Current base editors use DNA deamination enzymes, such as cytidine deaminase (CBE) or adenine deaminase (ABE), to facilitate transition nucleotide substitutions. Combining CBE or ABE with glycosylase enzymes, including CGBE and AYBE, can induce limited transversion mutations. Nonetheless, a critical demand remains for base editors capable of generating alternative mutation types, such as T>G corrections, which could address over 17% of monogenic SNVs responsible for human genetic diseases. In this study, we leveraged protein language models to engineer a uracil-N-glycosylase (UNG) variant with altered substrate specificities to thymines (eTDG). Notably, after only two rounds of testing fewer than 50 predicted variants, more than 50% exhibited a 1.5-11-fold enhancement in enzymatic activities, a success rate much greater than random mutagenesis. When eTDG was fused with Cas9 nickase without deaminases, it effectively induced programmable T to G or T to C substitutions in cell lines and precisely corrected db/db diabetic mutation in mice (up to 55%). Our findings not only establish orthogonal strategies for developing novel base editors, but also demonstrate the capacities of protein language models for optimizing enzymes without extensive task-specific training data or laborious experimental procedures.