QSAR and machine learning methods for prediction of albumin-ligand binding

Serum albumin is the most abundant protein found in human and mammalian blood plasma. It plays a crucial role in the transport and delivery of various molecules, including drugs. The affinity of small molecule drugs to albumin affects their ADME properties and overall efficacy. Hence, predictive models for albumin-ligand binding constants are of interest for computer-aided drug design. Due to the vast number of potential binding modes to albumin, the application of structure-based approaches relying on calculations of the binding free energy is challenging, while ligand-based QSAR and machine learning approaches can be trained to achieve rather good performance even for structurally diverse molecule sets. We consider the existing albumin affinity datasets and the numerous models designed to predict the affinity, which are based on various machine learning techniques and use different types of molecular descriptors. The transferability of the models to plasma protein binding values is discussed.