Aureka Launches OpenDDE, an Open-Source Engine for Drug Discovery

The engine utilizes biomolecular co-folding to model interactions between proteins, nucleic acids, and small-molecule ligands. Rather than treating structure prediction as a static endpoint, the system functions as a shared reasoning layer for sequence-structure-function modeling. This architecture supports future development in de novo design and affinity estimation, allowing researchers to explore molecular space with greater precision.

Technical Performance and Scaling

OpenDDE features 655 million parameters and required 414,000 GPU-hours to train, highlighting the shift toward large-scale computational infrastructure in biology. In testing, the model demonstrated robust antibody-antigen co-folding capabilities, achieving a 51.0% success rate on the PXMeter-AB benchmark and up to 80.1% under oracle selection on the 2026ARK-AB dataset. According to Will Hua of Aureka AI Research, these results represent an early foundation for a system capable of connecting structure prediction with experimental feedback loops.

The release, licensed under Apache-2.0, includes training code, inference pipelines, and benchmarks to facilitate community-driven validation. Aureka intends to pair this computational foundation with its high-throughput automated wet-lab platform, creating a closed-loop system where AI agents propose candidates that are subsequently refined through experimental screening. This integrated approach is intended to support the discovery of complex modalities, such as multispecific and pH-switch antibodies, by bridging the gap between digital modeling and physical validation.