Prediction of Antimicrobial Peptides Using Deep Neural Networks

Ümmü Gülsüm Söylemez

, Malik Yousef

and Burcu Bakir-Gungor

Department of Software Engineering, Faculty of Engineering, Muş Alparslan University, Muş, Turkey

Department of Information Systems, Zefat Academic College, Zefat 13206, Israel

Department of Computer Engineering, Faculty of Engineering, Abdullah Gul University, Kayseri, Turkey

Keywords: Antimicrobial Peptides, Deep Neural Networks, Physicochemical Properties.

Abstract: Antimicrobial peptides (AMPs) are crucial elements of the innate immune system; and they are effective

against bacteria that cause several diseases. These peptides are investigated as potential alternatives of

antibiotics to treat infections. Since wet lab experiments are expensive and time-consuming, computational

methods become crucial in this field. In this study, we suggest a computational technique for AMP prediction

using deep neural networks (DNN). We trained a DNN classifier using physicochemical features that include

a sequential model; and evaluated the model with 10-fold cross-validation on a benchmark dataset. We

compared our method with other machine learning approaches and demonstrated that the method we

developed generates higher performance (accuracy: 92%, precision: 92%, recall: 93%, f1: 93%, AUC: 98%).

In our experiments, we have realized that there is a strong positive correlation between the ‘Normalized

Hydrophobic Moment’ feature and ‘Angle Subtended by the Hydrophobic Residues’ feature; and strong

negative correlations between ‘Normalized Hydrophobicity’ feature and ‘Disordered Conformation

Propensity’ feature, and between ‘Amphilicity Index’ - ‘Disordered Conformation Propensity’ features. We

believe that the approach we proposed could guide further experimental studies and could facilitate the

prediction of other types of AMPs having anticancer, antivirus, antiparasitic activities.

1 INTRODUCTION

Antimicrobial peptides (AMP), which are known as

crucial elements of the innate immune system, are

shown to be effective against a variety of pathogenic

microorganisms, including bacteria, viruses, fungi,

parasites, and others (Erdem Büyükkiraz & Kesmen,

2022). In recent years, AMPs have drawn attention as

an alternative to chemical antibiotics due to the

developing resistance of microbial infections

(Thomas et al., 2010). AMPs have crucial roles

including quick microbial elimination and later

immunological regulation (Wang, 2014). These

effects come into the scene since AMPs cause

multiple bacterial harm, such as disruption of

bacterial membranes, inhibition of protein, or

interaction with specific intracellular targets (Bahar

& Ren, 2013; Malmsten, 2014). As a result, AMPs

became popular as novel medications.

https://orcid.org/0000-0002-6602-772X

https://orcid.org/0000-0001-8780-6303

https://orcid.org/0000-0002-2272-6270

Numerous computational techniques have been

proposed recently to advance the identification and

synthesis of antimicrobial drugs, and to accelerate the

candidate selection (Hammami & Fliss, 2010).

Sequence-based models have been trained using

machine learning algorithms as the primary way to

distinguish AMPs from non-AMPs. For example,

Thomas et al. used supervised learning methods such

as Random Forest (RF), Support Vector Machines

(SVM) and Discriminant Analysis (DA) for

prediction of AMPs based on physico-chemical

features (Thomas et al., 2010). Their prediction

models have the accuracy values of 93.2% for RF,

91.5% for SVM, and 87.5% for DA. Lata et al.

utilized an SVM based model using amino acid

composition as features (Lata et al., 2010). Their

model achieved 92.14% accuracy for the antibacterial

peptide classification problem. Joseph et al.

developed an algorithm called ClassAMP, a

combination of RF and SVM to predict the ability of

188

Söylemez, Ü., Yousef, M. and Bakir-Gungor, B.

Prediction of Antimicrobial Peptides Using Deep Neural Networks.

