2704400 and 2022YFC2704405), and the 14th Five-Year Project-Preclinical Research and Application of Fertility Protection in Patients with Gynecological Malignant Tumors (Topic 5, No. 2022YFC2704402).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-24-1233/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013).

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Lheureux S, Braunstein M, Oza AM. Epithelial ovarian cancer: Evolution of management in the era of precision medicine. CA Cancer J Clin 2019;69:280-304. [Crossref] [PubMed]
2. Webb PM, Jordan SJ. Epidemiology of epithelial ovarian cancer. Best Pract Res Clin Obstet Gynaecol 2017;41:3-14. [Crossref] [PubMed]
3. Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin 2024;74:12-49. [Crossref] [PubMed]
4. Werner H. BRCA1: An Endocrine and Metabolic Regulator. Front Endocrinol (Lausanne) 2022;13:844575. [Crossref] [PubMed]
5. Prat J. Ovarian carcinomas: five distinct diseases with different origins, genetic alterations, and clinicopathological features. Virchows Arch 2012;460:237-49. [Crossref] [PubMed]
6. Morand S, Devanaboyina M, Staats H, et al. Ovarian Cancer Immunotherapy and Personalized Medicine. Int J Mol Sci 2021;22:6532. [Crossref] [PubMed]
7. Hanahan D, Weinberg RA. Hallmarks of cancer: the next generation. Cell 2011;144:646-74. [Crossref] [PubMed]
8. Vasan K, Werner M, Chandel NS. Mitochondrial Metabolism as a Target for Cancer Therapy. Cell Metab 2020;32:341-52. [Crossref] [PubMed]
9. Li F, Wang Y, Zeller KI, et al. Myc stimulates nuclearly encoded mitochondrial genes and mitochondrial biogenesis. Mol Cell Biol 2005;25:6225-34. [Crossref] [PubMed]
10. Emmings E, Mullany S, Chang Z, et al. Targeting Mitochondria for Treatment of Chemoresistant Ovarian Cancer. Int J Mol Sci 2019;20:229. [Crossref] [PubMed]
11. Reyes-González JM, Armaiz-Peña GN, Mangala LS, et al. Targeting c-MYC in Platinum-Resistant Ovarian Cancer. Mol Cancer Ther 2015;14:2260-9. [Crossref] [PubMed]
12. Jung M, Russell AJ, Liu B, et al. A Myc Activity Signature Predicts Poor Clinical Outcomes in Myc-Associated Cancers. Cancer Res 2017;77:971-81. [Crossref] [PubMed]
13. Murakami R, Matsumura N, Brown JB, et al. Exome Sequencing Landscape Analysis in Ovarian Clear Cell Carcinoma Shed Light on Key Chromosomal Regions and Mutation Gene Networks. Am J Pathol 2017;187:2246-58. [Crossref] [PubMed]
14. Katzman JL, Shaham U, Cloninger A, et al. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Med Res Methodol 2018;18:24. [Crossref] [PubMed]
15. Chandra A, Pius C, Nabeel M, et al. Ovarian cancer: Current status and strategies for improving therapeutic outcomes. Cancer Med 2019;8:7018-31. [Crossref] [PubMed]
16. Odunsi K. Immunotherapy in ovarian cancer. Ann Oncol 2017;28:viii1-7.
17. Yang C, Xia BR, Zhang ZC, et al. Immunotherapy for Ovarian Cancer: Adjuvant, Combination, and Neoadjuvant. Front Immunol 2020;11:577869. [Crossref] [PubMed]
18. O'Malley DM, Krivak TC, Kabil N, et al. PARP Inhibitors in Ovarian Cancer: A Review. Target Oncol 2023;18:471-503. [Crossref] [PubMed]
19. Song P, Li W, Guo L, et al. Identification and Validation of a Novel Signature Based on NK Cell Marker Genes to Predict Prognosis and Immunotherapy Response in Lung Adenocarcinoma by Integrated Analysis of Single-Cell and Bulk RNA-Sequencing. Front Immunol 2022;13:850745. [Crossref] [PubMed]
20. Carbone DP, Reck M, Paz-Ares L, et al. First-Line Nivolumab in Stage IV or Recurrent Non-Small-Cell Lung Cancer. N Engl J Med 2017;376:2415-26. [Crossref] [PubMed]
21. Hellmann MD, Ciuleanu TE, Pluzanski A, et al. Nivolumab plus Ipilimumab in Lung Cancer with a High Tumor Mutational Burden. N Engl J Med 2018;378:2093-104. [Crossref] [PubMed]
22. McGrail DJ, Pilié PG, Rashid NU, et al. High tumor mutation burden fails to predict immune checkpoint blockade response across all cancer types. Ann Oncol 2021;32:661-72. [Crossref] [PubMed]
23. van der Velden BHM, Kuijf HJ, Gilhuijs KGA, et al. Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Med Image Anal 2022;79:102470. [Crossref] [PubMed]
24. Jiang Y, Yang M, Wang S, et al. Emerging role of deep learning-based artificial intelligence in tumor pathology. Cancer Commun (Lond) 2020;40:154-66. [Crossref] [PubMed]
25. Chen Y, Wang J, Wang C, et al. Deep learning models for disease-associated circRNA prediction: a review. Brief Bioinform 2022;23:bbac364. [Crossref] [PubMed]
26. Li S, Li W, Ma T, et al. Assessing the efficacy of immunotherapy in lung squamous carcinoma using artificial intelligence neural network. Front Immunol 2022;13:1024707. [Crossref] [PubMed]
27. Li W, Zhang M, Cai S, et al. Neural network-based prognostic predictive tool for gastric cardiac cancer: the worldwide retrospective study. BioData Min 2023;16:21. [Crossref] [PubMed]
28. Li W, Lin S, He Y, et al. Deep learning survival model for colorectal cancer patients (DeepCRC) with Asian clinical data compared with different theories. Arch Med Sci 2023;19:264-9. [Crossref] [PubMed]
29. Bajzikova M, Kovarova J, Coelho AR, et al. Reactivation of Dihydroorotate Dehydrogenase-Driven Pyrimidine Biosynthesis Restores Tumor Growth of Respiration-Deficient Cancer Cells. Cell Metab 2019;29:399-416.e10. [Crossref] [PubMed]
30. Yang S, Ji J, Wang M, et al. Construction of Ovarian Cancer Prognostic Model Based on the Investigation of Ferroptosis-Related lncRNA. Biomolecules 2023;13:306. [Crossref] [PubMed]
31. Li Y, Gong X, Hu T, et al. Two novel prognostic models for ovarian cancer respectively based on ferroptosis and necroptosis. BMC Cancer 2022;22:74. [Crossref] [PubMed]
32. Jiang J, Chen Z, Wang H, et al. Screening and Identification of a Prognostic Model of Ovarian Cancer by Combination of Transcriptomic and Proteomic Data. Biomolecules 2023;13:685. [Crossref] [PubMed]
33. Chen Y, Deng Z, Sun Y. Development of a risk model based on autophagy-related genes to predict survival and immunotherapy response in ovarian cancer. Hereditas 2023;160:4. [Crossref] [PubMed]
34. Wang H, Liu J, Yang J, et al. A novel tumor mutational burden-based risk model predicts prognosis and correlates with immune infiltration in ovarian cancer. Front Immunol 2022;13:943389. [Crossref] [PubMed]