Optimizing the first line therapy of canine B-cell lymphomas via machine learning

Presentation Detail

• Presenter: George Courcoubetis, PhD, Senior Data Scientist of ImpriMed, Inc.

‍

‍

Research Abstract

Authors: Jamin Koo¹, George Courcoubetis², Gyucheol Choi³, Ilona Holcomb⁴, Sungwon Lim⁵

^1#Jamin Koo, Ph.D., Department ofChemical Engineering, Hongik University, Seoul 04066, Republic of Korea; ImpriMed, Inc., CA 94043, U.S.A.

²George Courcoubetis, Ph.D., ImpriMed, Inc., 1130 Independence Ave, Mountain View, CA 94043, U.S.A.

³Gyucheol Choi, ImpriMedKorea, Inc.,Seoul 03920, Republic of Korea

⁴Ilona Holcomb, Ph.D., ImpriMed, Inc., 1130 Independence Ave, Mountain View, CA 94043, U.S.A.

⁵Sungwon Lim, Ph.D., ImpriMed, Inc.,1130 Independence Ave, Mountain View, CA 94043, U.S.A.

#Corresponding Author

Background: Treatment-naive canine B-cell lymphomas are commonly treated with CHOP therapy, yet approximately 20% of patients relapse early or become refractory, highlighting a critical gap in treatment efficacy. Leveraging machine learning to predict individualized outcomes for therapies with distinct mechanisms of action, such as CHOP and Tanovea, offers the potential to optimize treatment strategies and improve survival in this subset of patients.

Objectives: This study aims to develop and validate machine learning (ML) models to predict treatment outcomes in naive canine B-cell lymphomas treated with CHOP or Tanovea. By identifying the therapy likely to yield the best clinical outcome forindividual patients, we seek to provide a tool to guide personalized treatmentstrategies and improve survival rates in this population.

Methods: Our study employed a retrospective design using datasets from canine B-cell lymphoma patients treated with CHOP (n = 473) or Tanovea-based therapy (n = 130). Both datasets included biopsy samples and follow-up results. Predictive ML models for 3-month progression-free survival (PFS) were developed using input features consisting of clinical parameters, clonality, flow cytometry, and/or ex vivo drug sensitivity analysis results.

Results: XGBoost was selected as the algorithm for CHOP model development due to its superior performance in predicting disease progression, achieving an ROC-AUC of 0.81 among tested models. For the Tanovea model, a stacked model ensemble approach was developed, combining neural networks, Random Forest, Extra Trees, CatBoost and XGBoost algorithms with an ROC-AUC of 0.82. The ML models stratified patients based on the likelihood of early disease progression into likely optimal responders (OR) and suboptimal responders (SR) for each therapy. Patients classified as CHOP^SR had a hazard ratio (HR) of 2.6 (95% CI, 1.3–5.0) compared to CHOP^OR, while Tanovea^SR had an HR of 5.9 (95% CI, 2.0–17.1) relative to Tanovea^OR for PFS. Notably, 41% of patients predicted as CHOP^SR were classified as Tanovea^OR (HR of 2.5) demonstrating the potential to improve clinical outcomes through this personalized approach.

Conclusions: The ML models developed for predicting the PFS of naive canine B-cell lymphomas treated with CHOP or Tanovea demonstrated their utility in identifying patients likely to benefit most from each therapy. By stratifying patients as optimal or suboptimal responders, the models revealed significant differences in progression risks, highlighting their potential to improve outcomes by guiding therapy selection. These findings underscore the value of leveraging ML to personalize treatment strategies, aligning with the study's objective to enhance clinical decision-making and optimize patient outcomes.

Funding: This study was funded by ImpriMed, Inc.

Keywords: canine lymphoma; chemotherapy; precision medicine; machine learning; treatment selection