Using AI in drug development | Explained

Developing drugs is an expensive and time-consuming process. However, the advent of artificial intelligence (AI) has opened up a world of possibilities when it comes to speeding up drug development.

How does the process start?

The drug development process begins with target identification and validation. A target is a biological molecule (usually a gene or protein) that the drug binds directly to in order to function. The vast majority of targets are proteins. Only proteins with ideal sites where drugs can move and dock to achieve their goals are druggable proteins.

Target proteins are identified during the discovery phase, and the sequence of the target protein is entered into a computer, which searches for the best-matching drug from a library of millions of small molecules whose structures are stored in the computer. This process assumes that the target protein and drug structures are known. If not, the computer uses the model to understand which sites the drug can bind to. This discovery process avoids time-consuming laboratory experiments that require expensive chemicals and reagents and have high failure rates. Once a suitable protein target and its drug are identified, research advances to the preclinical stage, where potential drug candidates are tested for drug safety and toxicity outside of living systems using cells and animals. . After this, as part of the clinical phase, the drug is tested in a small number of human patients before being used in a larger number of patients to confirm efficacy and safety. Finally, the drug will go through regulatory approval, marketing, and post-market surveillance stages. Due to high failure rates, the discovery stage limits the number of drugs that move into the preclinical and clinical stages.

How can AI aid this process?

AI has the potential to revolutionize target discovery and understanding drug-target interactions by significantly reducing time, increasing the accuracy of predicting interactions between drugs and their targets, and saving costs. is hidden. The development of two AI-based predictive tools, AlphaFold and RoseTTAFold, developed by Google company DeepMind and researchers at the University of Washington, USA, respectively, has led to major scientific advances in the field of computing over the past four years. drug development. Both tools are based on deep neural networks. This tool's neural network uses a large amount of input data to produce the desired output, namely his three-dimensional structure of the protein. The recently released new avatars of AlphaFold and RoseTTAFold, called AlphaFold 3 (co-developed by DeepMind spin-off Isomorphic Labs) and RoseTTAFold All-Atom, respectively, take the functionality of these tools to a whole new level. The key difference between the upgraded version and the previous version is that it can predict not only the static structure of proteins and protein-protein interactions, but also the structure and interactions (including modifications) of any combination of proteins, DNA, and RNA. small molecules and ions. In addition, the new version uses a generative-diffusion-based architecture (a type of AI model) to predict structural complexes. In a test with 400 interactions between a target and its small molecule drug, AlphaFold 3 accurately predicted interactions 76% of the time, compared to 40% for RoseTTAFold All-Atom.

What are the disadvantages?

Despite all the promise and potential in drug development, AI tools have limitations. For example, this tool can provide up to 80% accuracy in predicting interactions (with significantly lower accuracy in predicting protein-RNA interactions). Second, this tool can only support a single step of drug development, target discovery, and drug-target interaction. They will still need to go through preclinical and clinical development stages, and there is no guarantee that AI-derived molecules will be successful at these stages. Third, one of the challenges of diffusion-based architectures is model illusion. In this case, the tool produces inaccurate or non-existent predictions due to insufficient training data. Finally, unlike previous versions of AlphaFold, DeepMind has not released the code for his AlphaFold 3, limiting its independent validation, widespread availability, and use for protein and small molecule interaction studies. doing.

What about India?

Developing new AI tools for drug development requires large-scale computing infrastructure, especially one with high-speed graphics processing units (GPUs) to perform multiple tasks in long sequences. GPU chips are expensive and have a short shelf life as hardware manufacturers produce newer, faster chips each year. India needs such massive computing infrastructure. This, in addition to the lack of skilled AI scientists unlike in the US and China, has meant that Indian researchers have not been ahead of the curve in developing AI tools for drug development, despite India's rich history with proteins. This is the second reason why profits could not be established. X-ray crystallography, modeling, and other areas of structural biology. However, with a growing number of pharmaceutical companies, India could take the lead in applying AI tools in target discovery, identification, and drug testing.

Binay Panda is a professor at JNU, New Delhi and posts at @ganitlabs.