AI tool predicts response to, survival after treatment with immune checkpoint inhibitors
Key takeaways:
- An AI tool outperformed current FDA-approved biomarkers for predicting survival after immune checkpoint inhibitor therapy.
- Researchers hope this noninvasive approach will help to “democratize precision oncology.”
An AI tool appeared superior to two FDA-approved biomarkers for predicting how people with cancer will respond to immune checkpoint inhibitors.
SCORPIO — a machine learning system developed by researchers at Tisch Cancer Institute at Mount Sinai in cooperation with Memorial Sloan Kettering Cancer Center —uses routine blood tests combined with clinical characteristics from 9,745 patients treated with immune checkpoint inhibitors (ICIs) across 21 cancer types to predict ICI response without the need for advanced genomic or immunologic assays.
An analysis of two internal test sets that included 2,511 patients showed SCORPIO attained median time-dependent area under the receiver operating characteristic curve (AUC) values of 0.763 and 0.759 for predicting OS at 6, 12, 18, 24 and 30 months. This performance surpassed results attained using tumor mutational burden, which achieved time-dependent AUC values of 0.503 and 0.543.
SCORPIO also outperformed tumor mutational burden for predicting clinical benefit in terms of tumor response or sustained stability (AUCs of 0.714 and 0.641 vs. AUCs of 0.546 and 0.573).
Investigators externally validated SCORPIO with data from 10 global phase 3 trials — which included 4,447 patients — and a real-world cohort from Mount Sinai Health System that included 1,159 patients.
The tool outperformed PD-L1 immunostaining in predicting outcomes from ICI treatment.
“Our long-term goal is to democratize precision oncology,” Diego Chowell, PhD, assistant professor at Icahn School of Medicine at Mount Sinai and director of the Chowell Lab, told Healio. “We envision this tool being used globally to guide patient selection for the treatment, and also possibly to monitor response once the patient receives the treatment.”
Healio spoke with Chowell about the motivation for this study, the implications of the findings and the next steps in this research.
Healio: How are patients currently selected for ICIs, and what are the drawbacks to these approaches?
Chowell: At the moment, two major biomarkers are approved by the FDA. One is tumor mutational burden, which represents the number of somatic mutations in the tumor tissue. The higher the number of mutations in the tumor, the higher the probability the patient will respond to immune checkpoint blockade therapy. The other is PD-L1 staining, which determines how much PD-L1 expression is in the tumor tissue.
Although these biomarkers are important, they have limitations. They don’t really accurately predict response — the correlation with response is very weak. Also, these biomarkers require a lot of infrastructure — a good deal of equipment and genomic resources — to use them correctly. Not everyone can do this globally. It also depends on the computational pipelines that people use to identify mutations.
Lastly, these biomarkers are invasive and costly, and getting these measurements from tumor tissue is not very straightforward.
Healio: How did your team develop SCORPIO and how does it work?
Chowell: We developed it based on our prior work published in 2022 in Nature Biotechnology. At that time, it was part of my postdoctoral work at Memorial Sloan Kettering Cancer Center. We developed a machine learning model that integrates clinical, demographic and genomic information. We showed that by using this multimodal approach, we could do much better than tumor mutational burden. In that particular study, we found that basic clinical information — such as levels of albumin — contains a lot of important predictive information related to checkpoint inhibitors. That prompted us to ask a question: If we could have access to all the clinical information and routine blood tests available in the electronic medical records of hospitals, could we make a much better prediction without needing genomic data? We then proceeded to do a comprehensive analysis.
Healio: How did SCORPIO perform?
Chowell: We compared the performance of the machine learning model with tumor mutational burden and PD-L1 expression in terms of clinical response, clinical benefit, and OS. We also compared it with other machine learning models that recently have been published and are based on more complex multi-modal data. For clinical benefit and OS, it outperforms the biomarkers and the more complex machine learning model. We did this analysis across 21 different cancer types.
Healio: Could your findings potentially change the way patients are selected for ICI treatment in the future?
Chowell: That is our hope. Our goal is for everyone to have access to precision oncology. This approach is less expensive, less invasive and more reliable. It comes from the blood of the patient, and it is based on easily accessible information. All this information is already available in the patient’s medical record. We also hope this could be used to improve clinical trial design.
Healio: What are the next steps in research?
Chowell: We are going to be working with different hospitals to prospectively validate the algorithm in different clinical settings.
Healio: Is there anything else you’d like to mention?
Chowell: These routine blood tests have been very important tools in modern medicine. I believe there is a lot of valuable knowledge to be gained from them.
References:
- Chowell D, et al. Nat Biotechnol. 2022;doi:10.1038/s41587-021-01070-8.
- Yoo SK, et al. Nat Med. 2024; doi:10.1038/s41591-024-03398-5.
For more information:
Diego Chowell, PhD, can be reached at diego.chowell@mssm.edu.
Collapse
Healio Interviews
Read more about