An advanced approach to identify antimicrobial peptides and their function types for penaeus through machine learning strategies

Abstract

Background

Antimicrobial peptides (AMPs) are essential components of the innate immune system and can protect the host from various pathogenic bacteria. The marine environment is known to be one of the richest sources for AMPs. Effective usage of AMPs and their derivatives can greatly improve the immunity and breeding survival rate of aquatic products. It is highly desirable to develop computational tools for rapidly and accurately identifying AMPs and their functional types, for the purpose of helping design new and more effective antimicrobial agents.

Results

In this study, we made an attempt to develop an advanced machine learning based computational approach, MAMPs-Pred, for identification of AMPs and its function types. Initially, SVM-prot 188-D features were extracted that were subsequently used as input to a two-layer multi-label classifier. In specific, the first layer is to identify whether it is an AMP by applying RF classifier, and the second layer addresses the multi-type problem by identifying the activites or function types of AMPs by applying PS-RF and LC-RF classifiers. To benchmark the methods,the MAMPs-Pred method is also compared with existing best-performing methods in literature and has shown an improved identification accuracy.

Conclusions

The results reported in this study indicate that the MAMP-Pred method achieves high performance for identifying AMPs and its functional types.The proposed approach is believed to supplement the tools and techniques that have been developed in the past for predicting AMPs and their function types.

Background

Antimicrobial peptides (AMPs) are crucial components of the innate immune system and can protect the host from various pathogenic bacteria and viruses. They are generally short peptides with 10–50 amino acids [1] and have very low sequence homology to one another. AMPs nowadays have attracted increased attention of research owing to their broad-spectrum antimicrobial activity and more importantly to the fact that AMPs may overcome the antimicrobial resistance, which makes it a potential alternative therapeutic agent for humans or a substitute to conventional antibiotics.

However, the mechanisms of action of AMPs, as well as their structure-activity relationships, are not completely understood [2]. Identification and optimization of AMPs can provide a theoretical basis for discovery and design of new and more effective antimicrobial agents. For instance, a multidimensional signature model was proposed in [3] that facilitates discovery of AMPs and offers insights into the evolution of molecular determinants. Experimental and computational studies are generally devoted to dealing with this challenging task. Computational methods were developed to accelerate the process of prediction and classification of AMPs. Recently, approaches based on machine learning techniques are commonly adopted due to their high efficiency, high speed, low cost and generalization abilities. They can sufficiently mine the intrinsic linear and non-linear relationship between antibacterial activity and biochemical attributes, which is suitable for dealing with large scale antimicrobial peptide prediction tasks with complex models.

Methods of choice include support vector machine (SVM) [4–7], nearest neighbor [8] or k-nearest neighbor algorithm [9], random forests (RFs) [10]), decision tree model [11], hidden Markov models (HMMs) [12], and neural network model [13] which seek for prediction power in a context of supervised classification. Most recent work includes a "deep" network architecture for chemical data analysis and classification together with a prospective proof-of-concept application proposed in [14]. Some predictors only apply binary classifiers to identify whether a query peptide sequence is AMP or not, such as [4, 5, 8]. Multi-class classifiers have also been developed which obtained more detailed quantitative results. Lira et al. [11] created a decision tree model to classify the antimicrobial activities of synthetic peptides into four classes. ClassAMP [4] has been developed to predict the propensity of a peptide sequence to have antibacterial, antifungal, or antiviral activity. However, it can be seen by a comparison of the sequences in APD database [15, 16] that a same sequence may occur in different subclasses, which in fact a very common phenomenon. Therefore, it is highly desirable to develop mechanisms for rapidly and accurately learning from multi-label datasets, for the purpose of helping design new and more effective antimicrobial agents. Considering various possible functional types of AMPs, Xiao et al. proposed a two-level multi-label classifier iAMP-2L, where an improved fuzzy K-nearest neighbour (FKNN) algorithm was applied, and after the AMPs are first identified, the positive samples are subjected to regular multi-label learning processing [9]. The prediction accuracy for 4 types of AMPs was further improved in [17]. Zhou’s method [18] has applied the LIFT multi-label learning algorithm to predict 5 types of AMPs and achieved 70% accuracy of prediction.

This paper aims to develop an advanced method, MAMPs-Pred, for classification and prediction of AMPs and their function types, which proves to achieve an improved prediction accuracy upon state of the art mechanisms. The marine environment is known to be one of the richest sources for AMPs. It is meaningful to predict the AMPs and their function types of penaeus by this method, which has helped us to understand the immune system of marine species. In addition, it eases subsequent mining and exploration of antimicrobial activity of other species.

