Connecting Peptide Physicochemical and Antimicrobial Properties by a Rational Prediction Model

1Antimicrobial peptides (AMPs) are molecules found in all biological kingdoms responsible for the fight against microbial infections in the first steps of the immunological response. New strategies developed by bacteria and other microorganisms to evade classical antibiotics have urged the pharmaceutical industry to develop new drugs in order to wipe out these resistant microorganisms. In particular, AMPs have been proposed as promising candidates against these pathogens,. Although AMPs show a low potency when compared with the small bioactive drugs used at present, they offer counterweigh advantages such as broad range of activity and low toxicity and are less prone to give rise to resistant strains,,. However, AMPs may present some drawbacks such as serum instability, degradation by proteases and high production costs in the case of large polypeptides.

2Hence, some in silico methods have been developed to find AMPs with potential therapeutic application. Several algorithms take advantage of data mining and high-throughput screening techniques and apply vector-like analysis to scan protein and peptide sequences,. Other bioinformatics' strategies include supervised learning techniques, such as artificial neural networks (ANN) or support vector machines (SVM), in order to evaluate easily and reliably a great amount of complex data,. In fact, most attempts have been centred in the prediction of highly active peptides using quantitative structure-activity relationships (QSAR) descriptors together with ANN,,, linear discriminant or principal component analysis. These systems use mainly 3D-QSAR descriptors to detail the antimicrobial properties of peptides. Recently, a QSAR-based ANN system was experimentally validated using SPOT high-throughput peptide synthesis, showing that this methodology can accomplish a reliable prediction by means of conventional and “inductive” QSAR descriptors. However, the datasets used contained only peptides with fixed length and the leads found were only populated in few amino acids (W, R and K and, more limitedly, L, V and I). Although AMPs are actually enriched in these residues, a wide diversity in the amino acid content can be found in natural AMPs.

3Despite the inherent complexity in designing a prediction system only by means of computational chemistry, the recent methods mentioned above have made a remarkable advance. Hence, the combined use of bioinformatics and experimental screening techniques will be essential for the discovery and refinement of new AMPs.

4We report here an ANN based method that is able to correlate a complete set of sequence-derived physicochemical properties with antimicrobial activity. The most pioneering feature in this method is the ability to translate the observed results in a model of action for AMPs taking into account some new concepts such as peptide aggregation. Additionally, we conclude that the amino acid sequence provides us with sufficient information to accurately predict antimicrobial peptides.

5In order to characterize the physicochemical properties of AMPs, we have selected eight parameters: isoelectric point (pI), peptide length, α-helix, β-sheet and turn structure propensity, in vivo and in vitro aggregation propensity and hydrophobicity. We have first evaluated these descriptors in antimicrobial and non-antimicrobial peptides to test its suitability to act as classifiers (See the Methods section for a complete description on the peptide datasets). The statistical distribution for each parameter is illustrated in Figure 1.

6It can be seen in Figure 1A that AMPs and non-AMPs have similar average isoelectric points of 9.26 and 9.20, respectively. However, the variances observed for both groups are significantly different (p-value <0.0001), being the data dispersion much greater for non-AMPs. Thus, variance analysis suggest that a high positive net charge is required for AMPs whereas it does not represent a distinctive feature in non-AMPs, probably due to the diverse functions exerted by these peptides.

7Similar results are observed for peptide length (Figure 1B). Non-AMPs tend to be larger than AMPs but no significant differences can be found in the mean value though variances in both groups differ significantly (p-value <0.0001). The length parameter can contribute to the antimicrobial mechanism of action by modulating the peptide insertion into the membrane. It has also been described to be an important parameter for the de-novo generation of antimicrobial peptides.

8Remarkable differences have been found in the analysis of the structural parameters of these two groups (Figures 1C, 1D, 1E). In this case, it can be observed that AMPs tend to be random coil in solution, with a low tendency to present any defined structure. In contrast, non-AMPs display a high proclivity to adopt α-helix or β-sheet structure (Figure 2). These results are consistent with the experimental data reported where it has been described that most AMPs have no structure in solution but acquire a defined secondary structure upon membrane interaction.

9We have also studied the aggregation propensity for both groups, as it could be an important modulator of the peptide function. Aggregation in solution has to be clearly distinguished from the capacity of many antimicrobial peptides to form aggregates upon interaction with cell membranes, a step required for the AMPs mode of action (e.g. in the “carpet –like” mechanism ). Besides, peptide aggregation on the bacteria surface has been observed in some cell-agglutinating AMPs,. This agglutinating activity could help bacteria clearance in the body at the infection focus.

