MultiPep: a hierarchical deep learning approach for multi-label classification of peptide bioactivities

Identifying bioactivities of peptides is becoming increasingly important. Illuminating the biological activity of peptides will help shedding light on their impact on various processes and help revealing peptides that can be used for therapies against diseases and disorders. A number of biologically available peptides have already now proven efficient for use as drugs for various diseases, including diabetes, osteoporosis, hypertension, cancer, and a variety of endocrine disorders [ 1 , 2 ]. Though peptide-based drug development is still hampered by certain challenges, such as physiochemical instability and short half-life [ 3 , 4 ], it can be expected that the future with its technical advances will contain more therapies involving peptides [ 2 , 3 ]. At the same time, mass spectrometry-based peptidomics technologies are rapidly developing, allowing for the discovery of previously unknown molecules [ 5 , 6 ]. Although a subset of these peptides likely represent potent metabolic signaling molecules, the field currently lacks an efficient approach to scan for specific bioactivities, which further allow meaningful follow-up experiments for target validation. Thus, accurately predicting the bioactivity of peptides is important for identifying novel drug candidates as well as new and unknown peptides. Many tools have been developed for predicting the bioactivity of peptides. Table 1 shows a selection of important state-of-the-art binary classifiers that can group peptides into different bioactivity classes. Table 1: State-of-the-art binary classifiers for bioactivity prediction Prediction class Algorithm type(s) Name/description of tool General bioactivity Neural network PeptideRanker [ 7 ] Antimicrobial peptides Neural network Convolutional long short-term memory neural network [ 8 ] Deep-AmPEP30 [ 9 ] Anticancer peptides Support vector machine mACPpred [ 10 ] Neuropeptide peptides Various PredNeuroP [ 11 ] Toxic peptides Support vector machine ToxinPred [ 12 ] Hemolytic peptides Various HLPpred-Fuse [ 13 ] Antioxidant peptides Neural network AnOxPePred [ 14 ] Though many of the peptide bioactivity predictors presented in Table 1 offer a wealth of additional information next to a prediction score [ 9 , 13–15 ], a common denominator for all of them is that they are binary classifiers. For example, the classifier PeptideRanker can predict whether a peptide has a bioactivity or not, and Deep-AmPEP can distinguish short antimicrobial peptides (AMPs) from short non-AMPs. This automatically implies that peptides either have one or none bioactivity, which is not in line with reality [ 15–25 ]. Getting an overview of peptides’ bioactivities is important when seeking to unravel therapeutics-relevant peptides. Further, an overview of peptides’ bioactivity profiles is vital when analyzing mass spectrometry-based peptidomics data, where a key step is to rank peptide candidates for downstream functional studies. A few tools that can predict multiple bioactivity classes simultaneously exist in the literature. PEPred-suite [ 26 ] uses eight random forest (RF) models to predict how peptides belong to eight different classes. Peptipedia [ 27 ] is another tool that, among other things, can predict different bioactivities of peptides by using 44 RF models trained on data from a large number of peptide databases. A combination of several machine learning algorithm types has been utilized for the prediction of peptide bioactivities ( Table 1 ). However, through the years, deep learning-based models have delivered state-of-the-art results in the field of biological sequence analysis [ 9 , 8 , 28 , 29 ]. And though much have been achieved already regarding peptide bioactivity prediction ( Table 1 ), the possibilities with deep learning are endless, and allows for continuous improvements. Here, we present MultiPep, a deep neural network multi-label classifier that can predict the bioactivity of peptides based on 20 bioactivity classes. The tool can take peptides of length 2–200 amino acids that can consist of natural amino acids (A, R, N, D, C, Q, E, G, H, I, L, K, M, F, P, S, T, W, Y, V). MultiPep utilizes convolutional neural networks [ 30 ] for predicting the peptide class belonging, and can classify peptides into zero or more bioactivity classes based on their intrinsic amino acid patterns. The architecture of MultiPep is inspired by a dendrogram from a hierarchical clustering of our defined bioactivity classes’ mutual overlaps ( Fig. 1 ). An overlap occurs when two or more classes contain the same peptides. The 20 output nodes of the network are divided into “network class-clades,” which have outputs that are agglomeratively combined until two single binary predictions are made. The architecture ensures that extra penalties are added while training, if a sequence is predicted to be in, for example, a wrong network class-clade. Figure 1: Dendrogram template and overall architecture of the convolutional neural network. ( A ) Dendr

