BALTIMORE – Using artificial intelligence and genomic profiling, Mayo Clinic researchers and collaborators derived biomarkers that, if clinically validated, could provide prognostic and predictive information for the clinical management of gastric cancer patients.
In an early retrospective study published earlier this month in Nature Communications, Tae Hyun Hwang, cancer chair at Mayo Clinic Cancer Center in Florida, and colleagues identified a 32-gene signature in a South Korean population that could infer gastric cancer patients' five-year overall survival as well as response to chemotherapy and immune checkpoint inhibitors.
"Gastric cancer is one of the leading cancer types that cause a lot of deaths," said Hwang. While most gastric cancer patients receive chemotherapy as the standard of care, he said, the treatment can be toxic and bear side effects and other complications.
To address this, Hwang said the team spent eight years working on this study, seeking to build a model using genetic markers and artificial intelligence to predict a patient's likelihood to benefit from chemotherapy or immunotherapy as well as overall survival.
To build the model, the researchers developed a machine learning algorithm, called NTriPath, and used it to analyze pan-cancer data from over 6,600 patients published by The Cancer Genome Atlas, aiming to identify pathways that were specifically altered in gastric cancers. From the analysis, the authors identified 32 genes, including TP53, BRCA1, MSH6, PARP1, and ACTA2, that were enriched in the top three gastric cancer-specific pathways: DNA damage response, TGF-ß signaling, and cell proliferation.
To investigate these genes' prognostic utility, the authors conducted a retrospective analysis using microarray-based mRNA expression profiles generated from 567 patients at Severance Hospital, Yonsei University College of Medicine in South Korea. Based on the expression level of the 32 genes, they were able to categorize four distinct molecular subtypes among the patients. Specifically, tumors from Group 1 patients overexpressed genes associated with the cell cycle and DNA repair, while Group 4 cancers overexpressed genes found in TGF-β, SMAD, estrogen signaling, and mesenchymal morphogenesis pathways. Meanwhile, tumors from Group 3 patients overexpressed genes associated with apoptosis signaling and cell proliferation pathways, and Group 2 tumors did not show a distinct pattern of overexpressed genes.
The first goal with these molecular data, Hwang said, was to devise a risk scoring system that can allow clinicians to better predict patients' five-year overall survival. According to him, although cancer staging plays an important role in informing oncologists' treatment plan, the traditional cancer staging system, which categorizes cancers into stage I, II, III, or IV, is rigid and often not sufficient to describe a patient's individual risk and prognosis. In contrast, Hwang said their scoring scheme is "continuous" and "individualized," promising clinicians "a clearer picture" for assessing treatment options for patients.
Additionally, the team investigated whether molecular subtypes could predict response to systemic therapies in gastric cancers. They compared patients who only had surgery to the counterparts who received one of three adjuvant chemotherapy treatments: 5-fluorouracil (5-FU) monotherapy, 5-FU and platinum doublet, or 5-FU plus another class of systemic therapy. The results showed that the 18 patients in molecular subtype Group 3 treated with 5-FU plus platinum showed "significantly better" overall survival compared to the 28 peers in Group 3 without adjuvant chemotherapy. Notably, the dozen molecular subtype Group 1 patients treated with 5-FU plus platinum had worse survival than the 26 Group 1 patients who did not receive adjuvant therapy. This result could "imply that maybe these patients could avoid any unnecessary chemotherapy," Hwang said. In addition, the team found that receiving therapy did not result in survival differences in Group 2 and 4 patients.
When it comes to immune checkpoint inhibitors, the researchers observed a much higher response rate to pembrolizumab in Group 1 and 3 patients versus those in Groups 2 and 5, demonstrating that the molecular subtypes are also predictive to immune checkpoint blockade outcomes in patients with recurrent or metastatic gastric cancer.
"This study provides a promising prognostic and predictive biomarker to guide the clinical care of gastric cancer patients," Ilyas Sahin, a gastrointestinal oncologist at the University of Florida College of Medicine who was not involved with the study, wrote in an email. "It is an important step towards the ultimate goal of giving the right treatment for the right patient at the right time."