In this approach, a 188-D feature set constructed from SVM-Prot features [19, 20] were used to map the peptide sequences to numeric feature vectors, which were subsequently used as input to a two-layer multi-label classifier. The first layer is to identify whether a query peptides sequence is an AMP, and the second layer addresses the multi-type problem by identifying whether an AMP belongs to multiple function types. Different classification methods were compared, and the results were discussed and analyzed. In short, a combination of first-layer 188D-RF classifier and second-layer PS-RF or LC-RF classifier is proved to have achieved the best performance. The proposed approach achieved higher accuracy than existing approaches of best performance, while performed upon benchmark dataset. In addition, the quality of the prediction was verified when applied to penaeus sequences. The proposed method may play an important complementary role to the existing predictors in this area.

Materials and methods

Benchmark dataset

For the convenience of later description, the benchmark dataset is expressed by

$$\begin{array}{ll}s& =s^{\mathit{\text{AMPs}}}\cup s^{\mathit{\text{non}}-\mathit{\text{AMPs}}}\end{array}$$ 1

Where s^AMPs is the AMPs dataset consisting of AMPs sequences only, s^non−AMPs the non-AMP dataset with non-AMP sequences only, and ∪ is the symbol for union in the set theory. The peptide sequences in s^AMPs were fetched from the APD database [15, 16], which has collected all antimicrobial peptides from the PubMed, PDB, Google and Swiss-Prot databases. According to their different functional types, the AMP sequences can be further classified into 16 categories; i.e.,

$$\begin{array}{ll}s& =s_1^{\mathit{\text{AMPs}}}\cup s_2^{\mathit{\text{AMPs}}}\cup s_3^{\mathit{\text{AMPs}}}\cup \dots \cup s_{16}^{\mathit{\text{AMPs}}}\end{array}$$ 2

Where the subscripts 1, 2, 3,...,16 represent “Wound healing”, “Spermicidal”, “Insecticidal”, “Chemotactic”, “Antifungal”, “Anti-protist”, “Antioxidant”, “Antibacterial”, “Antibiotic”, “Antimalarial”, “Antiparasital”, “Antiviral”, “Anticancer/tumor”, “Anti-HIV”, “Proteinase inhibitor” and “Surface immobilized”. The lengths of AMPs are varying within the region from 5 to 100 amino acids. Note that among the original 2954 s^AMPs sequences, 278 sequences have unknown antibacterial activity.

Furthermore, to reduce homology bias and redundancy, the program CD-HIT [21] was utilized to winnow those sequences that have ≥ pairwise sequence identity to any other in a same subset. The alignment bandwidth of the CD-HIT field is set to 5 according to the shortest length of AMPs. To ensure that each subset has enough samples for statistic processing, and to ensure that all categories are covered, the CD-HIT only performs redundancy removal to a subset of samples with sequence numbers larger than 180, which means that the de-redundancy processing are only performed for antifungal, antibacterial, antiviral and anti-cancer polypeptides. Finally, we obtained 2618 AMPs as the current benchmark dataset s^AMPs as shown in Table 1.

Table 1.

Preprocessed benchmark dataset

Function	Dataset	Function type	Sequence
AMPs	$s_1^{\mathit{\text{AMPs}}}$	Wound healing	18
	$s_2^{\mathit{\text{AMPs}}}$	Spermicidal	13
	$s_3^{\mathit{\text{AMPs}}}$	Insecticidal	28
	$s_4^{\mathit{\text{AMPs}}}$	Chemotactic	57
	$s_5^{\mathit{\text{AMPs}}}$	Antifungal	593
	$s_6^{\mathit{\text{AMPs}}}$	Anti-protist	4
	$s_7^{\mathit{\text{AMPs}}}$	Antioxidant	22
	$s_8^{\mathit{\text{AMPs}}}$	Antibacterial	1297
	$s_9^{\mathit{\text{AMPs}}}$	Antibiotic	32
	$s_{10}^{\mathit{\text{AMPs}}}$	Antimalarial	25
	$s_{11}^{\mathit{\text{AMPs}}}$	Antiparasital	101
	$s_{12}^{\mathit{\text{AMPs}}}$	Antiviral	125
	$s_{13}^{\mathit{\text{AMPs}}}$	Anticancer	125
	$s_{14}^{\mathit{\text{AMPs}}}$	Anti-HIV	109
	$s_{15}^{\mathit{\text{AMPs}}}$	Proteinase inhibitor	26
	$s_{16}^{\mathit{\text{AMPs}}}$	Surface immobilized	43
	s AMPs		2618
non-AMPs	s non− AMPs		4371

The negative samples s^non−AMPs contains polypeptide sequences s^{non−AMPs−Pept}, and protein fragments s^{non−AMPs−Prot}.

