Five ways in which artificial intelligence is accelerating the discovery of new medicines
Antimicrobial resistance to existing antibiotics is a growing crisis. Not only are new antibiotics exceptionally costly to develop, but they are usually limited to a narrow spectrum of chemical diversity. In January, the Director-General of the World Health Organisation declared:
“Never has the threat of antimicrobial resistance been more immediate and the need for solutions more urgent.”
The process of bringing new medicines from the bench to the bedside, known as drug discovery, is expensive, difficult and sometimes inefficient. It has been estimated that the average drug discovery project takes over a decade and costs around $3 billion. This is not just the case for new antibiotics but also the search for other new drugs, such as an effective antiviral treatment for SARS-CoV-2, the virus that causes COVID-19.
But, earlier this year, a team from MIT and Harvard made headlines when they announced that they had discovered a powerful new antibiotic using artificial intelligence (AI). Deep neural networks had identified an existing drug, halicin, that was structurally divergent from existing antibiotics and exhibited potent bactericidal activity against a wide spectrum of pathogens.
Today, AI is a ubiquitous feature of modern life and is responsible for the “smart” technology behind email spam filters, social media feeds and plagiarism checkers. More advanced forms of AI, such as machine learning and deep learning, solve complex problems by breaking them down into layers of data, similar to the neurons in our brains.
Consequently, it is hoped that integrating AI methods with traditional drug discovery will improve time- and cost-efficiency, accelerating the transition of new drugs from the laboratory to the shelf. Here are five exciting developments at the interface of medicine and machine.
Classifying and sorting cells by image analysis
A drug discovery project usually begins by identifying a disease area and a target related to that disease, such as a metabolic pathway or a protein. In the case of an infectious disease, such as malaria or tuberculosis, it is important that the pathway or protein is unique to the disease-causing organism to minimise potential interactions with the cells of the host (the patient).
Typically, suitable compounds are identified from vast molecular libraries using methods such as high-throughput screening. Visual inspection of these data is a tedious task and quickly becomes inefficient for the analysis of big data sets. Fortunately, AI is an excellent tool for recognising images and can be trained to rapidly classify and even sort various cell types.
These algorithms are also finding application in the interpretation of screening mammograms for breast cancer, in which an AI system successfully reduced the number of false positive and false negative cases and, in an independent study, outperformed six human radiologists.
Predicting the three-dimensional structure of a target protein
Understanding the target protein is a crucial piece of the drug discovery puzzle. If the binding pocket of a target protein is well understood, new molecules can then be designed to fit into this pocket, analogous to the way in which a key (drug) precisely fits into a lock (protein).
Predicting the three-dimensional structure of a protein is notoriously challenging but, with the development of AI-based tools, these calculations have become more accurate and sophisticated. For example, a tool called AlphaFold relies on deep neutral networks trained to predict the distances between pairs of amino acid residues. In a blind competition, AlphaFold was able to predict the three-dimensional structure of a target protein 24 of 43 times. This was significantly better than the runner-up software, which correctly predicted only 14 out of 43 test sequences.
Predicting physical properties
Some molecules are more “drug-like” than others, which means that they are more likely to be absorbed by tissues or are less likely to be metabolised by liver enzymes. These parameters are often linked to the physical properties of the molecule, such as its melting point or its partition coefficient. Computers are getting better at predicting these physical properties, saving precious time in the laboratory.
The toxicological profile of a compound is also an important parameter. The DeepTox algorithm, which is based on deep learning methods, gave outstanding results in the Tox21 data challenge in which the participating groups attempted to predict the toxic effects of 12 000 environmental chemicals and drugs. Once a deep neural network has “learned” to detect toxic features of molecules, it can be used to assess the toxicity of new compounds at an early stage in the drug discovery pipeline.
Planning a chemical synthesis pathway
Once it has been decided that a molecule holds promise for development into a potential medicine, the search for an optimal chemical synthesis pathway begins. A technique called retrosynthetic analysis recursively searches for “backward” reaction pathways until a set of simpler, readily-available precursor molecules is obtained. AI performs this task much more efficiently than its human counterparts, typically taking a matter of seconds.
Digitising chemical synthesis
Finally, to make the molecule of interest, a medicinal chemist will spend most of her time in the laboratory manipulating flasks and funnels to react, purify and characterise compounds. Recently, the Chemputer platform has been developed to codify standard “recipes” for robotic synthesis. The system has been validated by synthesising several pharmaceutical compounds, without any human intervention, with comparable (or increased) product yields than those achieved manually.
Preliminary efforts to accelerate drug discovery by incorporating AI, such as the discovery of halicin above, are promising. Last year, Takeda Pharmaceuticals and “digital biology” company Recursion announced that they had identified potential drug candidates for more than 60 unique indications in just one-and-a-half years — much faster than the traditional drug discovery pipeline of approximately a decade.
The prospect of accelerated drug discovery is enticing, but what does it mean for traditional research scientists? Will they eventually be replaced by algorithms and digital platforms such as the Chemputer? Perhaps this is the time for laboratories and organisations to encourage the evolution of a new type of research scientist — a scientist for the 21st century, who is equally skilled at handling pipettes and harnessing technology to make useful predictions.