Based on the downloaded peptides and their annotations, 20 bioactivity classes were defined (see Supplementary Information and Supplementary Table S1 ). The complete list of classes and how the used databases have contributed to their generation can be seen in Table 2 . Some database class contributions were based on overlaps. For instance, the class “ACE inhibitor” stems only from the database BIOPEP-UWM, however, the database LAMP2 contributed with 39 peptides, because it had peptides from another class that overlapped with the peptides from the “ACE inhibitor” class. Table 2: Classes and databases. Classes CAMP3 LAMP2 APD3 SATPdb DBAASP BIOPEP-UWM PeptideDB NeuroPedia CancerPDD BioDADPep NeuroPep Total class size 0 ACE inhibitor 1 39 1 687 29 973 4 1 1 65 3 973 1 Antibacterial 2104 11 516 2858 4533 10310 512 1059 16 265 31 37 13 538 2 Anticancer 245 1575 419 1177 1919 104 121 1 440 14 6 2426 3 Antidiabetes 11 46 17 125 41 189 9 3 2 1091 3 1112 4 Antifreeze 0 0 0 0 0 0 192 0 0 0 0 192 5 Antifungal 1457 4747 1955 2818 4172 219 559 11 197 22 27 5342 6 Antihypertensive 1 73 6 1664 51 783 10 3 5 107 8 1672 7 Antimicrobial 1882 13 446 2125 8887 3681 239 2544 15 225 9 57 14 362 8 Antioxidative 10 38 28 81 35 649 6 4 1 15 5 675 9 Antiparasite 133 479 190 342 388 46 94 5 20 6 10 503 10 Antivirus 735 4219 990 3918 1611 90 196 8 36 6 14 4500 11 Cellcellsignaling 11 38 15 679 30 22 378 392 0 3 391 682 12 Cytokines_ growthfactors 46 66 64 24 30 11 3729 0 1 0 1 3760 13 Dipeptidyl peptidase inhibitor 0 58 2 169 50 459 6 4 1 158 6 459 14 Drugdelivery 11 108 11 1484 96 17 44 29 5 1 45 1484 15 Hemolytic 183 1299 342 1279 1045 61 111 0 51 1 2 1339 16 Neuropeptide 18 70 26 473 33 108 2101 552 2 4 3822 3926 17 Opioid 0 5 0 27 1 117 12 9 2 0 16 117 18 Peptidehormone 22 149 28 499 33 34 6943 495 1 2 2077 6943 19 Toxic 411 1746 523 3840 1416 310 2404 2 76 1 10 5793 Bioactivity classes were only created if 100 or more peptides were available for that class, and if the class had more than two peptides that were unique for that specific class. We believed that if the classes did not live up to these criteria, the classes would be to information-poor for a classifier to use. Some of the classes had major overlaps, while other classes had none ( Supplementary Fig. S2 ). The largest class was the Antibacterial class with 14 362 peptides and the smallest was the Opioid class with 117 peptides. Analyses to see how the defined bioactivity classes differed in terms of overall peptide lengths and amino acids distributions were made ( Supplementary Figs S3 and S4 ). Some of the bioactivity classes are very general (antimicrobial, peptide hormone, and antihypertensive) while other classes are more specific (ACE inhibitor, opioid, and dipeptidyl peptidase inhibitor). No negative class with peptides without bioactivity was defined, as the many different classes generated negative classes for each other. If required, the applied databases can be used to acquire more information about the classes.

