A Data-Driven Approach to Predicting Tablet Properties after Accelerated Test Using Raw Material Property Database and Machine Learning.

The purpose of this study was to develop a model for predicting tablet properties after an accelerated test and to determine whether molecular descriptors affect tablet properties. Tablets were prepared using 81 types of active pharmaceutical ingredients, with the same formulation and three different levels of compression pressure. The tablet properties measured were the tensile strength and disintegration time of tablets after two weeks of accelerated test. The material properties measured were the change in tablet thickness before and after the accelerated test, maximum swelling force, swelling time, and swelling rate. The acquired data were added to our previously constructed database containing a total of 20 material properties and 3381 molecular descriptors. The feature importance values of molecular descriptors, material properties and the compression pressure for each tablet property were calculated by random forest, which is one type of machine learning (ML) that uses ensemble learning and decision trees. The results showed that more than half of the top 25 most important features were molecular descriptors for both tablet properties, indicating that molecular descriptors are strongly related to tablet properties. A prediction model of tablet properties was constructed by eight ML types using 25 of the most important features. The results showed that the boosted neural network exhibited the best prediction accuracy and was able to predict tablet properties with high accuracy. A data-driven approach is useful for discovering intricate relationships hidden within complex and large data sets and predicting tablet properties after an accelerated test.

Modeling of quantitative relationships between physicochemical properties of active pharmaceutical ingredients and tensile strength of tablets using a boosted tree.

OBJECTIVES: The aim of this study was to explore the potential of boosted tree (BT) to develop a correlation model between active pharmaceutical ingredient (API) characteristics and a tensile strength (TS) of tablets as critical quality attributes. METHODS: First, we evaluated 81 kinds of API characteristics, such as particle size distribution, bulk density, tapped density, Hausner ratio, moisture content, elastic recovery, molecular weight, and partition coefficient. Next, we prepared tablets containing 50% API, 49% microcrystalline cellulose, and 1% magnesium stearate using direct compression at 6, 8, and 10 kN, and measured TS. Then, we applied BT to our dataset to develop a correlation model. Finally, the constructed BT model was validated using k-fold cross-validation. RESULTS: Results showed that the BT model achieved high-performance statistics, whereas multiple regression analysis resulted in poor estimations. Sensitivity analysis of the BT model revealed that diameter of powder particles at the 10th percentile of the cumulative percentage size distribution was the most crucial factor for TS. In addition, the influences of moisture content, partition coefficients, and modal diameter were appreciably meaningful factors. CONCLUSIONS: This study demonstrates that BT model could provide comprehensive understanding of the latent structure underlying APIs and TS of tablets.

Application of machine learning to a material library for modeling of relationships between material properties and tablet properties.

This study investigates the usefulness of machine learning for modeling complex relationships in a material library. We tested 81 types of active pharmaceutical ingredients (APIs) and their tablets to construct the library, which included the following variables: 20 types of API material properties, one type of process parameter (three levels of compression pressure), and two types of tablet properties (tensile strength (TS) and disintegration time (DT)). The machine learning algorithms boosted tree (BT) and random forest (RF) were applied to analysis of our material library to model the relationships between input variables (material properties and compression pressure) and output variables (TS and DT). The calculated BT and RF models achieved higher performance statistics compared with a conventional modeling method (i.e., partial least squares regression), and revealed the material properties that strongly influence TS and DT. For TS, true density, the tenth percentile of the cumulative percentage size distribution, loss on drying, and compression pressure were of high relative importance. For DT, total surface energy, water absorption rate, polar surface energy, and hygroscopicity had significant effects. Thus, we demonstrate that BT and RF can be used to model complex relationships and clarify important material properties in a material library.

In silico predictions of tablet density using a quantitative structure-property relationship model.

The purpose of this study was to explore the potential of a quantitative structure-property relationship (QSPR) model to predict tablet density. First, we calculated 3381 molecular descriptors for 81 active pharmaceutical ingredients (API). Second, we obtained data that were merged with a dataset including powder properties that we had constructed previously. Next, we prepared 81 types of tablet, each containing API, microcrystalline cellulose, and magnesium stearate using direct compression at 120, 160, and 200 MPa, and measured the tablet density. Finally, we applied the boosted-tree machine learning approach to construct a QSPR model and to identify crucial factors from the large complex dataset. The QSPR model achieved statistically good performance. A sensitivity analysis of the QSPR model revealed that molecular descriptors related to the average molecular weight and electronegativity of the API were crucial factors in tablet density, whereas the effects of powder properties were relatively insignificant. Moreover, we found that these descriptors had a positive linear relationship with tablet density. This study indicates that a QSPR approach is possibly useful for in silico prediction of tablet density for tablets prepared using more than a threshold compression pressure, and to allow a deeper understanding of tablet density.