1 Introduction

The β2 adrenergic receptor (ADRB2) is one of the most studied targets in the G protein coupled receptors family (GPCRs), with multiple resolved crystallographic structures and many known ligands with different pharmacological profiles. GPCRs play a key biological role mediating the interaction between the cell and the environment via signalling transduction pathways. The binding of a molecule in the extracellular binding pocket regulates the signalling by enhancing it (agonist), or inhibiting it (antagonists, and inverse agonist). ADRB2 agonists are an important class of drugs for the treatment of respiratory diseases such as asthma and chronic obstructive pulmonary disease (COPD) 1, 2. Based on their duration of action they have been classified as short acting (SABA), long acting (LABA), and ultra-long acting (Ultra-LABA) β-agonists 3. Each class has a specific therapeutic role: SABAs are used as quick relief medications, while LABAs are used as long-term control medications 4. In the past twenty years pharmaceutical companies have developed new LABAs and Ultra-LABAs 3, 5, aiming for a once-daily dosing agonist, whose duration of action is optimal for patient management 6.

The exact mechanism regulating the duration of action of β2-agonists is still unclear. The most supported hypothesis is the diffusion microkinetic theory 7, 8, which links the duration of action to the partition of the ligand between the aqueous phase and the lipid membrane. More lipophilic compounds accumulate in the membrane, forming a ligand's reserve, increasing the probability of rebinding and consequently the duration of their effect. Other mechanisms could explain the difference in duration of action: The binding kinetics of the ligands was proposed as a determining factor, however experiments examining this aspect gave conflicting results 9, 10. The existence of an additional binding site, called exosite, was proposed to explain the long duration of action of salmeterol and its derivatives 11. A secondary binding pocket accommodating the ligand long aliphatic chain was identified by crystallography in the region between the extracellular ends of TM6 and TM7, and ECL2 12. Mutations in this region however only affected the binding affinity of salmeterol but not its duration of action 13.

In this work we propose to explore different computer aided drug design (CADD) methods, compatible with high-throughput virtual screening, to qualitatively predict the duration of action of 22 well characterized β2-agonists: 10 LABAs (including both LABAs and Ultra-LABAs) and 12 SABAs (Figure 1).

The significant value of the dataset is enhanced through rigorous validation by a pharmacologist, involving extensive exploration of the literature concerning pharmacodynamics and pharmacokinetics (Table S1–2).

The evaluated approaches can be divided in two categories: structure-based, and ligand-based. Ligand based drug design (LBDD) uses information from known binders to select new potential active molecules. In this work we explore different molecular representations of increasing complexity: calculated physicochemical properties, molecular and pharmacophoric fingerprints, and 3D ligand structures.

The aim is to identify the properties and structural features regulating the duration of action. Structure based drug design (SBDD) uses the 3D structure of the target protein to guide drug discovery. The most used technique is protein-ligand docking, which predicts the pose of a molecule in the target's binding site and evaluates its binding affinity using a scoring function. Pharmacophore models, based on the interactions observed in crystallographic structures 14 or molecular dynamic simulations 15, were often used to post-process docking results, especially to identify ligands with a desired pharmacological profile. SBDD strategies were applied using both crystallographic structures of the receptor-agonist complex 16, 17, and structures generated by an in silico pipeline combining AlphFold2 (AF) 18 and molecular docking. The development of deep learning methods for protein structure prediction carries the promise of access to SBDD for any target, although the use of such structures in virtual screening is still limited by the quality of the predictions 19-22. Protein-ligand docking, and molecular dynamics were used to model the complex between ADRB2 and the agonists formoterol and salbutamol. The modelled and crystallographic binding modes were used for pose rescoring in a retrospective docking exercise using the dynamic representation obtained combining interaction graphs (IGs) and one class support vector machine (IG-OCSVM) 15.

The objective is to determine the key molecular properties regulating the duration of action of known β2-agonists and evaluate different strategies for its prediction. This analysis was performed considering different descriptors, different machine learning methods, as well as various degrees of information (one reference, two references with different duration, a curated dataset). We also explored the use of interaction-based methods in SBDD for such task, since they were already successfully applied for the detection of ligands with a specific pharmacological profile, even with limited available experimental data 15. The presence of a bias towards ligands with a specific duration of action in SBDD interaction-based methods could be linked to the used representation indirectly encoding some ligand properties. Comparing the results of the two approaches might explain the origin of such bias.