Benchmarking AlphaFold‐enabled molecular docking predictions for antibiotic discovery

“We used a Python script that employs the DockPrep function from UCSF Chimera (Pettersen et al , 2004; Allen, 2018) to add hydrogens and partial charges.”

“This is consistent with previous assessments of AlphaFold’s fidelity to experimentally determined protein structures (Jumper et al , 2021) and suggests that the weak performance of our model arises from the docking method, and not the quality of protein structures.”

“As positive controls for our docking simulations, we re-docked 11 experimentally evidenced protein-ligand complexes from the RCSB Protein Data Bank (Dataset EV2).”

“Different software for performing molecular docking are widely available and commonly used platforms include AutoDock Vina (Eberhardt et al , 2021 ) and DOCK (Allen et al , 2015 ).”

“We shortlisted genes identified in at least two of these three studies, resulting in a total of 295 genes out of ∼ 4,000 total genes in E. coli (Blattner et al , 1997).”

“As shown in the workflow here, in order to leverage AlphaFold 2 for docking, we first downloaded all 296 AlphaFold 2 -predicted E. coli essential protein structures from the AlphaFold Protein Structure Database (Jumper et al , 2021 ; Varadi et al , 2022 ).”

“The advent of AlphaFold2 and other deep neural networks for protein folding, such as RoseTTAFold (Baek et al , 2021; Humphreys et al , 2021), has been widely anticipated and celebrated in structural biology.”

“Various approaches to identifying molecular drug targets have been developed, including those based on biochemical assays, genetic interactions, and molecular docking (Kitchen et al , 2004; Schenone et al , 2013).”

“Model predictions based on the following rescoring functions were ensembled in sequence: RF-Score, NNScore, PLEC score, and RF-Score-VS.”

“These genes were manually mapped to corresponding UniProt identifiers based on the E. coli K-12 reference proteome (UniProt: UP000000625) and used to obtain the corresponding AlphaFold2-predicted protein structures.”

“For each protein, the protein-averaged pLDDT value shown in Fig EV3 was calculated as the average pLDDT value across all residues.”

“Model predictions based on the following rescoring functions were ensembled in sequence: RF-Score, NNScore, PLEC score, and RF-Score-VS.”

“Various approaches to identifying molecular drug targets have been developed, including those based on biochemical assays, genetic interactions, and molecular docking (Kitchen et al , 2004; Schenone et al , 2013).”

“These assessments of model performance indicate that certain machine learning-based scoring functions improve prediction accuracy.”

“In view of these observations, it is important to note that docking has been widely used in “one target, many compounds” approaches to enrich for screening hits (Lyu et al , 2019; Bender et al , 2021).”

“These assessments of model performance indicate that certain machine learning-based scoring functions improve prediction accuracy.”

“In view of these observations, it is important to note that docking has been widely used in “one target, many compounds” approaches to enrich for screening hits (Lyu et al , 2019; Bender et al , 2021).”

“The advent of AlphaFold2 and other deep neural networks for protein folding, such as RoseTTAFold (Baek et al , 2021; Humphreys et al , 2021), has been widely anticipated and celebrated in structural biology.”

“Model predictions based on the following rescoring functions were ensembled in sequence: RF-Score, NNScore, PLEC score, and RF-Score-VS.”

“Con-comitantly, limitations to the development of more accurate docking methods are the use of rigid protein docking in this and other benchmarking studies (Durrant et al , 2013; Pereira et al , 2016) and the scarcity of benchmarking datasets.”

“…dnaE , rpoABCEZ , murA , and murF ) of the 12 essential proteins tested for the inclusive binding affinity threshold of − 5 kcal/mol and for 9 essential proteins ( gyrAB , dnaG , dnaB , ligA , rpoABCEZ , glmU , murA , murD , and murF ) for the stringent binding affinity threshold of − 7 kcal/mol…”

“…in integrating machine learning with docking have resulted in machine learning-based scoring functions, and their use to rescore poses generated by other docking platforms (Ballester & Mitchell, 2010 ; Durrant & McCammon, 2010 ; Pereira et al , 2016 ; W (cid:2) ojcikowski et al , 2017 , 2019 ).”

