(M1230-07-46) Machine Learning Predicts Drug Release Profiles Based on Tablets’ Formulations

Purpose: Direct compression remains a widely used manufacturing technology for producing solid dosage forms. However, formulation optimisation is a laborious, costly, and time-consuming process. The purpose of this study was to investigate whether machine learning (ML) can accelerate formulation development by predicting drug release profiles under dynamic conditions based on formulation composition. A total of 377 formulations of APIs with diverse chemical structures were produced in-house, and their release profiles were measured under dynamic dissolution conditions from 0 to 480 minutes across 11 time points. Subsequent ML analysis was performed to predict the APIs entire release profile.

Methods: Principal Component Analysis (PCA) was applied to decompose high-dimensional space data into two-dimensions. The analysis was performed also using Python, utilizing the scikit-learn package (v.1.5.2). All feature inputs were scaled prior to PCA fitting. For generating the physicochemical space, the RDKIT (v.2023.3.2) package was used to generate 200 physico-chemical features. Six MLTs were investigated, which were adaptive boosting (ADA), k-nearest neighbour (kNN), light gradient-boosted machines (LGBM), random forest (RF), support vector regressor (SVR) and extreme gradient boosted (XGB). The analysis was performed also using Python, where the scikit-learn package (v.1.5.2) was used to develop ADA, kNN, RF and SVR. The lightgbm package (v.4.5.0) and xgboost (v.2.1.2) were used for LGBM and XGB, respectively. The input data was left untransformed. A five-fold cross-validation (CV) was applied to ensure model reproducibility. In 5-fold CV, the dataset is divided into five equal-sized subsets, ensuring they are approximately of the same size. Each fold is sequentially used as the test set (20% of data), while the remaining serve as the training set (80% of data). The model is trained on the training data and then evaluated on the test fold, with the resulting performance score recorded [1]. The model’s overall performance is calculated as the average score across all repetitions. Specifically, in this research, the 377 formulations produced in-house correspond to 75 formulations for each fold, meaning that 302 formulations were used for training and 75 for testing each time. For training, the (%) composition of each excipient/API, the release profile at 11 time points up to 480 minutes, and the pH value at each time point were used. The goal was to use the (%) composition of each excipient/API in the formulations and the pH value at each time point as inputs, and, as output, have the cumulative release (%) at all 11 time points. The performance of the ML results was evaluated using the coefficient of determination (R²) and the root mean squared error (RMSE) of predictions.

Results: The 20 APIs physicochemical features were compared to over 10,000 drugs extracted from DrugBank, which includes both commercial drugs and drugs’ candidates under investigation, and found to occupy a reasonably wide drug space, revealing that the formulations were composed of diverse ingredients. Exploratory Data Analysis (EDA) was performed on all 377 drug release profiles, which included a mixture of immediate, sustained, and delayed release forms. At 30 minutes, cumulative release ranged from 0% to 100%, reflecting the diversity and complexity of the release profiles (Figure 1). Using the formulation and pH data as input, six different MLTs were investigated to predict the release profiles using five-fold cross-validation analysis. Each MLT possessed a different learning mechanism. The five-fold analysis allows each MLT to sample all 377 release profiles and for each profile to be part of the test data, as well as providing insight into the technique’s reproducibility. The ML analysis revealed that RF and XGB were the better performers, with an R² of 0.635 ± 0.047 and 0.601 ± 0.091, respectively. Their respective RMSEs were 13.67 ± 0.99% and 14.16 ± 0.66%. This means that on average the prediction is off by ±13.67% and ±14.16%, respectively, across the entire release profile, which are relatively small errors given the datasets distribution. No significant increase in performance was observed when optimising the MLTs parameters (Figure 2). These findings suggest that ML has the potential to prospectively predict dynamic release profiles, and the potential to minimise the need for performing time-consuming, laborious and costly dissolution studies. In addition, MLT performance was evaluated by providing initial cumulative release values and predicting the remaining profile. By training with values between 60–240 minutes it was shown that including initial release data improved predictive performance across all MLTs. The highest R² values (above 0.8) and lowest RMSEs (below 10%) were achieved when including all release points up to 240 minutes (Figure 3). Although this approach requires dissolution testing up to 240 minutes, it still reduces the study duration compared to a full profile, offering significant time savings.

Conclusion: These findings demonstrate that ML can be used to predict entire drug release profiles during dynamic dissolution studies, thereby making the modeling process more informative for pharmaceutical researchers, while saving both time and money in a highly competitive industry.

References: 1. Abdalla Y, Elbadawi M, Ji M, Alkahtani M, Awad A, Orlu M, et al. Machine learning using multi-modal data predicts the production of selective laser sintered 3D printed drug products. International Journal of Pharmaceutics. 2023;633:122628.

Acknowledgements: The research work was supported by the Hellenic Foundation for Research and Innovation (HFRI) under the 5th Call for HFRI PhD Fellowships grant to C.P. (Fellowship Number: 20610)


The complexity in the release profile data. (A) all 377 release profiles superimposed. (B) Boxplot of the distribution at each time point. (C) PCA biplot of each release profile, colour-coded by the type of release and (D) the corresponding loading score.


A) & (B) the metrics for the untuned models and (C) & (D) the metrics for the tuned model (n = 5).


(A) The R2 and (B) RMSE metrics as a function of adding 0-240 min release profile data points as inputs. The data reveals a general improvement in predictive performance as more time points were used in the input (n = 5).