Where s^{non−AMPs−Pept} were constructed according to following procedures:

1. Collected all the polypeptide sequences s^{UNP−Peptide} with length 1 to 15483, in total 79378, from the UniProt database.

2. Removed any sequence that already exists in s^AMPs, any sequence that contains any code other than the 20 native amino acid codes, and any sequence with length less than 5 or larger than 100.

3. The process is described by following equation, and at this point 10503 sequences s^{non−AMPs−Pept} were obtained.

   $$\begin{array}{lll}{s}^{\mathit{\text{non}}-\mathit{\text{AMPs}}-\mathit{\text{Pept}}}& =s^{\mathit{\text{UNP}}-\mathit{\text{Peptide}}}-s^{\mathit{\text{AMPs}}}-\mathit{\text{se}}{q}_{\mathit{\text{illeg}}}& \\ (\mathit{\text{len}}\in [5,100\left]\right)\end{array}$$ 3

On the other hand, s^{non−AMPs−Prot} were constructed according to following procedures:

1. Obtained Pfam families that s^AMPs belong to. Because some AMPs are homologous and have the same family number, we remove duplicate family numbers from Pfam and get de-redundant families posPfam.

2. Removed posPfam from the Pfam families and obtained negPfam. Fetched a random protein sequence with the length between 5 and 100 from each negPfam family.

3. The process is described by following equation. In total 109 short protein sequences s^{non−AMPs−Prot} were obtained.

   $$\begin{array}{lll}{s}^{\mathit{\text{non}}-\mathit{\text{AMPs}}-\mathit{\text{Prot}}}& =\mathit{\text{Ran}}(\mathit{\text{Pfam}}-\mathit{\text{posPfam}})& \\ (\mathit{\text{len}}\in [5,100\left]\right)\end{array}$$ 4

The s^non−AMPs were constructed by following equation.

$$\begin{array}{ll}{s}^{\mathit{\text{non}}-\mathit{\text{AMPs}}}& =s^{\mathit{\text{non}}-\mathit{\text{AMPs}}-\mathit{\text{Pept}}}\cup s^{\mathit{\text{non}}-\mathit{\text{AMPs}}-\mathit{\text{Prot}}}\end{array}$$ 5

The CD-HIT [21] program was then applied to winnow s^non−AMPs. Finally, 4371 sequences were constructed, which were used to form the negative samples dataset s^non−AMPs as shown in Table 1.

Feature extraction

In machine learning, choosing informative, discriminating and independent features is a crucial step for the success of a prediction method. The optimal feature set shall be able to capture the distribution patterns of the dataset.

In this study, we have adopted two feature extraction algorithms for comparison, which are SVM-Prot 188-D based on 8 types of physical-chemical properties and amino acid composition, and Pseudo amino acid composition features (Co-Pse-AAC) based on 5 types of physical-chemical properties respectively.

SVM-Prot is a web server for protein classification. It constructs 188-D features for protein sequences description and classification [19, 20]. The features have been applied successfully in several protein identification works, such as cytokines [22, 23] and enzymes [24, 25]. The extracted features include hydrophobicity, normalized van der Waals volume, polarity, polarizability, charge, surface tension, secondary structure and solvent accessibility [19]. For each of these 8 types of physical-chemical properties, some feature groups were designed to describe global information of protein sequences. These feature groups contain composition (C), transition (T) and distribution (D) [19, 26]. Thus, the dimension of each feature vector is 21. In addition, considering amino acid composition (AAC), the protein structure is composed of 20 amino acids. The dimension of 188-D features is therefore expressed as below formula:

$$\begin{array}{ll}{D}_{188-D}& =\sum _{i=1}^L{D}_{21\mathit{\text{Vct}}}+D_{\mathit{\text{aac}}}\end{array}$$ 6

Where L is the number of features, which is 8 in this context. Take Cecropin A as an example. The 188-D features of Cecropin A is showed in Table 2. To the best of our knowledge, it is the first attempt in literature to apply SVM-Prot 188-D feature set composition in AMPs and non-AMPs classification and identification.

Table 2.

