Methods and applications of machine learning in structure-based drug discovery

Before and after AlphaFold2: An overview of protein structure prediction

“…Efforts are in progress to improve AlphaFold performance and models. Johansson-Akhe and Wallner (2022) demonstrated that randomly perturbing the neural network weights, forcing it to sample more conformational spaces can improve AlphaFold Multimer performance. Terwilliger et al (2022) suggested that inclusion of new experimental information can improve parts of the models, showing that application of experimental density maps used iteratively allows the rebuilding of models that can be used as templates by AlphaFold new prediction.…”

“…Third, AlphaFold fails to predict defects in protein folding due to point mutations. As demonstrated by Buel and Walters (2022), the differences between mutated and wild-type models predicted by AlphaFold are very small, represented by backbones RMSD lower than 1 A. Moreover, Pak et al (2021) also demonstrated that there is no correlation between AlphaFold accuracy metrics (pLDDT) and the impact of mutations on protein stability change (ΔΔG), neither with the side chain size change.…”

“…Today, AlphaFold DB has over 200 million entries from the human and 47 other organism proteomes, with the structure predictions and their respective analyses freely available to the scientific community. Porta-Pardo et al (2022) demonstrated that AlphaFold2 increases the structural coverage from 48% to 76% of all human protein residues, dropping the number of human protein without structural coverage from 5027 to 29. Moreover, they quantified that, among the 5027 of the proteins without structure previously, 4459 had structure prediction for over 50% of the protein's length (88,7%), in which 1408 with highaccuracy (28%).…”

“…Its application has been considered to design better protein expression experiment; To solve experimental structures faster, overcoming tedious model building, especially for X-ray crystallography, and to facilitate the interpretation of lowresolution cryo-EM; To protein design and drug development; To examine the effect of mutation in protein function, elucidating their potential impact on human diseases; To provide novel insights in poorly known molecular mechanisms (Perrakis and Sixma, 2021;Porta-Pardo et al, 2022). In this context, Noone et al (2022) has demonstrated that indeed AlphaFold offers shortcuts to solve protein structures experimentally, predicting the remaining N-terminal region of PTX3 complex, which was, then, validated with cryo-EM class averaging. The hybrid cryoEM/ AlphaFold structure allowed the mapping of the putative sites and regions of interaction, giving insights of the functions of PTX3.…”

Deep learning-based molecular dynamics simulation for structure-based drug design against SARS-CoV-2

“…In addition, some new programs of DL such as AlphaFold and its 2nd version (AlphaFold2) have been utilized in 3D structure prediction of protein which show ultra-high accuracy comparable to data collected by cryo-EM [24] , [25] , [26] . Artificial intelligence (AI) has attracted tremendous interests in the research of SBDD [27] , [28] , [29] , [30] , [31] . However, limitations of DL models used in drug design are that they learn through observations, and fail to consider the dynamic interactions of protein-protein/ligand.…”