Antimicrobial peptides (AMPs) are widely distributed defense molecules and represent a promising alternative for solving the problem of antibiotic resistance. Nevertheless, the experimental time required to screen putative AMPs makes computational simulations based on peptide sequence analysis and/or molecular modeling extremely attractive. Artificial intelligence methods acting as simulation and prediction tools are of great importance in helping to efficiently discover and design novel AMPs. In the present study, state-of-the-art published outcomes using different prediction methods and databases were compared to an adaptive neuro-fuzzy inference system (ANFIS) model. Data from our study showed that ANFIS obtained an accuracy of 96.7% and a Matthew's Correlation Coefficient (MCC) of0.936, which proved it to be an efficient model for pattern recognition in antimicrobial peptide prediction. Furthermore, a lower number of input parameters were needed for the ANFIS model, improving the speed and ease of prediction. In summary, due to the fuzzy nature ofAMP physicochemical properties, the ANFIS approach presented here can provide an efficient solution for screening putative AMP sequences and for exploration of properties characteristic of AMPs.

Abstract. The cysteine knot motifs are widely spread in several classes of peptides including those with antimicrobial functions. These motifs offer a major stability to the protein structure. Nevertheless, the antimicrobial activity is modulated by physicochemical properties. In this paper, we create a model of support vector machine to predict antimicrobial activity from sequences with similar motifs, based on physicochemical properties: net charge, ratio between hydrophobic and charged residues, average hydrophobicity and hydrophobic moment. The support vector machine model was trained with 146 antimicrobial peptides with six cysteines from the antimicrobial peptides database and an equal number of random sequences predicted as transmembrane proteins. The polynomial kernel shows the best accuracy (77.4%) on 10-fold cross validation. Testing in a blind dataset, we observe an accuracy of 83.02%. Through this model, proteins of varied size with a cysteine knot motif can be predicted with good reliability.

No abstract

No abstract

2007), Doctorate in Biotechnology (Bioinformatics), Catholic University of Brasilia (2012). He is currently professor at Federal Institute of Brasilia. His scientific works concern bioinformatics, artificial intelligence, data mining, data science, machine learning and informatics in health.