AI-Designed De Novo Proteins — Galux Engineering Therapeutic Antibodies from Scratch

$29 million. That is what a South Korean biotech called Galux announced in its Series B in February. The number matters less than what it buys: the ability to design therapeutic antibodies from scratch, without starting from anything nature built.

Antibody discovery has followed the same recipe for decades. Immunize an animal, screen millions of candidates, pick the best binder, then spend months humanizing and affinity-maturing it. The process works. It also wastes enormous time and capital on molecules that fail late. Probabilistic screening, however sophisticated, remains a numbers game.

It is trying to replace the numbers game with precision design.

How de novo antibodies are designed

The design pipeline combines three layers of computation. First, a generative model proposes antibody sequences conditioned on a target epitope. It builds an entirely new amino acid sequence around the binding geometry rather than mutating a known framework. Second, a structure prediction model (analogous to AlphaFold but specialized for antibody-antigen interfaces) estimates how each candidate will fold and interact. Third, a scoring function ranks candidates by predicted affinity, stability, and developability.

What distinguishes this from earlier computational antibody design is the tight experimental feedback loop. It maintains an internal wet lab that synthesizes and tests the top-ranked designs within weeks. Results for binding affinity, structural validation by cryo-EM, and expression yield are fed back into the AI models, closing the design loop. This build-measure-learn cycle is what turns a one-shot prediction into a reliable engineering platform.

The November 2025 study is the best evidence that the loop works. Across 8 target epitopes covering diverse structural classes (PD-L1, HER2, EGFR S468R mutant, ACVR2A/B, FZD7, ALK7, CD98hc, IL-11), GaluxDesign produced binders for every target. The overall hit rate of 31.5% means nearly one in three designed sequences bound the intended target. The 10.5% rate for therapeutically meaningful affinity (sub-100 nM EC50) means one in ten was immediately viable as a drug lead.

These numbers are not incremental improvements over library screening. They represent a different regime entirely. A typical phage display campaign screens 10⁹ variants to find a handful of leads. The platform achieves comparable or better results from 50 computationally designed sequences.

Why cryo-EM validation matters

The most common criticism of AI-designed proteins is that computational models hallucinate structures that look plausible but do not actually fold that way in solution. Cryo-EM validation addresses this directly. The company resolved a designed anti-PD-L1 antibody by cryo-EM and found the interface matched the computational prediction with 1.1 Å RMSD. The designed and measured structures were virtually identical at the atomic level.

This is not a trivial result. In earlier attempts at computational antibody design by academic groups, the gap between predicted and actual structure was often several angstroms. Enough to lose binding specificity entirely. Achieving sub-2 Å consistency across multiple targets is the evidence that the physics models underlying the platform are capturing real protein behavior, not overfitting to training data.

The cryo-EM data also confirmed that the designed antibodies were structurally novel. The AI did not simply rediscover known antibody frameworks but generated new folds with comparable stability. For therapeutic development, structural novelty matters because it expands the patentable space beyond existing antibody IP.

From targets to partnerships

Traditional antibody development starts with a library of 10⁹ to 10¹¹ candidates. Phage display or yeast display screens reduce that to hundreds of hits, which then go through iterative rounds of affinity maturation, humanization, and developability optimization. Each round takes weeks. Most candidates that look good at the hit stage fail at the IgG stage: poor expression, aggregation, or off-target binding that only shows up when you move to full-length format.

De novo design inverts this. Instead of searching a massive library for a needle, you compute the needle's exact geometry from first principles. GaluxDesign, the company's AI platform, integrates deep learning with atomic-level physics: protein folding prediction, binding site complementarity, and structural stability. It generates antibody sequences that are designed, not found.

In a study published on bioRxiv in March 2025, it demonstrated de novo antibody design across six therapeutic targets, including cases where no experimentally resolved target structure was available. The designed antibodies matched cryo-EM structures with an interface RMSD of 1.1 Å, effectively atomic resolution.

A follow-up study in November 2025 went further. The team generated only 50 AI-designed antibodies per epitope across eight targets and achieved a 31.5% binder rate, with 10.5% showing therapeutically meaningful affinity. Several candidates reached picomolar binding strength. All were validated in full-length IgG format. The designed sequences worked as drugs without additional engineering.

From Seoul National to a global pipeline

The company traces its technology to 15 years of molecular design research at Seoul National University, led by CEO Cha-ok Seok. It was founded in 2020 and raised an $18 million Series A in 2022, which moved its AI systems from academic theory to a production platform. The Series B, announced in February 2026, brought total funding to $47 million and added institutional backers including Korea Development Bank, Yuanta Investment, and Mirae Asset Securities.

The money is funding expansion on two fronts: harder targets and more partners.

Capital is concentrating on platform proof

The $3.1 billion that flowed into AI drug discovery between Q1 2025 and Q1 2026 was not distributed evenly. The bulk went to companies that could show real data: resolved structures, validated candidates, or pharma partnerships. Earendil Labs raised $787 million for its small-molecule platform. Isomorphic Labs raised $600 million on the strength of AlphaFold-derived candidates entering human trials. Profluent Bio closed a $106 million round for its generative protein design engine.

