ML Model Predicts Cancer Patient Response to Immunotherapy

For some cancer patients, immunotherapy may be a far better treatment alternative than undergoing chemotherapy or radiotherapy. Unfortunately, current diagnostic methods for identifying patients who can benefit from immunotherapy treatment and drugs are unreliable. However, a new study suggests that deployment of machine learning can dramatically improve accuracy in predicting how a patient will respond to immunotherapy.

The Diagnostic Challenge

Immunotherapy activates the body’s immune system to fight cancer cells without use of chemotherapy or radiotherapy. In addition to significantly improving survival rates, use of immunotherapy causes fewer side effects than conventional cancer drugs. And, in leveraging the immune system’s memory and adaptability, immunotherapy’s therapeutic benefits also tend to be longer lasting and more sustainable. For these reasons, the list of cancers for which immunotherapy treatment is utilized has expanded in recent years.

The cloud to this silver lining is that only approximately 30 percent of cancer patients actually experience these benefits; toxicity may also occur after immunotherapy. That makes it essential to come up with a reliable method of identifying biomarkers capable of detecting which patients will respond positively to immune checkpoint inhibitors (ICIs) and other immunotherapy drugs across multiple cohorts. However, existing biomarkers used for predicting cancer drug response tend to be contradictory and unreliable—with an accuracy of only 60 percent, according to one study (Litchfield, K. et al. Meta-analysis of tumor- and T cell-intrinsic mechanisms of sensitization to checkpoint inhibition. Cell, 184, 596–614.e14 (2021)).

The New Machine Learning Model

With this in mind, a research team led by professor Sanguk Kim of Pohang University of Science and Technology (POSTECH) in South Korea has set out to improve the accuracy of predicting patient response to ICIs via the use of network-based machine learning. Network biology is based on the observation that genes with similar phenotypic roles tend to co-localize in a specific region of a protein-protein interaction (PPI) network. Network-based machine learning leverages this tendency to identify gene modules that are much more robust than single gene-based approaches in predicting phenotypic outcomes.

The team analyzed the clinical results of more than 700 patients with three types of cancer (melanoma, bladder cancer, and gastroesophageal cancer), along with the transcriptome data of their cancer tissues to discover new network-based biomarkers. They then used those biomarkers to develop artificial intelligence (AI) capable of accurately predicting the response to anticancer treatment. Treatment response prediction based on the newly discovered biomarkers was not only accurate but superior to conventional anticancer treatment biomarkers including immunotherapy targets and tumor microenvironment markers, the study, which was published in Nature Communications on June 28, concludes.

The study builds on the team’s previous work supporting the effectiveness of machine learning in predicting drug responses to chemotherapy in patients with gastric or bladder cancer. This study demonstrates that AI using the interactions between genes in a biological network can successfully predict the patient response to not only chemotherapy, but also immunotherapy in multiple cancer types.

Takeaway

The POSTECH study is part of a growing body of evidence showing that AI and machine learning can be used effectively to predict cancer patient response to immunotherapy and design personalized treatment plans to generate optimal results.