188-D feature of cecropin A

Sequence	KWKLFKKIEKVGQNIRDGIIKAGPAVAVVGQATQIAK
Property	Value of feature vector
Amino acid composition	13.5	0.0	2.70	2.70	2.70	10.8	0.0
	135	00	27.0	27.0	27.0	108	00
	13.5	18.9	2.70	0.00	2.70	2.70	8.10
	135	189	27.0	0.00	27.0	27.0	81.0
	2.70	0.00	2.70	10.8	2.70	0.00
	27.0	00	27.0	108	27.0	00
Hydro-phobic	37.8	29.7	32.4	19.4	30.5	19.4	2.70
	378	297	324	444	555	444	27
	16.2	35.1	45.9	100.	32.4	48.6	64.8
	162	351	459	000	324	486	648
	81.0	97.2	5.40	13.5	40.5	70.2	94.5
	810	972	54	135	405	702	945

On the other hand, Pseudo amino acid composition features (Co-Pse-AAC) [27] as an efficient computation tool has been diffusely leveraged for protein sequences in predicting protein structures and functions, as well as DNA and RNA sequences [28]. The 40-dimension Co-Pse-AAC features were extracted and sufficiently incorporate the effects of sequence order. This method has taken 5 types of physical-chemical properties into consideration.

Data balancing

Most machine learning classification algorithms are sensitive to the imbalanced data sets [29]. The classifiers tend to have a higher recognition rate for the majority class, which makes it difficult to identify the minority class correctly [30–32]. In this study, there were 2718 AMPs samples and 4371 non-AMPs samples, which were highly imbalanced. In order to eliminate the over fitting problem caused by imbalanced data, we have applied two sampling mechanisms to construct the training dataset.

Firstly, we have implemented a random-under-sampling method to down sample the large class set s^non−AMPs, so that the sample number of large class set equals the small class set, and the resulting training dataset is defined as s_train. Another method we have applied is weighted random sampling [33], which has balanced the dataset by applying different weights to the unbalanced samples. Given that the ratio of s^AMPs and s^non−AMPs is approximately equal to 3:5, weight factor 5 and 3 were applied to s^AMPs and s^non−AMPs respectively, and the obtained train dataset is defined as s_weight−tr.

Test dataset

The test dataset was constructed by following method. Firstly we randomly pick up 1382 negative samples from the sequences that have been deleted from s^non−AMPs in the CD-HIT process, and noted it by s^{non−AMPs−DEL}. Further, in the phrase of acquiring benchmark dataset from APD (The Antimicrobial Peptide Database) database, there are 278 sequences with unknown antibacterial activity among the original 2954 s^AMPs sequences, which is defined by s^{non−AMPs−NOACT}.

The 278 s^{non−AMPs−NOACT} sequences, together with the 1382 s^{non−AMPs−DEL}, form the independent test dataset S_test for the first layer of our two-layer multi-label classifier, which is in total 1660 samples.

The 278 s^{non−AMPs−NOACT} sequences were also applied as prediction dataset for the second layer of our two-layer multi-label classifier, which will be illustrated in following chapters.

Two-layer multi-label classifier

In machine learning, multi-label classification is the problem of finding a model that maps inputs x to binary vectors y, i.e., assigning a value of 0 or 1 for each label in y. In the multi-label problem there is no constraint on how many of the classes the instance can be assigned to. An overview of multi-label classification is available at [34].

In general, the methods to study multi-label classification can be divided into two categories: adapted algorithm methods and problem transformation methods. Some classification models have been adapted to the multi-label task, without requiring problem transformations. For instance, AdaBoost.MH and AdaBoost.MR are extended versions of AdaBoost for multi-label data. And the ML-kNN algorithm extends the k-NN classifier to multi-label data. Examples also include decision trees, neural networks adapted for multi-label learning.

Problem transformation methods fall into another category of multi-label classification. With converting multi-label problems into one or more single-label problems, literally existing single-label classifier can be used to meet the multi-label classification requirements. Representative algorithms include Binary Relevance (BR), Classifier Chains (CC), Label Combination Method (LC/LP), Integrated LP Method Rakel, and Pruned Sets Method (PS). BR amounts to independently training one binary classifier for each label; CC is similar to BR, except that it takes into account label dependencies; LC/LP treats each label combination as a new label and implicitly considers the label.

A polypeptide can be a non-AMP that does not have any antimicrobial activity. It is actually a prediction problem with negative samples, which cannot be handled directly by traditional multi-label classification. Incorporating non-AMPs rationally into predictive models is an essential issue for multi-label classification to predict function types of AMPs. To address this issue, we improve upon the state of the art in multi-label classification and make several contributions.