The CNNs from the different network class clades have the same architecture, which is shown in Fig. 2 . Each of the CNNs are connected to the input layer. The input layer passes n input examples as a 3D tensor of size n × 4000 × 1 to seven 1D convolutional layers within the different CNNs. The input data are passed as a 3D tensor, as this is required by Keras’ 1D convolutional layers. See the section “Generating data for training and testing” for a more detailed explanation of the input data size. Figure 2: Architecture of network class-clade convolutional neural networks. All layers with dark-gray backgrounds (except for the input layer) have weights, whereas the layers with white backgrounds are mathematical operations or dropout layers. Below the name of each layer (except for the dropout layers), the sizes of the layers’ output tensors are written. Arrows show the flow of information and how the layers are connected. What constitutes the convolutional neural network (CNN) is encapsulated by a light-gray box. “Conv4” means a 1D convolutional layer with a kernel size of four one-hot-encoded amino acids. The same logic applies to the remaining convolutional layers. “Dense500” means a dense layer with 500 nodes. The convolutional layers have 40 kernels with no bias parameters and they convolve over 4, 6, 10, 16, 22, 30, and 40 one-hot encoded amino acids, respectively. Their strides are 20, which is the length of 1 one-hot encoded amino acids. The convolutional layers perform valid convolutions. The 3D tensors that are outputted by the convolutional layers are max pooled along their second axes. The max pooled convolutional output values are concatenated, such that a ( n × 280 ) tensor is created. Each of the n input examples are then normalized by a division of their summed absolute values. This is done to ensure that the network focuses on amino acid patterns and does not predict based on sequence length. Dropout with a rate = 0.2 is applied on the normalization layer. The normalization layer is connected to a chain of three dense layers with 500 nodes where dropout is applied on each of them (rate = 0.5). See the section “Training settings and network initialization” to read more about the used regularization techniques. In the top of Fig. 2 , the CNN is connected to an output layer. This could, for instance, be “CNN1” that is connected to “Output 4_1” in Fig. 1B . All layers with weights that are not output layers consist of Parametric Rectified Linear Units (PReLU) [ 36 ].

The start parameters of the convolutional layers were initialized using orthogonal weights. The initial weights of the dense non-output layers were sampled from a uniform distribution with min = - 1 and max = 1 . The initial weights of the output layers were sampled from a uniform distribution with min = 0.001 and max = 0.05 . The bias parameters of the dense layers and output layers were initialized as zeros. The Adam [ 37 ] update function was used with a learning rate of 0.0005. The remaining hyper-parameters for the Adam update function were set to the Keras default values. The batch size was 128 and dropout with a rate of 20% and 50% on the normalization layer and the dense layers, respectively, was applied ( Fig. 2 ). As a final regularization technique, early stopping after 20 epochs was used. All data were shuffled before an epoch was started. To acquire optimal models, weights for the network class clades were saved whenever their performances on the validation set had improved. Additionally, the weights of the entire network were saved whenever the overall performance had increased. The early-stopping count-down was set to zero whenever the parameters of a “network class clade” were saved. The performance was measured in terms of loss.

The entire loss function for a single output node given all instances of mini-batch is defined as (7) Loss = - 1 N ∑ i = 1 N [ y i · log ⁡ y ^ i + 1 - y i · log ⁡ 1 - y ^ i ] + MC C loss , where N is the number of examples in the mini-batch. The loss function can be seen as two-step function where the MC C loss values are calculated for the entire batch in the first step and then added to the BCE loss values, which then are averaged. During the updating steps while training, the loss is summed across all output nodes.

A Numpy [ 41 ] version of MultiPep based on the trained weights was implemented. It is available on https://agbg.shinyapps.io/MultiPep/ . Users can choose between using the average of all CV models’ prediction as the final output or to get the max of all CV models’ predictions of the different classes as the final output. This is true for the webtool and the stand-alone-program. The webtool rounds the predictions to three decimals, while the stand-alone-program provides a higher number of decimals. The stand-alone-program is available at https://github.com/scheelelab/MultiPep . If predicting more than 500 in one batch of peptide sequences, we recommend using the stand-alone-program.

Architecture rationale We designed the architecture of MultiPep with inspiration from a dendrogram from a hierarchical clustering of our defined bioactivity classes’ mutual overlaps. The performed clustering of the max-normalized class overlap values provided a proxy for the classes’ peptide similarities. We hypothesized that the network would make more exact predictions when we grouped similar classes together and connected them to the same CNN. The CNNs would then be forced to learn how to distinguish between similar peptide patterns using the same network parameters. Further, the hierarchical structure of MultiPep ensures that an extra penalty is added while training, if a peptide is predicted to be, for example, in a wrong class clade or if the output layer of the network class clade produces only false negatives or positives. This happens because a, for example, false positive in an erroneous class clade will propagate through all connected upper output layers and thus produce more loss via these layers.

