AI protein folding, drug screening, and drug discovery

“Drug and vaccine discovery have historically been extremely time-consuming and costly. It took over a hundred years to develop and perfect a vaccine for meningitis. Pharmaceutical companies were able to move much faster in developing vaccines for COVID-19, spurred on by unprecedented spending (the U.S. government alone spent $10 billion just in 2020) to run multiple clinical trials and manufacturing efforts on parallel tracks. Had COVID-19 been as contagious or as lethal as the worst pandemics, however, even waiting a year for a vaccine would have been too long. So we need to continue to accelerate the speed of vaccine and drug development.

Drug discovery requires an understanding of how virus proteins, which are sequences of amino acids, fold into unique 3D shapes. Understanding these 3D structures is essential to understanding how viruses work, as well as how to fight them.

Today, it costs $1 billion and takes many years to get a successful drug or vaccine through the development process. I believe that AI will significantly accelerate drug discovery and reduce its cost, making available many more effective drugs at lower prices. This will help us live longer and healthier lives.

AI can greatly accelerate the speed and reduce the cost of drug and vaccine discovery. For determining protein folding, in 2020, DeepMind developed AlphaFold 2, which is AI’s greatest achievement for science to date. Proteins are the building blocks of life, yet one aspect of proteins that has remained a mystery is how a sequence of amino acids will fold into a 3D structure to carry out life’s tasks. This is a problem with profound scientific and medical implications and appears well-suited for deep learning.

DeepMind’s AlphaFold, trained on a large database of previously discovered 3D protein structures, has demonstrated that it is able to simulate the 3D structure of unseen proteins with similar accuracy to traditional techniques (such as cryo-electron microscopy, mentioned on page 156), which are expensive and can take years for each protein. For this reason, traditional methods have solved less than 0.1 percent of all proteins; thus, AlphaFold may offer a way to rapidly grow the number of solved proteins. AlphaFold has been heralded by the biology community as having solved a “fifty-year-old grand challenge in biology.”

Once the protein’s 3D structure is known, one expeditious way to discover effective treatment is repurposing, or trying every existing drug that has been proven safe for some other ailment, to see if one of them can fit into this 3D structure. Drug repurposing may be a quick fix that could stop the spread of a serious pandemic at its onset. Because established drugs have already been tested for adverse effects, they can be used without the extensive clinical trials required with new drugs.

Scientists can also work with AI symbiotically to invent new compounds. AI can be used to propose targets on which a treatment molecule would be attached. Then, given a target, AI models can narrow the search for a drug by identifying patterns within the data and proposing lead candidates. In 2021, Insilico Medicine announced the first AI-discovered drug for idiopathic pulmonary fibrosis, by first finding a target on the 3D structure and then proposing leads and selecting from among them the best biomolecule. Insilico’s AI saved 90 percent of the cost of these two steps in drug discovery.

Many types of knowledge can be used by AI to optimize drug discovery. For example, Natural Language Processing (NLP) can be used to mine an avalanche of academic papers, patents, and published data to extract new insights that can help propose targets or rank possible new molecules. And based on past outcomes of clinical trials, AI can predict the likelihood of each lead candidate, and rank them accordingly. These experiments are called “in silico,” as silicon-based software simulates the actual effect of the drugs and clinical trials. After in silico efforts produce high-confidence candidates, scientists can work from the AI-ranked list.

Besides the in silico approach, in vitro wet-lab experimentation, which involves testing the proposed drugs on human cells in petri dishes, can also expedite drug discovery. Nowadays these experiments could be conducted more efficiently by robotic machines than by lab technicians to generate massive data. A scientist can program these robots to iterate a series of experiments 24/7, without human intervention. This will accelerate the speed of drug discoveries greatly.”

Lee, Kai-Fu; Qiufan, Chen. AI 2041 (pp. 155-158). Crown. Edición de Kindle. #avltda01