Hybrid Random Forest and Support Vector Machine Model for Protein Sequence Classification

Protein plays an important role in human life and the effective use of biological sequence information is essential to the development of biomedicine. This study uses skip-gram model to vectorize protein sequences, and then performs the tasks of protein sequence classification. Considering the protein family classification is a multi classification problem, we attempt to use the random forest (RF) and multi class SVM model algorithm. And we use weighted accuracy as the standard to measure the classification effect, since the family with more instances should have greater impact on the overall accuracy. The results show that the accuracy of RF is 76 ± 0.06%, and the weighted average accuracy of SVM model is only 63 ± 0.03%, which indicates that RF model has obvious advantages in protein family classification task. As for the ordered and disordered protein classification, we use decision tree model (DT). To validate the model, we calculate the precision, recall and accuracy through 10 fold cross validation. It is found that the average accuracy of DT model is 93±0.06
^%
which is better than the result of binary SVM (76±0.08
^%
). At the same time, in order to visualize the classification effect, we reduce the dimension of protein sequence to better observe its distribution in two-dimensional space after classification.