Using data derived from multiple peptide databases, we conducted 10-fold CVs and tested the models on their associated test sets. We trained several models using two different loss functions. The neural network architecture and all hyper-parameters and regularization settings were always the same. First, we trained models using our loss function with the MC C loss penalty, where we applied our training scheme that saved the parameters of the best “network class clades” whenever performance on the validation sets increased. Secondly, while training these models, we saved the parameters of the entire network whenever the overall performance had increased. Thirdly, we ran a 10-fold CV using the weighted BCE loss function. We used the training scheme, where the parameters of the “network class clades” were saved whenever their performances increased and the training scheme where we saved all network parameters whenever the overall performance had increased. This produced 4 × 10 different models, which we used for a comparative study to find the best way of training the network. Tables 4 and 5 show the mean and standard deviations of the performances of the trained models on the CV test sets using MCC and the F1 score as performance metrics. In the tables, the mean values have been compared and the standard deviations are stated to provide readers with an overview of the robustness of the models. Similar tables showing the performance on the test sets, but using precision, recall, and accuracy as performance metrics are available in the Supplementary Information ( Supplementary Tables S2–S4 ). Additionally, plots of receiver operating characteristic (ROC) curves of MultiPep models’ predictions of the test sets can be found in the Supplementary Information (Supplementary Figs S5–S14). Tables showing the performance on the validation sets can as well be found in the Supplementary Information ( Supplementary Tables S5–S9 ). The classification threshold was 0.5 and all mean values in the tables have been rounded to three decimals, whereas all standard deviation values have been rounded to two decimals. We did not use the upper layers’ predictions for the comparisons, as we were only interested in finding how the models performed on the bioactivity classes. The performances on the test sets and validation sets of the individual CV models trained using the “save the individual network class-clade” training scheme can be found in Supplementary Tables S10–S29 . Table 4: Mean and standard deviation of MCC of CV models on test sets Bioactivity class/output name BCE + MC C loss BCE + MC C loss − overall lowest loss Weighted BCE Weighted BCE—overall lowest loss Output 1_1 0.839 ± 0.01 0.843 ± 0.01 0.803 ± 0.01 0.799 ± 0.01 Output 1_2 0.856 ± 0.01 0.859 ± 0.01 0.81 ± 0.02 0.847 ± 0.01 Output 2_1_1 0.782 ± 0.01 0.788 ± 0.01 0.764 ± 0.02 0.778 ± 0.01 Output 2_1_2 0.849 ± 0.01 0.853 ± 0.01 0.813 ± 0.01 0.802 ± 0.01 Output 2_2_1 0.895 ± 0.01 0.894 ± 0.01 0.875 ± 0.01 0.878 ± 0.01 Output 2_2_2 0.826 ± 0.01 0.83 ± 0.01 0.751 ± 0.02 0.811 ± 0.01 Output 3_1 0.798 ± 0.01 0.798 ± 0.01 0.711 ± 0.03 0.787 ± 0.02 Output 3_2 0.766 ± 0.03 0.765 ± 0.02 0.749 ± 0.02 0.751 ± 0.02 Hemolytic 0.531 ± 0.04 0.539 ± 0.03 0.514 ± 0.03 0.52 ± 0.03 Toxic 0.785 ± 0.01 0.789 ± 0.01 0.768 ± 0.02 0.782 ± 0.01 Antimicrobial 0.675 ± 0.01 0.674 ± 0.01 0.617 ± 0.01 0.587 ± 0.01 Antivirus 0.611 ± 0.02 0.614 ± 0.02 0.593 ± 0.01 0.569 ± 0.03 Antiparasite 0.281 ± 0.05 0.173 ± 0.1 0.299 ± 0.06 0.289 ± 0.07 Anticancer 0.51 ± 0.03 0.488 ± 0.02 0.513 ± 0.03 0.47 ± 0.03 Antibacterial 0.702 ± 0.01 0.703 ± 0.01 0.659 ± 0.01 0.644 ± 0.01 Antifungal 0.487 ± 0.02 0.471 ± 0.02 0.411 ± 0.03 0.335 ± 0.06 Cell–cell signaling 0.553 ± 0.02 0.552 ± 0.02 0.544 ± 0.03 0.548 ± 0.04 Neuropeptide 0.78 ± 0.02 0.771 ± 0.02 0.74 ± 0.02 0.757 ± 0.03 Peptide hormone 0.845 ± 0.01 0.844 ± 0.01 0.82 ± 0.01 0.828 ± 0.01 Antifreeze 0.966 ± 0.04 0.976 ± 0.03 0.976 ± 0.03 0.977 ± 0.02 Cytokines/growth factors 0.932 ± 0.01 0.936 ± 0.01 0.851 ± 0.02 0.923 ± 0.01 Antioxidative 0.467 ± 0.06 0.461 ± 0.06 0.378 ± 0.09 0.483 ± 0.07 Drugdelivery 0.598 ± 0.05 0.588 ± 0.04 0.248 ± 0.13 0.556 ± 0.05 Opioid 0.701 ± 0.11 0.676 ± 0.11 0.67 ± 0.11 0.733 ± 0.1 ACE inhibitor 0.511 ± 0.02 0.511 ± 0.02 * 0.502 ± 0.03 0.498 ± 0.03 Antihypertensive 0.668 ± 0.03 0.664 ± 0.02 0.652 ± 0.02 0.655 ± 0.02 Antidiabetes 0.596 ± 0.03 0.58 ± 0.03 0.587 ± 0.03 0.587 ± 0.03 Dipeptidyl peptidase inhibitor 0.564 ± 0.05 0.564 ± 0.04 0.557 ± 0.03 0.559 ± 0.08 Average of bioactivity classes: 0.638 0.629 0.595 0.615 Best total 8 7 2 3 Notes : Bold values indicate the model with the best performance. The asterisk symbols indicate that more than three decimals are needed to reveal the highest values. At the two bottom rows, “Best total” shows the number of times the average of the models are the best and “Average of bioactivity classes” shows the average of the columns above, but only for the bioactivity classes. Table 5: Mean and stand

