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

Area of Science:

• Pharmacology
• Computational Chemistry
• Biotechnology

Background:

• Optimizing pharmacokinetic (PK) profiles is critical for developing therapeutic peptides and proteins.
• Predicting protein PK properties is challenging due to complex influencing factors and limited data.
• Insulin analogs, modified with small molecules or peptides, represent a key area for PK optimization.

Purpose of the Study:

• To develop and evaluate machine-learning models for predicting PK parameters of insulin analogs.
• To investigate the impact of novel molecular descriptors, including attached small molecules, on PK prediction accuracy.
• To identify key molecular features driving PK properties in modified insulin analogs.

Main Methods:

• Utilized a dataset of 640 structurally diverse insulin analogs with various modifications.
• Employed classical machine learning models like Random Forest and Artificial Neural Networks.
• Tested multiple molecular representations: global physicochemical descriptors, amino acid composition, small molecule descriptors, protein language model embeddings, and NLP-inspired embeddings (mol2vec).

Main Results:

• Achieved accurate predictions for clearance (CL), half-life (T1/2), and mean residence time (MRT) with Random Forest and Artificial Neural Networks.
• Root-mean-square errors for CL were 0.60 and 0.68 log units for RF and ANN, respectively.
• Encoding attached small molecules significantly improved prediction accuracy; combining protein and small molecule features was key for PK prediction.

Conclusions:

• Machine learning models can effectively predict PK parameters for modified therapeutic proteins like insulin analogs.
• Integrating descriptors for both the protein and its attached modifications enhances predictive performance.
• Molecular size of the protein and protraction moiety are critical determinants of PK properties.