DOI: 10.5220/0011690000003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 3: BIOINFORMATICS, pages 188-194

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

a sequence to have antibacterial, antifungal or

antiviral mode of action (Joseph et al., 2012). Several

other computational tools have been developed for

this purpose (Lata et al., 2007; Thakur et al., 2012;

Xiao et al., 2013).

During recent years, deep learning techniques

have been used for antimicrobial peptide prediction.

Bhadra et al. developed an approach for sequences

shorter than 30 amino acids (Bhadra et al., 2018).

They achieved 77% accuracy with their deepAMP

method, which combined a convolutional neural

network with an ideal feature set with reduced amino

acid composition. In addition, they evaluated their

outcomes using RF and SVM algorithms. The SVM

model reached 72% accuracy while the RF reached

75% accuracy. Schneider et al. presented a

feedforward neural network using self-organizing

maps as input layers on AMP data and achieved 92%

accuracy (Schneider et al., 2017). Lin et al. developed

a method called AI4AMP using 6 different

physicochemical properties for encoding and a deep

learning model (Lin et al., 2021). Witten et al.

proposed a convolutional neural network (CNN)

method for regression of AMP (Witten & Witten,

2019). Minimum Inhibition Concentration (MIC)

values are used for regression. k-Nearest Neighbour

(kNN) and Ridge Regression models are used for

comparison with CNN. They showed that CNN

model has better root mean square error performance

(0.501) than other models used. Veltri et al. utilized a

deep neural network model based on convolution and

lstm layers using sequence to vector model for input

layers (Veltri et al., 2018). They showed that their

model has better performance when compared with

state-of-the-art models.

In this study, we propose a deep neural network

prediction model based on physico-chemical

properties of sequences for identifying AMP. Various

classification models are applied on the dataset, and

the outcomes are compared using various

performance metrics.

2 METHODS

2.1 Dataset

The independent dataset used in this study was

obtained from (Xiao et al., 2013). The dataset consists

of 920 AMP sequences and 920 non-AMP sequences,

which forms a two-class data set.

2.2 Sequence Representation

The key to solving operational difficulties effectively

is understanding how to formulate peptides

mathematically into the fixed length features to create

a robust AMP classification model. Table 1 shows

example sequences included in the existing dataset.

Each sequence is represented with 11 features

obtained from DBAASP web server and labeled as

“pos” if the sequence is AMP and “neg” if the

sequence is Non-AMP.

2.2.1 Generation of Physicochemical

Features

Normalized Hydrophobic Moment, Normalized

Hydrophobicity, Isoelectric Point, Penetration Depth,

Tilt Angle, Disordered Conformation Propensity,

Linear Moment, Propensity to in vitro Aggregation,

Angle Subtended by the Hydrophobic Residues,

Amphiphilicity Index, Propensity to PPII coil are

used as physicochemical features. DBAASP

(Vishnepolsky et al., 2018) web server is used to

calculate these features.

2.3 Deep Neural Network

Deep Neural Network is a machine learning method

that allows us to train a model and to predict outputs

for a given dataset. The artificial intelligence model

can be trained using both supervised and

unsupervised learning techniques. Neurons make up

artificial neural networks, similar to the human brain.

All neurons are linked together and have an impact on

the result. There are three layers that make up

neurons:

1. Input Layer; 2. Hidden Layer(s) and 3. Output Layer

The first hidden layer receives the input data from

the input layer and transmits it. On our inputs, hidden

layers run mathematical computations. Choosing the

amount of hidden layers and neurons for each layer is

one of the challenges in creating neural networks. The

output layer returns the output data.

Dense layer allows neurons from one layer to be

connected to the next layer as an input.

Batch Normalization (BN) is an algorithmic

technique that speeds up and improves the stability of

Deep Neural Network’s training. BN is a

normalization technique used in multilayer deep

neural networks to reduce the covariance between

hidden layers. The input of each layer is used by

normalizing the output vector of the previous layer.

