The advent of AlphaFold2 (AF2) has brought AI applications in drug development to unprecedented heights. Its groundbreaking ability to accurately predict protein structures is transformative for identifying new drug targets and designing effective drug molecules. G protein-coupled receptors (GPCRs) are major drug targets, yet their complex and dynamic structures pose significant challenges to traditional structural analysis methods. By using machine learning, AF2 can accurately predict the 3D structures of GPCRs with atomic-level accuracy and remarkable scalability, providing crucial structural insights for drug development. The AlphaFold database encompasses over 200 million proteins, aiding structural biology, protein design, and function prediction. However, the impact on structure-based ligand discovery remained uncertain due to the necessity for accurately modeled binding sites. This study examines these concerns by comparing AF2's predictions with experimental structures for docking experiments.
Lyu et al. prospectively docked ultra-large libraries of molecules against unrefined AF2 models of the σ2 and 5-HT2A serotonin receptors, comparing the results to docking against experimental structures. They found that AF2 models achieved accurate side-chain predictions and successfully docked high-affinity ligands. For the σ2 receptor, AF2 predicted side-chain conformations with an RMSD of 1.1 Å from the crystal structure. Docking 490 million molecules against the σ2 receptor's AF2 model yielded a 54% hit rate, comparable to the 51% hit rate using the crystal structure. For the 5-HT2A receptor, AF2 models showed some variations at key residues, and docking 1.6 billion molecules resulted in high hit rates. Of 161 molecules tested, 42 substituted more than 50% of [³H]-LSD at 10 μM, achieving a 26% hit rate. The highest affinity compounds (15 to 24 nM) were identified from AF2 docking. Functional activity of selected compounds was assessed across 5-HT2A, 5-HT2B, and 5-HT2C receptors, with several compounds demonstrating subtype selectivity and high potency, some with sub-nanomolar EC50 values, indicating strong and selective receptor binding. This validates AF2's potential in enhancing drug discovery precision and efficiency.
Despite its transformative potential, AI in drug development faces several challenges. One significant concern is the reliance on data quality and quantity, where inaccuracies or biases in training data can lead to flawed predictions. Moreover, advanced AI models like AlphaFold3, which can predict complex protein-molecule interactions and post-translational modifications, are now accessible only via a restricted server, unlike its predecessor AlphaFold2. This proprietary nature limits direct access to the model and imposes a cap on daily predictions. The lack of transparency surrounding its operations and usage constraints restricts extensive academic scrutiny. Addressing these challenges is crucial to fully harness AI's potential in accelerating drug discovery and optimizing therapeutic outcomes.
Looking ahead, AI offers significant advantages in drug development, such as the ability to tackle complex targets and accurately predict protein-molecule interactions. To fully harness these benefits, it is essential to keep updating and refining the databases with high-quality data. This effort will help resolve issues related to data accuracy and bias, leading to more dependable predictions. Additionally, transparency in AI models and their operations is crucial for building trust and allowing for academic review. Furthermore, AI's role in structural biology and drug design is set to spark innovation in automated screening and large-scale data analysis, paving the way for a new era in precision medicine and medical research.
References:
Lyu, J. et al. AlphaFold2 structures guide prospective ligand discovery. Science https://doi.org/10.1126/science.adn6354Â (2024)
Abramson, J., Adler, J., Dunger, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024). https://doi.org/10.1038/s41586-024-07487-w
Comments