The use of AlphaFold2-enabled molecular docking simulations to predict protein-ligand interactions for antibiotic discovery

The use of AlphaFold2-enabled molecular docking simulations to predict protein-ligand interactions for antibiotic discovery

In a recently published study in Molecular systems biology the researchers showed that advances in modeling protein-ligand interactions using machine learning-based approaches are needed to better utilize AlphaFold2 for antibiotic discovery.

Lernen: Benchmarking AlphaFold-fähiger molekularer Docking-Vorhersagen für die Entdeckung von Antibiotika. Bildnachweis: Jaromond/Shutterstock

background

A major challenge in drug discovery is the identification of drug-target interactions. Researchers have developed several approaches to address this problem, including biochemical assays, genetic interactions, and molecular docking. However, only molecular docking has proven useful for identifying protein-ligand interactions and the mechanism(s) of action of a drug.

Although versatile, docking requires prior knowledge of protein structures. The number and quality of target protein structures further limit their application for identifying drug targets.

About studying

In the present study, researchers utilized the recently released AlphaFold2 protein structure prediction database to enable reverse docking approaches that span the essential proteome of Escherichia coli (E. coli) and enable comprehensive prediction of antibiotic binding targets. These experiments could help evaluate the performance of the modeling platform and reveal the prediction accuracy of AlphaFold2-enabled molecular docking simulations.

The predicted protein-ligand interactions between antibiotics and essential proteins could be partially interrogated experimentally using biochemical assays that measure enzymatic activity, with binding interactions supported by enzymatic inhibition.

The researchers conducted high-throughput screens of 39,128 growth inhibition compounds against wild-type E. coli through. These compounds were natural products, antibiotics, and structurally diverse molecules with molecular weights between 40 Daltons (Da) and 4,200 Da. All compounds that inhibited relative growth by 80% were considered active, and each active compound was computationally docked to 296 AlphaFold2-predicted E. coli essential proteomes.

As a control, a subset of the inactive compounds was docked in the same way. The researchers used AutoDock Vina, a widely used open-source docking program with benchmarking, to dock all 218 compounds against the 296 AlphaFold2-predicted essential proteomes. These simulations predicted both specific and widespread protein-ligand interactions and protein promiscuity. Finally, researchers evaluated predictions using four machine learning-based scoring functions (SFs), namely RF score, RF score-VS, protein-ligan extended connectivity (PLEC) score, and neural network (NN) score.

Study results

Genetics & Genomics eBook

Compilation of the top interviews, articles and news from the last year.

Download a copy today

A total of 218 antibacterial compounds were active against E. coli, and approximately 80% were antibiotics of the structural classes β-lactam, aminoglycosides, tetracyclines, quinolones, and polyketides. The remaining active ingredients consisted of toxins and antineoplastic compounds. The study also identified an additional set of compounds whose antibacterial properties against E. coli have not previously been documented.

Likewise, the study analysis predicted the binding position and binding affinity of 64,528 protein-ligand pairs. An additional 100 randomly selected inactive compounds led to binding pose and affinity predictions for 29,600 protein-ligand pairs via analog docking simulations. In addition, the researchers measured the enzymatic activity of several E. coli proteins involved in deoxyribonucleic acid (DNA) replication, transcription and cell wall synthesis.

Interestingly, several protein molecules enzymatically inhibited each identified antibacterial compound, confirming extensive promiscuity. This phenomenon also enabled benchmarking of model performance on a statistically significant scale.

The researchers extensively compared experimental data of protein-ligand interactions with in silico predictions to show that this modeling approach had a prediction accuracy between 41% and 73%, depending on the binding affinity threshold used. Regardless of the binding affinity threshold, the area under the receiver operating characteristic curve (auROC) across the essential proteins averaged 0.48.

Remarkably, model performance remained similar even when researchers used specific protein structures experimentally. Assuming that a random model corresponds to an auROC of 0.5, these results showed that molecular docking simulations performed poorly.

The authors found a profound improvement in model performance as measured by the auROC, with RF-score, RF-score-VS and NN-score. Conversely, model performance did not improve when DOCK6.9 was used and rescoring was performed using the PLEC score. Furthermore, consensus models that included multiple machine learning-based SFs improved the true positive rate to false positive rate ratio and prediction accuracy.

Conclusions

The current study demonstrated that using AlphaFold2 to predict drug targets is a promising method but is still in its early stages. Accordingly, realizing its potential for drug discovery will require significant improvements in modeling protein-ligand interactions. Benchmarking the performance of molecular docking simulations is one of the viable ways to improve prediction accuracy; however, it requires the accompanying use of machine learning-based approaches. Overall, the study results could inform the appropriate use of AlphaFold2 in drug research.

Reference:

• Benchmarking von AlphaFold-fähigen molekularen Docking-Vorhersagen für die Entdeckung von Antibiotika, Felix Wong, Aarti Krishnan, Erica J. Zheng, Hannes Stärk, Abigail L. Manson, Ashlee M. Earl, Tommi Jaakkola, James J. Collins. Molekulare Systembiologie. doi: https://doi.org/10.15252/msb.202211081 https://www.embopress.org/doi/full/10.15252/msb.202211081

.