Disclosed are a machine learning gene mining method and a phosphinothricin dehydrogenase mutant for amino translocation. The phosphinothricin dehydrogenase mutant for amino translocation is obtained by mutation of a wild-type phosphinothricin dehydrogenase with an amino acid sequence as shown in SEQ ID No.2 at one of the following sites: (1) E263D-K134R-H96A-R290V; (2) E263D-K134R-H96A; (3) E263D-K134R; (4) E263D; (5) E263N; (6) E263C; and (7) E263G. The present invention utilizes the site-saturation mutagenesis technology to mutate a phosphinothricin dehydrogenase gene as shown in SEQ ID No. 1, finds that the 263rd, 134th, 290th and 290th positions are the key sites affecting enzyme activity and stereoselectivity, and obtains a mutant with enzyme activity and ee value much higher than those of the parent phosphinothricin dehydrogenase.