Protein Classification using Modified N-Gram and Skip-Gram Models: Extended Abstract

Machine Learning (ML)-based classification of protein characteristics from primary sequences is an important tool for exploring candidate proteins in targeted drug discovery, mutational analysis, and functional identification. However, ML feature selection requires extensive manual curation and knowledge of protein chemistry, interactions, and micro-environment of the proteins of interest. Current approaches include amino acid composition strategies, specific motif analysis or Quantitative Structure-Activity Relationship (QSAR)-based feature generation methods. In contrast, we propose an automated generalized feature generation method based on Natural Language Processing (NLP), using a modified combination of N-Gram and Skip-Gram models (m-NGSG). Optimal parameters are selected using an adapted grid search algorithm, enabling a high-throughput global application of our approach. A meta-comparison of logistic regression mediated classification approaches exploiting m-NGSG with other published models illustrates enhanced functional and structural binary and multi-class classification accuracy in every instance. The lack of dependence on detailed physicochemical knowledge makes the m-NGSG approach ideal for the exploration of protein characteristics recalcitrant to previous approaches without any loss in predictive accuracy. A further test on prediction quality of m-NGSG on cationic channel blockers with 70% sequence identity from Arthropods demonstrated 94.10% and 92.30% accuracy on the training and test set, respectively. The latter study demonstrates the applicability of m-NGSG model on a functional classification of proteins employing a novel dataset.Thus, without the requirement of expert intervention for optimal feature selection, it is hoped that this automated feature generation approach will reduce the time needed to employ ML classification strategies for prediction of protein characteristics.