For the first-layer classifier in identifying a query peptide sequence as an AMP or non-AMP, the random forest (RF) algorithm was applied as a base classifier because of its good performance and simple-to-use feature. Random forest is an ensemble method in which a classifier is constructed by combining several independent base classifiers. The individual predictions are aggregated to combine into a final prediction, based on a majority voting on the individual predictions. By averaging several trees, there is a significantly lower risk of over fitting.

For the second layer classifier in identifying which functional type(s) the query AMP peptide sequence belongs to, a task of multi-label classification was launched. We choose Meka/Mulan open source framework to implement our second layer multi-label classifier. Meka is based on the Weka machine learning toolkit, one of the well-known data mining platforms (http://www.cs.waikato.ac.nz/ml/weka/), and integrates the open-source Java library Mulan framework for providing the capability of multi-label datasets learning. Meka proposed a trimming set method and a Classifier Chains (CC) method, and uses logarithmic loss to punish misplaced tags to prevent partial misprediction in the overall label distortion. For the second-layer prediction, PS-RF or LC-RF is applied as a base multi-label classifier due to its performance.

Measurement metrics

The metrics Sensitivity (SN), specificity (SP), overall accuracy (Acc) and Matthew’s correlation coefficient (Mcc) were applied to measure the performance of the first-layer classifier [18, 35–40], where TP_i,FP_i,TN_i,FN_i denote the numbers of true positive instances, false positive instances, true negative instances and false negative instances respectively.

$$\begin{array}{ll}\mathit{\text{SN}}& =\frac{T{P}_i}{T{P}_i+{\mathit{\text{FN}}}_i}\end{array}$$ 7

$$\begin{array}{ll}\mathit{\text{SP}}& =\frac{T{N}_i}{F{P}_i+{\mathit{\text{TN}}}_i}\end{array}$$ 8

$$\begin{array}{ll}\mathit{\text{Acc}}& =\frac{T{P}_i+{\mathit{\text{TN}}}_i}{T{P}_i+{\mathit{\text{FP}}}_i+{\mathit{\text{TN}}}_i+{\mathit{\text{FN}}}_i}\end{array}$$ 9

$$\begin{array}{cc}\mathit{\text{Mcc}}& =\frac{T{P}_i\times {\mathit{\text{TN}}}_i-{\mathit{\text{FP}}}_i\times {\mathit{\text{FN}}}_i}{\sqrt{({\mathit{\text{TP}}}_i+{\mathit{\text{FP}}}_i)\times ({\mathit{\text{TN}}}_i+{\mathit{\text{FN}}}_i)\times ({\mathit{\text{TP}}}_i+{\mathit{\text{FN}}}_i)\times ({\mathit{\text{TN}}}_i+{\mathit{\text{FP}}}_i)}}\end{array}$$ 10

The metric Exact-Match Ratio (EMR), Hamming-Loss (H-Loss), Accuracy (Acc), Precision (Precison, Recall), Ranking-Loss (RL), Log-Loss, One-error (OE), F1-Measure (F1-Mic, F1-Mac) were applied for evaluation the second-layer multi-label classifier.

$$\begin{array}{ll}\mathit{\text{EMR}}\left({\Lambda }_t\right)& =\frac{1}{K}\sum _{i=1}^K(\widetilde{y_i}=y_i)\end{array}$$ 11

$$\begin{array}{ll}H-\mathit{\text{Loss}}\left({\Lambda }_t\right)& =\frac{1}{\mathit{\text{KL}}}\sum _{i=1}^K\frac{|\widetilde{y_i}\cup y_i|-|\widetilde{y_i}\cap y_i|}{L}\end{array}$$ 12

$$\begin{array}{ll}\mathit{\text{Acc}}\left({\Lambda }_t\right)& =\frac{1}{K}\sum _{i=1}^K\frac{|\widetilde{y_i}\cup y_i|}{|\widetilde{y_i}\cap y_i|}\end{array}$$ 13

$$\begin{array}{ll}\mathit{\text{Precision}}\left({\Lambda }_t\right)& =\frac{1}{K}\sum _{i=1}^K\frac{|\widetilde{y_i}\cap y_i|}{\widetilde{y_i}}\end{array}$$ 14

$$\begin{array}{ll}\mathit{\text{Recall}}\left({\Lambda }_t\right)& =\frac{1}{K}\sum _{i=1}^K\frac{|\widetilde{y_i}\cap y_i|}{y_i}\end{array}$$ 15