10It has been described that TANGO is a good predictor of aggregation in solution, as it uses sequence-derived structural parameters to forecast aggregation. On the other hand, AGGRESCAN is a good analyst of aggregation in bacteria, as it has been developed in an E. coli system,. We have used TANGO software to predict in vitro (or “in solution”) aggregation, whereas AGGRESCAN has been used to predict in vivo aggregation. As we can see in Figures 1F and 1G, AMPs display a low in vitro aggregation propensity when compared with non-AMPs while the in vivo aggregation parameter is considerably higher for AMPs. Additionally, great dispersion in the in vitro aggregation propensity has also been observed for non-AMPs.

11If TANGO and AGGRESCAN values are plotted together for AMPs and non-AMPs we can observe an interesting pattern (Figure 3). Whereas AMPs present high in vivo and very low in vitro aggregation propensity, non-AMPs can be divided in two main groups, presenting either high or low values for both descriptors. These results suggest that AMPs may minimize its aggregation in solution but promote aggregation in a more hydrophobic environment (i.e. the bacteria cell membrane). On the contrary, non-AMPs exhibit a dispersed aggregation pattern, probably related to their diverse biological functions. Thus, the results suggest that peptide's aggregation behaviour could be useful for classification purposes.

12Finally, the hydrophobicity analysis (Figure 1H) shows that both the means and the variances differ significantly between AMPs and non-AMPs (p-value <0.0001). AMPs need to insert, partially or totally, into the membrane hydrophobic core in order to destabilize the bilayer and/or promote the cell depolarization. Thus, a higher mean hydrophobicity value is expected for AMPs.

13Using the physicochemical properties described above, an ANN system has been constructed in order to classify peptides in two groups: AMPs and non-AMPs. We have used the CAMP peptide database to build the positive dataset and the Uniprot database to construct the negative dataset as described in the Methods section. After subtracting peptides containing non-standard amino acids the two datasets together contained 2148 peptides, 1157 AMPs and 991 non-AMPs. The ANN system (see the Methods section for further details) was evaluated using a training dataset containing 1074 peptides. The validation and testing datasets were populated each one with 537 peptides.

14The receiver-operating curves (ROC) obtained (Figure 4) show that the method is able to correctly classify peptides in the two groups considered. The overall accuracy of the method was 90%, as shown in the corresponding confusion plot (Supporting Information, Figure S1) and is similar in both the validation and testing datasets, meaning that there was no over fitting. The use of 50 hidden neurons was optimum for the system designed. Less than 20 neurons consistently reduced the overall accuracy (<85%) and more than 50 nodes did not improve the result but might promote over convergence and a loss of model generality.

15QSAR-based ANN methods described, using a set of 44 descriptors, obtained accuracy values near 80%. Our system can achieve a similar or even better accuracy using only 8 parameters. In any case, the comparison of accuracy values between the two methodologies has to be taken with caution, as different databases have been used. As the QSAR-based methods use two and three-dimensional structure estimations and similar descriptors for pI and hydrophobicity, we conclude that peptide aggregation could be of crucial importance in our classification system. Moreover, we have observed that structure and aggregation parameters have the most impact when deleted in the ANN analysis (data not shown).

16A support vector machine (SVM) approach has also been tested in order to classify antimicrobial and non-antimicrobial peptides. In this case, we have only obtained a 75% correct classification using a 5-degree polynomial kernel (Supporting Information, Figure S2). Thus, in our system, a SVM approach has been found to be less accurate than the ANN approach. It is possible that the high amount of data and the reduced space dimensionality could favour ANNs over SVMs.

17To go a step further in this study we have tested our model for antimicrobial activity prediction using two independent and non-redundant data sets. We have used the data published by Cherkasov et al. where the antimicrobial activity of different CAMEL variants was experimentally determined. The dataset employed is homogeneous, as all peptides were assayed in the same conditions and thus is of great value to test the parameters used in this study. From the 101 peptides described, we have used an ANN to compute the antimicrobial activity. The results obtained (Figure 5) show that our descriptors are closely correlated with antimicrobial activity (r2 = 0.72, q2 = 0.65).

18We have also inspected the database (named RANDOM database) described recently by Fjell et al. that contents a set of 189 peptides randomly synthesized and experimentally tested, again in uniform conditions. The results confirm (Figure 6), once more, a good correlation between our model and the experimental data (r2 = 0.85, q2 = 0.72).

19The successful results obtained both for peptide classification and antimicrobial activity prediction may suggest that our set of parameters is complete and can appropriately describe the antimicrobial mechanism of action of peptides.

