NEW YORK — A new gene signature may predict which gastric cancer patients will respond to chemotherapy or to immunotherapy.
After surgery, gastric cancer patients typically undergo chemotherapy. But as Mayo Clinic's Tae Hyun Hwang noted, not all patients benefit from chemotherapy treatment and many experience side effects. This, he added, underscores the need for better biomarkers to guide therapeutic decisions.
As Hwang and colleagues recently described in a Nature Communications paper, they used a machine learning approach to scour the data in The Cancer Genome Atlas to uncover disrupted molecular pathways in gastric cancer. Based on those pathways, they identified a set of 32 genes whose expression patterns not only predict overall survival but also patient response to chemotherapy or immune checkpoint inhibitors.
After validating their findings in additional patient cohorts from South Korea, the researchers are now further refining their gene signature, adapting it for use in resource-limited areas, and prospectively testing it in gastric cancer patients.
Such a signature might be able to discern responders from non-responders more precisely so oncologists can avoid prescribing patients treatments they are unlikely to benefit from and recommend drugs that have a good chance of treating their gastric cancer. "If our signature is right … then some patients just need to receive surgery and actually don't need to receive chemotherapy because if they receive chemotherapy, it's highly likely there will be an adverse event," Hwang said.
When Hwang and his colleagues applied their NTriPath machine learning algorithm to TCGA data, the top three molecular pathways they found altered in gastric cancer were DNA damage response, TGF-beta signaling, and cell proliferation pathways, which encompass 32 genes including TP53, BRCA1, and PARP1, among others.
When they examined microarray-based mRNA expression profiles of those 32 genes within pre-treatment samples from 567 gastric adenocarcinoma patients seen at Severance Hospital at Yonsei University College of Medicine in Seoul, the researchers uncovered four distinct molecular subtypes that corresponded with prognosis.
Group one tumors, for instance, overexpressed cell cycle and DNA repair genes and had the best survival; tumors from group three overexpressed apoptosis and cell proliferation pathway genes and had lower survival than groups one and two; group four tumors overexpressed genes in the TGF-beta, SMAD, estrogen signaling, and mesenchymal morphogenesis pathways and had the worst survival. Tumors from group two did not have a distinct pattern of overexpressed genes and had an intermediate survival.
Using the Yonsei data as a training set, Hwang and his colleagues developed the 32-gene signature into a binary classifier to predict the probability of five-year overall survival among gastric cancer patients and validated it using data from the Asian Cancer Research Group, TCGA, and another previously published dataset from a group at MD Anderson Cancer Center.
After establishing that their risk score was prognostic of five-year overall survival, the researchers were interested to see if the score could also predict treatment benefit. Since the Yonsei cohort included samples from patients before they underwent treatment but also had data on how patients fared on the therapies they eventually received, the researchers were able to compare the treatment outcomes from patients across the four prognostic subgroups.
Patients with group three tumors stood out to the researchers because they did better on 5-fluorouracil plus platinum chemotherapy than those who did not receive adjuvant chemo. Meanwhile, group one patients had worse survival on 5-FU and platinum chemo than those who did not receive adjuvant chemo.
Hwang and his colleagues likewise found that patients in these four groups responded differently to immune checkpoint inhibitors by analyzing a previously published cohort of patients from Samsung Medical Center and additional patients from Yonsei and Seoul St. Mary's Hospital. Patients in groups one and three responded better to Merck's Keytruda (pembrolizumab) compared to patients in groups two and four.
Hwang said his group's gene signature could identify patients who might respond to treatments they might not otherwise be eligible to receive. For instance, Hwang noted that microsatellite instability is currently a standard biomarker oncologists consider when deciding whether to prescribe patients checkpoint inhibitors. But Hwang and colleagues' gene signature is predictive of immunotherapy benefit independent of other biomarkers like MSI or Epstein-Barr virus status and identified MSI-high and other tumors that respond to such treatment.
Hwang and his colleagues are now refining their gene signature. In particular, Hwang said he wants to whittle it down to only the genes strictly necessary to classify tumors to make it a more viable clinical tool. To that end, in addition to continuing to test their 32-gene set in both retrospective and prospective studies, he and his colleagues are evaluating whether a leaner 10-gene subset can predict chemotherapy and immunotherapy response as well as prospectively testing whether a single gene can predict immunotherapy response.
At the same time, recognizing that transcriptomic analysis is not always easily accessible in the developing world, Hwang's group is developing a way to translate its findings to tools that are readily available in lower-income areas. For example, Hwang and his colleagues found that their gene signature subgroups correlated with certain morphological features on H&E slides, which are commonly used to diagnose cancer. They are investigating whether scanning these features and applying AI-based algorithms could predict treatment response in patients.
Hwang further noted that gastric cancer is a heterogeneous disease that manifests differently across populations and his group hopes to study the signature in diverse cohorts. Their signature, he noted, was isolated from TCGA, which is a cohort of largely white individuals, and then further analyzed in Korean cohorts. The researchers are now examining the signature's generalizability to Hispanic and Eastern European populations, which like Asian populations, have a high incidence of gastric cancer.