Multimodal Machine Learning–Based Clinical and Genomic Prognostic Factors for Progression-Free Survival in Metastatic Breast Cancer Using Phase III MONALEESA Trials
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Purpose
Cyclin-dependent kinase 4/6 inhibitors in combination with endocrine therapy have become widely adopted as a first-line treatment approach for hormone receptor-positive (HR+)/human epidermal growth factor receptor 2-negative (HER2−) advanced breast cancer (ABC). Through a collaboration between Novartis and the US Food and Drug Administration, we integrated multimodal data from phase III MONALEESA trials to identify novel prognostic factors for HR+/HER2−ABC.
Experimental Design
Baseline clinical and genomics data were collected from patients in the MONALEESA-2, −3, and −7 trials. Circulating tumor DNA (ctDNA) from plasma samples and gene expression in archival tumor biopsies were analyzed. Machine learning models were trained ( n = 1,355) on progression-free survival (PFS) using multimodal features via fusion models. In a test set ( n = 382), prognostic performances were evaluated using the Kaplan-Meier, Cox, and Weibull models.
Results
The late fusion model achieved a prognostic performance (C-index: 0.727, 95% CI: 0.641–0.799) slightly outperforming early fusion (C-index: 0.718, 95% CI: 0.631–0.791) and single-modality models. Model-based risk stratification identified distinct low- and high-risk groups (PFS HR, 0.28; 95% CI: 0.20–0.38). Among the top 40 features for progression risk (18 clinical, 20 gene expression, and 2 ctDNA), these were associated with higher risk: liver involvement, higher histological grades, early relapse; lower STC2 and RBBP8 expressions and higher FBX05 expression; and higher tumor mutational burden and ctDNA purity.
Conclusions
Integrating clinical and genomic modalities improved progression risk prediction compared with single-modality models, while highlighting salient features that inform the underlying biology and support hypothesis generation.
