Bioinformatics platform predicts optimal combinations of cancer treatments
Click Here to Manage Email Alerts
The development of targeted cancer therapies has led to improved clinical outcomes for many patients. However, monotherapies against a single target can result in treatment resistance.
Co-occurring alterations, such as mutations in genes associated with two signaling pathways, often are essential in driving tumor progression.
“We know that the individual driver genomic alterations are most likely critical for tumor physiology, but we also know that multiple oncogenic alterations in human genome, transcriptome and proteome all function together and drive oncogenesis,” Anil Korkut, PhD, assistant professor of bioinformatics and computational biology at The University of Texas MD Anderson Cancer Center, told Healio. “So, on one side, you want to choose the right combination therapy for the right patient, but we also know that the alteration landscape is complex in each patient.”
To address these complexities, Korkut and his colleagues at MD Anderson developed a bioinformatics platform that predicts favorable treatment combinations for specific patient groups based on co-occurring tumor alterations. The platform, Recurrent Features Leveraged for Combination Therapy (REFLECT), has performed well in retrospective validation studies, selecting combinations that yielded improved therapeutic outcomes in clinical and preclinical studies.
Korkut spoke with Healio about the complex mechanisms driving tumor progression, the efficacy of REFLECT, and the need to proceed carefully in its continued development.
Healio: What inspired you to develop the REFLECT tool?
Korkut: Precision oncology, sequencing and other profiling technologies have contributed significantly to cancer treatment in the past 2 decades, but a major problem is resistance to therapy or lack of initial response. There is a consensus that combination therapies could be used to overcome resistance to therapy or achieve more durable responses in larger patient cohorts, compared with monotherapies. Of course, the major goal is to be able to identify the combination therapies that benefit specific cancer patient groups. This is a challenging problem, because molecular landscapes that may indicate response to combination therapies are complex.
Healio: How does the REFLECT platform work?
Korkut: This is a highly multidisciplinary approach that involves a bioinformatics and machine-learning component, a clinical component and a genomics component. We have access to a large amount of data from diverse resources and different modalities at the DNA-to-protein level. We first generated a large compendium of multiomic data from patient-derived xenografts and cell lines. We’ve generated a resource of more than 10,000 tumors of patients with cancer.
From this meta-cohort, we generated 200 cohorts of patients with individual alterations such as EGFR-mutated cancers, and then we applied to each cohort our machine-learning algorithm that identifies a smaller number of critical, highly discriminate features that co-occur in patient subcohorts. Those co-alterations in patient subcohorts form what we call the REFLECT signatures. These are sets of composite features that define smaller groups of patients and may indicate important oncogenic events that are co-occurring in a group of patients.
Once we had the signatures, we mapped them to potential combination therapies using the information stored in precision oncology knowledge bases. We then validated the predictions on combination therapies in independent data sets, from cell lines, patient-derived xenografts and tumors of patients. Through retrospective analyses of data from a small clinical trial, we demonstrated that combination therapies selected with our method are leading consistently to more significant responses as improvements in OS and PFS compared with those combinations that were not matched by our method.
I want to note that the validations, particularly the clinical study, is preliminary in a small patient group. Our knowledge of such combination therapies remains limited. There will be more data available for further validation.
Healio: Does this tool hold promise for predicting outcomes?
Korkut: Possibly. I think data science, bioinformatics and artificial intelligence approaches will become increasingly important to precision oncology. I believe our tool may one day deliver meaningful benefit to patients with cancer. That is our ultimate goal.
Healio: What is next in your research on this?
Korkut: There are some critical future goals — one is that much of these data are from bulk samples. We are already working to identify such co-occurring features using single-cell RNA data to address the role of tumor heterogeneity. I think there is a very important signal hidden in the tumor microenvironment in the vicinity of the tumor cells, and we are studying those effects using single-cell methodologies that monitor the spatial features in the tumor microenvironment.
We are also very interested in understanding and hopefully predicting drug-induced toxicities. We have a therapeutic window, and on one side we look at efficacy, but REFLECT does not address toxicity. We want to extend our approach to better understand toxicities, but that is in very early stages.
Healio: Is there anything else you’d like to mention?
Korkut: We are in the early stages of integrating these bioinformatics tools to actual applications. We hope this puts the concept on the table, but we need to move forward very carefully, considering major challenges including toxicity.
Lastly, this was a highly collaborative work, and I’d like to acknowledge the major contributions of my lead author, Xubin Li, PhD, as well as my other colleagues at MD Anderson and my external collaborators at Harvard.
For more information:
Anil Korkut, PhD, can be reached at Korkut Lab, The University of Texas MD Anderson Cancer Center, 1515 Holcombe Blvd., Houston, TX 77030; email: akorkut@mdanderson.org.
Collapse
Li X, et al. Abstract LB119. Presented at: American Association for Cancer Research Annual Meeting; April 8-13, 2022; New Orleans.
Click Here to Manage Email Alerts