AlphaFold2 in Molecular Discovery

Abstract

A dvanced machine learning methods impacted not only the prediction of properties for small molecules and quantitative structure-activity relationships 1 but also the prediction of protein structures.Since its unveiling at the 14th Critical Assessment of protein Structure Prediction (CASP14), AlphaFold2 (AF2) 2 is widely considered as a major breakthrough in protein structure prediction due to the high accuracy it achieved.AF2 uses the primary sequence of the protein as input and, employing a combination of multiple sequence alignment and neural networks, it returns as output a prediction of the protein structure and a confidence score per residue, indicating how reliable the model is.Apart from the code available online, there is also the AlphaFold database, 3 with predictions available for human proteins and proteins from other organisms.Thanks to AF2, computational chemists can now have access to structures of proteins which had no experimental structure available and which were not suitable for modeling by other strategies to predict structures, such as homology modeling.This virtual issue showcases the multiple uses of AF2 in computational chemistry, and it consists of a collection of 15 selected papers published in JCIM from 2021 onward.It covers different applications of AF2, grappling with significant issues for drug discovery, such as whether the models obtained from AF2 are of sufficient quality for docking, virtual screening, and free energy calculations.It also highlights different applications enabled or enriched by the ability to model proteins not suitable for other modeling approaches, such as the identification of new druggable binding sites or the prediction of protein function and enzyme stability upon mutation.The Viewpoint from Skolnick and colleagues 4 explains some reasons why AF2 is so accurate in its predictions.The authors highlight that training was helped by the fact that the library of single domain protein structures is complete and that the combination of deep learning with residue covariation led to significant improvement in the predictions.Baek and Kepp 5 evaluated the performance of AF2 in the prediction of human proteins using the local solvent exposure of a residue as a quality metric.AF2 had excellent performance for monomeric proteins.However, challenges and points of improvement were identified for multidomain structures, prolines, and exposed residues.Before AF2 many structure-based drug design campaigns were hampered by the lack of reliable information about protein structure.Now, many researchers are testing how useful AF2 models are in applications such as virtual screening and free energy perturbation.The Perspective from Schauperl and Denny 6 argues that AF2 can assist drug design by providing additional information about protein structure.The

Topics & Concepts