“We used AutoDock Vina 1.2.0 (Eberhardt et al , 2021) to dock each of our 218 antibacterial compounds with each of the 296 essential E. coli proteins.”

“Model predictions based on the following rescoring functions were ensembled in sequence: RF-Score, NNScore, PLEC score, and RF-Score-VS.”

“…RF-Score (Ballester & Mitchell, 2010), RF-Score-VS (W (cid:2) ojcikowski et al , 2017), PLEC score (W (cid:2) ojcikowski et al , 2019), and NNScore (Durrant & McCammon, 2010) — we found improvements in performance, as measured by the auROC, with three of the four scoring functions (RF-Score,…”

“Long molecular dynamics simulations that focus on a specific protein of interest could account for protein conformational changes that, in certain cases like AcrB, might be important for ligand binding (Vargiu & Nikaido, 2012).”

“As computational docking approaches are known to predict many false positives (Adeshina et al , 2020), the predicted protein-ligand interactions could be experimentally interrogated, in part, using biochemical assays that measure enzymatic activity, with binding interactions supported by enzymatic…”

“Essential genes in E. coli were compiled from previous studies based on transposon-directed insertion site sequencing (Goodall et al , 2018) and CRISPR interference screening (Rousset et al , 2018, 2021).”

“The mean predicted local distance difference test (pLDDT) value of each protein was extracted and calculated from the B-factor column containing the pLDDT value per atom and per residue (Tunyasuvunakool et al , 2021).”

“Although versatile, reverse docking requires a priori knowledge of the protein structures of interest, and its application to drug-target identification has been limited by the number and quality of target protein structures (Chen & Zhi, 2001; Kharkar et al , 2014; Lee et al , 2016).”

“Although versatile, reverse docking requires a priori knowledge of the protein structures of interest, and its application to drug-target identification has been limited by the number and quality of target protein structures (Chen & Zhi, 2001; Kharkar et al , 2014; Lee et al , 2016).”

“We proceeded to dock all 218 active compounds against the 296 AlphaFold2-predicted essential protein structures using Auto-Dock Vina, a widely used and benchmarked open-source program for docking (Pereira et al , 2016; Vieira & Sousa, 2019; Eberhardt et al , 2021; Fig EV1).”

“Essential genes in E. coli were compiled from previous studies based on transposon-directed insertion site sequencing (Goodall et al , 2018) and CRISPR interference screening (Rousset et al , 2018, 2021).”

“To enable docking of this compound, we replaced its boron atom with a carbon atom, as is often done in molecular docking (Tiwari et al , 2009).”

“Similar to previous work determining growth inhibition (Stokes et al , 2020; Wong et al , 2021a, 2021b), E. coli BW25113 was grown overnight in liquid LB medium in 14 ml Falcon tubes at 37 ° C with shaking at 300 rpm in a light-protected incubator, then diluted 1:10,000 in fresh liquid LB, and…”

“The binding affinity ( pK d ) predictions generated by all rescoring methods are reported in Dataset EV5.”

“Model predictions based on the following rescoring functions were ensembled in sequence: RF-Score, NNScore, PLEC score, and RF-Score-VS.”

“The predicted binding pose of each complex was superimposed with the experimentally determined structure using PyMol.”

“A clear limitation to AlphaFold2 is that it is unable to distinguish between the active and inactive conformations of a protein (Mullard, 2021).”

“…AutoDock Vina, DOCK6, Schro¨dinger’s Glide, Surflex, and internal coordinate mechanics (ICM) have shown that performance is similar on the directory of useful decoys (DUD) dataset (Durrant et al , 2013; Pereira et al , 2016), suggesting that our findings do not depend on the docking software used.”

“The binding affinity ( pK d ) predictions generated by all rescoring methods are reported in Dataset EV5.”

“…targets lacking experimental structures, especially for target classes with less obvious structural homologies in the ligand-binding pocket, the performance of AlphaFold2 models in small-molecule docking showed disappointing results in recent assessments for GPCR and antibacterial targets 107,108 .”

“Wong et al. 102 combined molecular docking simulations with AF2 for protein-ligand interactions prediction.”