One of the exciting aspects of the new signature, Sahin pointed out, is that it is not only predictive for chemotherapy response but also for immunotherapy response, making the signature even more valuable "as immunotherapy seems a viable therapeutic option for gastric cancer with many ongoing studies." Moreover, he said, by generating a risk score with prognostic information, the model described in the study will also help oncologists and patients with treatment decisions.
According to Sahin, current gastric cancer treatments usually include various combinations of surgery, radiation therapy, chemotherapy, targeted therapy, or immunotherapy, depending on the cancer location and stage. And in most cases, clinicians follow general guidelines, rather than individualized therapies to treat patients. "[E]ven if we perform individualized therapy based on some of the markers of their tumors, it is difficult to predict how their tumor will respond to therapy," he pointed out. "Unfortunately, many gastric patients ultimately do not get benefit from these intense therapies with debilitating side effects."
The genetic markers described in this study may eventually help gastric cancer patients and their physicians be better informed about the potential course of their disease and help with the therapy decision process, Sahin said. Assuming the tool will be validated for clinical use, he added, "patients will be able to see the probability of their individualized treatment response information which will help them mak[e] decisions aligning with their goals for lifestyle, quality of life, length of treatment, and other priorities."
Despite its promises, Sahin said the new signature will not drastically change physicians' decision process just yet. "As a physician, it is our job to combine multiple layers of data to provide the best treatment option for our patients," he explained. "This starts with the patient's clinical characteristics, then gets combined with the pathology and extent of the tumor, and different biomarkers such as the ones presented in this paper." Sahin thinks that, ultimately, a predictive test will become "one of the multiple layers of data helping cancer patients and their doctors weigh the benefits of chemotherapy against quality-of-life concerns."
Wafik El-Deiry, associate dean for oncologic sciences at the Warren Alpert Medical School of Brown University and the director of Brown's cancer center, also considers this study "significant." El-Deiry, also not involved with the study, said that although existing biomarkers such as HER2 or microsatellite instability may influence treatment in some gastric cancer patients, these subgroups are still "pretty rare."
"One of the significant findings of this paper is that based on the molecular profile of the tumors, [the authors] found a predictor of doing worse or doing better," El-Deiry added. "If you're a patient who has gastric cancer, you would want to know which group you're in — maybe this affects the risk-benefit discussion of how you should be treated."
However, El-Deiry cautioned that people "have to be careful" when extrapolating the prognostic data in this study. "It's important to stay focused on the relevance of specific data in the paper as it relates to stage of gastric cancer," he noted. For instance, because 89 percent of the patients included in the Yonsei University cohort had stage II or III cancer, the prognostic data about chemotherapy outcome "is really applicable to early stages — stage II, stage III — in the adjuvant setting, which means chemo after surgery," while less applicable to stage IV patients, El-Deiry said. Similarly, since this study only focused on patients with advanced disease when predicting immunotherapy outcomes, results should primarily be used to inform similarly advanced patients.
Furthermore, because this study was conducted in the Korean population, El-Deiry said more validations are still needed for the model to serve other populations. There are "a lot of differences" among different populations, he said, including differences in genomic, diet, microbiome, and environmental exposure.
To that, Hwang said his team is working to recruit more participants from different populations for future studies. Although the authors have validated the model in multiple independent cohorts in this paper, they still aim to further validate the model in both retrospective and prospective studies. Additionally, Hwang said they are working on reducing the number of gene markers from 32 to "hopefully one." Although a 32-gene signature is actionable, he said, his team is currently trying to minimize the number of genes for the model without losing predictive power.
Hwang also hopes to leverage the predictive power of the biomarkers for drug development, noting that he and other co-authors have applied for a patent in multiple jurisdictions for the 32-gene marker and associated technology, and he is hoping to license the patent to his cofounded AI cellular therapy startup, Kure AI, to develop novel therapeutics. “Now [that] we know who is not going to respond to immunotherapy,” he said, “we really want to make new drugs,” adding that his lab is currently working to develop novel cellular therapy or combinatory therapy, which harnesses the synergy between immunotherapy and specific antibodies.