A.I., Imaging Shown to Predict Immunotherapy Success

(Lawrence Livermore National Lab)

27 August 2018. Researchers in France developed a system for computing data from CT scans and genomic analysis to predict the success for therapies that invoke the immune system in patients with solid-tumor cancers. A team from the Gustave Roussy cancer center near Paris, as well as other institutions and the start-up company TheraPanacea in Paris, describe their system in the 14 August issue of the journal The Lancet Oncology (paid subscription required).

The researchers are seeking ways to determine in advance if patients can benefit from immunotherapy treatments that harness the immune system to fight cancer. The team focused on two of the more established immunotherapy types: anti-programmed cell death protein 1, or anti-PD1, and anti-programmed cell death ligand 1, or anti-PDL1. In these therapies, signals to a tumor’s protective cells are blocked, allowing the activation of T-cells, white blood cells in the immune system, to attack the tumor. The treatments can also, in some cases, block the ligand or binding molecule in PD1 proteins, also allowing an effective immune response against the tumor.

While the overall mechanism is understood, only a fraction of cancer patients, from 15 to 30 percent, receiving these therapies respond to the treatments. A factor correlating with immunotherapy success is the richness of lymphocytes, white blood cells in the immune system, surrounding the tumor. Therefore the researchers hypothesized that calculating a numeric value to characterize the extent of lymphocytes surrounding the tumor could help predict the success of immunotherapy with those patients.

Calculating this value, called a radiomic signature, however, is anything but simple. The team of medical researchers, cancer physicians, and bioinformatics specialists found they had to account for a range of variables, including expression of genes related to the immune system as well as the concentration of lymphocytes surrounding the tumor. The researchers designed a process with machine learning to inspect computed tomography, or CT, scans of tumors. At the same time, their algorithm also accounts for expression of the CD8b gene found on the surface of tumor-killing T-cells that regulates efficient cell to cell interactions with the immune system.

To train the algorithm, researchers used data from a clinical trial that screened the genomics of solid tumor cancer patients for treatments targeting their genomic profiles. Participants in this trial, individuals with several types of solid tumor cancers, had CT scans taken of their tumors as well as molecular profiles of their tumors. The algorithm includes 8 variables to calculate a radiomic signature that predicts expression of tumor-killing CD8b proteins, when treated by anti-PD1 or -PDL1 immunotherapies.

The team validated their algorithm with data from the Cancer Genome Atlas that collects data on cancer genomics, then tested the algorithm with 3 independent data sets from early-stage clinical trials testing anti-PD1 and -PDL1 immunotherapies. The results show patients in whom immunotherapies succeeded within 3 or 6 months, as well as participants with longer overall survival times, have higher radiomics scores than their trial counterparts.

The researchers next plan to refine their algorithm to include tissue analysis data from biopsies, as well as validate the radiomics scores in prospective clinical trials. TheraPanacea, the start-up company taking part in the team, is a spin-off enterprise from CentraleSupélec in Paris, developing artificial intelligence applications for radiology.

More from Science & Enterprise:

• Biotechs Collaborate on More Precise Gene, Cell Cancer Therapies
• A.I. System Catches Tiny Lung Tumors Missed by Humans
• Big Data, A.I. Applied to Precision Medicine for Lung Disease
• Blood Tests Shown Able to Identify Early Lung Cancer
• Data Tools Designed for Genomics-Based Precision Cancer Care

*     *     *