Align a (pretrained) protein language model to alpha-amylase variants with improved activityon using DPO.