NEW YORK – After publishing a validation study last week in the Journal of Clinical Oncology, Io9 is seeking additional funding and a pharma partner to develop its deep learning-based classifier DeepHRD into a companion diagnostic that can identify patients eligible for PARP inhibitors.
The La Jolla, California-based firm views DeepHRD, developed on Io9's OncoGaze AI platform, as an alternative or potentially a replacement for next-generation sequencing (NGS)-based tests for cancer patients, which can be costly and, in some cases, take weeks to return results, delaying the initiation of treatment.
Io9 anticipates pairing DeepHRD with a PARP inhibitor approved for treating ovarian cancer as its first regulatory submission. The US Food and Drug Administration has approved two companion diagnostics that gauge homologous recombination deficiency (HRD) in the ovarian cancer setting: Myriad Genetics' MyChoice CDx and Foundation Medicine's FoundationOne CDx. The MyChoice CDx test is a next-generation sequencing-based in vitro diagnostic test that measures alterations in BRCA1 and BRCA2 genes and genomic instability to determine an ovarian cancer patient's eligibility for treatment with AstraZeneca's PARP inhibitor Lynparza (olaparib). Myriad has recently signed deals to expand global access to MyChoice CDx and develop the laboratory-developed test as a kit that can be performed by other labs.
Similarly, FoundationOne CDx is an in vitro diagnostic test that detects alterations in 324 genes and certain genomic signatures, such as microsatellite instability and tumor mutational burden. It's approved for identifying patients eligible for targeted therapies across a range of solid tumor indications, including ovarian cancer patients with BRCA1/2 alterations who can get Lynparza.
Both tests require DNA extracted from formalin-fixed, paraffin-embedded tumor samples. Myriad touts a turnaround time of 14 days or less from receipt of sample for MyChoice CDx, while Foundation Medicine reports a median turnaround time of 8.8 days. In real-world use, genomic testing can cost thousands of dollars per patient and take as long as six weeks after surgery to return results including the time required to process tumor tissue and ship the sample to a laboratory for analysis.
Researchers from Io9 and the University of California, San Diego saw the potential for AI to offer an alternative to NGS to evaluate biomarker status in cancer patients so they could begin treatment sooner and save on costs, particularly for patients outside the US who have limited access to NGS testing.
"While we've made great progress in precision medicine, the data is showing that we're falling short," Io9 CEO Greg Hamilton said. For example, a survey published in the Journal of Clinical Oncology in 2022 of nearly 38,000 patients with stage IV non-small cell lung cancer showed that only 22 percent had molecular test results in their medical records and only 3 percent received targeted therapies, even though guidelines at the time recommended testing all patients for EGFR, ALK, and ROS1 mutations.
Hamilton said his group's goal was to prove they can identify biomarkers in the clinic based on imaging and determine an appropriate treatment at the time of cancer diagnosis. "Cancer is a race against time," he said. "Especially for aggressive cancers like triple-negative breast or ovarian, waiting three to six weeks is not tenable."
The team developed two DeepHRD breast cancer models by training and validating a neural network algorithm to predict HRD and differences in outcomes from more than a 1,000 each of flash-frozen and formalin-fixed paraffin-embedded slides from The Cancer Genome Atlas. Within the training process, the researchers assigned each whole slide image an HRD label calculated using NGS testing or genotyping data according to conventional methods. They calculated the HRD scores from sequencing or genotyping data using the open source software application scarHRD. Typically, in breast cancer, an HRD cutoff of 42 or more has been used to identify patients for treatment with platinum-based chemotherapy or PARP inhibitors in triple-negative breast cancer, while a score of 63 or more has been the threshold for ovarian cancer.
Senior author Ludmil Alexandrov, a professor in the department of cellular and molecular medicine at UCSD and Io9 CSO, said the company's OncoGaze platform is a "completely new AI framework" that allows researchers to train a model that behaves like a pathologist. The model, like a human, first identifies regions of interest in an image at low resolution and then zooms in within those regions at higher resolution to make a prediction. Alexandrov said the resulting model is a "very practical approach" that allows doctors to skip "laborious and expensive" NGS testing by generating the biomarker based on a hematoxylin and eosin (H&E)-stained slide instead. "You can get the prediction, and an oncologist can start treating [the patient] immediately," he said.
