The first entirely AI-developed drug candidate, DSP-1181, entered Phase I clinical trials in Japan earlier this year, having been discovered and readied in record time. The drug, a full agonist for the serotonin 5-HT1A receptor, was developed to treat obsessive compulsive disorder. Taking 12 months from discovery to optimisation, the AI used was coined Centaur ChemistTM. In other recent AI homeruns, a team from MIT and Harvard discovered a new antibiotic, halicin, using the technology, that can kill previously untreatable bacteria strains.
But first, lets take a step back. How can algorithms lead us to new drugs? As a top-level explanation, AI generates millions of potential novel molecules that fit the criteria needed by drug developers. Machine learning is then deployed to predict the compounds that will be active against target proteins. A third layer of algorithm called active learning is then applied to automatically prioritise the most appropriate compounds for researchers to experiment on.
But now let’s take a look at the five most significant ways in which AI can increase productivity in the drug development process:
1. Classification and sorting of cells using image analysis
Traditionally, high-throughput screening identifies compounds from huge molecular libraries. AI can be designed to quickly classify and sort various cell types that display the biological activity that is needed.
2. Prediction of 3-D structure of a target protein
The binding pocket of a target protein has to be well understood in order for the designed drug to bind in the correct orientation to exert its effects. Predicting the structure is challenging but AI tools such as AlphaFold have been used to predict the distances between pairs of amino acid residues.
3. Prediction of physical properties
Understanding the physical properties of molecules, such as their melting points and partition coefficients, is important to estimate their eventual metabolism in the body. Algorithms such as DeepTox can also predict the toxicological profile of a molecule, ensuring its safety for clinical trials.
4. Understanding a chemical synthesis pathway
Once the molecule has been decided, an optimal chemical synthesis pathway for the molecule must be discerned in order to search for a set of simpler, readily available precursor molecules. AI is able to perform this task in seconds.
5. Digitising chemical synthesis
AI can replace the role of the medicinal chemist by purifying and characterising the compound of interest. For example, the Chemputer platform is programmed for robotic synthesis of molecules.
These advanced algorithms significantly speed up the pre-clinical development, as digitalised candidate identification and optimisation can occur at a much faster rate than traditional methods. Evidently, accelerating candidate development should increase the overall productivity of pharma R&D, and allow for the savings to be reinvested into more diverse clinical trials. It’s a big ask, but many optimists believe that such technologies will allow biopharma to invest in more diverse drug pipelines and spur more competitive R&D strategies.
It has been estimated that an average drug discovery project costs approximately $3 billion and takes over a decade. Currently, only one out of ten drugs are approved after clinical trials due to efficacy and safety issues. An improvement in just 10% of the current accuracy of molecule prediction could save millions globally spent on drug development.
Image credit: Freepix