Walk into any drug discovery conversation right now and you can feel it: AI is the wave breaking over the field. It's in every conference keynote, every board deck, every vendor pitch. And depending on who you ask, it's either about to cure everything or about to disappoint everyone.
We've been here before. If you've spent enough years in this industry, you've watched waves roll through — computational chemistry, combinatorial chemistry, high-throughput screening, the -omics. Each one arrived promising to upend the field. Each one settled, eventually, into something more useful and less dramatic: a good tool in the workshop. So it would be easy to be cynical, to say we've seen this movie. But that's only half right, and the other half is the half that matters.
Here's the thing about waves. Most of their motion happens far offshore and underneath, where you can't see it — slow, building, traveling enormous distances before they ever reach the beach. Most of the time they wear things down incrementally. But every once in a while, one gathers enough force to crash through everything at once. The skeptics are right that AI isn't magic. The enthusiasts are right that something real just hit shore. Both are true, because that is how waves work. The trick is to stand back far enough to see the whole ocean: to understand where this wave came from, so you can make a better guess about what the next one will do.
First, What Do We Even Mean by “AI”?
Part of the confusion is that “AI” has become a single word for a dozen different things. It helps to categorize them not by what's under the hood, but by what they're for. The same large language model can show up wearing three different hats. Ask it to answer a question and book a meeting, and it's an agent. Point it at ten thousand papers to pull out structured findings, and it's perceiving.
In our sorting by purpose, today's AI in drug discovery falls into roughly four buckets. Predictive models estimate a property or an outcome — will this compound bind, will it be toxic, what shape will this protein fold into. (Simulation lives here too; more on that shortly.) Generative models produce something new — a candidate molecule, a synthetic route, a paragraph of text. Agentic systems take actions and chain steps toward a goal, calling tools and other models as they go. And perceptive systems read unstructured input — images, spectra, free text, a pathology slide — and turn it into something structured. No taxonomy is perfectly clean, and you can argue the edges all day. But sorting by purpose keeps you honest about what a tool actually does, instead of being dazzled by what it's built from.
Where the Wave Came From
Take the predictive bucket. It feels brand new. It isn't.
In 1972, a medicinal chemist named John Topliss published something deceptively humble: a decision tree.1 Faced with the question every chemist faces — which analog do I make next? — Topliss drew a branching chart. Make the unsubstituted compound and the 4-chloro version; depending on which is more potent, go left or right; repeat. It was a way to navigate structure-activity relationships by hand, on paper, no computer required. He built it on the quantitative work Hansch and Fujita had published a decade earlier.2 It was, in the most literal sense, a model for prediction.
Now follow that branch forward. By the late 1990s, the same decision-tree logic had been automated and turned loose on real screening data — recursive partitioning, carving thousands of compounds into active and inactive. In 2003, a team at Merck published Random Forest for compound classification,3 and a random forest is, precisely, an ensemble of decision trees voting together. Then in 2015, much of the same group showed us that deep neural networks could beat the forests.4 Topliss's hand-drawn chart had become a forest had become a deep network. The object kept evolving; the lineage never broke.
That's the part people miss when they treat today's models as if they sprang fully formed from nowhere. They didn't. They are the latest swell of a wave that has been traveling since at least 1972.
Prediction and Simulation Are the Same Wave
The same story plays out with protein structure — and here the metaphor gets literal, because the swell was building underwater for half a century.
The premise goes back to Christian Anfinsen, who showed in the early 1970s that a protein's sequence determines its folded shape: the structure settles wherever the energy is lowest.5 For decades, that pointed to an obvious path — if you could just compute that energy landscape, you could simulate folding directly. Physics-based methods chased exactly this, with mixed success, because the search space is astronomical.
Meanwhile, a second idea was building offshore: amino acids that touch in the folded structure tend to mutate together across evolution.6 Line up enough related sequences and you can read the contacts straight out of the correlations. This approach predates the deep-learning era — but it was quietly accumulating power.
Then the wave broke. AlphaFold fused the evolutionary signal with deep learning, and at the CASP competitions of 2018 and 2020 it cracked a problem the field had wrestled with for fifty years.7 And here's the punchline for anyone keeping score on “prediction versus simulation”: the analysis suggests AlphaFold didn't merely pattern-match the alignments — it appears to have learned the energy function the physics simulators were trying to compute. The latest version is now described explicitly as a move from prediction to simulation.8 The two ideas were never rivals. They were the same wave, arriving from different directions.
What Comes Next Isn't a Smarter Model
So where is the next swell?
Our bet is that it isn't a bigger algorithm. Look at those four buckets again. We've spent fifteen years sharpening each one in isolation — better predictors, better generators, more capable agents. They are remarkably good on their own. But a discovery program doesn't need four brilliant tools sitting in four corners of the room. It needs them working together: a perceptive model reading the assay data that feeds a predictive model that prioritizes what a generative model designs and an agent then helps you synthesize and test. That coordination is the hard part — and it doesn't run on cleverness. It runs on two unglamorous things.
The first is the organization of data and information. You cannot orchestrate anything on top of fragmented, undocumented, inconsistent data. AlphaFold worked in large part because the Protein Data Bank existed — roughly 170,000 carefully curated structures, the product of decades of patient organization. Every team dreaming of AI-driven discovery is, whether they say so or not, facing a data infrastructure problem first.
The second is the orchestration of workflow and decisions — getting these tools to hand off data and options to one another reliably, in the right order, with people making the decisions and driving the process, not copying from spreadsheets and pasting into request tools. Neither of these makes a flashy demo. Neither will headline a keynote. But they are the unglamorous groundwork that lets the tools actually work together, and they are where the next wave is gathering. They are also, as it happens, where we're spending our time. More on that soon.
The Bottom Line
The AI wave is real, and it is breaking now. But it didn't start last year, and it won't end with a bigger model. It started with a chemist drawing a decision tree on paper, and it continues with the patient work of organizing data and orchestrating decisions. Stand back far enough, and you can see the whole ocean.
