A three-year-old biotech startup has built an AI model that predicts antibody structures 58% more accurately than Google DeepMind's AlphaFold 3, and it just gave the model away for free under a commercial-use license. Aureka Biotechnologies closed a $100 million Series B funding round on August 10, 2026, backed by this counterintuitive strategy: release the algorithm openly while keeping the proprietary wet-lab infrastructure that trained it locked behind closed doors.

The model, called OpenDDE, achieved a 76% success rate on antibody-antigen structure prediction in independent testing by Tamarind Bio, a bioinformatics platform, compared with AlphaFold 3's 48% on the same benchmark. That gap represents a meaningful leap in a task that matters enormously to drug discovery: designing antibodies that bind tightly to disease targets without triggering unwanted side effects.

Why Would a Company Give Away Its Best Work?

The answer reveals a shift in how AI-driven biotech companies think about competitive advantage. Aureka's thesis is that the model itself is not the moat; the moat is the closed-loop laboratory system that trained it. By open-sourcing OpenDDE under the Apache 2.0 license, the company is betting that every computational biologist and pharmaceutical researcher who uses the free model will generate demand for the proprietary infrastructure that created it.

This strategy mirrors what successful software companies call the "open core" model: release the algorithm to build ecosystem adoption and scientific credibility, then monetize the infrastructure no one else can replicate. In Aureka's case, that infrastructure is a proprietary single-cell functional screening system integrated directly into the model's training pipeline. When the model proposes a molecular design, the experiment that tests it generates functional data that flows back into both the foundation model and project-specific fine-tuned versions. The result is a flywheel: data feeds models, models drive experiments, experiments generate better data, and better data trains better models.

What Makes OpenDDE Technically Different From AlphaFold 3?

OpenDDE uses the same underlying Pairformer architecture as AlphaFold 3, the deep learning framework that powers Google DeepMind's reference model. But Aureka scaled it substantially upward. The model contains 655 million trainable parameters across 48 Pairformer blocks, 4 multiple sequence alignment (MSA) module blocks, and 24 diffusion transformer blocks. Training it required approximately 414,000 GPU-hours, equivalent to a single computing unit running continuously for roughly 47 years.

The key architectural change is width scaling: the model's hidden and channel dimensions are set to 384, larger than the AlphaFold 3-style baseline. Aureka's research team argues that this width scaling, not just the parameter count, drives the performance gain on complex multi-body interactions. Antibody-antigen co-folding benefits most from this approach because the binding geometry between an antibody's variable regions and an antigen's epitope involves highly variable loop conformations that require dense pairwise reasoning across the entire complex.

How Does This Fit Into the Broader AI Drug Discovery Landscape?

Aureka is not alone in racing to compress drug discovery timelines. Over the past two years, several companies have created real medicines using artificial intelligence, and some of these drugs are already being tested on people. Insilico Medicine, a biotechnology firm at the forefront of AI-driven drug development, created an experimental drug called Rentosertib designed to treat pulmonary fibrosis, a chronic and often fatal condition that scars lung tissue. The company completed this discovery process in roughly 30 months, about half the time of traditional methods.

Rentosertib is currently being evaluated in Phase IIa clinical trials in the United States and China. A 12-week randomized, double-blind study recently showed dose-dependent improvement in lung function, with some patients experiencing an increase in lung capacity. Insilico Medicine used its proprietary Pharma.AI platform, which includes PandaOmics for analyzing biological datasets and Chemistry42 for generative AI molecular design, to identify the drug target and optimize the molecular structure.

DeepMind's AlphaFold project solved one of biology's greatest challenges by accurately predicting protein folding structures, an achievement that significantly accelerated biomedical research worldwide. AlphaFold 2, released in 2020, achieved over 90% accuracy on protein structure prediction, a problem that had resisted solution for fifty years. AlphaFold 3 expanded its capability to model interactions across a wide array of biomolecules, including DNA, RNA, and ligands.

