How DECAF Speeds Up Protein Structure Predictions by 20 Times

Researchers at Genesis Molecular AI have developed a new framework called DECAF. It speeds up protein cofolding predictions drastically. This means it can predict how proteins and ligands fit together much faster than before.

DECAF works by turning slow diffusion-based models into faster flow-based models. These flow-based models can create high-quality protein structures in just a few steps. Traditional diffusion models need hundreds of thousands of neural function evaluations (NFEs). DECAF cuts this down to only 10-20 NFEs. That’s about 20 times faster.

The speed boost doesn’t come with a loss in quality. DECAF beats the Boltz-1x model on benchmark tests while using far less computing power. It also matches the top-performing Pearl model but uses five times fewer NFEs. Pearl is a state-of-the-art cofolding model that DECAF learns from.

Why Speed Matters in Protein Folding

Protein folding predictions help scientists understand biomolecular structures. This is key for drug discovery and virtual screening. But current AI diffusion models, like AlphaFold 3, require millions to billions of computations per step. This makes them slow and expensive to run.

DECAF changes that by reducing the inference cost dramatically. Dr. Joey Bose, assistant professor at Imperial College and co-author of the research, said, “Our paper shows this inference cost can be dramatically reduced for state-of-the-art cofolding models like Pearl without a trade-off in performance—unlocking much faster virtual screening capabilities which are critical in AI-based drug programs.”

What DECAF Means for the Future

DECAF inherits some limitations from Pearl, including issues with generating non-planar sp2 bonds. These minor flaws remain in the new model. However, it uses a denoiser formulation in the sigma (σ) space. This technique isn’t limited to protein cofolding. It could extend to other biomolecular systems.

The bigger takeaway is that more compute power alone doesn’t guarantee better results. Smarter strategies like DECAF’s flow-based approach can unlock faster and more efficient predictions. This is a big step for AI in healthcare and biotech. It makes virtual screening faster, cheaper, and more accessible to researchers.

Genesis Molecular AI’s CTO Sergey Edunov and co-founder Evan Feinberg are leading the charge. Their work shows that rethinking model design can push AI beyond just scaling up compute. DECAF proves that speed and quality can go hand in hand.

As protein structure prediction grows in importance, tools like DECAF could transform drug discovery pipelines. Faster predictions mean quicker insights and faster responses to health challenges. That’s a future worth watching closely.