“The Chai 2 model maintained a consistent success rate even when tested on targets with as low as 25% sequence similarity to its training data, indicating strong generalization capabilities.”