Drug development is expensive, time-consuming, and risky. Typical new property costs Billions of dollars to develop and require more than ten years From work – only so far About 0.02% of the drugs under development reach the market.
Some claim that artificial intelligence, or AI, will revolutionize drug development by ushering in much shorter development times and drastically reducing costs. Many scholars and business consultants are in particular Optimistic about AI being able to predict the shapes of nearly every known protein Using AlphaFold from DeepMind, an artificial intelligence tool developed by Google subsidiary Alphabet; Predicting this information in great detail would be a quick key to drug development. As one artificial intelligence company brag“We… firmly believe that AI has the potential to transform the drug discovery process for time and cost efficiencies.”
This type of speech may be dismissed as a fake mock-up until you make it. but the Washington PostPresumably a neutral observer, he also promoted this novel. “New research on proteins promises breakthroughs in drugs and much more” Said one of the opinion articles they published. The The New York Times He agreed in a similar article titled “We need to talk about how good AI is.” One thing is definitely true: “Investors are betting on AI start-ups to boost drug development,” such as financial times mentioned.
Moderna claims that AI computer simulations have helped analyze genetic sequencing data in an “optimal way”. It doesn’t matter that anything “optimal” is rarely known. In addition, most of the time and effort in drug development is not spent in computer simulation, but in Clinical trials – a Uncertain and risky process It is not made faster or less expensive by artificial intelligence algorithms. a temper nature An article titled “The Rapid Search for COVID Vaccines – And What This Means for Other Diseases” didn’t even mention artificial intelligence among the reasons for the rapid development of COVID-19 vaccines. Instead, he asserts that,
The world was able to develop COVID-19 vaccines very quickly due to years of previous research on related viruses and faster ways to manufacture vaccines, massive funding that allowed companies to run multiple trials in parallel, and regulators moving faster than usual. Some of these factors may translate into other vaccine efforts, particularly faster manufacturing platforms.
Determined to get the COVID-19 vaccine to the public before the November 3, 2020 presidential election, the United States government has committed $14 billion to support the vaccine efforts of pharmaceutical companies. The government agreed to pay Pfizer $5.87 billion for 300 million doses if Pfizer develops an FDA-approved vaccine — regardless of whether the vaccine is still needed. Moderna $954 million was awarded for research and development and $4.94 billion in federal guaranteed purchases of 300 million doses. Johnson & Johnson He was awarded $456 million for research and development and promised $1 billion for 100 million doses.
Want more health and science stories in your inbox? Subscribe to the salon’s weekly newsletter vulgar world.
The relative irrelevance of AI in the development of COVID 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 most expensive part of the process, and AI cannot replace actual trials. Even the impact of AI on drug discovery may be limited. a Science Magazine He recently argued that “[AI] It doesn’t make as much of a difference in drug discovery as it has in many stories and press releases…. Predicting protein structure is a challenging problem, but even more difficult problems remain.”
The deluge of data has made the number of promising patterns – but chance waiting to be discovered – far greater than the number of useful relationships – which means that the probability that a pattern discovered is truly useful is very close to zero.
Often an important part of drug development is determining if the drug binds to a candidate protein, something MIT researchers show that AlphaFold can’t do it and DeepMind admits AlphaFold can’t: “Predicting drug binding is probably one of the most difficult tasks in biology: these are polyatomic interactions between complex molecules with many potential conformations, and the goal of docking is to identify just one of them.”
Similarly, the CEO of the pharmaceutical company Verseon very skeptical:
People say, “Artificial intelligence will solve everything.” They give you great words. “We will absorb all this longitudinal data and do latitude analysis.” It’s all rubbish. It’s just hype.
One big hurdle for all AI data mining algorithms is that the deluge of data has made the number of promising patterns – but chance waiting to be discovered – far greater than the number of useful relationships – which means that the probability of a detected pattern that is really useful is very close to zero. Verseon’s CEO says the total number of potential chemical compounds in the universe is on the order of 10 to the 33rd. Companies can’t conduct clinical trials on every possible compound, nor can they rely on artificial intelligence to find the needles in this massive haystack.
It will take real intelligence – not artificial intelligence – to determine which compounds are most likely to generate gains that justify the enormous costs of testing. Likewise, it would take real intelligence – not artificial intelligence – to conduct the clinical trials necessary to measure efficacy and potential side effects.
on artificial intelligence