Application of machine learning in combination with mechanistic modeling to predict plasma exposure of small molecules

Abstract

Prediction of a new molecule's exposure in plasma is a critical first step toward understanding its efficacy/toxicity profile and concluding whether it is a possible first-in-class, best-in-class candidate. For this prediction, traditional pharmacometrics use a variety of scaling methods that are heavily based on pre-clinical pharmacokinetic (PK) data. We here propose a novel framework based on which preclinical exposure prediction is performed by applying machine learning (ML) in tandem with mechanism-based modeling. In our proposed method, a relationship is initially established between molecular structure and physicochemical (PC)/PK properties using ML, and then the ML-driven PC/PK parameters are used as input to mechanistic models that ultimately predict the plasma exposure of new candidates. To understand the feasibility of our proposed framework, we evaluated a number of mechanistic models (1-compartment, physiologically based pharmacokinetic (PBPK)), PBPK distribution models (Berezhkovskiy, PK-Sim standard, Poulin and Theil, Rodgers and Rowland, and Schmidt), and PBPK parameterizations (using in vivo, or in vitro clearance). For most of the scenarios tested, our results demonstrate that PK profiles can be adequately predicted based on the proposed framework. Our analysis further indicates some limitations when liver microsomal intrinsic clearance (CLint) is used as the only clearance pathway and underscores the necessity of investigating the variability emanating from the different distribution models when providing PK predictions. The suggested approach aims at earlier exposure prediction in the drug development process so that critical decisions on molecule screening, chemistry design, or dose selection can be made as early as possible.

Keywords: PBPK; QSAR; artificial intelligence; drug discovery; machine learning; model-based drug development; pharmacokinetics.

FIGURE 1

Schematic of the framework studied in this work. Molecules were represented as molecular fingerprints (bit vector), descriptors, or graphs. Machine learning was then used to predict pharmacokinetic (PK) and physicochemical (PC) parameters based on molecular representations. PK/PC parameters were finally inserted to either 1-compartment or physiologically based models to predict rats’ exposure.

FIGURE 2

Scatter plots of observations (experimentally measured data) and predictions for test and train data in log scale for (A) Clearance (CL), (B) Volume of distribution (Vdss), (C) Intrinsic Clearance (CLint), (D) pka (most acidic), (E) pka (most basic), (F) Fraction unbound (Fu). Model predictions were based on best model as described in Table 2.

FIGURE 3

Comparison between observed rat plasma exposure and exposure predicted by the 1-compartment model using as input ML-derived CL, Vdss parameters. (A) Observed plasma exposure vs. 1-compartment model predictions. The solid line indicates the identity line. (B) Ratio between observed AUC until the final time point (AUC_{last_observed}) and 1-compartment model-predicted AUC (AUC_{last_predicted}). (C) Ratio between observed maximum concentration (C_{max_observed}) and 1-compartment model-predicted maximum concentration (C_{max_predicted}).

FIGURE 4

Comparison between observed PK profiles and profiles predicted using PBPK modeling with ML-driven PC/PK parameters and in vivo CL for the individual compounds tested. Different subplots indicate different compounds tested, and different colors indicate different distribution models.

FIGURE 5

Ratio between observed and predicted AUC and C_max based on PBPK prediction using ML-driven PK/PC parameters and in vivo clearance (CL). (A) Ratio between observed AUC until the final time point (AUC_{last_observed}) and PBPK model-predicted AUC (AUC_{last_predicted}). (B) Ratio between observed maximum concentration (C_{max_observed}) and PBPK model-predicted maximum concentration (C_{max_predicted}). Different boxplot colors indicate different distribution models.

FIGURE 6

Ratio between observed and predicted AUC and Cmax based on PBPK prediction using ML-driven PK/PC parameters and intrinsic clearance (CL_int). (A) Ratio between observed AUC until the final time point (AUC_{last_observed}) and PBPK model-predicted AUC (AUC_{last_predicted}). (B) Ratio between observed maximum concentration (C_{max_observed}) and PBPK model-predicted maximum concentration (C_{max_predicted}). Different boxplot colors indicate different distribution models.