Hyunjung Lee† (1), Hyeonseok Kang† (2), Sungwoo Goo(1), Hyuk Namgoong(2), Jung-woo Chae*(1,3), Hwi-yeol Yun*(1,3), Sangkeun Jung*(1,2)
(1) Department of Bio-AI convergence, Chungnam National University, Daejeon, Republic of Korea, (2) Department of Computer Science and Engineering, Chungnam National University, Daejeon, Republic of Korea, (3) College of Pharmacy, Chungnam National University, Daejeon, Republic of Korea, †These authors contributed equally to this work.,*Those authors contributed equally as correspondence.
Introduction: In the clinical field or during the development of small molecule drugs, it is important to predict clearance to optimize dosage regimens and drug efficacy, and to determine the rate of elimination from the body depending on the degree of organ tissue binding. By predicting clearance, it is possible to forecast the change in time versus blood drug concentration profile, that is, the pharmacokinetic profile, and through this, the first-in-human (FIH) dose can be estimated. The in vitro-in vivo extrapolation (IVIVE) method has been used to estimate clearance. Though recent advances in computational modeling have developed promising in silico tools, current models only focus partially on predicting ADMET and pharmacokinetic results by considering only the cytochrome P450 (CYP450) family, a representative protein in metabolic processes or compound properties.
Objectives: The main aim of this study was to develop a new clearance prediction methodology that utilizes deep learning-based predictions on an existing drug-target interaction (DTI) model and 23 metabolic proteins. This methodology aimed to address the limitations of existing clearance prediction methods, particularly the limited predictive power of intrinsic clearance using the widely used IVIVE method.
Methods: The study incorporated data imported from eight distinct databases, which were subsequently organized and standardized. The experimental process entailed constructing a comprehensive model encompassing drug-target binding interactions, physicochemical properties, and clearance prediction. Integration of Simplified Molecular Input Line Entry System (SMILES) structural data with metabolic and physicochemical information was utilized to forecast drug clearance. Additionally, due to the small size of the training dataset, an approach was employed to augment the training data and mitigate data bias. To evaluate the accuracy and precision of hepatic intrinsic clearance prediction, the IVIVE deep learning-based prediction methods were compared by calculating average fold error (AFE), absolute average fold error (AAFE), and root mean square error (RMSE).
Results: In the study, employing the Multiple-Layer Perceptron (MLP) architecture within the clearance prediction model alongside a feature set consisting of SMILES, binding interaction data, LogP, and Fup led to the highest performance metrics. Notably, when the MLP structure was used, the feature data outperformed the default data by a margin of 0.023 in terms of r^2 m. Additionally, compared to the RDkit data and all data, enhancements of 0.033 and 0.004, respectively, were observed. Similarly, using feature data with an encoder structure also showed superior performance, maintaining a consistent trend across different training dataset compositions. A comparative analysis of the most effective feature data configurations across both MLP and Transformer encoder clearance prediction models indicated that the MLP model outperformed the encoder model by a margin of 0.02 in performance efficacy. The Concordance Index (CI) of the model predicted using all features of the MLP method was 0.6451 before augmentation. After augmentation, the CI increased to 0.6834, indicating a 5.6% improvement. Notably, external validation results in predictions without in vitro or in vivo experiments revealed that more than 50% of the results fell within a 2-fold range. The MLP and Transform encoder precision results were 2.413 and 2.419, the AFE values were 1.116 and 1.148, and the AAFE values were 1.507 and 1.488, respectively.
Conclusions: The study successfully developed a novel clearance prediction methodology utilizing deep learning technology. The methodology showed promising results in improving predictive accuracy, particularly after data augmentation. These predictions can aid in forecasting pharmacokinetic parameters, facilitating the generation of FIH dosage predictions. Overall, the study suggests that integrating deep learning-based approaches with existing drug-target interaction models can enhance clearance prediction in drug development, potentially improving the efficiency and accuracy of the drug development process.
References:
[1] Smith, D.A., Beaumont, K., Maurer, T.S., Di, L.: Clearance in drug design: miniperspective. Journal of Medicinal Chemistry 62(5), 2245–2255 (2018)
[2] Hingorani, A.D., Kuan, V., Finan, C., Kruger, F.A., Gaulton, A., Chopade, S., Sofat, R., MacAllister, R.J., Overington, J.P., Hemingway, H., et al.: Improving the odds of drug development success through human genomics: modelling study. Scientific reports 9(1), 18911 (2019)
[3] Pantale˜ao, S.Q., Fernandes, P.O., Gon¸calves, J.E., Maltarollo, V.G., Honorio, K.M.: Recent advances in the prediction of pharmacokinetics properties in drug design studies: a review. ChemMedChem 17(1), 202100542 (2022)
[4] Yang, H., Lou, C., Sun, L., Li, J., Cai, Y., Wang, Z., Li, W., Liu, G., Tang, Y.: admetsar 2.0: web-service for prediction and optimization of chemical admet properties. Bioinformatics 35(6), 1067–1069 (2019)
[5] Dong, J., Wang, N.-N., Yao, Z.-J., Zhang, L., Cheng, Y., Ouyang, D., Lu, A.-P., Cao, D.-S.: Admetlab: a platform for systematic admet evaluation based on a comprehensively collected admet database. Journal of cheminformatics 10, 1–11 (2018)
[6] Shao, C.-Y., Su, B.-H., Tu, Y.-S., Lin, C., Lin, O.A., Tseng, Y.J.: Cyprules: a rule-based p450 inhibition prediction server. Bioinformatics 31(11), 1869–1871 (2015)
[7] Dhanda, S.K., Singla, D., Mondal, A.K., Raghava, G.P.: Drugmint: a webserver for predicting and designing of drug-like molecules. Biology Direct 8(1), 1–12 (2013)
[8] Lagorce, D., Sperandio, O., Galons, H., Miteva, M.A., Villoutreix, B.O.: Faf- drugs2: free adme/tox filtering tool to assist drug discovery and chemical biology projects. BMC bioinformatics 9(1), 1–9 (2008)
[9] Stork, C., Embruch, G., Sˇ´ıcho, M., Bruyn Kops, C., Chen, Y., Svozil, D., Kirch- mair, J.: Nerdd: a web portal providing access to in silico tools for drug discovery. Bioinformatics 36(4), 1291–1292 (2020)
[10] Pires, D.E., Blundell, T.L., Ascher, D.B.: pkcsm: predicting small-molecule phar- macokinetic and toxicity properties using graph-based signatures. Journal of medicinal chemistry 58(9), 4066–4072 (2015)
[11] Zaretzki, J., Bergeron, C., Huang, T.-w., Rydberg, P., Swamidass, S.J., Bren- eman, C.M.: Rs-webpredictor: a server for predicting cyp-mediated sites of metabolism on drug-like molecules. Bioinformatics 29(4), 497–498 (2013)
[12] Kang, H., Goo, S., Lee, H., Chae, J.-w., Yun, H.-y., Jung, S.: Fine-tuning of bert model to accurately predict drug–target interactions. Pharmaceutics 14(8), 1710 (2022)
Reference: PAGE 32 (2024) Abstr 10925 [www.page-meeting.org/?abstract=10925]
Poster: Methodology – AI/Machine Learning