Protein features are often complex, and they are challenging to classify. In identifying the most discriminatory features in protein sequences, we propose a new feature-selection strategy by integrating the multivariate filter and Particle Swarm Optimisation (PSO) algorithms. Experimental results, based on the number of reducts and classification accuracy, were analysed in both the filter and wrapper phases. For our dataset, the proposed method statistically significantly improves the obtained classification accuracy and reduces the number of feature subsets. In the filter phase, the accuracy is improved more than 4% in three out of four multivariate feature selection methods compared to a model without feature selection. In the second phase, the accuracy is increased from 97.51% to 100%. We also demonstrate the importance of the correct parameter settings in the PSO to guarantee good performance.