A Boston biotech where AI designs molecules, robots synthesize them, and patient tissue gets the final say.
Above: the wordmark of a company that thinks a chemistry set belongs next to a GPU cluster.
Walk into DeepCure and you will not find the usual biotech split - chemists on one floor, data scientists on another, politely ignoring each other. Here the machine that dreams up a molecule sits a short walk from the machine that builds it. In 2024 the company did the thing most AI-drug startups only promise: it picked an actual clinical candidate that its software designed, called DC-9476, and pointed it at autoimmune disease.
That is the company today - small, roughly 29 people, but holding two programs on a clinical track and a chemical search space large enough to embarrass a search engine. The pitch is unfashionably concrete in a field full of slideware: design it, make it, test it, repeat.
Drug discovery has a quietly absurd math problem. The number of small, drug-like molecules you could in theory build runs to something like 10^60. Chemists have made a rounding error of that. So the good molecule for a stubborn target is almost certainly out there - it has just never been imagined, let alone synthesized.
For decades the answer was to guess, make a few hundred compounds, test them, and guess again. Slow, expensive, and biased toward chemistry that was easy to make rather than chemistry that worked. DeepCure's founders saw the real chokepoint: not a shortage of ideas, but the painfully slow loop between designing a molecule and learning whether it does anything.
DeepCure was founded in 2018 by Kfir Schreiber, Thrasyvoulos Karydis, and Joseph Jacobson - researchers out of the MIT Media Lab, a place better known for inventing gadgets than for inventing drugs. Their bet was specific. If you ground a generative model in a database of compounds that can actually be made, and you wire it directly to automated chemistry that makes and tests them, the loop collapses from months to days.
Schreiber, the CEO, framed the company around a single discipline: don't let the AI imagine molecules no one can synthesize. So DeepCure built a proprietary library of up to 10^18 synthesizable, drug-like compounds and taught its models to design within reach. The ideas were free; making them real was the hard part, and that is where they spent the money.
The platform stacks three things that usually live in three different companies. Deep learning proposes molecules. Physics-based and quantum-mechanical simulation scores how they will behave. An in-house automated wet lab - robots, essentially - synthesizes the promising ones and runs them against the target. The output is not a prediction. It is a vial.
Deep learning proposes novel molecules from a database of up to 10^18 compounds that can actually be synthesized.
Quantum-mechanical simulation and molecular dynamics rank candidates before a single atom is built.
An in-house robotic wet lab makes the top molecules - no waiting weeks on an outside vendor.
Real assays, including ex vivo patient tissue, decide what advances. The data feeds back to the model.
Kfir Schreiber, Thrasyvoulos Karydis, and Joseph Jacobson spin out of MIT Media Lab research on AI for drug design.
Round led by Morningside Venture, bringing total raised to about $47M and funding the automated chemistry build-out.
Led by IAG Capital Partners to push the lead programs toward the clinic.
DC-9476, a selective BRD4 (BD2) inhibitor for autoimmune disease, selected. Leeds Institute collaboration announced.
In vivo and ex vivo results for DC-9476 presented at the EMBO Conference and ACR Convergence 2024.
DC-9476 advances; pipeline expands to a STAT6 program for asthma and atopic dermatitis.
A platform is only as good as the molecule it produces. At ACR Convergence 2024, DeepCure reported that DC-9476 outperformed three established classes of arthritis therapy in a collagen-induced arthritis mouse model. The chart below sketches the relative funding milestones that paid for getting there - useful context for a company that put its money into making compounds, not just predicting them.
DC-9476 targets the BD2 domain of BRD4 with the selectivity that older pan-BET inhibitors lacked, aimed at rheumatoid arthritis, Still's disease, and macrophage activation syndrome. Behind it sits DC-15442, an oral STAT6 inhibitor reported to fully suppress pSTAT6 in vivo with efficacy comparable to dupilumab in models - pointed at asthma and atopic dermatitis. To pressure-test the lead, DeepCure partnered with the Leeds Institute of Rheumatic and Musculoskeletal Medicine to study DC-9476 on real RA patient tissue.
Selective BRD4 (BD2) inhibitor. First AI-generated candidate, for autoimmune disease.
Oral STAT6 inhibitor for asthma and atopic dermatitis; dupilumab-comparable efficacy in models.
Collaboration testing DC-9476 on ex vivo rheumatoid arthritis patient tissue.
DeepCure's stated aim is plain: reimagine small-molecule therapies for immune diseases, with a focus on the people current medicine fails. Over a million rheumatoid arthritis patients respond poorly to today's standard of care. The company's wager is that targets long written off as undruggable were not impossible - just too hard to search by hand.
The interesting question in AI drug discovery is no longer whether a model can sketch a plausible molecule. It can. The question is whether anyone closes the loop all the way to a compound a patient could take. That is the bar DeepCure set for itself, and in 2024 it cleared the first rung by naming a candidate its own software designed.
So return to that room - the one where the dreaming machine sits beside the building machine. A few years ago it held an idea and a database. Now it holds DC-9476, a STAT6 program, patient-tissue data, and a chemical search space larger than most people can picture. The drug still has to survive the clinic, where AI gets no special treatment. But the gap between imagining a molecule and holding it has narrowed to something close to a walk across the lab.