Nehaveigur

AI will soon design effective and safe drugs for any ailment. At least that’s commonly assumed, and on the surface it’s a reasonable prediction. After all, AI can already predict protein structures and sometimes even the effects of genetic variants.

Having worked in drug development for a decade and having done research in computational biology and human genetics for twice that, I think this optimism is misplaced. I believe that AI can shorten some aspects of drug development, but that it will not fundamentally transform drug discovery anytime soon.

Some of the most prominent people in my field would disagree with me. Aviv Regev heads Research at Genetech, which is the company where I started my career. She has bet many millions of dollars on AI and machine learning. Demis Hassabis is one of the most accomplished researchers and entrepreneurs in AI. He could have done anything, but he chose to start Isomorphic Labs, a company that uses AI for drug discovery. Both of them clearly believe that AI can cut the cost and time required for developing drugs, but as far as I know neither of them has been on the record making a prediction of what those savings will amount to. I don’t think that the likely savings justify the investment they’ve sunk into their companies.

My pessimism is founded on first principles. Many problems faced by drug discovery, and by biological science more generally, are computationally irreducible. This means that no amount of computation can replace real-world experiments. AI can help design and interpret those experiments, but since the experiments themselves, including clinical trials, are the most expensive part of drug discovery, the savings will be modest.

Another problem is the lack of training data. AI is able to predict protein structures because it was able to train on a high-quality training set, the Protein Data Bank. There is no comparable data for drug discovery as a whole. Dennis Hassabis dismisses this by claiming that “lack of training data” is an excuse, but he doesn’t provide any evidence for why he believes that.

In my own field of genomics, there is a huge amount of DNA sequencing data, as well as terabytes other “omics” data such as transcriptomics, proteomics, metabolomics and so on. This data can help solve individual problems in drug discovery, but it’s not in any meaningful way training data for developing drugs.

My prediction that AI will help on the margins, but that we will not see a massive increase in FDA approvals due to AI in the next ten years.