Activation function determines whether or not a

neuron should be activated by generating a weighted

Prediction of Antimicrobial Peptides Using Deep Neural Networks

189

Table 1: Representation of the physico-chemical characteristics of AMP and Non-AMP peptides that are included in the

dataset, obtained from DBAASP (Vishnepolsky et al., 2018).

Sequence

Norm.

Hyd.

Moment

Norm.

Hydrophobicity

Isoelectric

Point

Penet.

Depth

Tilt

Angle

Dis. Conf.

Propensity

Linear

Moment

Prop. to in

vitro

Aggregation

Angle Subt.

by the Hydr.

Residues

Amph.

Index

Prop. to

PPII coil

class

MPIAQIHILEG

RSDEQKETLIR

EVSEAISRSLD

APLTSVRVIIT

EMAKGHFGIG

GELASKVRR

0.42 -0.53 7.66 30 110 0.20 0.15 4.77 20.00 0.63 0.97 neg

GTLPCESCVWI

PCISSVVGCSC

KSKVCYKN

0.23 -0.70 7.83 30 44 0.33 0.39 12.84 40.00 0.81 1.14 pos

TPCGESCVYIP

CISGVIGCSCT

DKVCYLN

0.27 -0.99 3.94 30 136 0.49 0.32 7.67 60.00 0.52 1.11 pos

IWGIGCNP 0.93 -1.27 3.50 14 15 0.42 0.56 0.79 110.0 0.87 1.05 ne

… …

total and then including a bias with it. The activation

function's objective is to add non-linearity to a

neuron's output.

2.4 Machine Learning Models for

Comparison

In this study we have considered different Machine

Learning algorithms for comparisons. The Support

Vector Machines (SVM) find the optimal hyperplane

based on the super vectors from each class (Cortes &

Vapnik, 1995). k-Nearest Neighbour is defined as the

classification of the data that has not yet been

assigned to a class by setting it in the optimal class

based on the distance which is calculated by

comparing the data of the unknown class with the

other data in the training set (Fix & Hodges, 1989).

The bagging approach seeks to retrain the basic

learner by creating new training sets from an old

training set. By using estimators on the bootstrapped

samples collected from the original dataset, an

ensemble is created (Breiman, 1996). Gradient

Boosting algorithm is a machine learning technique

that creates prediction models similar to decision

trees for regression and classification problems

(Friedman, 2002). A decision tree is built using the

data set as it is, and a new decision tree is created

based on its errors. Thus, a large number of decision

trees are created. The gradient boosting technique

gives the final decision by finding the sum of the

decisions made by all these trees.

2.5 Model Construction

As the details are illustrated in Figure 1, a deep neural

network algorithm is utilized to classify a sequence as

AMP or non-AMP. In our experiments, we used 10-

fold cross validation. We use 90% of data as training

and 10% as testing. For the first two activation

functions ‘relu’ and for the last one ‘sigmoid’ are

applied for classification. ‘Binary Cross Entropy’ is

used as a loss function and ‘adam’ is applied as an

optimizer.

2.6 Performance Metrics

Several performance evaluation criteria, including

accuracy, precision, recall, AUC and F1 measure,

were computed to evaluate the performance of our

models. These measures are defined as follows:

(1)

(2)

(3)

(4)

where TP stands for “true positive”, TN for “true

negative”, FP for “false positive”, and FN for “false

negative”.

BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms

190

Figure 1: The architecture of the network consists of three blocks. First block consists of a dense layer, a batch normalization

layer, an activation layer and it is connected to the following one. The second block does not have a batch normalization layer

and it is connected to the previous and following blocks. The last block consists of a dense layer and an activation function

layer and it is connected to the previous block and output section.

