Physics-grounded AI is getting Galux hits on the hardest antibody targets

In results released in November, Galux generated fifty designs per epitope across eight target epitopes and reported an overall binder rate of thirty-one and a half percent. Roughly ten percent of all designs reached therapeutically meaningful affinity, with several candidates binding at picomolar strength. Every candidate was tested as a full length IgG rather than as a fragment, which means the outputs behaved like drug leads without a round of downstream engineering to rescue them.

Conventional antibody campaigns screen somewhere between a million and a trillion candidates. Working from dozens changes the cost structure of discovery. It also changes who decides where an antibody binds and what it does once it gets there.

That inversion is the company's thesis. Galux starts from the medical unmet need, defines what a successful therapy has to achieve, and translates that into what chief executive Chaok Seok calls a molecular specification, "a description of what we want the molecule to do, not simply what we want it to bind."

The approach has a long runway behind it. Seok, a chemistry professor at Seoul National University, has worked on protein structure and interaction since 2004, building the GALAXY modeling suite by combining the physical chemistry of molecular interactions with data-driven bioinformatics. She founded Galux in 2020.

"Around 2020, AlphaFold2 showed that a data-driven AI model could achieve both practical accuracy and remarkable generality across proteins," Seok said. "To me, that was the fundamental change."

Her reference point is older than the current wave.

"Linus Pauling has always been an inspiration to me. He predicted the alpha helix before the first protein structures were experimentally determined, based on the principles governing chemical bonds and molecular geometry," she said. "It showed what becomes possible when we understand the underlying principles of nature."

Galux is often described as a physics-based alternative to sequence-pattern models, and Seok declines the framing.

"I would not draw a sharp distinction between learning physical principles and recognizing patterns in data," she said. "I see modern AI as an opportunity to bring these two approaches together."

What the company tests for instead is generalization. Models are evaluated on whether they extend to proteins and design problems they have never seen, which Seok treats as evidence that something more general than training data has been captured.

The performance story also starts earlier than the November results. In 2024, Galux ran library screening across roughly a million variants, and the finding that mattered was not the binders themselves.

"What surprised us was not simply that we found binders, but how precisely the experimental results followed what we had intended to design," Seok said. "That was when I felt we had crossed a major barrier."

By the time the fifty-design experiment ran, she expected it to work. She did not expect it to work that well.

"I personally did not expect the results to be as good as they were," she said. "Interestingly, our technical team seemed less surprised than I was."

In favorable cases, the company can now find a binder within a month, though harder targets and tighter molecular specifications need more design and test cycles. Seok is unwilling to let speed carry the argument.

"Discovering a binder quickly has limited value if the molecule later fails in development," she explains.

That physics foundation is aimed at the targets that have resisted structural design. GPCRs and ion channels carry flexible extracellular regions that adopt many conformations, which makes precise design far harder than it is against rigid epitopes. Co-founder Taeyong Park has described the approach as refusing to treat a target as a fixed structure, and instead building molecular motion and interaction dynamics into the design step itself. Galux reports confirmed binders against a GPCR target from a set of only 50 de novo VHH designs, again reaching picomolar affinity.

The validation runs deeper than affinity numbers. Earlier library-scale work spanning 8 therapeutic targets produced binders as tight as 9 pM, and the company resolved a designed PD-L1 antibody complex by cryo electron microscopy with interface accuracy near 1 angstrom. Galux built wet lab and dry lab capability from the start, so designs are tested internally and the results feed back into the platform.

Galux has also published a head-to-head comparison against reported Nabla Bio and Chai Discovery results on shared targets, finding binders for 8 of 9 where it places the other platforms at 5 and 4. The analysis is internal and compares across datasets generated under different conditions, so it is not a controlled bake-off. It is still the only such comparison anyone in this field has put on paper.

Commercially, the company works with more than twenty domestic pharmaceutical partners and two global ones in Boehringer Ingelheim and AstraZeneca, including joint research partnerships built on the GaluxDesign platform.

"Many partners begin with focused collaborative projects to evaluate how AI protein design can be applied to real drug discovery challenges," said Nicole Noh of Galux. "As these projects generate data and build confidence, we expect them to expand into deeper partnerships and mega deals over time."

Galux raised 42 billion won, about 30 million dollars, in a Series B in February, and has named Mirae Asset Securities and Korea Investment and Securities as joint lead managers for a planned listing. Seok describes two priorities for the next 12 months, pushing the platform past binder discovery toward hard therapeutic problems, and proving the molecules survive contact with drug development. Several pharmaceutical companies are already testing Galux-designed molecules.

"Our next important step is to advance our first assets into preclinical development and toward the clinic," Seok said.

The attention economy around AI protein design still points at Boston and San Francisco. Galux is evidence that the capability is more widely distributed than the coverage suggests, and that the interesting question now is not who can generate a binder but who can specify one. Seok says the same thing in her own register.

"As AI becomes capable of satisfying increasingly precise requirements, defining the right molecule becomes both more important and more feasible," she said. "Pushing that frontier, rather than simply making today's discovery process faster, is where we want Galux to go next."