About the Article
Article Files
[This article belongs to Volume - 70, Issue - 12]
Published on : 2025-12-16 21:51:29
Article Code: AMJ-16-12-2025-12358
Title : Machine Learning Approaches to Predict Gene Therapy Response in Rare Sarcomas. Real-World Evidence from Basket Trial Data
Author(s) : Kenji Tanaka, MD, PhD, Elena Volkov, MD, PhD, Sarah Chen, MSc, PhD, Ahmed Al-Jabri, MBBS, FRCPath, M MD
Abstract :
Background: Rare sarcomas represent a heterogeneous group of malignancies with limited therapeutic options.
Gene therapies, including oncolytic viruses and chimeric antigen receptor (CAR) T-cell approaches, show promise but
exhibit variable response rates. Basket trials offer an opportunity to evaluate these agents across multiple histologies,
yet predictive biomarkers remain elusive.
Objectives: To develop and validate machine learning models for predicting objective response to gene therapy in
rare sarcomas using basket trial real-world data.
Methods: We analyzed 892 patients with rare sarcomas (synovial, myxoid liposarcoma, alveolar soft part) enrolled
in seven basket trials (2019-2023). Genomic, transcriptomic, and clinical data were integrated. Five algorithms were
compared: random forest, gradient boosting (XGBoost), support vector machine, neural network, and logistic
regression. Model performance was assessed using area under the receiver operating characteristic curve (AUC),
calibration plots, and decision curve analysis.
Results: The XGBoost model achieved the best performance (AUC 0.87, 95% CI: 0.84-0.90). Key predictors included
baseline platelet count, tumor mutational burden (TMB >10 mut/Mb), PD-L1 expression, and circulating interleukin
6 levels. The model stratified patients into high-response (ORR 68.2%) and low-response (ORR 12.4%) groups
(p<0.001). External validation on 156 patients showed consistent performance (AUC 0.83).
Conclusions: Machine learning can reliably predict gene therapy response in rare sarcomas, potentially guiding
patient selection and trial design. Integration of circulating biomarkers with genomic features offers superior
predictive accuracy over traditional classifiers.