Molecular Representations in Machine-Learning-Based Prediction of PK Parameters for Insulin Analogs

Therapeutic peptidesand proteins derived from either endogenoushormones, such as insulin, or de novo design via display technologiesoccupy a distinct pharmaceutical space in between small moleculesand large proteins such as antibodies. Optimizing the pharmacokinetic(PK) profile of drug candidates is of high importance when it comesto prioritizing lead candidates, and machine-learning models can providea relevant tool to accelerate the drug design process. PredictingPK parameters of proteins remains difficult due to the complex factorsthat influence PK properties; furthermore, the data sets are smallcompared to the variety of compounds in the protein space. This studydescribes a novel combination of molecular descriptors for proteinssuch as insulin analogs, where many contained chemical modifications,e.g., attached small molecules for protraction of the half-life. Theunderlying data set consisted of 640 structural diverse insulin analogs,of which around half had attached small molecules. Other analogs wereconjugated to peptides, amino acid extensions, or fragment crystallizableregions. The PK parameters clearance (CL), half-life (T1/2), and meanresidence time (MRT) could be predicted by using classical machine-learningmodels such as Random Forest (RF) and Artificial Neural Networks (ANN)with root-mean-square errors of CL of 0.60 and 0.68 (log units) andaverage fold errors of 2.5 and 2.9 for RF and ANN, respectively. Bothrandom and temporal data splittings were employed to evaluate idealand prospective model performance with the best models, regardlessof data splitting, achieving a minimum of 70% of predictions withina twofold error. The tested molecular representations include (1)global physiochemical descriptors combined with descriptors encodingthe amino acid composition of the insulin analogs, (2) physiochemicaldescriptors of the attached small molecule, (3) protein language model(evolutionary scale modeling) embedding of the amino acid sequenceof the molecules, and (4) a natural language processing inspired embedding(mol2vec) of the attached small molecule. Encoding the attached smallmolecule via (2) or (4) significantly improved the predictions, whilethe benefit of using the protein language model-based encoding (3)depended on the used machine-learning model. The most important moleculardescriptors were identified as descriptors related to the molecularsize of both the protein and protraction part using Shapley additiveexplanations values. Overall, the results show that combining representationsof proteins and small molecules was key for PK predictions of insulinanalogs.