Tuesday, September 27, 2022

Why AI won’t revolutionize drug development

Jeffrey Funk - Gary N. Smith, No, AI probably won’t revolutionize drug development, Salon, September 24.

The opening paragraphs:

Drug development is expensive, time consuming, and risky. A typical new drug costs billions of dollars to develop and requires more than ten years of work — yet only about 0.02% of the drugs in development ever making it to market.

Some claim that AI, or artificial intelligence, will revolutionize drug development by ushering in much shorter development times and drastically lower costs. Many scientists and business consultants are especially optimistic about AI's ability to predict the shapes of nearly every known protein using DeepMind's AlphaFold, an artificial intelligence tool developed by Google parent company Alphabet; predicting this information with great detail would be key to quickly developing drugs. As one AI company boasts, "We … firmly believe that AI has the potential to transform the drug discovery process to achieve time and cost efficiencies."


The relative unimportance of AI in COVID development is consistent with the conclusion of many scientists that AI is not about to revolutionize drug development. The biggest problem is that clinical trials are the longest and typically most expensive part of the process, and AI cannot replace actual trials. Even AI's impact on drug discovery may be limited. A Science op-ed recently argued that, "[AI] doesn't make as much difference to drug discovery as many stories and press releases have had it…. Protein structure prediction is a hard problem, but even harder ones remain."

An often crucial part of drug development is to determine whether a drug binds to the candidate protein, something that MIT researchers have shown that AlphaFold cannot do and DeepMind admits AlphaFold can't do: "Predicting drug binding is probably one of the most difficult tasks in biology: these are many-atom interactions between complex molecules with many potential conformations, and the aim of docking is to pinpoint just one of them." [...]

One huge hurdle for all AI data mining algorithms is that the data deluge has made the number of promising-but-coincidental patterns waiting to be discovered far, far larger than the number of useful relationships — which means that the probability that a discovered pattern is truly useful is very close to zero.

There is more at the link.

No comments:

Post a Comment