(W1030-04-21) Computational Pharmaceutics Approaches to Guide Lipid Based Formulation Development: How Data Mining Can Improve the Process

Purpose: Computational pharmaceutics is emerging as key area of research that can support structured guidance in formulation strategy to rectify the declining productivity in drug development. The first aim of this study was to develop a computationally informed decision tree (figure 1) to streamline the development of lipid-based formulations (LBF). A number of previously published machine learning techniques were integrated to form the basis of this decision tree [1–4]. Predicted criteria which was deemed important in the decision tree include; dose loading in type-I LBF (miglyol-812), solubility gain upon dispersion and digestion of the LBF in vitro and the presence of a pharmaceutical food effect. The second aim was to perform these predictions on a dataset containing drugs licenced by the FDA from 2010-2023 and apply the results to the LBF developability decision tree, to assess if any currently marketed drug could be considered a ‘missed opportunity’ for LBFs. Computational predictions of solubility gain upon LBF dispersion and digestion also informed a biopharmaceutical dose number in simulated intestinal fluids, which can be incorporated within the developability classification system (DCS) framework to further inform drug developability. Strategies to address poor drug solubility in lipid excipients such as lipophilic salt synthesis was also explored. Four candidate drugs were selected based on these predictions and their suitability for formulation as a LBF was assessed in vitro.

Methods: A database of drugs licenced by the FDA from 2010-2023 was constructed. Physicochemical properties and SMILES strings were mined for all drugs using PubChem and ADMET Predictor version 9.0 (Simulations Plus, USA). Computational predictions on the dataset of drugs was performed using artificial neural networks (SPSS version 28.0, IBM, USA), partial least square (SIMCA version 18.0, Sartorius, Germany) and machine learning descriptor analysis (scikit learn, France). Solubility studies in type-I LBF and both medium and long chain type III-A LBF were conducted for dapagliflozin, mavacamten, ospemifene and flibanserin at 25°C, over 72 hours. In vitro lipolysis was carried out in a method adapted from Williams et al. to assess the impact of dispersion and digestion of LBF in simulated intestinal fluid [5]. Drug content in both solubility and lipolysis studies was quantified via HPLC.. Lipophilic salt synthesis of flibanserin was performed using docusic acid and dodecyl sulfate acid in a biphasic system containing dichloromethane and water. The resulting salt was dried under vacuum and chemically analysed using NMR, HPLC and DSC.

Results: The models were successfully applied to the drug database and a rank order was determined for both the predicted dose number in type-I LBF and the predicted solubility ratio gain following dispersion and digestion of a type-IIIA LBF for each drug. Food effect was also predicted for each drug to act as a pre-cursor for LBF suitability. It was predicted that 45% of drugs displayed a positive food effect, based on their physicochemical properties. Dapagliflozin and mavacamten were selected based on their predicted dose number in type-I LBF’s. Solubility studies in miglyol-812 demonstrated that both drugs could be successfully formulated as a type-I lipid solution at a clinically relevant dose (figure 2). The models employed under predicted the lipid solubility of dapagliflozin, while the lipid solubility of mavacemten was within the range of the root mean square error of the model. Ospemifene was predicted to display high solubility gains after dispersion and digestion of a type-IIIA LBF containing the drug. This was experimentally confirmed, with up to 15-fold increases in biorelevant solubility compared to unformulated ospemifene. The results demonstrated that ospemifene benefited from formulation as an LBF as a reduction in the biopharmaceutical dose number saw ospemifene transition from a DCS class IIB drug to a class I. Finally, flibanserin displayed attributes of a successful LBF based on the results of the computational predictions. Dose loading challenges however resulted in the realisation that an alternative strategy was required to increase flibanserin solubility in lipid excipients. Through synthesis of flibanserin as a lipophilic salt, up to 4-fold increases in solubility in type III-A formulations were observed compared to the free base form of the drug, allowing flibanserin to be formulated as an LBF despite an initial poor dose loading (figure 3).

Conclusion: The ability of computational pharmaceutics to facilitate structured formulation decisions was demonstrated in this work. Four molecules which were not commercially available as LBFs were successfully formulated as LBF’s demonstrating that through integration of computationally informed decisions into a streamlined framework a broader ‘chemical space’ of drugs can be explored for use as LBF’s. The proposed refining of the drug substance to drug product development framework demonstrates the significance of both computationally informed and experimentally confirmed aspects of drug developability decision-making. Finally, this work highlights the importance of a collaborative, accessible and overall user-friendly approach to reinforce the appeal of machine learning informed formulation design.

References: 1. Bennett-Lenane H, Koehl NJ, O’Dwyer PJ, Box KJ, O’Shea JP, Griffin BT. Applying Computational Predictions of Biorelevant Solubility Ratio Upon Self-Emulsifying Lipid-Based Formulations Dispersion to Predict Dose Number. J Pharm Sci. 2021;110:164–75.
2. Lange JJ, Anelli A, Alsenz J, Kuentz M, O’Dwyer PJ, Saal W, et al. Comparative Analysis of Chemical Descriptors by Machine Learning Reveals Atomistic Insights into Solute-Lipid Interactions. Mol Pharm. 2024;
3. Ejskjær L, Holm R, Kuentz M, Box KJ, Griffin BT, O’Dwyer PJ. Predictions of biorelevant solubility change during dispersion and digestion of lipid-based formulations. European Journal of Pharmaceutical Sciences [Internet]. 2024;200:106833. Available from: https://linkinghub.elsevier.com/retrieve/pii/S0928098724001453
4. Bennett-Lenane H, Griffin BT, O’Shea JP. Machine learning methods for prediction of food effects on bioavailability: A comparison of support vector machines and artificial neural networks. European Journal of Pharmaceutical Sciences. 2022;168.
5. Williams HD, Sassene P, Kleberg K, Bakala-N’Goma JC, Calderone M, Jannin V, et al. Toward the establishment of standardized in vitro tests for lipid-based formulations, Part 1: Method parameterization and comparison of in vitro digestion profiles across a range of representative formulations. J Pharm Sci. 2012;101:3360–80.

Acknowledgements: This publication has received funding from the Science Foundation Ireland (SFI) through the SSPC, The SFI Research Centre for Pharmaceuticals under grant number (12/RC/2275_P2).

Figure 1: Decision tree of the lipid-based formulation developability guide which combines computationally informed and experimentally confirmed aspects of drug developability.

Figure 2: Bar chart illustrating the computationally predicted and experimentally confirmed lipid solubility of dapagliflozin and mavacamten in type-1 LBF containing miglyol-812.

Figure 3: Bar chart illustrating the solubility of flibanserin free base and lipophilic salts in type-IIIA LBF composed of long-chain formulations (olive oil 40%, tween 85 40% and RH40 20%) and medium-chain formulations (miglyol-812 40%, tween 85 40% and RH40 20%).