3 RESULT

In our study, we have used 11 physicochemical

features that we mentioned above. We applied

different machine learning methods to classify each

peptide according to being AMP or not. We used

different performance metrics such as accuracy,

precision, recall, F1 measure, Area Under Curve

(AUC) for comparison. As shown in Table 2, our

methods achieve higher performances for all above-

mentioned metrics. Figure 2 represents the AUC

values for all tested classifiers. Our prediction method

yielded the best AUC result with 98%. In this study

we used a 10 fold cross validation technique;

calculated physicochemical properties and used them

as features. Other studies that use the same dataset do

not use these features and they use different peptide

properties (Meher et al., 2017; Veltri et al., 2017;

Xiao et al., 2013). They also apply different cross

validation techniques. For these reasons no

comparative evaluation could be made with these

above-mentioned studies.

Table 2: Comparison of our model with different machine

learning algorithms on the dataset in use.

Model Accuracy Precision Recall F1 AUC

SVM 0.77 0.71 0.93 0.80 0.90

kNN 0.80 0.87 0.79 0.80 0.86

Bagging 0.84 0.88 0.81 0.82 0.91

Gradient

Boostin

0.87 0.90 0.88 0.87 0.97

DNN

Model

0.92 0.92 0.93 0.93 0.98

Figure 2: Comparison of classifiers based on Area Under

Curve (AUC) metric.

The Pearson correlation values between all feature

pairs have been calculated using the Python Seaborn

Library in order to assess the pairwise correlations of

the features. These relations are depicted in Figure 3.

According to the correlation heatmap (shown in

Figure 3), there is a strong positive correlation

between ‘Normalized Hydrophobic Moment’ and

‘Angle Subtended by the Hydrophobic Residues’

features; and strong negative correlations between

‘Normalized Hydrophobicity’ and ‘Disordered

Conformation Propensity’ and between ‘Amphilicity

Index’ and ‘Disordered Conformation Propensity’

features. Except these, there are no other significant

correlations found between other feature pairs.

3.1 Feature Scoring and Feature

Ranking

Since deep neural networks do not allow us to

calculate importance scores for the features, we refer

to the Gradient Boosting model, which generated the

second highest performance in our experiments.

Prediction of Antimicrobial Peptides Using Deep Neural Networks

191

Hence, we scored each feature based on the Gradient

Boosting model. We have realised that Angle

Subtended by the Hydrophobic Residues, Propensity

to in vitro Aggregation and Isoelectric Point are more

important features than others when we examine the

score for each feature.

Figure 3: Correlation heatmap of physicochemical features, extracted from dataset in use.

Figure 4: Feature scoring according to their importances in classification using Gradient Boosting model.

BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms

192

4 DISCUSSION

Antimicrobial peptides (AMPs) are crucial elements

of the innate immune system and they are efficient

against bacteria that cause several diseases. These

peptides are investigated as potential alternatives to

antibiotics in order to treat infections. Since wet lab

experiments are expensive and time-consuming,

computational methods become crucial. In this study,

we suggest a precise computational technique for

AMP prediction using deep neural networks (DNN).

We evaluated the DNN classifier using

physicochemical features. Physicochemical

properties are one of the most frequently used

features for this problem(Lin et al., 2021; Moretta et

al., 2020; Vishnepolsky et al., 2019). In our previous

work, we have demonstrated that these features

perform better in predicting and describing the dataset

than other features and these properties should be

taken into account while developing novel models

(Söylemez et al., 2022). In this respect, we focused on

these features for this study and obtained satisfactory

results for different performance metrics (Table 2).

Additionally, it was found that Angle Subtended by

the Hydrophobic Residues is the greatest

distinguishing factor for antimicrobial peptide

prediction using the feature significance attribute of

the Gradient Boosting model.

5 CONCLUSION

AMPs are essential components of the innate immune

system and gaining importance in drug development.

Identification of AMPs has emerged as a critical

research area. The findings of this study suggest that

the model designed offers a reliable and practical

method.

