Imma Perfetto
Cosmos science journalist
Scientists want to use computer models to help reduce the cost and time associated with drug discovery, to develop new antibiotics to fight the growing crisis of antimicrobial resistance. But a new study shows that using the latest tools together are little better than guesswork at the moment.
This is a barrier to drug development – at least as the computer models exist now – according to the study a new study published in Molecular Systems Biology.
Researchers from Massachusetts Institute of Technology (MIT) explored whether existing computer programs could accurately predict the interactions between antibacterial compounds and bacterial protein structures generated by Google’s new tool called AlphaFold – an artificial intelligence program that generates 3D protein structures from their amino acid sequence.
AlphaFold is exciting the science world.
But the MIT team found that the predictions of existing models, called molecular docking simulations, performed little better than chance.
“Breakthroughs such as AlphaFold are expanding the possibilities for in silico (ie by computers) drug discovery efforts, but these developments need to be coupled with additional advances in other aspects of modelling that are part of drug discovery efforts,” says senior author James Collins, professor of Medical Engineering and Science in MIT’s Institute for Medical Engineering and Science (IMES) and Department of Biological Engineering.
“Our study speaks to both the current abilities and the current limitations of computational platforms for drug discovery.”
The hope is that scientists could use modelling to perform large-scale screening for new compounds that affect previously untargeted bacterial proteins. The end result being the development of new antibiotics that work in unprecedented ways.
Read more: Google’s protein-folding AI AlphaFold has nearly cracked them all
The team studied the interactions of 296 essential proteins from Escherichia coli with 218 antibacterial compounds, using molecular docking simulations that predict how strongly two molecules will bind together based on their shapes and physical properties.
Previously, these simulations have been used successfully to screen large numbers of compounds against a single protein target to identify compounds that bind the best. But here, the predictions became much less accurate when attempting to screen many compounds against many potential protein targets.
In fact, the model produced false positive rates similar to true positive rates when simulating interactions between existing drugs and their targets.
“Utilising these standard molecular docking simulations, we obtained an auROC value of roughly 0.5, which basically says you’re doing no better than if you were randomly guessing,” Collins explains.
But this wasn’t due to some fault of AlphaFold, as similar results occurred when they used the same modelling approach with protein structures that had been experimentally determined in the lab.
“AlphaFold appears to do roughly as well as experimentally determined structures, but we need to do a better job with molecular docking models if we’re going to utilise AlphaFold effectively and extensively in drug discovery,” adds Collins.
One explanation for this poor performance is that the protein structures fed into the model are static, but in real biological systems proteins are flexible and often shift their configurations.
The researchers were able to improve the performance of the molecular docking simulations by running them through four additional machine-learning models trained on data that describe how proteins and other molecules interact with each other.
Read more: TB drug discovery paves way
“The machine-learning models learn not just the shapes, but also chemical and physical properties of the known interactions, and then use that information to reassess the docking predictions,” says co-lead author Felix Wong, applied physicist and postdoctoral fellow in Collins’ lab at MIT.
“We found that if you were to filter the interactions using those additional models, you can get a higher ratio of true positives to false positives.”
“We’re optimistic that with improvements to the modelling approaches and expansion of computing power, these techniques will become increasingly important in drug discovery,” concludes Collins. “However, we have a long way to go to achieve the full potential of in silico drug discovery.”