The New Race to Build AI-Designed Medicine

It’s good for us that Gertrude Elion didn’t wait around for luck.

She worked in an era when drug discovery still had one foot in alchemy, and then decided to help drag it toward method. 

Along with George Hitchings, Gertrude leaned on biochemistry and logic to design compounds with a purpose, instead of throwing chemicals at a wall and praying something stuck. It was that mindset that produced real medicines, and quietly rewired how we approached the drug development pipeline. 

You see, the pharmaceutical industry has been inching, decade by decade, away from that “throw everything against a wall and see what sticks” and toward “let’s start testing smarter.” 

The only problem is that biology got a lot bigger and noisier than people could’ve imagined — more than any human team can fully digest.

So now the method needs a machine.

What’s interesting today is that companies aren’t simply discovering AI; they’re starting to build around it the way they once built around high-throughput screening, sequencing, and cloud compute. 

In other words, Big Pharma isn’t treating AI drug development as a novelty but more as an infrastructure.

That’s where this story starts.

The AI Factory Era Begins

Recently, we talked a little about Big Pharma’s iPhone moment, when Eli Lilly flipped the switch on something the industry has been hinting at for years — a true in-house AI factory built for drug discovery and genomics.

When a top-tier drugmaker installs this much compute on campus, it’s a clear signal that the bottleneck has shifted. 

The constraint isn’t just wet-lab capacity anymore, either. It’s how quickly you can generate and evaluate hypotheses across messy biological systems, then decide which experiments deserve real time, real samples, and real money.

Let’s not mince words here…

Drug discovery is an elimination contest. And the drug development pipeline is a long hallway of dead ends that burns billions of dollars as you walk it. 

However, AI is turning “design” into a repetitive loop instead of a slow relay race.

First, you start with a target and a hypothesis.

Models help propose molecules, predict binding modes, and flag toxicity or metabolic issues sooner than traditional cycles can. Then, automated labs generate data that flows back into the model, tightening the loop. 

Suddenly, a win doesn’t come from one miracle prediction, and scientists are able to crank them out like clockwork. 

There’s a few advantages here to keep in mind, too:

• Faster target triage before the chemistry starts.

• Better hit-to-lead selection, so fewer weak programs enter the expensive part of the funnel.

• Smarter optimization, where you don’t “improve potency” by accidentally wrecking solubility or safety.

• More informed go/no-go calls earlier, when the check is still small.

The ability to churn out new drug candidates is why the deal flow has gotten more serious lately.

You might recall last month when Takeda struck a multi-year collaboration with Iambic Therapeutics, valued at more than $1.7 billion in potential milestones with royalties. The focus will be on AI-enabled small-molecule discovery across oncology, gastrointestinal, and inflammatory disease areas. 

The interesting part is what the partners emphasize. You see, it wasn’t just about speed, but also better molecular design through improved prediction of how drugs bind to proteins. This is the central heartbreak of small-molecule discovery when you get it wrong. 

Meanwhile, Big Pharma isn’t treating AI like a single vendor relationship anymore.

AstraZeneca agreed to acquire Modella AI to strengthen oncology research, with the stated goal of deploying foundation models and agents into areas like biomarker discovery and clinical development. 

For the record, that’s a different kind of bet which shows us that AI is becoming an internal capability and not just rented software. 

Another sign of “platformization” showed up in January, when a collaboration began to make Lilly’s TuneLab available through Schrödinger’s LiveDesign environment, which already sits inside many discovery workflows. 

We’re still in the first inning, folks. 

Generate:Biomedicines has moved toward an IPO pitch centered on AI-designed protein therapeutics, backed by big-name partners and a narrative that computational design isn’t just for small molecules anymore. 

As I mentioned last week, even Sam Altman is looking to enter the fray after announcing he’s ready to dive into AI models for drug discovery.

So what changed?

Well, a few things, actually. 

For starters, compute got cheaper per unit of insight, models got better at representing biology, and the economics of the pipeline got tighter. 

Patent cliffs are a very real phenomenon, and the timelines can be brutal. 

Of course, the old approach of having to run thousands of experiments to learn what the first ten could have told you is getting harder and harder to justify to shareholders. 

The industry isn’t replacing chemists or biologists.

It’s trying to give them a smarter first draft, and a faster feedback loop, so the lab spends more time confirming good ideas and less time disproving bad ones.

AI Drug Discovery: The Quiet Edge

There’s a reason the smartest teams are building “platform” capability instead of chasing one-off wins.

A single drug can pay for decades of research, but one drug is also a fragile story.

The platform itself is the repeatable mechanism. 

Think about it…

It’s the part of this system you can run again tomorrow and go after an entirely different target, a different disease, or a different dataset.

Naturally, the best platforms are starting to look less like “AI companies” and more like compact discovery shops with a computational advantage baked into every decision. They don’t just predict and blindly throw ideas at the wall — they repeat, learn from their mistakes, and then tighten the loop.

A simple way to judge the next wave of AI drug discovery is to ignore the slogans and look for three things — a proprietary data flywheel, a lab that can validate quickly, and a pipeline that shows the engine is actually producing candidates, not just graphics.

You want to find the players already pushing fresh drug candidates through the funnel.

We’ve moved beyond searching for a concept, dear reader. What we want is a working system.

Perhaps it’s time you check one of them right now.