We proposed a deep neural network based model

using different physicochemical features. We

demonstrated that our model outperformed its

competitors when we compared with regular machine

learning models such as SVM, kNN, Bagging and

Gradient Boosting. We believe that the approach we

proposed could guide further experimental studies

and could facilitate the prediction of other types of

AMPs having anticancer, antivirus, antiparasitic

activities.

REFERENCES

Bahar, A., & Ren, D. (2013). Antimicrobial Peptides.

Pharmaceuticals, 6(12), 1543–1575. https://doi.org/

10.3390/ph6121543

Bhadra, P., Yan, J., Li, J., Fong, S., & Siu, S. W. I. (2018).

AmPEP: Sequence-based prediction of antimicrobial

peptides using distribution patterns of amino acid

properties and random forest. Scientific Reports, 8(1),

1697. https://doi.org/10.1038/s41598-018-19752-w

Breiman, L. (1996). Bagging predictors. Machine

Learning, 24(2), 123–140. https://doi.org/10.1007/

BF00058655

Cortes, C., & Vapnik, V. (1995). Support-vector networks.

Machine Learning, 20(3), 273–297. https://doi.org/

10.1007/BF00994018

Erdem Büyükkiraz, M., & Kesmen, Z. (2022).

Antimicrobial peptides (AMPs): A promising class of

antimicrobial compounds. Journal of Applied

Microbiology, 132(3), 1573–1596. https://doi.org/10.

1111/jam.15314

Fix, E., & Hodges, J. L. (1989). Discriminatory Analysis.

Nonparametric Discrimination: Consistency Properties.

International Statistical Review / Revue Internationale

de Statistique, 57(3), 238. https://doi.org/10.23

07/1403797

Friedman, J. H. (2002). Stochastic gradient boosting.

Computational Statistics & Data Analysis, 38(4), 367–

378. https://doi.org/10.1016/S0167-9473(01)00065-2

Hammami, R., & Fliss, I. (2010). Current trends in

antimicrobial agent research: Chemo- and

bioinformatics approaches. Drug Discovery Today,

15(13–14), 540–546. https://doi.org/10.1016/j.drudis.

2010.05.002

Joseph, S., Karnik, S., Nilawe, P., Jayaraman, V. K., &

Idicula-Thomas, S. (2012). ClassAMP: A Prediction

Tool for Classification of Antimicrobial Peptides.

IEEE/ACM Transactions on Computational Biology

and Bioinformatics, 9(5), 1535–1538.

https://doi.org/10.1109/TCBB.2012.89

Lata, S., Mishra, N. K., & Raghava, G. P. (2010). AntiBP2:

Improved version of antibacterial peptide prediction.

BMC Bioinformatics, 11(S1), S19. https://doi.org/

10.1186/1471-2105-11-S1-S19

Lata, S., Sharma, B., & Raghava, G. (2007). Analysis and

prediction of antibacterial peptides. BMC

Bioinformatics, 8(1), 263. https://doi.org/10.1186/

1471-2105-8-263

Lin, T.-T., Yang, L.-Y., Lu, I.-H., Cheng, W.-C., Hsu, Z.-

R., Chen, S.-H., & Lin, C.-Y. (2021). AI4AMP: An

Antimicrobial Peptide Predictor Using

Physicochemical Property-Based Encoding Method

and Deep Learning. MSystems, 6(6), e00299-21.

https://doi.org/10.1128/mSystems.00299-21

Malmsten, M. (2014). Antimicrobial peptides. Upsala

Journal of Medical Sciences, 119(2), 199–204.

https://doi.org/10.3109/03009734.2014.899278

Meher, P. K., Sahu, T. K., Saini, V., & Rao, A. R. (2017).

Predicting antimicrobial peptides with improved

accuracy by incorporating the compositional, physico-

Prediction of Antimicrobial Peptides Using Deep Neural Networks

193

chemical and structural features into Chou’s general

PseAAC. Scientific Reports, 7(1), 42362.

https://doi.org/10.1038/srep42362

Moretta, A., Salvia, R., Scieuzo, C., Di Somma, A., Vogel,

H., Pucci, P., Sgambato, A., Wolff, M., & Falabella, P.

