Start with a number: roughly 1060. That’s a conservative estimate of the number of drug-like molecules that could theoretically exist. The entire history of pharmaceutical chemistry has explored maybe 108 of them. We have been, by any honest accounting, searching a beach by examining individual grains of sand.
That situation is changing fast, and the system at the center of it isn’t doing what most people imagine when they picture “AI drug discovery.” It isn’t screening existing compounds or predicting which known molecules might bind to a known target. It’s doing something stranger and more ambitious: learning the deep grammar of chemistry itself, then using that grammar to compose molecules that have never existed before — designed from first principles to do a specific job inside a human body.
The technical approach that’s producing the most striking results right now is structure-based generative design, supercharged by what researchers call “multi-objective molecular optimization.” Systems like those coming out of Insilico Medicine, Recursion Pharmaceuticals, and several well-funded academic labs are simultaneously optimizing across binding affinity, ADMET properties (absorption, distribution, metabolism, excretion, toxicity), synthetic accessibility, and selectivity against off-target proteins — not sequentially, as chemists have historically done, but all at once, in a continuous latent space where molecular structure and biological function are jointly encoded.
The results are genuinely weird, in the best possible way. Molecules that emerge from these processes often look nothing like anything a medicinal chemist would have proposed. Bond arrangements that seem counterintuitive. Scaffolds outside the standard pharmacopoeia. Ring systems that aren’t in any textbook. And yet when they get synthesized and tested, they work — sometimes spectacularly. Insilico’s INS018_055, an AI-designed molecule targeting IPF (idiopathic pulmonary fibrosis), has made it into Phase II clinical trials, and it represents a genuinely novel chemical series that no human chemist had on their radar. That’s not a small thing. That’s the system finding real signal in a part of chemical space no one was looking at.
What’s accelerating all of this is the convergence with accurate protein structure prediction. Once AlphaFold made high-confidence structures available for essentially the entire human proteome, generative molecular design suddenly had a vastly richer target landscape. You can now design against proteins whose structures were unknown five years ago — including, critically, the “undruggable” ones. Transcription factors, intrinsically disordered proteins, protein-protein interaction surfaces. The targets that conventional small-molecule chemistry couldn’t touch are slowly becoming tractable.
The next frontier is even more audacious: co-design of the target and the molecule together. Rather than treating the protein structure as fixed, newer models are beginning to account for conformational flexibility and induced fit at design time — generating molecules that work with the protein’s natural dynamics rather than trying to catch it in a single snapshot. This is computationally brutal, and the models capable of doing it well are only now becoming mature enough to be useful, but the direction is clear.
There’s also a quiet revolution happening in synthesis planning. Designing a novel molecule is one thing; being able to actually make it is another. AI systems trained on vast reaction databases can now predict viable synthesis routes for generated structures in real time, feeding back into the generation process to bias it toward molecules that are chemically accessible. The loop is tightening: design, predict, synthesize, test, learn, redesign — running faster than any human team could manage alone.
The part that should make you lean forward is what this implies for the next decade. We are at the very beginning of a systematic exploration of chemical space, guided by AI that can hold more constraints in mind simultaneously than any human chemist, searching regions of that space that were previously invisible to us. The 1060 isn’t shrinking — but for the first time, we have a tool that can actually navigate it.