cancer development
Credit: CIPhotos/Getty Images

SAN DIEGO—With the exhibition floor about to close on the third day of AACR 2024, most presenters had deserted their posters. But in one of the dozens of rows, several clusters of conversations were happening all around one poster. Several members from Genialis were explaining krasID—an RNA-based biomarker that uses machine learning to model fundamental aspects of KRAS cancer biology to predict tumor response to and clinical benefit of KRAS inhibitors.

My luck of the draw were Aditya Pai, head of business development, and Sami Takriti, business development manager, who walked me through their poster that culminated in results demonstrating accurate predictions of clinical benefit in real-world non-small cell lung cancer (NSCLC) patient cohorts treated with sotorasib (Lumakras). But Pai and Takriti’s guidance, while helpful, wasn’t necessary because the data is so clear—krasID predicts clinical response to the KRAS G12C inhibitor sotorasib in a real world cohort of non-small cell lung cancer patients.

“Responder” or “not responder” 

Genialis launched in 2017 with a seed round of $2.3 million to advance the visual exploration of next generation sequencing data. This led to the development of Genialis Expressions software, which enables machine-learning driven biomarker discovery by aggregating consistently analyzed and annotated data.

But over the years, Genialis morphed into a computational precision medicine company dedicated to unraveling complex biology to find new ways to address disease. Genialis began to focus its efforts on developing next-generation patient classifiers using machine learning and high-throughput omics data to capture underlying disease biology and predict how patients will likely respond to targeted therapies.

The first iteration of this endeavor was ResponderID. This machine learning platform was designed to identify new biomarkers for drug development and discovery programs as well as diagnostic tests. ResponderID has algorithms that can predict how well treatments will work in the area around a tumor, sort microsatelitte instability (MSI), tumor mutational burden (TMB), and other cutting-edge immune signatures, and help the creation of new drugs like KRAS inhibitors. Behind ResponderID, Genialis raised more than $13 million in Series A in 2023.

Within the ResponderID framework, Genialis developed krasID based on a computational tool called a “classifier.” These machine learning algorithms automatically order or categorize data into one or more of a set of “classes.” One of the most common examples is an email classifier that scans emails to filter them by class label: “Spam” or “Not Spam.” Classifiers have even shown up in pop culture and entertainment, most notably making an appearance in the show Silicon Valley as a tool to classify images as either “hotdog” or “not hotdog.”

Personalizing KRAS Inhibitor Therapy 

The krasID classifier was designed to categorize patient’s RNA sequencing data into “responder” or “not responder”—in reality, krasID doesn’t provide a binary answer but instead generates a probability of response to KRAS inhibition.

To do this, Genialis used RNA sequencing to measure gene expression in tumor tissue, including FFPE samples, and found a number of biological “modules” that were related to or close to KRAS. For example, one biological module represented cellular dependency, a gene expression signature that distinguishes tumors dependent on KRAS for survival and pathway activation. The krasID classifier was trained using five of these models to predict the probability of response to KRAS inhibition.

After being trained, krasID was used on gene expression data from a real-life group of people with NSCLC before they were treated with sotorasib (Lumakras) to guess how the patients would respond to the treatment. Grouping patients by krasID score—either “krasID-high” or “krasID-low”—splits the average Kaplan-Meier survival curve into two survival curves, one shifted towards increased survival and the other towards decreased survival. Specifically, the median survival for patients classified as “krasID-high” (338 days) was more than double that of those classified as “krasID-low” (158 days).

Also of Interest