Protein Secondary Structure Prediction using N-Grams and 1-Dimensional Convolutional Neural Network

In this study, we initiated an approach to predict Protein Secondary Structure Prediction (PSSP) by using n-grams modeling, namely n adjacent amino acid sequences, to represent the amino acid sequence and 1-Dimensional Convolutional Neural Network (CNN). We compare n-grams with a one-hot encoding approach and find that n-grams provide better performance. The analysis results show that bigrams are more effective than trigrams in describing protein amino acid sequence patterns. We also combine it with the PSSM profile feature to improve model performance. The results of this study reveal that the n-grams approach, especially bigrams, in combination with PSSM, can produce more accurate protein secondary structure predictions than previous models, with superior accuracy evaluations of Q8 68.1 % and Q3 82.75% on the CB513 dataset.