SAN ANTONIO – Researchers have developed a machine learning (ML) model to predict whether hormone receptor (HR)-positive, HER2-negative breast cancer patients are likely to respond to CDK4/6 inhibitors and endocrine therapy.
At the San Antonio Breast Cancer Symposium on Friday, Pedram Razavi, a breast medical oncologist at Memorial Sloan Kettering Cancer Center, discussed the ML model's predictive performance using data from 1,000 patients with HR-positive, HER2-negative metastatic breast cancer who received first-line treatment with a CDK4/6 inhibitor with endocrine therapy.
"The use of CDK4/6 inhibitors have completely evolved our practice and improved the outcomes for our patients, but there is an extensive amount of heterogeneity in the outcomes of patients," Razavi said at the meeting. "In clinic, there's a group of patients who derive no benefit from the addition of CDK4/6 inhibitors to endocrine therapy, and there's a group of patients who do outstanding and remain on these combinations for years."
There are several CDK4/6 inhibitors approved with endocrine therapy for the frontline treatment of HR-positive, HER2-negative breast cancer, including Pfizer's Ibrance (palbociclib), Novartis' Kisqali (ribociclib), and Eli Lilly's Verzenio (abemaciclib).
Verzenio with endocrine therapy is also approved in an even earlier treatment setting, as an adjuvant treatment for HR-positive, HER2-negative early breast cancer patients at high risk of recurrence and, earlier this year, the US Food and Drug Administration expanded Kisqali's indication as an adjuvant treatment of HR-positive, HER2-negative early breast cancer. To identify the eligible population for Verzenio, doctors consider patients at high risk of recurrence based on their tumors' clinical features such as nodal status, tumor size, and tumor grade, while when considering treatment with Kisqali, doctors identify patients' high-risk tumors by nodal status and their scores on molecular assays, such as Exact Sciences' Oncotype DX, Veracyte's Prosigna PAM50, Agendia's MammaPrint, or Myriad Genetics' EndoPredict.
Razavi noted that previous studies, such as the INAVO120 trial, identified certain clinical characteristics that were associated with improved outcomes on CDK4/6 inhibitors, such as whether patients had progressed during or within 12 months of adjuvant endocrine therapy or if they had measurable disease. Razavi also highlighted research that has characterized mechanisms of response and resistance to CDK4/6 inhibitors.
"We are using 1990s risk scores to stratify patients, that's the reality now," Razavi said at a press conference. "We look at treatment-free interval, we count the number of [affected] lymph nodes, we measure the size of the tumor. A lot of the features that we are using for risk are the features that are outdated at this point. We can do much better. The science has moved on, but the risk stratification has not."
Hoping to do better, Razavi and his team developed a machine learning model based on patients' clinical and genomic data collected at baseline. They drew on data from 1,087 patients included in the MSK Breast Translational Program data platform to develop the ML model. They used 70 percent of the data to train the model and 30 percent to test its performance.
The goal was to use the ML model to predict patients' progression-free survival and compare the model's performance to risk assessments based on either just clinical criteria or just genomic data from MSK's OncoKB precision oncology database. Razavi said that the clinical model and the genomic model alone did a good job stratifying patients by risk, but that the combined ML model, using both clinical and genomic data to predict outcomes, "performed significantly better than the clinical or genomic model alone," Razavi said. The combined model identified four distinct categories of risk: low, intermediate-low, intermediate-high, and high risk.
In the training set, patients in the low-risk category had a median progression-free survival of 29 months while high-risk patients had a median progression-free survival of 5.3 months. Clinical features such as liver metastases, disease-free interval, and progesterone receptor (PR)-negative status along with genomic features such as TP53 alteration, high tumor mutation burden, and high genomic loss of heterozygosity were associated with risk of progression.
The performance of the ML model on the test dataset was "almost identical" to the training set, Razavi said. In the test set, the median progression-free survival was 24.6 months for the low-risk group and 3.9 months for the high-risk group. Median progression-free survival was 19.4 months and 9.1 months for the intermediate low-risk group and the intermediate-high group, respectively. Razavi added that there was no statistically significant difference between the hazard ratios of the training and test sets.
The researchers then explored whether the ML model could improve upon the conventional clinical risk criteria used for selecting patients, which includes treatment-free interval of more than 12 months and measurable disease. When applying the ML model to the high- and low-risk groups determined by clinical criteria, it determined that 20 percent of the patients deemed low risk based purely on clinical features were actually high risk. Similarly, when assessing patients in the high-risk group based on clinical features, the model determined that 32 percent were actually low risk.
"With this kind of clinical-genomic model we may be able to both identify more high-risk patients that may need treatment escalation, as well as better stratify endocrine-resistant cases to avoid overtreatment," Frederick Howard, an assistant professor of medicine at the University of Chicago Medicine, said in a discussion of the study results.
One biomarker in the combined model that had a significant association with higher risk was TP53 mutations. In the high-risk group identified by the ML model, 75 percent of patients harbored a TP53 mutation.
Going forward, Razavi's team aims to validate the ML model and, in particular, use external cohorts to validate its predictive ability. They also hope to integrate more data, such as circulating tumor DNA genomic features and radiology and digital pathology images, to further refine the model. The researchers' ultimate goal is to develop a validated and clinically practical single-patient outcome predictor tool.
"The increased prevalence of certain targetable alterations, such as BRCA1/2, PTEN, ESR1, and p53, in high-risk groups suggests an opportunity for first-line treatment escalation approaches," Razavi said. "I would hope we could see these type of models in adaptive clinical trials for escalation and de-escalation [of treatment] for our patients."
Howard, who was not involved in developing the ML model, said it could be used to better select patients who need more intense upfront treatment, such as the addition of a PIK3CA inhibitor to CDK4/6 and endocrine therapy for patients with PIK3CA, PTEN, or AKT alterations or the addition of CDK2 inhibitors to the combination for those with TP53 mutations.
"There is a lot happening in breast cancer at this point, but everything comes with a price: We add to the toxicity of the patient, and there is a financial toxicity to what we do," Razavi said. "If you want to escalate treatment for patients, you should identify the group that benefits from this."
He added that the model needs further validation and testing in patient populations outside of MSK to ensure there is not a single-institution bias in the model.
"There's a huge amount of data in everything we do these days, and we are not using it to predict outcomes for the patients," Razavi added. "We can easily develop prediction models that include multiple features, this is the present, it's completely doable."