“To ensure generalization, Chai Discovery's model evaluation excluded any targets that had more than 70% sequence identity with the training data.”