NEW YORK – In spite of promising results showcased at multiple scientific meetings this year, Nucleai has made a strategic decision not to further develop its deep learning model for predicting which non-small cell lung cancer patients are likely to have a durable benefit with Merck's checkpoint inhibitor Keytruda (pembrolizumab).
However, the firm will pursue its spatial biology approach to help pharmaceutical companies personalize next-generation immunotherapies in early development.
Nucleai's classifier is trained to predict immunotherapy benefit based on the spatial arrangement of immune cells in the tumor microenvironment. Patients with a positive score had not yet reached a median overall survival after two years, while overall survival for patients with a negative score was 17.8 months. Among the positive score group, two-year overall survival was 70.8 percent versus 33 percent in the negative score group.
The research that Nucleai presented at the American Society of Clinical Oncology's annual meeting and the European Society for Medical Oncology Congress this year is part of a larger initiative to identify predictive biomarkers for immunotherapy, according to Ori Zelichov, VP of clinical development at Nucleai.
While immune checkpoint inhibitors are highly effective in the short term, only about 20 to 30 percent of patients experience a durable response. "We don't have good predictive biomarkers to know in advance which [patients] should be treated and which should not be treated with those drugs," said Zelichov. That's a problem, he added, not just for patients and physicians, but also for pharma companies developing the next generation of immuno-oncology therapies and for the entire healthcare system that spends billions of dollars on drugs that end up being ineffective.
As opposed to targeted therapies that act directly on the tumor to inhibit a specific cancer driver, patients receiving immuno-oncology drugs can't be stratified based on a genomic test. And immunohistochemistry-based PD-L1 testing has not proven to be a reliable method for predicting which patients will benefit from checkpoint inhibitors.
Nucleai researchers investigated whether, by using spatial biology to analyze interactions between the immune system and tumor microenvironment, it is possible to find biomarkers predictive of sensitivity to immuno-oncology drugs.
In a retrospective multi-center study conducted in the US and Israel, where the firm is based, Nucleai analyzed biopsy slides and clinical outcomes data from about 200 patients treated with first-line, single-agent Keytruda. They then trained their machine learning algorithm using 100 slides from patients who responded and 100 slides from those who didn't and generated a spatial signature from the images that distinguished between the two groups. Factoring in features such as the interactions between the cells and the densities of different cell populations, they built a classifier to score patients positive or negative for response to immunotherapy.
The classifier was able to separate responders and nonresponders independent of other approved biomarkers such as PD-L1 expression. Nucleai is now expanding the study to include more patients and hopes to publish the results.
The company is not disclosing all of the features used to build its classifier, but Zelichov said one of them is the proximity of tumor-infiltrating lymphocytes to tumor cells. "We know that hot tumors are associated with being more infiltrated. What we found in this study is that if you look at a patient that is a responder and a patient that is not responding, the absolute number of TILs infiltrating the tumor is the same," Zelichov explained. "But if you look at the specific interaction between the cells, you can see that for the resistant patient that had a lot of TILs, the two subpopulations are more separated. There is some kind of barrier between them."
In the sensitive patients, however, the TILs are in closer proximity with the tumor cells, leading to more frequent interactions. "And that's very interesting because even hot tumors like lung cancers or melanoma can be further stratified by that method," said Zelichov.
Some similar studies have been done linking spatial biology signatures to immune checkpoint inhibitor response. However, Zelichov said that those studies included a lot of "mixed information," such as combining first-line, second-line, and third-line patients, or patients on single-agent and combination therapies.
In contrast, Nucleai was able to carry out a "very clean" study using only patients who received first-line single-agent Keytruda due to its access to medical centers with longitudinal data on hundreds of patients, according to Zelichov. "That provided us with a very unique dataset," he said.
South Korean biotech Lunit has also developed an AI-based classifier to predict checkpoint inhibitor response based on an analysis of TILs in tumor tissue. That test is only available for research use at the moment.
Theoretically, this sort of test can be used to stratify patients for a clinical trial or to guide therapy for patients for an approved drug. The biopsy slide would be scanned at the hospital, then uploaded into the cloud, and the algorithm, running in the cloud, would assign the score.
Although it would be possible to do so, Nucleai isn't planning to commercialize its classifier for NSCLC patients receiving Keytruda. "Strategically, Nucleai today is not working on taking these specific findings further into the clinic," Zelichov said. "The work we did is more of a proof-of-concept that is aiming to show the power of spatial biology in understanding immuno-oncology and the way immuno-oncology drugs work."
Nucleai further doesn't see value in pursuing a clinical trial and regulatory approval of the classifier as a tool for guiding NSCLC treatment with checkpoint inhibitors already on the market, where standard practices are already well established. Keytruda is already approved as a single-agent for first-line treatment of PD-L1-positive NSCLC patients and for PD-L1-positive metastatic NSCLC patients after progression on platinum chemotherapy, and the drug is approved for all comers in multiple other tumor types.
Instead of launching its classifier for predicting response to marketed immunotherapies, Nucleai is aiming to develop similar biomarkers for next-generation immune checkpoint inhibitors currently in early clinical trials. The company's strategy is to partner with pharmaceutical companies first and get in on the ground floor by developing a companion diagnostic alongside their drugs in the early development phase. That way, the risk is lowered for Nucleai, and it can generate revenue through the development process.
Many leading pharma companies "struggle … to understand how to enroll patients into the [immunotherapy] trial," Zelichov said, because they lack a biomarker to select patients most likely to benefit from the investigational drug. Nucleai is working to apply spatial biology to that problem by developing histopathology signatures that can identify patients who will be sensitive to a therapy, thus enabling pharmas to carry out more stratified clinical trials with a greater likelihood of advancing to Phase III and registration.
While immuno-oncology diagnostics are the centerpiece of Nucleai's pipeline, the company is also working on similar tests for antibody-drug conjugates in hematology indications and for CAR T-cell therapies. In some cases, rather than fully realized biomarkers, the tests are envisioned as an improvement on quantification or scoring for an already-developed assay, such as an IHC assay for a drug target. The software could then aid in quantification of the test in an automated manner.