20As pointed before, a major interest of the present method lies in the ability to connect these physicochemical parameters with a global overview on the peptide mechanism of action (Figure 7). It is widely known that most of AMPs act mainly at the membrane level, destabilizing the bilayer structure by creating pores (toroidal or barrel-stave pores) or modifying its permeability by a mechanism known as “carpet –like”. Both models promote the membrane depolarization and eventually the bacteria cell death.

21The first step is the membrane binding, driven by the electrostatic interactions between the positive peptide charges and the negative charges located in the phospholipid polar heads. It is known that, in certain conditions, some antimicrobial peptides can form aggregates in solution before they can interact with membranes,. Most of these aggregates have been described to be inactive because they lack the ability to effectively insert totally or partially in the bilayer and thus are unable to promote the cell depolarization,. For example, the aggregation of temporins prior to membrane interaction can avoid its activity,.

22The following steps in the antimicrobial mechanism involve membrane destabilization and diverge depending on the particular action exerted by AMPs. In a “carpet-like” mechanism, as observed in dermaseptin S and cecropin, formation of transmembrane structures is not necessary, but aggregation on the membrane surface and partial insertion is critical in order to promote destabilization and even bilayer micellization, as depicted in Figure 7. On the contrary, formation of toroidal pores, as observed for melittin, or barrel-stave pores, as observed for alamethicin, require a deep insertion in the membrane and a more ordered distribution of peptides across the bilayer.

23In summary, the results presented here suggest that AMPs can be distinguished from non-AMPs by their physicochemical properties. We have also shown that peptide aggregation propensity must be included in the antimicrobial mechanism of action of peptides in order to correctly describe (and predict) its antimicrobial capacity. In addition, we have demonstrated that our method is able to correlate sequence-derived physicochemical peptide properties with the antimicrobial activity and thus is a fairly good approach to a general model able to describe and predict antimicrobial activity.

24The CAMP database was used to construct the positive data set, containing 1157 peptides, where 95% of the peptides ranged from 10 to 50 residues. The Uniprot database was used to build the negative data set selecting peptides ranging between 10 and 50 amino acids length and filtered by Uniref50 in order to avoid over representative sequence similarities. None of the 991 selected peptides were reported as antimicrobial and/or toxic.

25The data described by Cherkasov et al. on antimicrobial CAMEL peptides (named CAMEL database) was used to assess the antimicrobial activity prediction. This database is composed of 101 peptides with experimentally tested antimicrobial potency. The data described by Fjell et al. (named RANDOM database) on different antimicrobial peptides was also independently used to test the method. This database is composed of 189 peptides after discarding inactive peptides.

26In vitro aggregation and secondary structure prediction were accomplished by using the TANGO software. Tango calculates the partition function of the conformational phase space assuming that every segment on the protein populates one state: random coil, β-turn, α-helix, α-helix aggregation and β-sheet aggregation. Therefore, TANGO software can predict aggregation in solution, taking into account only structural parameters determined by the peptide sequence.

27In vivo aggregation was computed using AGGRESCAN, an algorithm based on an amino acid aggregation-propensity scale derived from in vivo experiments and on the assumption that short and specific sequence stretches modulate protein aggregation. The algorithm can actually predict the aggregation propensity of peptides in the presence of cell material.

28The isoelectric point (pI) was computed using the Expasy reference values and the peptide hydrophobic mean character using the GRAVY scale (http://expasy.org).

29To study the differences in these parameters between AMPs and non-AMPs a two-tailed unpaired t-test analysis with a confidence interval of 95% has been used. All the parameters computed are considered to be independent as a correlation lower than 0.9 is observed between them in all the databases described above.

30Artificial neural networks (ANN) were computed using the Matlab software (Natick, MA). To predict antimicrobial peptides, a two-layer feed-forward network with sigmoid hidden and output neurons has been used. The network was trained with scaled conjugate gradient backpropagation. Hidden layer was populated with 50 neurons using 1074 peptides in the training database and 537 peptides on validation and testing databases. Both, matrix confusion plots and receiver-operative characteristic (ROC) curves were plotted for training, validation and testing databases.

31To predict the antimicrobial potency of CAMEL and RANDOM antimicrobial peptides, a two-layer feed-forward network with sigmoidal hidden neurons and linear output neurons was used. The network was trained with Levenberg-Marquardt backpropagation algorithm. Regression plots were computed for training, validation and testing databases. The results obtained were cross-validated using the leave-20%-out methodology (q2 values).

32Support vector machine (SVM) models were computed using Gist SVM (http://svm.sdsc.edu). A training database of 1611 peptides and a testing database, containing 537 peptides, have been used. As the model contains variables with a heterogeneous scale, the data was adjusted to give a 0 mean value and 1 variance value in order to enhance performance. A two norm soft margin has been used and different polynomial kernels were tested in order to increase accuracy.