Steps to Understanding Lab-in-the-Loop Drug Discovery

• Traditional Workflow: A computational model predicts a protein structure or candidate molecule, and a wet lab then validates or rejects that prediction downstream. The laboratory serves as a final checkpoint after computational work is complete.
• Aureka's Inverted Approach: The experimental platform is integrated directly into the model's training pipeline. When the model proposes a design, the experiment testing it generates functional data that flows immediately back into both the foundation model and project-specific fine-tuned versions.
• The Flywheel Effect: Better data trains better models, better models drive more targeted experiments, and more targeted experiments generate higher-quality data. This continuous loop accelerates the entire discovery process compared to sequential, disconnected stages.
• Commercial Validation: Aureka has established partnerships with multiple major global pharmaceutical companies for antibody therapeutic programs and reported tens of millions of dollars in commercial revenue over the past two years, providing partial verification that the system produces usable drug candidates.

Aureka's strongest current advantage sits specifically in the antibody-antigen domain. On protein-ligand co-folding tasks, where a protein and a small molecule drug candidate are predicted together, OpenDDE currently scores approximately 0.601, below AlphaFold 3 (0.649) and Protenix-v1 (0.625). This finding is confirmed in independent technical reviews, meaning the model has not yet matched competitors on small molecule work.

The company has produced differentiated antibodies at scale for target classes that conventional approaches struggle with: G protein-coupled receptors (GPCRs), a family of membrane receptors that regulate heart rate, blood pressure, pain, and neurotransmission and account for roughly one-third of all FDA-approved drugs but are notoriously difficult to target with antibody-based therapeutics; and bispecific antibodies (BsAbs), engineered molecules that simultaneously engage two different antigens, with substantially higher manufacturing complexity than conventional monoclonals.

What Does This Mean for the Future of Drug Discovery?

The shift toward AI-native drug discovery models is accelerating. In the future, AI may allow physicians to prescribe treatments tailored specifically to a patient's genetic makeup, lifestyle, and medical history. Rather than prescribing identical medications to millions of patients, doctors could use AI systems to develop individualized therapies with greater effectiveness and fewer side effects.

AI may also transform how clinical trials are conducted. Traditional clinical trials often take years because researchers struggle to recruit enough participants and monitor results manually. In the future, AI systems could rapidly identify ideal candidates by scanning electronic health records and genetic databases. Machine learning algorithms may also monitor patients in real time through wearable devices that track heart rate, oxygen levels, blood sugar, and other vital signs, making clinical trials faster, safer, and more accurate.

Another important advancement involves rare diseases. Many rare illnesses receive little research funding because pharmaceutical companies often prioritize drugs with larger markets and higher profits. However, AI significantly lowers research costs and accelerates discovery timelines, making it more financially realistic to develop treatments for smaller patient populations. As a result, diseases that were previously ignored may finally receive effective therapies.

The broader context matters here. In August 2026, Demis Hassabis stepped down as CEO of Google DeepMind to focus on long-term artificial general intelligence (AGI) research. The team that built AlphaFold, which won the Nobel Prize in Chemistry in October 2024, was disbanded by July 2026. John Jumper, a co-author of the AlphaFold papers, left for Anthropic, and two other co-authors followed. The lab that won a Nobel Prize is now organized around making Gemini, Google's large language model, better and catching up to competitors.

This organizational shift at DeepMind reflects a tension between scientific ambition and commercial pressure. DeepMind's founding philosophy was that the path to transformative AI ran through hard science: pick a problem that has resisted solution for decades, build a purpose-built system to crack it, and let the breakthrough speak for itself. Google Brain's philosophy was structurally different: build a general architecture that could be trained on anything and applied to everything. The 2023 merger of DeepMind and Google Brain forced a choice, and the disbandment of the AlphaFold team revealed which philosophy won.

Aureka's strategy suggests a different path forward: open-source the general capability to build ecosystem adoption, but keep the proprietary infrastructure that generates competitive advantage locked behind closed doors. Whether that model proves more sustainable than either pure scientific ambition or pure commercial pressure remains to be seen.