"First AI-created drug" is ambiguous, meaningless, and dangerous

Despite this headline, I'm an optimist. Artificial intelligence will transform drug discovery. Indeed, it is already.

I live it every day. At BenchSci we use AI to quantifiably cut the cost and time of drug discovery by helping scientists run more successful experiments. And we're not alone. I track relevant startups, pharma companies, trends, and drugs. Many great people and companies are driving productivity gains in life science R&D with machine learning.

It's for these people and companies that I'm compelled to write this post. Because claims of "first AI-created drug," which have come up a lot lately, threaten the field. Allow me to elaborate:

Ambiguous

"First AI-created drug" is ambiguous. And it's ambiguous in many ways.

For starters, what do we mean by "first?" I've found more than 100 drugs that were, according to their developers, created with AI. And these are at each stage of the pipeline, from preclinical through phase 3. So I don't see how anyone can claim a "first." Not a first compound, nor a phase 1, phase 2, or phase 3.

Now, someone might say, "none of those used real AI." And by this they might mean that they didn't use some form generative AI. Data in, drug out.

But even then, "AI-created is ambiguous." Is the only criteria that it's generative? Shouldn't novelty matter? For example, if a machine generates a drug that's 5% different from an approved compound, do we care? Even if it moves to a phase 1 trial? And what if a machine generated a compound and then a human optimized it? Or vice-versa? How much AI is enough to make it a "first?"

Bottom line: Who gets to define what we mean by "AI" in "AI-created?" We have no standard for what that even means. So everyone interprets it in different ways. And they can then make any claim they want by choosing the interpretation that suits them best.

But that's not the only problem.

Meaningless

"First AI-created drug" is also meaningless. This is true even if the claim seems big, like getting the first AI-created drug into a human. This is because the drug might not be important, novel, or more likely to be safe and efficacious than average.

What if, for example, I used AI to generate a me-too drug for a well-treated condition? And what if I did this by iterating on an approved compound only enough to make it patentable? And what if I did all this and it still wasn't more likely to make it through phase 1, 2, and 3 to approval? Do we even care?

For this reason, the simple phrase "first AI-created" drug is meaningless.

Even worse than that, it's dangerous.

Dangerous

"First AI-created drug" is dangerous. Yes, dangerous.

For starters, the excessive hype and inevitable letdown will create skepticism. Skepticism in turn will dampen interest and investment. This will slow real progress.

But the other reason it's dangerous is that it's a distraction. What we should focus on is importance, novelty, and productivity. Importance means under-treated conditions. Novelty means new treatment approaches (which are increasingly biologic, genetic, and sometimes curative) with significantly differentiated efficacy, safety, or both. Productivity means faster progression through preclinical and clinical to approval at lower cost.

To do that, we can't rely on generative AI alone. At least not yet, because we don't have enough knowledge of biology in machine readable form for a machine to generate an important and novel compound with a significantly greater likelihood of progressing to approval at a maximally low cost. (We probably also don't yet have enough computing power to build models from all of the relevant features. Quantum computing might help here, but we don't know yet.) There will still be lots of work along the way. AI can help a lot in this work, but not if everyone's focused only on being "first."

So let's shift the conversation. Let's focus on using AI it to develop important, novel treatments with greater productivity. There are many under-treated conditions, many promising new therapeutic approaches, and many bottlenecks in pharma R&D. AI can help with all of this. That should be the goal.