Article: AI Can't Crack Drug Discovery (Yet)

Photo Source: Unsplash

Artificial Intelligence has accelerated the time it takes to get new drugs from the laboratory to the clinic. Still, recent AI-generated drugs have faltered in clinical research, underscoring drug discovery's complex challenges. AI has permeated every aspect of the industry to speed up and automate manufacturing, decision making, and information-processing processes, including research and development. The pharmaceutical industry's quest to discover new drugs to treat and cure diseases afflicting patients faces a growing problem--complexity.

The complexity of drug discovery is reflected in the staggering amount of time and money it costs to bring new drugs to market. It takes, on average, 12-15 years and costs $1-2 billion to bring a new drug to market today. (nih.gov). Drug discovery and development must start with the identification of a target. A target in drug discovery represents an aspect of the biology of a disease that a drug can change for a therapeutic effect. For example, the process of making firm skin for a bacterium is the target of the famous antibiotic penicillin. Penicillin prevents the production of a stable cell wall by blocking the production of molecules called peptidoglycans. Without peptidoglycans the bacterium's cell wall will collapse and kill the bacterium when it tries to reproduce. Target, as the name implies, gives researchers something to aim for.

After finding a target, researchers need to invent a molecule that can affect the target while remaining safe for the patient to take. This medicinal chemistry process involves great creativity and takes years of trial and error. In medicinal chemistry, researchers in academia and industry take three to four years to find a suitable molecule that performs as expected and has a safety profile in animals that suggests it is safe for testing in humans. This process has captured the imagination of AI researchers who have applied techniques to virtually design and test molecules against targets. The technique called generative adversarial neural networks, or GANs, can run trial and error experiments thousands of times faster than a chemist in the laboratory. GANs work with two AI models battling each other to find new molecules. One model called the generator invents new molecules and sends them to the discriminator model. Based on the properties of good, safe molecules, the discriminator model then decides if the proposed molecule could work as a drug. Startup companies such as Exscientia, BenevolentAI, and In Silico Medicine have demonstrated a sharp acceleration in time from target to molecules ready for the clinic by shaving the time in half or better compared to traditional methods. In Silico claims that their clinical-level drug for idiopathic pulmonary fibrosis took only 18 months from target to molecule. (medcitynews.com) 

The rapid increase in time from target to molecule looks promising, but the past few months have seen several AI-generated compounds fail in human clinical trials. Sumitomo Chemical's schizophrenia drug failed in late-stage clinical trials, as well as a drug from BenevolentAI's drug for inflammation. Exscientia has discontinued its lead drug for cancer to divert resources to other, more promising candidate molecules. So far, no AI-generated medicines have been approved, but that does not mean the AI approach will not work. AI-driven drug discovery helps speed up the process of molecule generation. However, the complexity of drug discovery and the risk of failure in human clinical trials has not been overcome by AI. Hopefully, better models and more data will open the door to faster and cheaper therapies for patients, but the discriminator in GANs learns from known molecules and their safety profiles. From this property, GANs work in the halo of known molecules, but the next generation of great drugs may be outside of the known space and, therefore, out of reach of GANs at all. 

Dr. Smith’s career in scientific and information research spans the areas of bioinformatics, artificial intelligence, toxicology, and chemistry. He has published a number of peer-reviewed scientific papers. He has worked over the past seventeen years developing advanced analytics, machine learning, and knowledge management tools to enable research and support high-level decision making. Tim completed his Ph.D. in Toxicology at Cornell University and a Bachelor of Science in chemistry from the University of Washington.

You can buy his book on Amazon in paperback and in kindle format here.