NEW YORK – Researchers from the Icahn School of Medicine at Mount Sinai and Memorial Sloan Kettering Cancer Center aim to test a new artificial intelligence tool for predicting response to immune checkpoint inhibitors in prospective clinical trials following a positive validation study published in Nature Medicine earlier this month.
Immune checkpoint inhibitors are approved for a range of cancer indications. However, just 20 percent to 40 percent of patients respond to treatment. Drugmakers have sought to identify patients most likely to benefit from these therapies using immunohistochemistry-based PD-L1 expression testing, as well as genomic profiling to gauge tumor mutational burden or other types of immune system analysis. However, oncologists don't think PD-L1 testing is very accurate in its identification of responders and non-responders, and other tests can be costly and are not available in every clinical setting. The researchers set out to develop an AI model that could make predictions about immunotherapy response by drawing on information collected from every patient in the course of their treatment using simple, inexpensive, and routine tests.
The tool, dubbed SCORPIO, which stands for Standard Clinical and LabOratory features for Prognostication of Immunotherapy Outcomes, makes predictions based on complete blood count and comprehensive metabolic profile plus other clinical data like sex, age, lines of prior therapy, and smoking history. Researchers trained SCORPIO on data from more than 1,600 patients with 17 types of cancer who were treated at Memorial Sloan Kettering. Then, as described in the recently published study, they developed and evaluated the model using real-world data collected on nearly 10,000 patients with 21 cancer types at Memorial Sloan Kettering, Mount Sinai, and within 10 global Phase III clinical trials.
Patients in the study had been treated with PD-1, PD-L1, or CTLA-4 inhibitors or various combination regimens and had been followed for a median of more than two years. Clinical trials included in the analysis, such as IMbrave150, IMspire150, and others, involved patients treated with Roche's PD-L1 inhibitor Tecentriq (atezolizumab) alone or in combination with other agents. Among the clinical trial datasets, the model performed particularly well when predicting overall survival among patients in the IMvigor211 and IMspire150 trials. The former tested Tecentriq with chemotherapy in patients with advanced bladder cancer, and the latter tested a Tecentriq combination regimen in advanced BRAF V600-mutated melanoma.
To gauge the performance of the model in predicting overall survival in these studies, researchers relied on area under the curve (AUC) analyses. Using data from the IMvigor211 group, the model had an AUC of 0.782, and with data from IMspire150, it was 0.684. An AUC of 0.5 suggests a model's predictive performance is no better than random chance, while an AUC of 1 indicates perfect predictive performance.
Scorpio did even better in the real-world cohorts, where it achieved a median AUC of 0.809 for predicting overall survival. It also outperformed tumor mutational burden and PD-L1 testing when compared head-to-head for predicting immune checkpoint inhibitor efficacy.
While SCORPIO performed well for predicting overall survival in many of the trial cohorts, it did not perform as well at predicting tumor shrinkage. "SCORPIO's more modest performance in predicting tumor response is expected in the context of immunotherapy, especially [immune checkpoint inhibitors], where the link between tumor response and patient survival can be weak, influenced by factors like pseudo-progression, delayed responses, and development of new lesions followed by responses," the study authors wrote in the paper, noting that future iterations of the model will have improved prediction capabilities for surrogate endpoints in addition to overall survival.
Diego Chowell, an assistant professor at Mount Sinai and senior author of the study, noted that in addition to SCORPIO's ability to predict survival and tumor response, the model also uncovered an interesting relationship between the routine blood tests and immune cell populations. "A lot of the information that is [collected] in the routine blood tests is highly correlated with immune phenotypes in the tumor microenvironment," he said.
Unlike some other AI-driven platforms for predicting immunotherapy response such Pangea Biomed's ENLIGHT-DP, SCORPIO does not make use of imaging data such as hematoxylin and eosin-stained (H&E) slides. Chowell said such data is not as easily obtained for training and testing purposes from external sources, so his team didn't use them to develop this iteration of the model. He did not rule out evaluating in the future whether H&E slides could improve SCORPIO's predictions.
Currently Chowell and his colleagues are focused on validating SCORPIO in a prospective trial. "One of the limitations of the [published] study is that all of the data used was retrospective," Chowell said. "It will be very important to do prospective studies using the algorithm to demonstrate applicability in different clinical contexts."
Pangea CEO Tuvik Beker noted that a blood-based test has the advantage of being easier to carry out and less invasive than an image-based test requiring a tumor biopsy. However, he added that "the advantage of using H&E slides is that they can teach us about the tumor, its environment, and the surrounding immune system."
Another difference between an approach like Pangea's and SCORPIO is that SCORPIO focuses on features that relate to the potential efficacy of any immune checkpoint inhibitor, regardless of the specific checkpoint being blocked. Meanwhile, Pangea's ENLIGHT zeroes in on interactions between the tumor and its environment, allowing the model to make predictions about the potential efficacy of different checkpoint inhibitors for a patient.
"SCORPIO could provide a quick and affordable test to stratify patients according to their chance of responding well to immune checkpoint inhibition," Beker said. "It could improve upon widely used markers like tumor mutation burden and PD-L1 expression. The problem is that current drug labels are tied to the old biomarkers, making their replacement quite difficult even when convincing evidence for better biomarkers is presented."
In developing SCORPIO, Chowell and his collaborators envisioned that physicians could use it to prioritize treatment options for patients, including immune checkpoint inhibitors, cytotoxic therapies, and biomarker-driven therapies, and even assess the risks and benefits of treatment or enrich clinical trials with patients more likely to benefit from immune checkpoint therapies.
Toward that end, Chowell has cofounded a company, IOPredict, to advance SCORPIO as a commercial tool that can enable precision oncology and is planning to work with the US Food and Drug Administration to gain clearance for the model as a laboratory-developed test.