Machine Learning in Predicting Longitudinal Platelet Counts: Applications in Dose Optimization

Introduction: INCB057643 is a small molecule bromodomain and extra-terminal (BET) inhibitor under clinical investigation for treatment of solid tumors and hematological malignancies. Accurate prediction of platelet counts during BET treatment is important to determine patient dosages. While conventional exposure-response models are used to predict platelet counts during patient treatment and inform dosing regimens (Blanchet B. Br J Cancer. 2024; Liu C. Clin Pharmacol Ther. 2017; van der Bij S. Cancer Causes Control. 2013) inherent modeling assumptions may exist. The advent of machine learning (ML) provides a novel opportunity to inform the potential unknown covariate search and selections and fill in some current gaps in platelet count modeling (Dou Y. Comput Biol Med. 2024; Feng D. J Biomol Struct Dyn. 2024; Pozdnyakova O. J Clin Pathol. 2023). ML is solely driven by data and offers the benefit of analyzing nonlinear exposure-response relationships, while also considering unknown factors that could impact physiological responses to drugs. Previously, XGBoost was selected as one of the final ML models used to describe grade ≥2 thrombocytopenia in patients treated with INCB057643 (Raheem E. PAGE 2024). The objectives of this work are to use XGBoost to 1) predict grade ≥2 thrombocytopenia following administration of INCB057643 at 4-12 mg once daily (qd) in patients with advanced solid tumors or hematological malignancies, and 2) predict the probability of grade ≥2 thrombocytopenia associated with various INCB057643 dosing regimens and treatment frequencies, including different intermittent dosing intervals.

Methods: Maximum exposures at steady state (C_max,_ss) of INCB057643 and its metabolite were simulated using a previously developed population pharmacokinetic (PPK) model for 3 dosing regimens; 7-day continuous qd (7/0) dosing, or intermittent qd dosing with either 5-days-on-2-days-off (5/2) or 4-days-on-3-days-off (4/3). Grade ≥2 thrombocytopenia was defined as platelet counts <75×10⁹/L. Clinical data were collated for 118 patients with solid tumors or hematological malignancies receiving INCB057643 monotherapy at 4, 8, 10, and 12 mg qd during the phase 1 INCB057643-101 study (NCT02711137). Firstly, the XGBoost model was applied to data from the 7/0 dosing regimen from INCB057643-101 to predict occurrences of time-varying grade ≥2 thrombocytopenia for each patient. Time points were rounded to the nearest hour. The last observation carried forward method was employed to address missing platelet values. Secondly, occurrences of grade ≥2 thrombocytopenia predicted by the XGBoost model were compared with, and validated against, clinically observed outcome data (NCT04279847; Watts J. ASH 2023). Thirdly, the XGBoost model was then used to predict grade ≥2 thrombocytopenia events in patients administered INCB057643 with 5/2 and 4/3 intermittent dosages using landmark analysis.

Results: The XGBoost model adequately predicted time-varying grade ≥2 thrombocytopenia occurrences in the 7/0 dosing regimen dataset, as indicated by the area under the curve values of the receiver operating characteristic curve, sensitivity, specificity, and accuracy scores. These scores were 0.897, 0.809, 0.984, and 0.960, respectively, when compared with clinically observed data. Consistently, classification error rate was 0.042 when predicted values were compared with clinically observed data (Watts J. ASH 2023). Landmark analyses indicate the predicted grade ≥2 thrombocytopenia occurrence following initiation of INCB057643 treatment is similar across the different dosing regimens at 1 week (~23%). However, predicted prevalence at 2 weeks is 37%, 33%, and 30% for the 7/0, 5/2, and 4/3 dosage regimens, respectively. Further landmark exposure-safety analyses suggest a similar trend of higher grade ≥2 thrombocytopenia events with increased dosing and duration of treatment, as seen at 1-month (47% vs 39% vs 34%), 3-month (49% vs 40% vs 35%), and 6-month (49% vs 40% vs 34%) time points.

Conclusions: An ML framework was established to predict the exposure-safety (grade ≥2 thrombocytopenia) relationship for INCB057643 and was validated with clinically observed data. Our ML framework predicts that safety events vary with INCB057643 dosage and dosing regimens. This information can be used to inform dose optimization for patients.