Protopapa Chrystalla, Siamidi Angeliki, Eneli Amelia Adibe, Elbadawi Moe, Vlachou Marilena
Section of Pharmaceutical Technology, Department of Pharmacy, National and Kapodistrian University of Athens, 157 84, Athens, Greece.
School of Biological and Behavioural Sciences, Faculty of Science and Engineering Queen, Dept W, Mary University of London, 81 Mile End Rd, London, E1 4UJ, UK.
AAPS J. 2025 Jul 28;27(5):124. doi: 10.1208/s12248-025-01101-1.
Direct compression (DC) remains a popular manufacturing technology for producing solid dosage forms. However, the formulation optimisation is a laborious process, costly and time-consuming. The aim of this study was to determine whether machine learning (ML) can be used to accelerate developments by predicting the drug release profiles under dynamic conditions given the composition of formulations. A total of 377 formulations were produced in-house and their release profile under dynamic dissolution conditions was measured from 0 to 480 min across 11 time points. A subsequent ML analysis involved predicting the entire release profile. Six different ML techniques were explored, where random forest (RF) and extreme gradient boosting (XGB) were found to achieve a fivefold cross-validation R of 0.635 ± 0.047 and 0.601 ± 0.091, respectively. A second ML strategy was developed, where the ML techniques predict the kinetic parameters of the Weibull and a modified first-order kinetic release model and subsequently use the predicted parameters to fit the release profiles. The R results using RF were comparable to the first strategy. These findings demonstrate that ML can be used to predict entire drug release profiles during dynamic dissolution studies, whilst simultaneously providing insight into kinetic parameters, thus making the modelling process more informative for pharmaceutical researchers. Future work will seek to investigate more 'kinetic-informed' ML models.
直接压片（DC）仍然是生产固体剂型的一种常用制造技术。然而，制剂优化是一个费力的过程，成本高且耗时。本研究的目的是确定机器学习（ML）是否可用于通过在给定制剂组成的情况下预测动态条件下的药物释放曲线来加速研发。在内部共制备了377种制剂，并在动态溶出条件下从0至480分钟的11个时间点测量了它们的释放曲线。随后的ML分析涉及预测整个释放曲线。探索了六种不同的ML技术，其中随机森林（RF）和极端梯度提升（XGB）分别实现了五重交叉验证R值为0.635±0.047和0.601±0.091。开发了第二种ML策略，其中ML技术预测威布尔和修正的一级动力学释放模型的动力学参数，随后使用预测参数拟合释放曲线。使用RF的R结果与第一种策略相当。这些发现表明，ML可用于预测动态溶出研究期间的整个药物释放曲线，同时提供对动力学参数的洞察，从而使建模过程对药物研究人员更具信息性。未来的工作将寻求研究更多“动力学知情”的ML模型。
