Machine Learning Navigated Allosteric Network to Unveil Biased Allosteric Modulation of GPCRs

Abstract

Biased allosteric modulators (BAMs) offer a promising avenue for developing safer and more selective therapeutics for G protein-coupled receptors (GPCRs). However, their molecular mechanisms remain unclear due to the complex combination of biased and allosteric characteristics. Motivated by the challenge, we proposed a machine learning navigated allosteric network strategy to address the issue. It consists of molecular dynamics simulation, a residue-level interpretable deep learning model, and allosteric network analysis, named as RMLNA. RMLNA first obtains biased conformation states through MD simulation and a density map. Then, an interpretable CNN-based classification model is utilized to identify important residues deciding the biased conformation. Navigated with these important residues, allosteric network analysis uncovers their regulation effects. With RMLNA, we revealed the biased allosteric modulation mechanism of a β-arrestin-biased modulator (SBI-553) for the clinically important target NTSR1. SBI-553 stabilizes a unique β-arrestin-biased state with an expanded intracellular binding site and the orthosteric ligand binding mode related to the β-arrestin-biased signaling. The interpretable deep learning model suggests that the middle and the lower parts of TM5 and TM6 are key determinants for the G protein/β-arrestin bias, while SBI-553 modulates the β-arrestin signaling mainly by H8 and the intracellular end of TM6 and TM7. Under the guidance of these results, the community network analysis underlines that the communication between TM5/6 and TM1/7 or TM2/4 is important for the β-arrestin-biased signaling, where SBI-553 redirects the communication between TM5/6 and TM1/7 via F8.50 of H8, inducing enhanced β-arrestin-biased signaling. NTS–NTSR1−β-arrestin complexes with and without binding of SBI-553 are constructed and simulated to further reveal the biased allosteric modulation mechanism to the β-arrestin and validate the reliability of the workflow. Collectively, this work provides molecular insights into the biased allosteric modulation of SBI-553 on NTSR1. More importantly, the novel computational workflow can be extended to other GPCRs.