(T1130-06-33) Investigating the Modelability of Human Oral Fraction Absorbed using Regulatory Data

Purpose: Unacceptable pharmacokinetics (PK) is a major driver of compound attrition during early drug development. (1) Poor absorption increases the risk of drug failure and may necessitate the use of bio-enabling technologies. Machine learning can mitigate this risk by prioritizing candidates with desirable PK profiles during multi-parameter optimization. Compound prioritization tasks often occur before the clinical dose, solid form, or formulation strategy have been finalized. Therefore, a cheminformatics-driven absorption model which can be incorporated at an early stage is desirable. Human oral fraction absorbed (F_a) is an attractive PK endpoint as it is unaffected by presystemic metabolism, thereby more precisely identifying drugs with potential absorption issues compared to oral bioavailability. Rath et al. recently reported binary classifiers for human oral bioavailability which achieved established metrics of acceptability using only Morgan fingerprints as inputs. (2) The use of regulatory data to support predictive modeling of oral PK parameters, such as food effect, has been described but warrants further exploration. (3) Specifically, programmatic data retrieval from regulatory documents may expedite data curation. This study describes the construction of oral F_a classifiers using data from the European Medicines Agency (EMA). The aims of this research were twofold: to investigate the modelability of oral F_a as a binary classification task, and to assess the practicality of leveraging regulatory data in developing PK prediction models.

Methods: A web scraper was built to download the Product Information File for all orally administered small molecules authorized centrally by the EMA up to 1^st January 2024. Regular expressions were used to isolate the drug name and PK properties of the active ingredient, from which F_a values were manually curated. The chemical structure of all drugs with an F_a value (n=123) was programmatically retrieved from PubChem and standardized using ChEMBL curation protocols. (4) Only the parent drug in a multicomponent active ingredient was considered. Binary classification models of oral F_a, using F_a ≥ 0.8 as the cutoff, were constructed with scikit-learn (v1.4.2). Data was partitioned using a random train test split stratified by F_a class with a test size of 0.2. The feature sets considered are summarized in Table 1, which also includes the modelability index of each set. (5) Energy minimization was performed using the MMFF94 force field before calculation of the three-dimensional descriptors. (6) A grid search of hyperparameters for random forest, adaptive boosting, and gradient boosting classifiers was performed using stratified five-fold cross validation. The best model was selected on the basis of average balanced accuracy across the five folds. All data in the training set was then used to refit the best performing model, and performance was assessed on the test set. The model was deemed acceptable if it had a balanced accuracy of ≥0.6 and all of positive predictive value, negative predictive value, sensitivity, and specificity≥0.5 on the test set.

Results: Of 897 orally administered drug products containing 394 unique parent drugs, only 123 valid F_a values were available in the regulatory data. Figure 1 demonstrates skewness in F_a values resulting in imbalanced F_a classes. 32 drugs fall within 0.1 of the decision boundary of 0.8, with three drugs having an F_a of 0.8 exactly. Further examination of the underlying data reveals errors in the values reported in the regulatory documents.  For example, opicapone is quoted to have an F_a of 0.2, but published clinical pharmacokinetic studies of opicapone suggest near-complete absorption. (7) The classification performance for each feature set is summarized in Table 2. Interestingly, the classifier with the fewest inputs and the second lowest cross validation performance had the highest balanced accuracy on the test set. The gradient boosting classifier which used Lipinski descriptors as input achieved a balanced accuracy of 0.75, a positive predictive value of 0.95, a negative predictive value of 0.6, a sensitivity of 0.95, and a specificity of 0.6, thus satisfying the acceptability criteria. While these results are promising, Table 2 also shows a tendency to overfit to the training data, and instability on cross-validation.

Conclusion: Despite the challenges associated with absorption as a PK endpoint, such as high inter- and intra-individual variability and its dependence on dose, delivery system, and solid state, this work demonstrates the feasibility of a binary F_a classifier to support compound prioritization tasks. Missing values and logical inconsistencies in the F_a values reported in regulatory data, however, impede modeling efforts and underline the need for cross-referencing and human curation. Nonetheless, the increased integration of programmatic data retrieval and machine learning enables efficient PK prediction and merits further investigation.

References: 1. Waring, Michael J., et al. "An analysis of the attrition of drug candidates from four major pharmaceutical companies." Nature reviews Drug discovery 14.7 (2015): 475-486.
2. Rath, Marielle, et al. “Pharmacokinetics Profiler (PhaKinPro): Model Development, Validation, and Implementation as a Web Tool for Triaging Compounds with Undesired Pharmacokinetics Profiles." Journal of Medicinal Chemistry 67.8 (2024): 6508-6518.
3. Bennett-Lenane, Harriet, Brendan T. Griffin, and Joseph P. O'Shea. "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 168 (2022): 106018.
4. Bento, A. Patrícia, et al. "An open source chemical structure curation pipeline using RDKit." Journal of Cheminformatics 12 (2020): 1-16.
5. Golbraikh, Alexander, et al. "Data set modelability by QSAR." Journal of chemical information and modeling 54.1 (2014): 1-4.
6. Tosco, Paolo, Nikolaus Stiefl, and Gregory Landrum. "Bringing the MMFF force field to the RDKit: implementation and validation." Journal of Cheminformatics 6 (2014): 1-4.
7. Loureiro, Ana I., et al. "Absorption, metabolism and excretion of opicapone in human healthy volunteers." British Journal of Clinical Pharmacology 88.10 (2022): 4540-4551.

Acknowledgements: The research in this abstract was funded by the Irish Research Council under grant number GOIPG/2022/1580.

This research was also financially supported by a Fulbright Irish Student Award from the Irish Fulbright Commission.

Figure 1. Histogram of Fa values for 123 orally administered drugs, as sourced from regulatory documents. Only the Fa value associated with the highest strength clinically approved orally administered dosage form is quoted. Each bin represents a 0.05 interval and is left-inclusive.

Table 1. Feature sets considered in this study. n_features: number of features; MODI: Modelability Index. The modelability index was calculated on all data. For RDKit Lipinski, RDKit 2D, and Mordred descriptor sets, the data was first scaled to zero mean and unit variance using values from all 123 drugs.

Table 2. Summary of performance of five descriptor sets in Fa classification. BA: Balanced Accuracy; SD: Standard Deviation. An acceptable model is one which has a balanced accuracy of ≥0.6 and all of positive predictive value, negative predictive value, sensitivity, and specificity ≥0.5 on the test set.