(W1030-12-77) Application of Machine Learning to the Development of Lipid Nanoparticle

Purpose: Lipid nanoparticle, LNP is a delivery carriers encapsulating nucleic acid. Ionizable lipid in LNP is a key component for nucleic acid delivery. In general, the development of LNP has following five steps. 1) To design ionizable lipid structure, 2) to synthesis the lipid, 3) to formulate LNPs, 4) to conduct experiments to evaluate various parameters of LNP such as potency, 5) to design ionizable structure and evaluate again. This one cycle takes a few months and requires a lot of costs and FTE. In this study, we built a machine learning model based on ionizable lipid structure and experimental results. We verified whether pretrained model could predict experimental results from new ionizable lipid structure. This approach has a possibility to reduce the time for each cycle of LNP development, costs and FTE as well as to find the optimum structure that is difficult for humans to design or identify. (Figure.1)

Methods: In proof of concept (POC) study, a target variable was predicted from ionizable lipid structure using machine learning model. Single and ensemble models were built with the target variable data (N=81), The chemical structures of ionizable lipid (N=81) were described using chemical descriptors such as RDKit, Mordred, ECFP4 and Graph convolution and algorithms such as Ridge, Lasso, k-NN, AdaBoost were used. Each model (total of 625 models) was evaluated, and the screened best performing model was selected based on the nested cross validation results. The performance of the selected model was further evaluated by testing new ionizable lipid (N=41). Similarly, in vitro and in vivo luciferase expressions were predicted from ionizable lipid structures. As for in vivo model, measured luminescence intensity in each whole image of In Vivo Imaging System (IVIS) were converted to number for modeling.

Results: For the POC study, the screening of models and descriptors identified ensemble model of LightGBM-RDKit and Ridge-ECFP4 as a model/descriptor pair with a reasonable predictive power (Figure 2). Using this ensemble model, the target variable of new ionizable lipid was predicted with 4% mean percent error (if the similarity of the structure with training data described using Tanimoto similarity was higher than 0.95). On the other hand, low performance was observed with dissimilar structures relative to the training data. With the same approach as above, luciferase expression was predicted from ionizable lipid. However, mean absolute error (MAE) was more than 1.2 x 10⁷.

Conclusion: POC was confirmed using structure of ionizable lipid and the target variable data. 4% mean percent error was observed if similarity of the structure with training data described using Tanimoto similarity was more than 0.95. This validated approach was applied to prediction of in vitro and in vivo luciferase expression. To improve the performances, countermeasures can be to increase data and to conduct different experiment.

Acknowledgements:COI: Ryoko Saeki, Tomohiro Maeda, Satoru Matsumoto, Yasutaka Hoashi, Hideto Hara, Tomomi Sato, Wataru Hoshina, Yumiko Mizuno and Nozomi Ogawa are employees of Takeda and hold stock in Takeda.

Figure.1 Purpose of this study.

Figure.2 Normalized predicted value and normalized observed value of a target variable data from nested cross validation. (POC study)