$$\begin{array}{lll}F1\left({\Lambda }_t\right)& =\frac{2.0\times \mathit{\text{Precision}}\left({\Lambda }_t\right)\times \mathit{\text{Recall}}\left({\Lambda }_t\right)}{\mathit{\text{Precision}}\left({\Lambda }_t\right)+\mathit{\text{Recall}}\left({\Lambda }_t\right)}& \\ \mathit{\text{OE}}\left({\Lambda }_t\right)& =\frac{1}{K}\sum _{i=1}^K\left\{\right[{\mathit{\text{argmax}}}_{y\in Y}h(x_i,y)]\notin y_i\}& \\ & =\frac{1}{K}\sum _{i=1}^K\frac{2|\widetilde{y_i}\cap y_i|}{\left|\widetilde{y_i}\right|+\left|y_i\right|}\end{array}$$ 16

$$\begin{array}{lll}\mathit{\text{RL}}\left({\Lambda }_t\right)& =\frac{1}{K}\sum _{i=1}^K\frac{1}{\left|\widetilde{y_i}\right|\times \left|y_i\right|}\left|\right\{(y_1,y_2)\left|f_{t}h\right((x_i,y_1))& \\ \le f_{t}h\left(\right(x_i,y_2\left)\right)\left\}\right|\end{array}$$ 17

$$\begin{array}{ll}\mathit{\text{Log}}-\mathit{\text{Loss}}\left({\Lambda }_t\right)& =\frac{1}{\mathit{\text{KL}}}\sum _{i=1}^K\sum _{j=1}^L\\ & \left\{\mathit{\text{min}}\left[-\mathit{\text{Log}}-\mathit{\text{Loss}}\left(\widetilde{w_j^i},y_j^i\right),\mathit{\text{ln}}\left(K\right)\right]\right\}& \end{array}$$ 18

Results

First classifier - Identifying AMPs or non-AMPs

Firstly, we extracted SVM-prot 188-D features and Co-Pse-AAC 40-D features for each peptide sequence. Then the first-layer classifier was followed for identifying if the sequence is AMPs or not. Several common classifiers, including Random Forest (RF), Bagging, J48, OneR, Naive Bayesian NB, KNN, and LibSVM, were chosen for performance comparison. The result showed that the Random Forest and Bagging classifiers based on decision trees have achieved the highest prediction accuracy rate that exceeded 84% for both SVM-prot 188-D and Co-Pse-AAC 40-D features (Fig. 1).

Fig. 1.

The main flowchart of the AMPs identification and prediction process

We further applied 1660 test dataset samples S_test to verify 5 RF and Bagging based classifiers (188D-RF–W, 188D-RF–R, 188D-Bagging–W, 188D-Bagging–R, 40D-RF-R), where W denotes weighted random sampling, and R denotes random-under-sampling, since the AMP dataset is highly imbalanced, whereas sampling methods might affect the prediction performance significantly.

Table 3 shows that the 188D-RF-W classifier based on weighted random sampling can guarantee good sensitivity and specificity on both training set and test set, which can efficiently identify AMPs and non-AMPs, where TPR represents true positive rate, FPR represents false positive rate, and AUC is area under the curve. Hence, we use it as the first-layer classifier of our proposed MAMP-Pred method. FPR TPR AUC

Table 3.

Performance comparison of first-layer classifiers on test dataset S_test

Classifier	AMPs			non-AMPs			Acc(%)
	TPR	FPR	AUC	TPR	FPR	AUC
188D-RF-W	0.831	0.156	0.900	0.844	0.169	0.900	84.157
188D-RF-R	0.892	0.205	0.897	0.795	0.108	0.897	81.145
188D-Bagging-W	0.888	0.205	0.899	0.795	0.112	0.899	81.084
188D-Bagging-R	0.921	0.220	0.897	0.780	0.079	0.897	80.361
40D-RF-R	0.874	0.194	0.890	0.806	0.126	0.890	81.747

a. Statements that serve as captions for the entire table do not need footnote letters

b. W = weighted random sampling, R = random-under-sampling, 188D = SVM-prot 188-D, 40D = Co-Pse-AAC 40-D

Second classifier - Identifying function types of AMPs

We investigated several multi-label classification methods on dataset s^AMPs in order to find the best classifier for identifying AMPs function types. We firstly evaluated different problem transformation methods, including Binary Correlation (BR), Classifier Chain (CC), Bayesian Classifier Chain (BCC), Tag Combination (LC), pruning set (PS), combined with representative single-label classifiers including J48, Random Tree, Random Forest, KNN and Bagging. We also investigated several adapted algorithm methods such as MLkNN, BRkNN, BP neural network, BPMLL, and DeepML, whereas the details were not illustrated in this paper due to the space limitations.

