Evaluating Different Descriptors for Model Design of Antimicrobial Peptides with Enhanced Activity Toward P. aeruginosa Ha ˚ vard Jenssen 1 , Tore Lejon 2 , Kai Hilpert 1 , Christopher D. Fjell 3 , Artem Cherkasov 3 and Robert E. W. Hancock 1, * 1 Centre for Microbial Diseases and Immunity Research, University of British Columbia, Vancouver, BC V6T 1Z4, Canada 2 Department of Chemistry, University of Tromsø, N-9037 Tromsø, Norway 3 Division of Infectious Diseases, Faculty of Medicine, University of British Columbia, Vancouver, BC V5Z 3 J5, Canada *Corresponding author: Robert E. W. Hancock, bob@cmdr.ubc.ca The number of isolated drug-resistant pathogenic microbes has increased drastically over the past decades, demonstrating an urgent need for new therapeutic interventions. Antimicrobial peptides have for a long time been looked upon as an interesting template for drug optimization. How- ever, the process of optimizing peptide antimicro- bial activity and specificity, using large peptide libraries is both tedious and expensive. Here, we describe the construction of a mathematical model for prediction, prior to synthesis, of pep- tide antibacterial activity toward Pseudomonas aeruginosa. By use of novel descriptors quanti- fying the contact energy between neighboring amino acids in addition to a set of inductive and conventional quantitative structure–activity rela- tionship descriptors, we are able to model the peptides antibacterial activity. Cross-correlation and optimization of the implemented descriptor values have enabled us to build a model (Bac2a- #2) that was able to correctly predict the activity of 84% of the tested peptides, within a twofold deviation window of the corresponding IC 50 val- ues, measured earlier. The predictive power, is an average of 10 submodels, each predicting the activity of 20 randomly excluded peptides, with a predictive success of 16.7 ± 1.6 peptides. The model has also been proven significantly more accurate than a simpler model (Bac2a- #1), where the inductive and conventional quantitative struc- ture–activity relationship descriptors were excluded. Key words: antimicrobial peptides, partial least square projections to latent structures, prediction of activity, Pseudomonas aeruginosa, quantitative structure–activity relationships, screening libraries Received 12 April 2007, revised 24 June 2007 and accepted for publica- tion 26 June 2007 The increasing use of antimicrobials has resulted in an increasing problem with drug-resistant bacterial and fungal pathogens (1,2). This has created the need for the discovery and design of new anti- microbial drugs. However, major difficulties have been experienced in discovering new chemical structures with low host toxicity and broad spectrum activity. We have suggested that cationic peptides serve as a good template for the design of such a new generation of antimicrobials (3). Several cationic antimicrobial peptides have been demonstrated to be quite effective in killing a wide selection of bacterial and fungal pathogens, including Pseudomonas aeruginosa, which is the third leading cause of hospital-associated infections and, as a result of chronic lung infections, the leading cause of morbidity and mortality in cystic fibrosis patients. Such peptides were initially demonstrated to target the bacterial cytoplasmic membrane but it is now recog- nized that many peptides translocate across the membrane and interact with cytoplasmic targets (3,4). A variety of modes of action have been ascribed to these peptides, but it has proven difficult to relate these modes of action to particular peptide sequences, as small changes can drastically affect structure (3,5). Thus, it has pro- ven challenging to systematically design and optimize new peptides with improved antimicrobial activity. Traditional design and optimiza- tion studies of peptides are also known to be expensive and time- consuming. However, production costs and the time required for evaluation of activity can be drastically reduced by synthesis of large peptide libraries on cellulose membranes and high-throughput antibacterial testing (6) as demonstrated through design of a single- substitution peptide libraries based on Bac2a, a linear peptide derivative of the 12-amino acid bovine neutrophil peptide bactene- cin (6,7). Another approach that has been used in streamlined pep- tide design has been to develop mathematical models to explain and predict the peptide activities. We have earlier demonstrated the use of principal component analysis (PCA) to explain the bio- logic activity of antimicrobial peptides (8,9). Partial least squares projection to latent structures (PLS) is another technique that has been used to build statistical models that can predict peptide activ- ity prior to synthesis (10). A peptide library containing more than 217 000 theoretical peptide sequences was screened for theoretical antimicrobial activity by the use of such a PLS technique, and the results were verified by synthesizing a limited number of these pep- tides and confirming their antimicrobial activity. However, larger 134 Chem Biol Drug Des 2007; 70: 134–142 Research Article ª 2007 The Authors Journal compilation ª 2007 Blackwell Munksgaard doi: 10.1111/j.1747-0285.2007.00543.x