As an additional analysis, we compared MultiPep with PeptideRanker. PeptideRanker is a binary classifier that can predict general bioactivity of peptides. Thus, we found it interesting to compare MultiPep against this tool. Using PeptideRanker, we predicted all peptides of the test set of the model with the lowest loss on the test trained using our “save the individual network class-clade” training scheme ( Supplementary Tables S15 and S25 ). The set contained 4585 peptides and we did not consider whether any of the used peptides from out test set were part of the training data of PeptideRanker. For both MultiPep and PeptideRanker, instead of using peptides with no bioactivity, we used all peptides of the test set as true positives and found the error of the tool’s predictions. We calculated the error of PeptideRanker’s predictions by ∑ i = 1 N 1 - round p N , where N is the number of examples in the test set, p is the predictions, and round ( ) is a function that rounds values to their nearest integer. The rounding function implies that the threshold for the predictions was 0.5. For MultiPep, we took the predictions on the Output_1 output layer and found how they diverged from the labels of that layer. We calculated ∑ i = 1 N ∑ j = 1 k | y j - round ( p j ) | N , where N is the number of examples in the test set, is an element of the row vector y with labels, is an element of the row vector p with predictions, and round ( ) is a function that rounds values to their nearest integer. Also, to make a more direct estimate of whether MultiPep can assign a prediction above 0.5 to bioactive peptides, we calculated an additional error percentage by ∑ i = 1 N 1 - round max ( p ) N , where N and p are the same as above and max ( ) is a function that finds the maximum value of the row vector p . With this approach, we calculated the prediction errors ( Table 7 ). This analysis shows that MultiPep, with a low prediction error, successfully identifies a higher number of bioactive peptide from a list of peptides with known bioactivity. Table 7: Prediction error of MultiPep compared with PeptideRanker Program names Rounded prediction error MultiPep (Output_1, max ) 0.074 MultiPep (Output_1) 0.140 PeptideRanker 0.299 Note : Final prediction errors rounded to three decimals.

We next aimed to compare our MultiPep deep learning models with linear logistic regression models. Therefore, we trained 10 logistic regression models using our peptide data and our 10-fold CV scheme. We used the logistic regression functionalities provided by the Python package Scikit-learn [ 40 ]. We initialized the following hyper-parameters as follows: random_state = 1234, max_iter = 1000, and class_weight = “balanced.” The remaining hyper-parameters were set to their default values. The peptides were one-hot encoded before introduced to the logistic regression models. When comparing the performances of the two model types, it can be seen that MultiPep outperforms the logistic regression models ( Tables 4 and 5 and Supplementary Tables S2–S9 , S33 , and S34 ). Nevertheless, the logistic regression models seem to have a generally better recall performance.

