Drug development is one of the slowest feedback loops cross-industry. The entire process (preclinical research, clinical trials, regulatory review) moves at the speed of biology, not technology.
The preclinical phase alone eats roughly 3 to 6 years, with some therapeutic areas requiring massive amounts of time for less than a 15% probability of success. In neuro, getting a compound across the blood-brain barrier is its own multi-year problem before you even know if the drug works. In every case, scientists are making the same sequence of high-stakes decisions: choose a disease target, decide whether to pursue it with a small molecule, an antibody, a biologic, a gene therapy, and then spend years validating that the modality they chose actually works against the target they picked. It’s an enormously expensive educated guess.
Chai Discovery’s latest model, Chai-3, produces antibodies that bind their targets with therapeutic affinity in roughly half of cases. What used to take months of iterative design now happens in weeks. Today, a biotech’s value proposition is straightforward: develop deep expertise in a therapeutic area, show proof of concept in a patient population, get acquired by pharma. But if AI reaches the point where you can, effectively, one-shot a molecule or antibody design, which Chai’s results suggest is not far off, the need for that expertise quickly lowers. It makes me wonder: will biotech companies as we know them still operate in the next two to three years?
The clinical phase has its own bottlenecks. Varying endpoints by therapeutic area make it difficult to determine clear timelines. Trial enrollment remains painfully slow. This is where tools like Claude Science are compressing the work. Anthropic launched their “workbench for biologists” this week, with skills to draft clinical trial protocols, convert instrument data to standardized formats, and automate time-consuming wet lab processes. I don’t think there’s a world in which AI will necessarily replace scientific discovery, but it will accelerate it.
There are a lot of moving pieces and areas of discovery, as noted by founders and members from research labs. Chai Discovery has proven that AI can model antibody design, but antibodies are one modality among many; small molecules, biologics, and gene therapies each carry their own layer of complexity. And Claude Science, for all it does on the computational side, doesn’t touch the hardware layer. The ambition, eventually, is to connect the model directly to physical lab instrumentation so it can design, execute, and interpret experiments end-to-end.
But maybe the right framing isn’t what’s solved and what isn’t. It’s what this trajectory implies. If a model can one-shot an antibody today, what does the preclinical phase look like in three years? If AI can draft a trial protocol in hours, how long before it’s designing the endpoints themselves? And if the entire discovery-to-clinic pipeline compresses from decades to years, do we actually have the regulatory infrastructure to keep up, or does the bottleneck just migrate from the lab to the FDA’s desk?