The Chemist’s Guide to Advances in Computational Drug Design
Drug development is an extremely expensive and time-consuming business. During the years 2000 to 2010, the average cost of developing a single new approved drug was USD 2.6 billion! On average, it takes 10 to 15 years to design a single successful drug and only around 12% of candidate molecules that do get tested on humans, are approved as safe and effective. Only a minority of approved drugs return the revenue that exceeds the research and development costs. As a result, there is a huge drive to make drug discovery cheaper and faster. Not too surprisingly, the new technologies of artificial intelligence and machine learning can come to help to drug developers.
A brief overview of drug function
Generally, drugs are molecules that bind to certain proteins in our bodies, or sometimes to proteins of microorganisms, as is the case with antibiotics, whose function is to kill infectious bacteria. The drug binds proteins such as an enzyme (it catalyses a certain chemical reaction) in order to prevent the designated natural molecule from binding to it, or a receptor (it produces a signal on binding to a special molecule) in order to suppress or amplify that signal. An important factor determining drug effectiveness in binding affinity – how strongly the drug molecule can bind to its designated binding site, which is commonly a pocket inside a protein. Drugs often resemble the natural molecules that bind to that site but have some modifications that make them bind much more strongly. Binding affinity is maximised by designing a molecule that has the right shape to fit inside the binding pocket and can form strong interactions with amino acid residues of the pocket. For example, if the binding pocket contains a charged amino acid residue, the drug should have an oppositely charged group adjacent to the charged pocket site to provide strong interaction.
Successful binding to the desired protein is not the end of the journey for a drug. The molecule needs to be successful at getting to the target protein in the first place. That means being soluble, able to get through the cell membrane, not binding to the wrong protein etc. Once it has performed its function, the drug molecule needs to be excreted or altered chemically (metabolised). It is important that the drug does not stay in the organism for too long or get metabolised into toxic compounds. There are just so many criteria that molecule must fulfil. Any possible target molecule needs to pass all the steps to be approved as a drug. That involves computational modelling of the binding interactions, called in silico testing, testing in microorganisms and animals such as mice, and finally clinical tests, where the drug is tested in humans. Only a few, if any, structures out of tens of thousands are approved.
How does Machine Learning (ML) aid drug discovery?
The process of drug development yields vast amounts of data. This data can be used to train artificial neural networks. One way ANNs can be used is to design new potential drug molecules. The molecules need to be written down as a code to be processed by the software. The neural network then uses large amounts of data to learn which fragments of the molecule may be important to the function of the drug. It can then suggest new structures with similar fragments. Often, the neural network may falsely attribute some fragments as being important to the drug, or important fragments as insignificant. These mistakes are normal to training the NN and are corrected data supplying more diverse data to gradually improve the accuracy of the predictions.
Recently, researchers at AstraZeneca have been using ML to train a neural network how to predict the properties of a molecule and its possible interactions with its target protein by feeding it data from already tested molecules. Not only can it predict the performance of a given molecule, it can also construct new molecular structures with desirable properties.
In addition, the new CRISPR technology is being used in cancer research: it is used to delete selected parts of the genome to observe how the modification changes the cell’s resistance to cancer medication. Machine learning is used to gain as much information from the CRISPR experiments as possible.
Predicting toxicity of drug compounds using ML
Drug-induced liver damage is a frequent issue in many drugs and a common reason for drugs being withdrawn from the market. Toxicity complicates the drug discovery process by increasing the number of animals used for in vivo testing. Furthermore, animal testing often provides different results than clinical human tests. Sometimes, alternatives to the harmful drug do not exist and treatment of a disease comes with the cost of liver damage. With ever-increasing computing power, machine learning can transfer much of the testing to the computer as well as help discover less toxic drugs. Scientists at AstraZeneca have used so called Bayesian neural networks to calculate, using existing toxicity data from other drugs, the probability that a certain new drug will be toxic to the liver.
In conclusion
As the processing power of computers is increasing, novel techniques such as machine learning are used to make more and more accurate predictions in various fields of science, including drug research. ML offers the possibility to reduce the time and costs of drug discovery, create safer and more diverse drugs to treat diseases that are still considered untreatable. With time, various stages of drug research offer increasingly more data needed to train accurate neural networks, making machine learning a promising tool in future pharmaceutical industry.
If you are interested in the application of machine learning in drug discovery, check out this video:
By MU Chemistry mentor, Domantas
Domantas is reading Cambridge Natural Sciences, with a specialisation in Chemistry. Domantas has a strong interest in chemical research, and a record of research experience, with current chemistry research interests in supramolecular chemistry. This deals with how intermolecular interactions can be manipulated in order to build large multi-molecule structures that can be used for applications such as drug delivery (e.g. designing molecular cages that deliver a drug to a diseased tissue without affecting the rest of the body).
Would you like to learn more from Domantas?
Domantas mentors students in their application for Natural Sciences to Cambridge, and hosts one of our MU chemistry/ biology masterclasses on the use of nanoscale control of our cells to cure diseases, among many other exciting applications. Get in touch if you are interested to learn more!