In the recently published study, Alexandrov and his collaborators tested the FFPE model in samples from an external cohort of 77 patients with metastatic breast cancer who had available whole-exome sequencing and digital H&E-stained whole slide images from Georges-François Leclerc Cancer Center in Dijon, France. All patients had received platinum-based chemotherapy, and 54 of those had first been treated with taxane chemo.
The researchers used a similar process to train and validate a separate DeepHRD model in 459 TCGA ovarian cancer samples, and they tested that model in a retrospective real-world cohort of 141 patients with advanced high-grade serous ovarian cancer generated by Memorial Sloan Kettering Cancer Center. Those patients had undergone treatment with first-line platinum-based chemotherapy and surgery. Samples from this cohort had available whole-slide images, overall survival outcome data, and targeted genomic sequencing.
In the platinum-treated metastatic breast cancer cohort from Leclerc, patients identified as having HRD cancers had a median progression-free survival of 14.4 months compared to 3.9 months for patients identified as homologous recombination proficient (HRP), and the algorithm discriminated between patients who had a complete response, partial response, or no response with an area under the curve of 0.76.
In comparison, separation of the groups based on BRCA1/2 mutations or HRD scores derived using genomic tests did not lead to significant differences in median progression-free survival. Stratifying patients based on specific pathogenic mutations they harbored and that were gauged using an HRD gene panel did show a difference in progression-free survival. However, DeepHRD classified 1.8-fold more patients than the HRD test.
In the real-world ovarian cancer cohort, HRD classification using DeepHRD yielded a three-year overall survival probability of 70.3 percent for HRD patients and 50.2 percent for HRP patients. DeepHRD identified 3.1-fold more patients with HRD than a test based on mutations in a set of HRD-related genes. Neither mutations in BRCA1/2 nor the COSMIC single base substitution SBS3 signature calculated by SigMA, a machine learning tool, showed any significant differences in overall survival in this cohort.
"We've proven we can identify clinically actionable biomarkers, and we can identify responders and non-responders," which was the team's primary goal in the study, Hamilton said.
With the validation study complete and published, Hamilton said Io9 is now raising funds through Series A financing to support a regulatory filing for DeepHRD, seeking its marketing authorization as a test to identify ovarian cancer patients eligible for PARP inhibitors. For that filing, Hamilton said the company hopes to partner with a pharmaceutical manufacturer of an approved PARP inhibitor and submit data to the FDA from a retrospective analysis of that company's drug trials.
Although Io9 is also planning to study DeepHRD in prospective clinical studies as a companion diagnostic, Hamilton doesn't expect those studies will be necessary to garner FDA approval for the test. "The precedent for [PARP inhibitor companion diagnostics] getting FDA approved is retrospective [data] because PARP inhibitors have been given to all of the patients in the Phase III trial, so you know the clinical outcomes," Hamilton said, pointing to the FDA's approval of MyChoice CDx and FoundationOne CDx. In regulatory submissions for those products, Hamilton explained, the sponsors only had to show that the tests differentiated between responders and non-responders by conducting a retrospective analysis of trials in which patients had received PARP inhibitors.
Io9 also aims to develop models similar to DeepHRD for other clinically actionable biomarkers such as KRAS, EGFR, and BRAF, partnering with pharma for those projects, too.
Hamilton said one advantage of DeepHRD is that it compares favorably to NGS-based tests in that it does not consume any of the tumor sample and its implementation doesn't necessitate a lab. "We're a software company," Hamilton said. "We'll have a cloud-based software [application] that we will sell to pathologists."
He also views DeepHRD as a more equitable test than an NGS test. "The US is the only place in the world where we're going to spend $3,000 or $5,000 on a test to figure out what drug to put you on," Hamilton said. "The rest of the world does not do that."
In other countries, for example, he said a patient with triple-negative breast cancer might only receive single-gene tests for BRCA1/2 mutations rather than one of the more expensive HRD tests. "That's a poor man's version of an HRD test, and they miss a ton of women that would respond to [HRD-targeted] therapies, whereas with a digital image we can [analyze] an image from anywhere in the world and eliminate a lot of the barriers that we see in clinical practice today."
Io9 believes that OncoGaze-based tests could eventually replace NGS testing as companion diagnostics for cancer drugs. "OncoGaze will be a faster, cheaper, better test that can be used as an alternative to their respective NGS tests," Hamilton said. Although some digital pathology companies are working on similar biomarker tests, Hamilton said that "to our knowledge, none have published clinically validated data or shown the capability to submit their data to the FDA for companion diagnostic approval."