Known for being expensive and not particularly efficient, current methods for evaluating drug mechanisms of action have not worked well when it comes to finding new antibiotics. Felix Wong and Aarti Krishnan, postdoctoral fellows in the Jim Collins Laboratory at MIT and members of the Broad Institute of MIT and Harvard, hope to address that with a new study focused on the ability of computer models to identify the mechanisms of drug action. . Here, Wong and Krishnan discuss AlphaFold, a promising AI program they have used to accurately predict the behavior of bacterial proteins in interaction with antibacterial compounds. But is AlphaFold ready for the big leagues?
Global deaths due to drug-resistant bacterial infections are projected to reach 10 million per year by 2050 (1), nearly double the number of global deaths from COVID-19 reported to date. The increased prevalence of AMR 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 into the clinic, the oxazolidinones in 2000, after the introduction of the quinolones in 1962 (2). There has been no shortage of innovation in antibiotic discovery, but finding clinically relevant antibiotics is difficult. The main classes that we discovered in the mid-20th century (which are still in use today) came from empirical analyzes of natural products, particularly soil bacteria. Now this pipeline has dried up and many of our efforts to invent new approaches have produced molecules that are either toxic to humans or to which resistance could easily develop. One way forward may be to vastly increase the chemical libraries we are exploring to better sample more aspects of chemical space (3).
One of the main reasons companies are hesitant to invest in new antibiotics is that bringing these drugs to market is generally not profitable. There is no financial incentive for a business that loses money. Governments, academia and industry must take immediate action so that we can discover new antibiotics against killer superbugs. A typical R&D cost for an antibiotic may be USD 1.5 billion, but the revenue it generates is only about USD 46 million per year (4). This situation is partly due to the possibility that not many people need that specific antibiotic, and partly to the fact that the price that can be charged for treatment is often limited by government regulations.
Definitely. Finding chemical compounds that kill bacteria isn’t difficult, but finding ones that kill bacteria without being toxic to humans, and have enough favorable medicinal chemical properties to inspire further testing, is rare. In addition to this, bacteria can quickly develop resistance to the compounds of interest and render them useless.
It wasn’t always that hard. The “golden age” of antibiotic discovery in the mid-20th century saw many new, selective, and effective antibiotics produced by many chemical screens. The problem now is that we have picked up much of this low hanging fruit. Meanwhile, bacteria have developed resistance. Our current chemical screens do not produce enough lead compounds, which may be due to the fact that we can only explore a limited amount of chemical space. Developing new computational detection pipelines is one approach that could help us navigate the chemical space and discover new lead compounds.
We now have computational approaches to virtual detection, so we can quickly and cheaply predict antibiotic activity from the chemical spaces of billions of compounds. We can review them in weeks and use our models to prioritize which molecules we get and test in the lab. The Collins lab pioneered this type of approach and resulted in the discovery of a new antibiotic candidate, halicin, two years ago (5).
However, predicting the activity of antibiotics is a coarse-grained approach. Toxic compounds often have antibiotic activity (which the model would recognize), but they are not very good drugs. In our study, we wanted to go a step further and predict drug binding targets. This means that, in principle, we could predict how exactly an antibacterial compound works and whether or not its mechanism of action might have toxic effects. If our in silicoApproaches could do this successfully, we could more easily select real antibiotics from large chemical spaces, characterize how they select against bacteria, and perhaps even design de novo antibiotics.
AlphaFold is an AI system developed by DeepMind that 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; those structures are freely available to the scientific community and can be used for simulations of molecular docking, an in silicoapproach similar to putting together a puzzle. This allows us to predict how a compound targets bacteria by simulating whether or not it binds to a specific protein of interest. Many antibiotics work this way; for example, quinolones specifically bind to bacterial DNA gyrase and topoisomerase, whereas beta-lactams specifically bind to bacterial penicillin-binding proteins.
In our research, our goal was to perform this type of prediction at scale. We examined the interactions between 296 proteins from Escherichia coli and 218 antibacterial compounds or ligands (6). By simulating each of the 64,310 pairwise protein-ligand interactions by molecular docking into the corresponding AlphaFold-predicted protein structures, we were able to predict which binding interactions were likely and which were unlikely. We then performed benchtop experiments for 12 different proteins, empirically testing their binding activity against each of the 218 antibacterial compounds. After comparing the predictions of our model with our experimental results, we found that the model did not perform better than chance; it correctly predicted a real interaction only about half the time. Therefore, one of the main conclusions of our study is that molecular docking needs to be improved so that we can correctly predict binding interactions and better leverage AlphaFold for antibiotic discovery. A known limitation of AlphaFold is that it only predicts static and rigid protein structures that are “stuck” in time, but the dynamic and disordered properties of these structures could be important for drug binding.
We may need more accurate ways to simulate protein-ligand interactions. In fact, when we used machine learning-based models trained with additional binding information to complement the reference coupling approach, we found that predictive accuracy increased. It is also likely that improvements in protein structures will allow us to better predict drug binding.
We believe that AlphaFold will be a useful resource for the drug discovery community. However, as Derek Lowe recently wrote: “It is very, very rare that knowledge of a protein’s structure is a limiting step in a drug discovery project” (7). Although AlphaFold can be improved to provide protein structure predictions that could better inform drug discovery, we believe that some of the most needed improvements are in molecular docking.
Molecular docking is inherently a difficult problem because it aims to predict only one binding conformation (the most energetically favourable) out of billions of potential conformations that a protein and a ligand can have. Setting aside 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 easily solved. Supercomputing resources, along with new ways of performing molecular docking that can take advantage of more information (for example, from basic truth protein-ligand interaction data sets), could help with this challenge.
To be clear, we think in silicoDrug discovery has great potential. Computational platforms, including the one our lab previously used to identify halicin, have often helped the field find interesting lead compounds with limited resources. Our study focuses on showing the limitations of one such computational approach, especially with regard to the challenging task of predicting drug targets.
As part of the Antibiotics-AI Project, we have been interested in applying other deep learning approaches to the newdesign of compounds that could have a strong selectivity against bacteria and favorable medicinal chemical properties (8). We believe that our study should help guide the drug and antibiotic discovery fields and provide empirical benchmarking data that may be useful in future efforts to predict protein-ligand interactions. We have also been exploring whether we could build on our study to develop molecular docking approaches that are better at identifying drug-target interactions or see if we can further study proteins for which the accuracy of this approach is encouraging. If so, we could begin to more precisely identify compounds that bind to specific bacterial proteins and may have the potential to become useful antibiotics.
After a BA in English Literature and an MA in Creative Writing, I entered the world of publishing as a proofreader, working my way up to editor. So far my career has taken me to some amazing places and I’m excited to see where I can go with Texere and TMM.