Artificial intelligence and machine learning (AI/ML) models that predict human intestinal absorption (HIA) can streamline pharmaceutical development by identifying promising compounds earlier and reducing reliance on costly and time-consuming lab and animal testing. This accelerates the screening process and helps optimize drug formulations before clinical trials begin. Several different types of AI/ML modeling approaches exist, including artificial neural networks, evolutionary computation, rule-based systems and automated machine learning therefore, it can be difficult to determine which approach to undertake for a given problem statement. This AAPS-hosted webinar will highlight and compare mechanistic and empirical AI/ML approaches to predicting human intestinal absorption (HIA) or fraction absorbed.

In the first part of the webinar, Aleksander Mendyk will describe empirical modeling via AI-based QSPR models for predicting HIA. AI/ML as purely empirical modeling techniques could be used in the same very manner both for predictive modeling of drugs dissolution and human intestinal absorption (HIA). The key elements are data representation and a supervised learning approach. The latter requires labeled databases with retrospective knowledge of the process of interest. The former is usually a linear vector with its elements, including labels, represented as real numbers, thus constituting a regression problem. The software, the modeling approaches and a procedure of selection of the best model are exactly the same, regardless whether dissolution profile or the HIA is modeled. There will be presented real modeling results concerning prediction of drugs dissolution profiles and HIA. Data representation using molecular descriptors, both failures and successes in the empirical modeling of these phenomena will be provided and commented. Specifically, modeling data domain partitioning resulting from numerical and domain knowledge insights will be described. It turned out that in fact empirical prediction of a whole range of HIA was impossible and therefore more specialized AutoML ensemble models were found to be a solution of poor data structure and quality. Moreover, a specific interplay between classification and regression-based models will be presented. Future directions in this line of research will be addressed.

In the second part of the webinar, Michael Lawless will discuss how machine learning and mechanistic models can be combined to predict HIA. High-throughput, mechanistic PBPK simulations were conducted in a human model for approximately 800 orally administered drugs, using an immediate-release tablet formulation. Each compound was simulated at two dose levels (1 and 1000 mg) to assess dose-dependent effects on absorption. The Advanced Compartmental Absorption and Transit (ACAT™) model was employed to predict the fraction absorbed (%Fa), representing human intestinal absorption. Input parameters (pKa, solubility, permeability etc.) were predicted directly from each compound’s 2D chemical structure using artificial neural network ensemble (ANNE) models. The simulations also generated time-course profiles of the amount dissolved in the GI tract and the amount absorbed into systemic circulation. This AI-driven approach enabled rapid, large-scale simulations without requiring experimental data. Post-simulation analysis identified physicochemical factors contributing to reduced %Fa at the higher dose. These insights support early-stage compound prioritization and risk assessment. As experimental data (e.g., solubility or intrinsic clearance) become available, they can be incorporated to refine predictions and enhance translational accuracy.

Key aspects of Alex’s talk: Prediction of both dissolution and HIA achieved through direct mapping of digital representation of chemical structure and the endpoint of interest. Autonomous behavior of automatic machine learning systems (AutoML) allows such early stage attempts to build meaningful relationships without a priori knowledge.

Key aspects of Michael’s talk: Prediction of fundamental phenomena driving HIA and bioavailability, including pKa, solubility in gastrointestinal fluids, intestinal permeability, and metabolism. Integration of these predicted properties into mechanistic simulations and parameter sensitivity analyses of HIA.

Panel:

• Sandip Tiwari
• Michael Lawless
• Aleksander Mendyk