Aureka released an open-source AI model that predicts how antibodies bind their targets more accurately than Google DeepMind's AlphaFold 3, and it closed a $100 million round to apply the tool to drug discovery. The task the model handles is narrower than it sounds: antibodies work by folding a loop-shaped tip into a shape that locks onto one patch of a target protein, and predicting that loop's exact shape from sequence alone has been the hard part for two decades, because the loop is floppy and its final form depends on the protein it meets. Aureka trained on a dataset built specifically for that loop geometry rather than the general protein-folding problem AlphaFold 3 was built to solve, which is the mundane reason a smaller, specialized model can out-predict a bigger general one on this slice of chemistry.

The model is a benchtop prediction tool, not a drug. What it changes is the front end of antibody discovery: instead of synthesizing thousands of candidate antibodies and testing which ones bind, a lab can rank candidates computationally first and only make the ones the model scores highest. Aureka has not disclosed a specific antibody program moving into animal testing off the back of this model; that is the gate that will show whether the accuracy gain on paper turns into fewer wet-lab cycles per drug candidate.