In this work, we construct and demonstrate the utility of a new multi-label classifier, MultiPep. Moreover, we test our novel loss function where BCE synergistically is merged with a customized version of the MCC function. Our results suggest that using our loss of function, instead of weighted BCE, is beneficial when training using large multi-label datasets containing an imbalanced class–size distribution. Whether this is a general tendency needs to be verified. Additionally, we use an innovative neural network architecture and a new training scheme to find optimal network parameters. We demonstrate that our novel training scheme on average produces models that are better than when saving all parameters of a network simultaneously whenever performance has increased. Further, we show that MultiPep on our data surpasses state-of-the-art peptide bioactivity classifiers, and that it can predict the general bioactivities of FDA-approved therapeutic peptides. It should be noted that the comparisons with the other tools were made using our data only; thus, the results do not suggest that MultiPep is a better tool as such. If the classifiers were trained on the same data the outcome might have been different. The results demonstrate that MultiPep is better on the currently used data. This was consistent with our goal for the comparisons, namely to show how the different pre-trained models performed on a previously unseen dataset. MultiPep is designed to be a tool that can grant scientists an overview of peptides’ bioactivities. It does not offer a wealth of additional predictive information like other peptide bioactivity classifiers [ 9 , 12–14 ]. Thus, we suggest that MultiPep can be used in a pipeline, where the predictions of MultiPep can be further evaluated using tools that have more narrow and specialized classification foci. All in the interest of the progress of research. MultiPep is not the only tool that simultaneously can predict more bioactivity classes of peptides. Previous tools in the literature, PEPred-Suite and Peptipedia, also predicts multiple classes for peptides. The Peptipedia models performed rather poorly on the applied test-subset. As implied above, the Peptipedia models were trained on other data, and their bioactivity classes were probably defined differently. For the comparison, we deemed that a range of MultiPep bioactivity classes and Peptipedia bioactivity classes were comparable ( Supplementary Table S32 ). We assumed that when the class names were similar or identical, then predictions of models trained on these classes would be somewhat overlapping. This did not seem to be the case. So, are the MultiPep and Peptipedia bioactivity classes just incomparable and of different origin? Though MultiPep’s performance was better, the binary classifiers that were tested against MultiPep ( Table 6 ) performed reasonably on our test subsets with our defined bioactivity classes. This indicates that MultiPep and these models agree on the definition of the used classes and that these bioactivity classes and included peptides contain essential class-specific features. Altogether, this suggests that the problem does not lie with the bioactivity classes defined in the data presented in this work. In addition, the peptide data used by MultiPep and Peptipedia stem from a number of identical databases [ 27 ]. Therefore, it is very unlikely that MultiPep’s and Peptipedia’s defined bioactivity classes are of completely different origin. The difference in performance might rather be explained by the algorithms used to create the tools. PEPred-Suite and Peptipedia utilize individual RF models to predict bioactivity classes of peptides. MultiPep on the other hand uses a single deep learning-based model with sub-models to predict bioactivities of peptides. In other words, PEPred-Suite and Peptipedia solve many binary classification problems, whereas MultiPep turns it into a multi-label classification problem. Another difference is that PEPred-Suite and Peptipedia use elaborate techniques to encode the peptide sequences, and MultiPep uses a simple one-hot encoding. Many peptide encoding techniques exist and peptide encoding in general is a field that is gaining a lot of attention [ 77 , 78 ]. PEPred-Suite uses adaptive feature representation strategy, where they, among other things, use 10 feature encoding algorithms, which together efficiently capture local and global compositional information and well as position-specific residue information and physiochemical information [ 26 ]. Peptipedia encodes peptides using representations of physicochemical properties and transforms them using Fourier transforms [ 27 , 79 ]. Although some general selection rules have been suggested, it is difficult to find a single universally optimal peptide encoding technique [ 77 , 78 ]. As it has been found that deep learning models require little encoding for the classification process [ 77 ], we chose to use a simple and, in

All data in this study are available either through https://github.com/scheelelab/MultiPep or via direct communication with the authors. The webtool is available at https://agbg.shinyapps.io/MultiPep/ .

C.S. and A.G.B.G. conceived and designed the project. MultiPep was implemented, trained, and tested by A.G.B.G. The MultiPep stand-alone program and webtool were implemented by A.G.B.G. All analyses and comparisons were carried out by A.G.B.G. All authors read and contributed to the manuscript.

bpab021_Supplementary_Data Click here for additional data file.