Regular readers will know that I have no problem believing that AI (in its various forms) will definitely have an impact on drug discovery. And regular readers will also know that I’m quite skeptical that it’s going to have an immediate impact on the high-level functions of drug discovery (what target to go after, what molecules to make, which one should be the clinical candidate, etc.) Not everyone doing AI/machine-learning in the field is even talking about going after these, but those who are tend to go all the way. A few months ago, I wrote about a hype-fest presentation on this very thing.
One of the companies mentioned on it was BenevolentAI, from the UK. They themselves announced just the other day that they’ve raised another $115 million, in a funding round that values the company at about $2 billion, so I guess it’s time to have a look at them in particular. They say that “BenevolentAI’s advanced technology is disrupting the pharmaceutical industry by lowering costs, decreasing failure rates and increasing the speed at which medicines are delivered to patients“. I had not noticed this happening, personally, but hey, lowering costs and decreasing failure rates are things we need very much in this business. How is this being done?
The company’s advanced technology has been shown to outperform human scientists in understanding the cause of disease and is capable of quickly generating drug candidates at scale. The technology is also able to decipher the molecular process of disease and link these disease signatures within patients to ensure that the best drug candidate is given to the best patient responders – the ‘right drugs in the right patients’.
“Understanding the cause of disease”. “Quickly generating drug candidates at scale”. “Decipher the molecular process of disease”. Well, I can accuse these folks of several things, but not lack of ambition. Nor lack of self-promotion. Rather than rant on for a paragraph about each of those statements, I will just say that I am extremely skeptical that they can do any such things – at least not to the level needed to “disrupt the pharmaceutical industry”. I base that opinion on my own human experience over the last 28 years or so.
We have a great deal of difficulty understanding the cause of disease, just to pick that first claim. And I’m not sure at all if AI is ready to help across the board with that problem, because that would imply that we have enough data and we need help interpreting it. And while we could certainly use such help, I think that bigger problem is that we just don’t know enough about cells, about organisms, and about disease. AI is going to be very good at digging through what we’ve already found, and the hope is that it’ll tell us, from time to time, “Hey guys, you’re sitting on something big here but you just haven’t realized it yet”. But producing new knowledge is something else again. Drawing correlations and connections is not really the same thing – new knowledge, in this field, comes from experimentation. More advanced AI could point us to the more fruitful areas to run experiments in: “Hey, there’s too much about long noncoding RNA that we really don’t understand and we need to shore up” (the gospel truth, by the way). Really impressive AI might predict pathways or connections that haven’t been confirmed by experiment, but look as if they should exist. But AI as it stands is (at best) just going to sift through and rearrange knowledge that we already have, not give us more of it, and I just don’t think we have enough on hand to “decipher the molecular process of disease”.
Lest you think that I’m taking those quotes out of context, here’s some more:
To achieve this, BenevolentAI has created a bioscience machine brain, purpose-built to discover new medicines and cures for disease. Proprietary algorithms perform sophisticated reasoning on over 50 billion ingested and contextualised facts to extract knowledge and generate complex insights into the cause of diseases that have, until now, eluded human understanding.
“Bioscience machine brain”, forsooth. Who, exactly, “contextualised” those fifty billion facts for it to chomp on? Fifty billion seconds is nearly 1600 years, guys, and putting biomedical facts in context can, at times, take even more than a second per fact. But what the hey, I’m not a bioscience machine brain. I hope that the PowerPoint deck that convinced people to part with $115 million is written at a less eye-rolling level than this press release; a person could pull a muscle.
Nonetheless, I honestly wish these folks luck. This is the way to find out if this stuff works. They are, I should note, some of the people behind the most recent retrosynthesis software that I wrote about, and I thought that was pretty interesting. But that is a perfect example of taking existing knowledge and using algorithms to root through it and rearrange it in new and useful ways. Drug discovery in general could be improved by more of that – but I don’t think there’s room to improve it as much as the BenevolentAI people are claiming. I’ll give them credit for rolling up their sleeves and coming on down to see if it works. But I don’t think it’s going to go the way that press release says. Not yet.
Update: I see that I’m not alone! And for a much more grounded look at AI in drug discovery, see this recent C&E News story.