Enzyme sequence design has always been a challenging task, particularly in optimising key properties such as enzyme solubility, stability, and activity. This study proposes an innovative approach by utilising a variational autoencoder (VAE) model integrated with the Gromov-Wasserstein (GW) distance for enzyme sequence optimisation. The GWAE model improves representation learning by using the GW distance, thereby generating functional variants with desired characteristics. We also introduce an innovative enzyme dataset construction method that incorporates multiple sequence alignment (MSA) techniques to address sequence length discrepancies, enhancing the accuracy of the optimisation process. Experimental results show that the GWAE model outperforms the traditional VAE on multiple metrics. The generated enzyme sequences demonstrate superior solubility, stability, and hydrophobicity. Additionally, by integrating AlphaFold3 for structural prediction, we verify the structural stability of the generated sequences, further enhancing their practical applicability.