All multi-label classifiers have adopted train/test dataset split and 10-fold cross-validation mechanisms based on s^AMPs for evaluation. The evaluation results of BR-RF, PS-RF, CC-RF, BCC-RF, LC-RF and BRkNN methods on dataset s^AMPs are shown in Table 4. It can be seen that PS-RF and LC-RF have achieved the highest overall accuracy, and 10-fold cross-validation performs better than train/test dataset split mechanism for all problem transformation methods.

Table 4.

Performance Comparison of Second-layer Classifiers (10 fold cross-validation)

Models	Acc	EMR	H-Loss	F1-Micro	F1-Macro	One-error	Rank-Loss	Log-Loss
BR-RF	0.839	0.785	0.021	0.920	0.941	0.122	0.019	0.076
PS-RF	0.856	0.825	0.020	0.923	0.939	0.138	0.052	0.056
CC-RF	0.844	0.794	0.021	0.922	0.942	0.165	0.051	0.057
BCC-RF	0.847	0.801	0.020	0.924	0.943	0.160	0.051	0.056
LC-RF	0.855	0.824	0.020	0.923	0.939	0.139	0.052	0.056
BRkNN	0.696	0.561	0.044	0.838	0.783	0.238	0.101	0.121

The second stage is to apply PS-RF and LC-RF classifiers for predicting the possible antimicrobial activities or function types of the 278 AMPs with unknown antibacterial activity s^{non−AMPs−NOACT}. Similar prediction results were obtained in PS-RF and LC-RF. As shown in Fig. 2, there is one wound healing activity, one spermicidal activity, one chemotactic activity, one antimalarial activity, 6 Insecticidal activities, 27 antifungal activities, 27 anti-HIV activities, 13 Antiparasital activities, 19 antiviral activities, 23 anticancer activities, 5 proteinase inhibitor activities, 223 antibacterial activities. In addition, none of the antimicrobial peptides may have anti-protist, antioxidant, antibiotics, and surface immobilized activities.

Fig. 2.

Predicting function types of s^{non−AMPs−NOACT}

Performance evaluation

To benchmark our method, we present a comparative analysis of our MAMPs-Pred method against other existing best-performing in literature. Most of the existing methods can only be used to identify a query peptide as an AMP or non-AMP.

To make the comparison feasible and applicable, we firstly compared the first-layer classifier of MAMPs-Pred with the first-level classifier of iAMP-2L. We have applied the independent test data sets $S_{\mathit{\text{test}}}^{\mathit{\text{Ind}}}$ in [9], which contains 920 AMPs and non-AMPs sequences. The overall accuracy rate of iAMP-2L was 86.32%. Our mechanism has achieved 87.14% classification accuracy, which shows better performance than iAMP-2L, as shown in Table 5.

Table 5.

Performance comparison of MAMPs-Pred and iAMP-2L first-layer on $S_{\mathit{\text{test}}}^{\mathit{\text{Ind}}}$ dataset)

Method	Acc	SN	SP	Mcc
MAMPs-Pred	93.91%	92.83%	94.99%	0.878
iAMP-2L	92.23%	97.72%	86.74%	0.845

The second-layer classifier of MAMPs-Pred was compared with the iAMP-2L method [9] and LIFT classification method proposed in [17]. It can be seen that our MAMPs-Pred method has gained an improved overall performance over iAMP-2L and LIFT as shown in Table 6.

Table 6.

Performance comparison of MAMPs-Pred and iAMP-2L, LIFT second-layer on $S_{\mathit{\text{test}}}^{\mathit{\text{Ind}}}$ data set

Method	Acc	EMR	Precision	Recall	H-Loss
MAMPs-Pred	0.856	0.825	0.918	0.929	0.020
iAMP-2L	0.669	0.43	0.833	0.75	0.164
LIFT	0.700	0.5365	0.838	0.741	0.1392

The first reason is that the amino acid composition and its eight physicochemical properties which are used for feature extraction in this study, can better express the relationship between structure and antimicrobial peptides function types thus yield significantly improved performance.

The second reason is that the pruning set method applied in the second-layer multi-label classification, which transforms the label set into a single label in the problem, and directly models the label correlation, can achieves an overall better prediction performance.