(2020). A bioinformatic study of antimicrobial peptides

identified in the Black Soldier Fly (BSF) Hermetia

illucens (Diptera: Stratiomyidae). Scientific Reports,

10(1), 16875. https://doi.org/10.1038/s41598-020-

74017-9

Schneider, P., Müller, A. T., Gabernet, G., Button, A. L.,

Posselt, G., Wessler, S., Hiss, J. A., & Schneider, G.

(2017). Hybrid Network Model for “Deep Learning” of

Chemical Data: Application to Antimicrobial Peptides.

Molecular Informatics, 36(1–2), 1600011.

https://doi.org/10.1002/minf.201600011

Söylemez, Ü. G., Yousef, M., Kesmen, Z., Büyükkiraz, M.

E., & Bakir-Gungor, B. (2022). Prediction of Linear

Cationic Antimicrobial Peptides Active against Gram-

Negative and Gram-Positive Bacteria Based on

Machine Learning Models. Applied Sciences, 12(7),

3631. https://doi.org/10.3390/app12073631

Thakur, N., Qureshi, A., & Kumar, M. (2012). AVPpred:

Collection and prediction of highly effective antiviral

peptides. Nucleic Acids Research, 40(W1), W199–

W204. https://doi.org/10.1093/nar/gks450

Thomas, S., Karnik, S., Barai, R. S., Jayaraman, V. K., &

Idicula-Thomas, S. (2010). CAMP: A useful resource

for research on antimicrobial peptides. Nucleic Acids

Research, 38(suppl_1), D774–D780. https://doi.org/

10.1093/nar/gkp1021

Veltri, D., Kamath, U., & Shehu, A. (2017). Improving

Recognition of Antimicrobial Peptides and Target

Selectivity through Machine Learning and Genetic

Programming. IEEE/ACM Transactions on

Computational Biology and Bioinformatics, 14(2),

300–313. https://doi.org/10.1109/TCBB.2015.2462364

Veltri, D., Kamath, U., & Shehu, A. (2018). Deep learning

improves antimicrobial peptide recognition.

Bioinformatics, 34(16), 2740–2747. https://doi.org/

10.1093/bioinformatics/bty179

Vishnepolsky, B., Grigolava, M., Zaalishvili, G.,

Karapetian, M., & Pirtskhalava, M. (2018). DBAASP

Special prediction as a tool for the prediction of

antimicrobial potency against particular target species.

Proceedings of 4th International Electronic Conference

on Medicinal Chemistry, 5608. https://doi.org/10.33

90/ecmc-4-05608

Vishnepolsky, B., Zaalishvili, G., Karapetian, M.,

Nasrashvili, T., Kuljanishvili, N., Gabrielian, A.,

Rosenthal, A., Hurt, D. E., Tartakovsky, M., Grigolava,

M., & Pirtskhalava, M. (2019). De Novo Design and In

Vitro Testing of Antimicrobial Peptides against Gram-

Negative Bacteria. Pharmaceuticals, 12(2), 82.

https://doi.org/10.3390/ph12020082

Wang, G. (2014). Human Antimicrobial Peptides and

Proteins. Pharmaceuticals, 7(5), 545–594.

https://doi.org/10.3390/ph7050545

Witten, J., & Witten, Z. (2019). Deep learning regression

model for antimicrobial peptide design [Preprint].

Bioinformatics. https://doi.org/10.1101/692681

Xiao, X., Wang, P., Lin, W.-Z., Jia, J.-H., & Chou, K.-C.

(2013). iAMP-2L: A two-level multi-label classifier for

identifying antimicrobial peptides and their functional

types. Analytical Biochemistry, 436(2), 168–177.

https://doi.org/10.1016/j.ab.2013.01.019

BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms

194