1. Wiley, Frieda PharmD

Article Content

The emergence of immune checkpoint inhibitors (ICIs) has given the oncology world new hope by offering targeted therapies to counter a continually growing list of cancers, including bladder cancer, gastroesophageal cancer, and melanoma. However, the purported benefits also come with three noted limitations.

artificial intellige... - Click to enlarge in new windowartificial intelligence. artificial intelligence

ICIs demonstrate limited efficacy, as less than one-third (30%) of patients with solid tumors respond to the treatment. Even then, those patients successfully treated run the risk of post-treatment toxicity. Additionally, ICI response-associated biomarkers treatment frequently proves unable to predict a response to treatment, hindering the fidelity and reliability of precision medicine. Moreover, previous research suggests that approximately 60 percent of biomarker response can be explained by ICI response-a finding alluding to the fact that many novel factors are pending discovery.


Now, researchers are exploring the potential of machine learning (ML) to bridge the gap by accurately predicting outcomes of immunotherapy treatment across numerous ICI datasets while identifying potential biomarkers. The findings were published recently in Nature Communications (2022;


The research builds upon the author's previous work, which suggests that biomarkers that contribute to the anti-cancer drug response are located near drug targets in a protein-protein interaction network. Their findings also indicate that scientists can use patient-derived organoid models to identify biomarkers associated with a therapeutic effect. This information can also help predict drug responses, such as in the case of 5-fluorouracil-treated colorectal cancer and cisplatin-treated bladder cancer.


Study Details

In this new study, the authors sought to identify specific pathways that contribute to the ICI response by concentrating on biological pathways proximal to the ICI targets. Investigators used NetBio as input features to carry out immunotherapy-response predictions generated by machine learning. Immunotherapy target genes, tumor microenvironment-based biomarkers, and pathways selected from data-driven machine learning approaches served as the negative controls. Researchers tested the AI's input features predictive performance based on its ability to predict drug response based on its ability to reduce tumor size following immunotherapy treatment or the patient's survival.


Training the ML model using supervised learning entailed engaging various combinations of treating and test datasets to evaluate prediction performance consistency. The tools involved within-study predictions, or studies in which training and test datasets were produced by a single cohort.


With all inputs in place, the tool was trained by applying logistic regression. Investigators then tested the fidelity of the tool's predictive performance, focusing on two outcomes to measure predictive performance. These were drug response, quantified by the measurement of tumor size reduction post-treatment, and patient survival.


After measuring the generalizability of biomarkers via extensive cross-validation and cross-study predictions, researchers found Net-Bio-based predictions proved more accurate than expression-based predictors of various biomarkers and outperformed other models by consistently delivering solid predictions. Examples include programmed cell death-1 (PD-1), programmed cell death-ligand-1 (PD-L1), or cytotoxic T-lymphocyte antigen-4 (CTLA-4), as well as markers found in the microtumor environment such as CD8 T cell, T-cell exhaustion, cancer-associated fibroblast, and tumor-associated macrophage markers.


The software, NetBio, consistently demonstrated the ability to predict the response to ICI treatment and overall survival. This includes an ability to consistently predict outcomes in melanoma datasets. Conversely, investigators noted that relatively weaker prediction performances regarding drug target expression, such as in the case of PD-1 for nivolumab and pembrolizumab, PD-L1 for atezolizumab, and CTLA-4 for patients treated with ipilimumab.


Some of the study's many additional important findings include observing that Net-Bio-based ML predicted outcomes with higher fidelity than other top-level tools, such as immunotherapy-response prediction and a deep neural network-based method. Comparatively, NetBio outperformed these two methods in predictions in 33 out of 34 comparisons. It also delivered stronger results in 17 of 18 comparative methods in across-study predictive performance.


While data-driven ML models for clinical applications routinely fail to consistently perform in new data sets regardless of the degree of comprehensive ML, NetBio consistently delivered improved predictions.


The paper's authors believe their most recent work sets the stage for new research opportunities in precision medicine exploring ICI treatment. They also believe that coupling a network-based ML model along with diverse omic layers will yield better clinical results-an immutable asset to the world of precision medicine.


Frieda Wiley is a contributing writer.