Predicting Palbociclib Outcomes in Breast Cancer Using Deep Learning

In this episode of Emerging Experts, Xiaojie Zhang, MD, a hematology/oncology fellow, and Akshat Singhal, PhD, a postdoctoral scholar, both at UC San Diego, shed light on innovative research leveraging deep learning to predict how patients with ER-positive/HER2-negative (ER+/HER2-) breast cancer will respond to palbociclib (Ibrance), a common first-line treatment for this patient population. Their work, fueled by the desire to improve precision oncology, demonstrates the significant potential of artificial intelligence (AI) in guiding cancer care.

In the episode, Singhal explains that the initial aim was to develop AI models capable of pinpointing genetic markers associated with either resistance or positive response to the drug. Collaborating closely with oncologists, including Zhang, ensured the research remained focused on real-world patient needs. Zhang highlights the clinical importance of this work, noting the lack of reliable predictive tools for CDK 4/6 inhibitors like palbociclib.

“The reason is that while this class of therapy is approved by the FDA in the first-line treatment setting for estrogen receptor-positive/HER2-negative advanced or metastatic breast cancers, there are still critical clinical gaps in these areas,” explains Zhang.

The study, a retrospective analysis of patients treated at UC San Diego between February 2016 and September 2023, included 139 patients with biopsy-proven ER+/HER2- advanced breast cancer who had available tissue-based genomic sequencing data.¹ The researchers used this genomic data as input for their "palbo-VNN" model, an ensemble of 5 deep learning models. Tumors were categorized as either palbociclib-sensitive (palbo-VNNsen) or palbociclib-resistant (palbo-VNNres) based on the model's predicted score.

The findings revealed a contrast in treatment duration and overall survival between patients deemed "sensitive" by the model and those predicted to be "resistant." In the first-line setting, where 54% of the patients received palbociclib, the median therapy duration for patients with palbo-VNNsen tumors was 49.6 months, compared with 15.6 months for those with palbo-VNNres tumors (HR, 0.63; P =.01). Moreover, patients with palbo-VNNsen disease had a median overall survival of 79.3 months, versus 45.7 months for those with palbo-VNNres disease (HR, 0.68; P =.03).

Beyond the predictive power, Zhang and Singhal emphasize the need for interpretable AI models in oncology. By understanding why a model makes a certain prediction, clinicians can gain valuable biological insights and build trust in these novel tools.

REFERENCE:

Zhang X, Singhal A, Hassan, S, et al. Using a machine learning model for prediction of palbociclib response in patients with ER+ HER2- advanced breast cancer a single institution, real-world analysis. Presented at: 2024 San Antonio Breast Cancer Symposium; December 10-14, 2024; San Antonio, TX. P2-06-23: