Aureka Opens Its Drug Discovery Model as AI Biotech Moves Toward Lab-Linked Systems

Why it matters

OpenDDE gives Aureka a public benchmark for its AI biology work, while the proprietary value remains in the lab data, feedback loops and antibody programs around the model.

Dr. Weian Zhao's Aureka announced in a July 6th PR Newswire release that it has released Open Drug Discovery Engine, an open-source, all-atom biomolecular foundation model for AI-driven therapeutic discovery.

The move gives Aureka a public technical artifact in a market where AI biotech companies often ask pharma partners and investors to trust proprietary claims about models, screening data and discovery loops. OpenDDE is not presented as a finished drug-discovery machine. Aureka says its current release focuses on biomolecular structure modeling and co-folding, with downstream capabilities such as molecular design, affinity prediction, active learning, experimental feedback and clinical translation still requiring further validation.

That caveat matters. The company is offering the community a model that can reason over proteins, nucleic acids, small-molecule ligands and other biomolecular components, while keeping the fuller company system around Aureka's wet lab, experimental feedback loops and antibody discovery programs inside the company. In AI biotech, that boundary is the business model: open enough to win technical attention, closed enough to preserve the data and workflow advantage.

Zhao, Aureka's founder and CEO, comes to that bet from academia. UC Irvine lists him as a professor whose research interests include stem cell therapy, diagnostics, biosensors, nano- and microtechnology and aptamers. His UC Irvine profile also lists postdoctoral appointments across Harvard Medical School, Brigham and Women's Hospital, the Harvard-MIT Division of Health Science and Technology and the Harvard Stem Cell Institute. Aureka's own materials describe the company as founder-led and interdisciplinary, with domain expertise across high-throughput synthetic biology, single-cell technologies, data science, computational biology, deep learning and therapeutic development.

What Aureka actually released

OpenDDE, according to Aureka, uses co-folding as the entry point for a broader drug discovery engine. In plain terms, the model is built to predict how biomolecules fit together at atomic resolution, then use that structural reasoning as a basis for later design and optimization tasks.

Aureka's OpenDDE README calls the release a preview, requires Python 3.11 or higher, lists an Apache-2.0 license and provides Docker support. The README also states that predictions are not guaranteed to be reproducible across releases and that OpenDDE is not yet intended for production pipelines. The model weights are available through Hugging Face.

That preview label should shape how operators read the announcement. OpenDDE is a serious technical release, with code paths, weights and installation instructions. It is also a research release from a company that is careful to say the system's broader drug-discovery ambitions still sit beyond the current package.

Will Hua of Aureka AI Research framed it that way in the announcement. "OpenDDE begins with open, all-atom co-folding and structural reasoning," Hua said, adding that Aureka views the release as an early foundation for connecting structure prediction, molecular design, affinity estimation and experimental feedback.

The numbers are company-reported, and still useful

Aureka says OpenDDE has about 655 million trainable parameters and required about 414,000 GPU-hours for training, which the company equates to roughly 54 years on a single computing unit. Those figures put the release squarely in the infrastructure-heavy phase of AI biology, where meaningful progress requires compute budgets, dataset engineering, evaluation infrastructure and long training runs.

The benchmark results are narrower. In Aureka's technical report, the company says OpenDDE reached top-ranked antibody-antigen co-folding success rates of 51.0% on PXMeter-AB, 70.0% on FoldBench-AB and 66.4% on 2026ARK-AB. Under oracle selection, Aureka reports success rates of 65.9%, 81.9% and 80.1% across those same benchmarks.

Those are Aureka-reported in silico results. They are not independent validation, clinical proof or evidence that OpenDDE can yet select better drug candidates in the lab. Aureka also frames the model as narrowing the gap with reported IsoDDE-level results, but the announcement does not provide enough context to treat that as a clean commercial comparison. The relevant takeaway is more concrete: Aureka is putting a co-folding model into the open and using antibody-antigen performance as the first public test of its structural reasoning claims.

Why open-source the model while building a proprietary company

Aureka's strategy becomes clearer when OpenDDE is read against the rest of the company's platform. Aureka's technology page says it combines scalable protein evolution, high-throughput single-cell functional screening, protein geometric language models and reinforcement learning. Its July 6th announcement says the computational foundation is being paired with high-throughput automated wet-lab systems, autonomous antibody-design agents and automated yeast evolution.

That setup follows a pattern across AI biologics: the model gets attention, but the durable asset is the loop between design, experiment, data capture and model update. Absci describes an AI drug creation platform that pairs wet lab and AI cycles for biologics design. BigHat Biosciences describes a design studio tied to an automated wet-lab characterization platform for sequential design-build-test cycles. Aureka is making the same class of bet, with its own emphasis on co-folding, structural reasoning and functional antibody discovery.

The timing also follows the money. OpenDDE is not a financing announcement. Aureka's funding push happened earlier. Cooley said on Nov. 10th, 2025 that it advised Aureka on a Series A co-led by 5Y Capital and Qiming Venture Partners, with NRL Capital and Agentic Ventures participating. On April 28th, 2026, PR Asia reported that Aureka completed a $35 million Series A+ led by HongShan, with MPCi, Matrix Partners China and BioTrack Capital joining, and 5Y Capital, Qiming Venture Partners and NRL Capital reinvesting. PR Asia reported that the Series A and Series A+ together totaled nearly $100 million. A valuation was not disclosed.

That capital gives Aureka room to release part of the stack without giving away the company. OpenDDE can help recruit researchers, attract external users, invite scrutiny and position Aureka inside the open scientific AI conversation. The internal value still depends on whether Zhao's team can generate proprietary experimental data, improve models through feedback and move antibody programs toward clinical relevance.

A public model raises the bar for the private platform

OpenDDE gives Aureka a sharper story than a standard AI biotech platform claim. Anyone can inspect the release materials, pull weights and test the preview within the limits Aureka has set. That openness also creates pressure. If the model is useful, researchers will expect maintenance, documentation, issue handling, reproducible examples and clearer governance. If it is mainly a technical calling card, the community will learn that too.

The harder question sits inside Aureka's own caveat. Structure prediction is useful because it can reduce blind search. Drug discovery still turns on biology, assay design, manufacturability, safety, pharmacology and clinical execution. Aureka's release acknowledges that by placing OpenDDE at the front of a longer chain: de novo design, affinity estimation, conformational ensemble modeling, structure-conditioned optimization, active learning and experimental feedback.

For Zhao, OpenDDE is a public stake in a larger founder thesis. Aureka is trying to turn academic biology, AI infrastructure and wet-lab automation into a repeatable antibody discovery system. The open-source release makes that thesis easier to examine. It also makes the next milestone less about whether Aureka can publish a model and more about whether the model improves the closed-loop work that actually produces drug candidates.