Notoriously expensive and not particularly effective, current methods for examining drug mechanisms of action have not fared well when it comes to the search for new antibiotics. Felix Wong and Aarti Krishnan, Postdoctoral Fellows Jim Collins Laboratory at MIT and members Broad Institute of MIT and Harvard UniversityI hope to address this with a new study that focuses on the ability of computer models to pinpoint a drug’s mechanisms of action. Here, Wong and Krishnan discuss AlphaFold, a promising AI program they used to accurately predict the behavior of bacterial proteins in interaction with antibacterial compounds. But is AlphaFold ready for the big leagues?
Global deaths from drug-resistant bacterial infections are expected to reach 10 million per year by 2050 (1), which is nearly double the number of global COVID-19 deaths reported to date. The increasing prevalence of antimicrobial resistance also means that there will be increased morbidity even for routine procedures such as surgeries and hospital care.
It took us 38 years to introduce a new class of antibiotics to the clinic — oxazolidinones in 2000 — after the introduction of quinolones in 1962 (2). There has been no shortage of innovation in antibiotic discovery, but finding clinically relevant antibiotics is challenging. The major classes that we discovered in the mid-20th century (and are still in use today) came from experimental screens of natural products, in particular of soil bacteria. Now, that pipeline has dried up and many of our efforts to devise new approaches have resulted in molecules that are toxic to humans or to which resistance could easily evolve. One way forward may be to significantly increase the chemical libraries we are exploring in order to better sample more aspects of the chemical space (3).
One of the main reasons companies are reluctant to invest in new antibiotics is that bringing these drugs to market is generally unprofitable. There is no financial incentive for a losing business. Governments, academia and industry must take immediate action so that we can discover new antibiotics against deadly superbugs. The typical research and development cost of an antibiotic can be as high as $1.5 billion, but the revenue it generates is only $46 million per year (4). This situation is due partly to the possibility that not many people will need a specific antibiotic and partly to the fact that the price one can charge for treatment is usually limited by government regulations.
Of course. Finding chemical compounds that kill bacteria isn’t difficult, but it’s rare to find ones that kill bacteria without being toxic to humans — and have enough medicinal chemical properties to inspire further testing. Furthermore, bacteria may quickly develop resistance to compounds of interest and render them useless.
It wasn’t always this difficult. The ‘golden age’ of antibiotic discovery in the mid-20th century saw many chemical screens yield new, selective and effective antibiotics. The problem now is that we picked too many of these low-hanging fruit. Meanwhile, the bacteria developed resistance. Our current chemical screens don’t produce nearly enough lead compounds, and that may be due to the fact that we can only explore so much chemical space. Developing new pipelines for computational screening is one approach that can help us navigate the chemical space and discover new lead compounds.
We now have computational methods for virtual screening, so we can quickly and inexpensively predict antibiotic activity from the chemical spaces of billions of compounds. We can go through these weeks and use our models to prioritize which molecules to buy and test in the lab. The Collins lab pioneered this type of approach and led to the discovery of a new candidate antibiotic, Halcin, two years ago (5).
However, predicting antibiotic activity is a coarse approach. Toxic compounds often have antibiotic activity (which the model will recognize), but they don’t make for very good drugs. In our study, we wanted to go further and predict drug binding targets. This means that we can, in principle, predict exactly how an antibacterial compound will act and whether or not its mechanism of action may have toxic liabilities. If we have it in silicoMethods can do this successfully, we can more easily select real antibiotics from large chemical spaces, determine how they select against bacteria, and perhaps even design antibiotics anew.
AlphaFold is an artificial intelligence system developed by Deep mind It uses the amino acid sequence of a protein to give us a three-dimensional structure. It can provide excellent predictions of the 3D structures of many proteins; These structures are freely available to the scientific community and can be used to simulate molecular fusion – in silicoSimilar approach to assembling a puzzle. This allows us to predict how a compound will target bacteria by simulating whether or not it binds to a specific protein of interest. Many antibiotics work this way. For example, quinolones bind specifically to bacterial DNA gyrase and topoisomerase, while beta-lactams bind specifically to bacterial penicillin-binding proteins.
In our research, we aimed to make this type of prediction on a large scale. We looked at the interactions between 296 proteins from Escherichia coli and 218 antibacterial compounds or ligands (6). By simulating each of the 64,310 protein-linker pairwise interactions using molecular docking on AlphaFold predicted protein structures, we can predict possible and unlikely binding interactions. We then performed bench experiments on 12 different proteins, experimentally testing them for binding activity with respect to each of the 218 antibacterial compounds. After comparing our model predictions with our experimental results, we found that the model performed no better than chance; It correctly predicted a real interaction only about half the time. Thus, one of the main findings of our study is that molecular docking needs to be improved so that we can correctly predict binding interactions and make better use of AlphaFold for antibiotic discovery. A known limitation of AlphaFold is that it only predicts stable and rigid protein structures that are ‘stuck’ in time, but the dynamic and disordered properties of these structures can be important for drug binding.
It is likely that we need more accurate ways of simulating the interactions between proteins and ligands. Indeed, when we used machine learning-based models trained on additional linkage information to complement the baseline docking approach, we found that predictive accuracy increased. It is also possible that improvements in protein structures may enable us to better predict drug binding.
We believe AlphaFold will be a useful resource for the drug discovery community. However, as Derek Lowe recently wrote, “it is very, very rare that knowing the structure of a protein should be any kind of rate-determining step in a drug discovery project” (7). Although AlphaFold could be improved to provide predictions about protein structure that could better aid drug discovery, we believe that some of the most needed improvements are in molecular docking.
Molecular docking is an intrinsically difficult problem because it aims to predict the conformation of only one binding (the most energetically favorable) out of the billions of possible conformations that a protein and linker might have. Aside from the dynamics and other biologically relevant factors, this involves simulating many-body interactions between large numbers of atoms. This task is computationally intensive, time-consuming, and not easy to solve. Supercomputing resources, together with new methods for performing molecular docking that can benefit from more information (for example, from protein–primary bonding interaction datasets) may help address this challenge.
To be clear, we’re thinking in silicoDrug discovery has great potential. Computational platforms, including those previously used by our lab to identify halocene, have often assisted the field in finding lead compounds of interest with limited resources. Our study focuses on demonstrating the limitations of one of these computational approaches, particularly with regard to the challenging task of predicting drug targets.
As part of the Antibiotics-AI project, we were interested in applying other deep learning approaches to de novoDesign of compounds that could have strong antibacterial selectivity and favorable medicinal chemo-properties (8). We believe that our study should help guide drug and antibiotic discovery fields and provide experimental reference data that could be useful in future efforts to predict the interactions of binding proteins. We also explored whether we could build on our study to develop molecular docking approaches that are better at identifying drug-target interactions or see if we could study proteins that would encourage the accuracy of this approach. If that’s the case, we can begin to more accurately identify compounds that bind to specific bacterial proteins and may have a chance of becoming useful antibiotics.
After earning a BA in English Literature and a MA in Creative Writing, I entered the publishing world as a proofreader, working my way up to becoming an editor. The career has so far taken me to some amazing places, and I’m excited to see where I can go with Texere and TMM.