Memorial Sloan Kettering Cancer Center researchers have created a multimodal machine learning (ML) tool that can identify patients with metastatic breast cancer most likely to respond to CDK4/6 inhibitors based on clinicopathologic and genomic features.
“There’s a huge need in clinic to identify patients who may or may not benefit from adding CDK4/6 inhibitors at the time of metastatic diagnosis so that we can think about escalation and de-escalation strategies in advance,” said Pedram Razavi, MD, PhD, scientific director of the Global Research Program at Memorial Sloan Kettering Cancer Center. “More accurate prediction of outcomes could also help some patients avoid unnecessary side effects and financial toxicity from escalated upfront approaches.”
Speaking at the 2024 San Antonio Breast Cancer Symposium, he explained that although adding of CDK4/6 inhibitors to endocrine therapy has led to marked improvement of outcomes in patients with hormone receptor (HR)-positive, human epidermal growth factor receptor (HER)2-negative metastatic breast cancer, responses can vary widely and some patients experience early treatment resistance.
At present, certain clinical features, such as treatment-free interval and measurable disease, are the main factors used to identify patients who may be at high risk for early progression on first-line CDK4/6 inhibitor combinations. Razavi and team wanted to explore whether a multimodal ML model that included additional clinical and genomic features could more accurately stratify patients.
Using OncoCast-MPM, a ML tool created at Memorial Sloan Kettering, they generated three models to predict progression-free survival (PFS) with CDK4/6 inhibitors: one based on clinicopathological features (CF), one that used genomic features (GF), and one that combined CF and GF (CGF).
The models were developed using a training cohort of 761 patients with HR-positive, HER2-negative metastatic breast cancer who received first-line endocrine therapy with CDK4/6 inhibitor combinations and had MSK-IMPACT targeted tumor sequencing performed prior to treatment or within two months of the start of treatment. The performance of the model was tested on a holdout test cohort of 326 patients.
Razavi reported that the models trained on CF and GF each identified three risk groups. Patients in the high-risk group had a median progression-free survival (PFS) of 6.3 months when categorized according to the CF model and 9.9 months using the GF model. PFS in the corresponding intermediate-risk groups was 15.2 and 18.1 months, while in the low-risk groups it was a respective 24.5 and 23.1 months.
The multimodal CGF model identified four risk groups. In this case, the median PFS was 5.3 months in the high-risk group, 10.7 and 19.8 months in the two intermediate risk groups, and 29.0 months in the low-risk group.
Of note, the hazard ratio for disease progression or death between the high- and low-risk groups was significantly higher in the CGF model (a 6.5-fold difference) compared with that for the CF and GF models (three- to four-fold difference), indicating a superior stratification of patients with the CGF model. The team observed similar results for the holdout test cohort yielded nearly identical PFS and hazard ratio results, confirming the robustness of the models.
“All three models performed really well, surpassing the conventional clinical risk models based on a single or a few clinical features. But the power of the analysis shone when we started combining the clinical and genomic features together,” Razavi said.
He also pointed out that the top 10 key features used by the multimodal model represent a combination of the top five clinical and molecular features of the unimodal models. These features, in order of importance were tumor mutational burden, fraction of genome altered, TP53 alteration, fraction of genome with loss of heterozygosity, presence of liver metastasis, adjuvant treatment free interval of less than one year, primary tumor grade III, presence of visceral metastasis, primary tumor progesterone receptor negativity, and whole genome doubling.
“All of these variables are potentially available when the patients are diagnosed with metastatic disease, making such ML models broadly applicable. The hope is to integrate these models in clinical trial design of escalation and de-escalation strategies potentially transforming how we approach treatment for newly diagnosed metastatic disease,” Razavi said. “Knowing that a patient on first-line CDK4/6 inhibitors is in the high-risk group could prompt the treating oncologists to implement closer disease monitoring and utilizing liquid biopsy and tumor-derived biomarkers to inform second-line treatment options and clinical trials. This could put us one step closer to staying ahead of breast cancer.”
Razavi and team also observed an increasing trend in the frequency of actionable mutations, particularly TP53, across increasing CGF risk groups. He said that TP53 could therefore “potentially be a biomarker in this group of patients.”
Limitations of this study include its single-institution design, retrospective data analysis, and potential referral bias associated with specialized cancer centers. To address these challenges, Razavi and co-investigators are validating the model using external data sets and aim to develop an online tool where physicians can input clinical and genomic data to receive patient-specific outcome predictions.
