DropXCell

Designing an antibody that actually binds its target takes months of lab work, and most candidates fail before they ever reach a patient.

DropXCell replaces months of manual experimentation with an agentic mutation loop that explores the molecular space in simulation and hands back 50 to 100 lab-ready antibody candidates, optimized for binding, variety, and manufacturability at once.

The situation

Antibodies are the workhorses of modern medicine, proteins engineered to latch onto a specific disease target. But designing one that actually binds, and holds up as a real drug, is brutally slow. A team picks a target, proposes candidates, and tests them at the bench, a cycle that runs for months and in which the overwhelming majority of candidates fail. The reason is scale. The number of possible antibody designs is astronomically large, far more than any lab can synthesize and screen by hand. So teams explore a tiny corner of the space, guided by intuition and luck, and hope a viable candidate is hiding in the small sample they can afford to test.

Why it was hard

The tempting shortcut is to let a model generate high-binding candidates and call it done. Binding is not the whole problem. A candidate has to be good on three fronts at once, and they fight each other. It needs high affinity, gripping the target tightly. It needs to be developable, stable and manufacturable enough to become an actual drug rather than just a winner on screen. And the pool needs diversity, because betting on one narrow design is how programs fail when that design stumbles later. Optimizing one dimension usually costs another. On top of that, an in-silico prediction is worthless if it does not survive contact with the bench, so the search has to be grounded in real biophysics, not just plausible-looking sequences. Exploring an astronomically large space against three competing objectives, and being right often enough to justify lab time, is the real problem.

The approach

DropXCell replaces the manual design cycle with an agentic mutation loop that does the exploration in simulation. Given a protein target, the system first identifies the binding pockets, the specific sites on the target where a candidate has to attach. From there, agents propose antibody candidates and iteratively mutate them, and system scores and steers each round across all three dimensions at once, affinity, developability, and diversity, so the pool improves without collapsing onto a single fragile design. The loop runs until it converges on a pool of 50 to 100 high-confidence candidates, ranked and ready for the lab to synthesize.

What happened

The shift is from months of manual trial and error to a curated shortlist in a fraction of the time. Instead of testing a handful of hand-picked designs and hoping, a team gets 50 to 100 candidates already optimized for binding, variety, and manufacturability, so the expensive lab step starts from a far stronger position and far fewer dead ends. Because the loop balances all three objectives together, the candidates that reach synthesis are the ones actually worth making. DropXCell is being built and validated in drug discovery, where a wrong candidate costs months of wet-lab time and the value is in getting the shortlist right.

What this means for you

If you run a discovery program, the bottleneck was never your science. It is the sheer size of the molecular space and the months it takes to explore even a sliver of it by hand. That is what DropXCell compresses, turning an open-ended search into a ranked, lab-ready shortlist, and it is the shape of what becomes possible when deep discovery expertise and agent infrastructure are built together.