(T1130-09-51) Accelerating SEDDS Formulation Development with Machine Learning

Purpose: Self-emulsifying drug delivery systems (SEDDS) and self-microemulsifying drug delivery systems (SMEDDS) have emerged as promising approaches in pharmaceutical formulations to enhance the solubility and bioavailability of poorly water-soluble drugs. These systems rely on the spontaneous formation of fine oil-in-water emulsions upon mild agitation, improving drug dissolution and absorption.^1-3 However, optimization of SEDDS and SMEDDS formulations is a complex and time-consuming process, often requiring numerous experiments involving variables such as type of excipients, ratios, and drug characteristics.⁴ Here, we aim to accelerate this process by using machine learning to predict the stability of binary and tertiary blends containing a wide range of excipients.

Methods: We prepared 1949 samples composed of a primary emulsifier (at ratios of 30-80%), a co-emulsifier (0-50%) and an oil (20%) and assessed their stability. For this, we used 14 distinct primary emulsifiers, 12 co-emulsifiers, and 7 oils. Of the 1949 samples, 1044 samples represent unique blend combinations. We gathered the respective HLB (Hydrophilic-Lipophilic Balance) value and acquired a ¹H-NMR spectrum in DMSO-d₆ for each raw material (see Figure 1 for an example). This data and the material ratios serve as input to our machine learning models. To test how our model will perform on unseen new data, we selected two primary emulsifiers and one co-emulsifier and moved all samples containing one of these materials to a holdout test dataset. The training+validation dataset contains 1544 samples (793 unique blends) and the test dataset 405 samples (251 unique blends). 62% of the samples in the test dataset are stable, making the dataset reasonably balanced. Because of its high dimensionality, it is advisable to preprocess NMR data. After removing some non-informative signal regions, we performed spectral binning with bin widths of 0.10, 0.15, and 0.20 ppm respectively. A relevant share of the obtained bins is highly correlated. Therefore, we additionally tested if decorrelation could improve model performance. For this, we tested two approaches: 1) performing a principal component analysis (with 2 to 5 components), and 2) removing bins from the cross-correlation matrix until only bins below a defined correlation threshold remain (with thresholds set to 0.9 and 0.95). To train models on our dataset, LightGBM, a gradient boosting framework that uses tree-based learning algorithms was used.⁵ To find optimal model hyperparameters, we performed Bayesian hyperparameter optimization with the optuna package and 50 trials.⁶ Each trial consisted of randomly splitting the training+validation dataset ten times into a training dataset containing 80% and a validation dataset containing 20% of the 793 unique blend combinations and using the mean of the ten binary validation log-losses as a criterion. Finally, we evaluated the models by scoring accuracy on the hold-out test data. This process was repeated five times for each model and the mean accuracy is reported.

Results: The performance of all models is summarized in Table 1. Overall, even the models with the lowest performance can predict the stability of binary and ternary blends with an accuracy of above 80%. Incorporating NMR data as input features improves accuracy in almost all cases (22 out of 24) – for the best model by 4.4%pt. The best model candidate achieves an accuracy of 86% on our test dataset that contains unique blend combinations absent in the training data. Best performances are observed for models that have both NMR data and HLB values as input features. Interestingly, models based on NMR data alone perform significantly better than models that rely solely on HLB values. For preprocessing NMR data, the best performances were observed with a binning width of 0.15 ppm. The use of principal component analysis to decorrelate the binned NMR data was found to decrease model performance.

Conclusion: The best model in this work can predict the stability of binary and ternary blends with a good accuracy of 86%. This demonstrates that a faster and more cost-effective approach to optimizing SEDDS formulations can be enabled through machine learning. Moreover, our work shows that incorporating NMR data of raw materials can improve models that predict blend stability – in our case by up to 4.4%pt. For decades, formulation chemists have heavily relied on HLB values to guide their experiments. However, the HLB concept falls short for complex mixtures that contain one or multiple unidentified components or that vary in their composition, e.g., due to their natural origin.⁷ Thus, this work presents a fast and promising alternative to traditional HLB values.

References: 1. Badadhe, Dalavi, Review on Self Nano Emulsifying Drug Delivery System. Sys. Rev. Pharm. 2022, 13, 63-68. DOI: 10.31858/0975-8453.13.1.63-68.
2. Anton, Vandamme. Nano-emulsions and micro-emulsions: clarifications of the critical differences. Pharm. Res. 2011, 28, 978-985. DOI:10.1007/s11095-010-0309-1
3. Pehlivanov, Self-Emulsifying Drug Delivery Systems (SEDDS): Excipient selection, formulation and characterization. J of IMAB 2020, 26, 3226-3233. DOI: 10.5272/jimab.2020263.3226
4. Zaslavsky, Allen, A dataset of formulation compositions for self-emulsifying drug delivery systems. Sci Data 2023, 10. DOI: 10.1038/s41597-023-02812-w
5. Ke, Meng, Finely, Wang, Chen, Ma, Ye, Liu, LightGBM: A Highly Efficient Gradient Boosting Decision Tree. NIP 2017 2017.
6. Akiba, Sano, Yanase, Ohta, Koyama, Optuna: A Next-generation Hyperparameter Optimization Framework. KDD '19 2019. DOI: 10.1145/3292500.3330701
7. Smejkal, Gross, Lazarev, Theoretical and Experimental Determinations of the Hydrophilic–Lipophilic Balance (HLB) of Representative Oils and Lecithins. Colloids Interfaces 2024, 8, 21. DOI: 10.3390/colloids8020021

Figure 1: Extract of the 1H-NMR spectra of the primary emulsifiers in DMSO-d6 at 400 MHz.

Figure 2: Principal component analysis (2 components) of the NMR data of the primary emulsifiers preprocessed with a bin width of 0.15 ppm. This highlights that NMR data provides information about the similarity of materials.

Table 1: Performance of all models as a function of input data (HLB values vs. NMR data vs. both) and preprocessing methods used for NMR data. The best 3 models are highlighted in bold green.