Galux's $47 million total is modest relative to those numbers, but the comparison misses the point. It is not raising to build a platform. The platform exists and is producing published, independently verifiable results. The Series B funds platform extension (GPCRs, ion channels) and preclinical validation, not infrastructure construction. The capital efficiency ratio, measured as results published per dollar raised, is among the best in the space.

The investor syndicate reinforces this. Returning backers InterVest (seed through Series B), DAYLI Partners, and PATHWAY Investment have followed the company for four years. New institutional names like Korea Development Bank, Yuanta Investment, and Mirae Asset Securities are not typical early-stage biotech investors. Their participation signals that Galux is being evaluated as a platform company with infrastructure-level potential rather than a binary drug developer.

The competitive field — several platforms, one validation gap

AI Proteins raised $41.5 million in Series A financing in November 2025 for its de novo miniprotein platform. The company operates at a smaller scale. Its miniproteins are roughly one-tenth the size of a full antibody, which gives them advantages in tissue penetration and manufacturing cost but limits their therapeutic surface area. AI Proteins has a partnership with Bristol Myers Squibb valued up to $400 million, providing both validation and revenue visibility.

Accipiter Biosciences emerged from stealth in November 2025 with $12.7 million in seed funding, pursuing multifunctional de novo protein therapeutics. The company is earlier stage and has not published structural validation data comparable to Galux.

Isomorphic Labs is the elephant in the room. Backed by DeepMind's protein folding breakthroughs, it has the deepest pockets and the strongest brand in AI biology. Its first AI-designed molecules entered human trials in 2025-2026, a lead that competitors will need to close. But Isomorphic Labs focuses primarily on small-molecule drug design, not antibodies. The two companies are not direct competitors in the near term.

What separates Galux from this field is not the AI architecture. Most platforms use some variant of diffusion models or transformers conditioned on structural data. The differentiator is the experimental validation density: cryo-EM at 1.1 Å, full-length IgG confirmation, picomolar affinity, and functional specificity across 8 targets. No other de novo antibody design platform has published comparable data across this many dimensions.

The open question is whether this validation edge translates into clinical candidates faster than competitors with larger balance sheets. That answer is 18-24 months away.

What a $47 million platform means for drug economics

Between Q1 2025 and Q1 2026, startups in the AI drug discovery space raised over $3.1 billion across 38 disclosed rounds. Protein and biologics design accounted for roughly $1.32 billion of that. It was the second-largest category behind small-molecule platforms, but growing faster. Earendil Labs raised $787 million in a single round. Isomorphic Labs took $600 million. Profluent Bio closed a $106 million round in late 2025.

Galux's $47 million total is modest by those standards. But the metrics that matter for de novo antibody design are not about capital raised. They are about whether the designed antibodies actually work as drugs. On that dimension, it has published more structural validation data than any comparable startup. The combination of cryo-EM confirmation, full-length IgG validation, and functional specificity across 8 targets puts it ahead of most peers in experimental rigor.

What remains unproven

AI-designed antibodies have not yet entered clinical trials. Galux's candidates remain at the preclinical stage. The transition from picomolar affinity in a dish to efficacy and safety in humans is the hardest step in drug development, and no de novo designed antibody has cleared it.

There is also the question of target scope. The 8 targets validated so far, while diverse, are all well-characterized proteins with known structures or close homologs. The platform's performance on truly novel, structurally uncharacterized targets (the "dark proteome") remains untested. GPCR and ion channel programs are underway but have not published results.

Competition is real. AI Proteins raised $41.5 million in November 2025 for its own de novo miniprotein platform, with a partnership at Bristol Myers Squibb valued up to $400 million. Accipiter Biosciences emerged from stealth with $12.7 million in seed funding. Isomorphic Labs, backed by DeepMind's AlphaFold team, has already advanced its first AI-designed molecules into human trials.

Galux's edge is not that it is the only company doing this. Its results are the most thoroughly validated at the structural level. Whether that translates into a clinical advantage is the open question for the next 18 months.

For an investor evaluating the space, the key distinction is between companies that are selling a platform story and those that have published platform proof. Galux falls into the second category. The funding is sufficient to reach an IND filing in at least one program. The partnership model reduces the capital burden of Phase 2/3 trials. The intellectual property around de novo designed sequences is structurally defensible: a cryo-EM map at 1.1 Å is a hard prior art barrier.

De novo antibody design is not yet a proven therapeutic modality. But the evidence base has shifted from "can AI design a protein that folds?" to "can AI design a protein that beats the best natural antibodies in the clinic?" That is a different question, and the industry is about to find out the answer.

Galux's trajectory over the next two years will tell us whether rational protein design can finally deliver on a promise the biotech industry has chased for two decades: writing the code for new medicines from scratch, with atomic precision, on demand. The tools are ready. The question is whether the molecules are too.