Performance on predicting Penaeus AMPs

In total 14298 protein sequences of shrimp (Penaeus) were fetched from the public UniProt database, including Penaeus monodon, Penaeus vannamei, etc. We then obtained 1452 sequences with a length between 5 and 100 from the 14298 sequences, followed by extracting SVM-prot 188-D features based on amino acid composition (AAC) and its 8 physicochemical properties for each penaeus protein sequence. The processed sequences were subsequently fed to the first-layer classifier of MAMP-Pred. A total of 126 AMPS/AMPS-like sequences were detected, accounting for 8.68% of the total sequence.

In the second-layer multi-label classification, we have predicted the possible antimicrobial activities or function types that an AMP belongs to. All 126 penaeus AMPs sequences had antibacterial activity, one with chemotactic activity, and four with antifungal activity, as shown in Fig. 3. MAMP-Pred can be regarded as an efficient data-mining method to predict the potential antimicrobial peptides and antibacterial activities of the query sequences.

Fig. 3.

AMPs activity prediction of 126 shrimp sequences

Discussion

Antimicrobial peptides are increasingly gaining considerable attention both from research and industry, as well as clinical interest. With the growing microbial resistance to conventional antimicrobial agents, the demand for unconventional and efficient AMPs has become urgent. Effective usage of AMPs and their derivatives can greatly improve the immunity and breeding survival rate of aquatic products.

The results reported in this study indicate that the MAMP-Pred method achieves high performance for identifying AMPs and its functional types. The proposed approach is believed to supplement the tools and techniques that have been developed in the past for prediction of AMPs. The primary reason is that the amino acid composition and its eight physicochemical properties which are used for the feature extraction in this study, can better express the relationship between structure and antimicrobial peptides function types. The second reason is that the pruning set method applied in the second-layer multi-label classification achieves an overall higher prediction performance.

As summarized in [41], the recognition accuracy of machine learning methods ranges from the upper 70 to the lower 90 percent. Reported recognition accuracy has steadily improved over the past decade, while there is room for improvement.

The current MAMP-Pred approach can be straightforwardly extended in following directions in future research work:

1. Construct a more reliable datasets of positive and negative samples to reduce potential bias of model training introduced by sequence homology. We also believe that with more data available in the future, the prediction accuracy can be significantly enhanced.

2. The two-level prediction requires learning and classification to be performed twice, which lowers down the prediction efficiency. An adaptive dynamic approach which possibly yields faster speed and higher efficiency is of definite interest in our future research.

3. In this approach, the overlay of prediction errors might incur significant drop of prediction accuracy. In future work, the current method shall be straightforwardly extended to address these issues.

4. Predicting the AMPs and their function types of penaeus by this method can help us to understand the immune system of marine species. In addition, it eases subsequent mining and exploration of antimicrobial activity of other species. The predictor holds very high potential to become a useful high throughput tool to predict antimicrobial activity of other species.

Conclusion

In this study, we made an attempt to develop an advanced machine learning based computational approach, MAMPs-Pred, for identification of AMPs and its function types. Initially, SVM-prot 188-D features were extracted that were subsequently used as input to a two-layer multi-label classifier. The first layer is to identify whether it is an AMP by applying RF classifier, and the second layer addresses the multitype problem by identifying the activities or function types of AMPs by applying PS-RF and LC-RF classifiers.

Abbreviations

Acc

Overall accuracy

AMPs

Antimicrobial peptides

APD

Antimicrobial peptide database

CD-HIT

Cluster database at high identity with tolerance

ClassAMP

A method to predict the propensity of a peptide sequence

Co-Pse-AAC

Pseudo amino acid composition

EMR

Exact-match ratio

FKNN

Fuzzy K-nearest neighbour

H-Loss

Hamming-loss

iAMP-2L

A two-level multilabel classifier

LC-RF

Label combination-random forests

LIFT

Zhou’s multi-label learning algorithm

MAMPs-Pred

Our method

Mcc

Matthew’s correlation coefficient

PS-RF

Pruned sets-random forests

RF

Random forests

HMMS

Hidden Markov models

SN

Sensitivity

SP

Specificity

SVM-prot 188-D

A web server for protein classification with 188-D feature

SVM

Support vector machine

Authors’ contributions

XL, YL and YC conceived and designed the experiments, YC collected the dataset, YL and YC performed the experiments, YL wrote the paper; XL,YL and CL analyzed the data, XL and YL discussed the results and improved the manuscript. All authors read and approved the final manuscript.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.