Classifying nitrilases as aliphatic and aromatic using machine learning technique

Abstract

ProCos (Protein Composition Server, script version), one of the machine learning techniques, was used to classify nitrilases as aliphatic and aromatic nitrilases. Some important feature vectors were used to train the algorithm, which included pseudo-amino acid composition (PAAC) and five-factor solution score (5FSS). This clearly differentiated into two groups of nitrilases, i.e., aliphatic and aromatic, achieving maximum sensitivity of 100.00%, specificity of 90.00%, accuracy of 95.00% and Mathew Correlation Coefficient (MCC) of about 0.90 for the pseudo-amino acid composition. On the other hand, five-factor solution score achieved a sensitivity of 96.00%, specificity of 84.00%, accuracy of 90.00% and Mathew Correlation Coefficient (MCC) of about 0.81. The total count of aliphatic amino acids, Ala (A), Gly (G), Leu (L), Ile (I), Val (V), Met (M) and Pro (P), was found to be higher, i.e., 42.7 in case of aliphatic nitrilases, whereas it was 40.1 in aromatic nitrilases. On the other hand, aromatic amino acids, Tyr (Y), Trp (W), His (H) and Phe (F) number, were found to be higher, i.e., 12.7 in aromatic nitrilases as compared to aliphatic nitrilases which was 10.7. This approach will help in predicting a nitrilase as aromatic or aliphatic nitrilase based on its amino acid sequence. Access to the scripts can be done logging onto GitHub using keyword ‘Nitrilase’ or ‘https://github.com/rover2380/Nitrilase.git’.

Electronic supplementary material

The online version of this article (10.1007/s13205-018-1102-9) contains supplementary material, which is available to authorized users.

Introduction

Nitrilases are the enzymes which catalyze the hydrolysis of various nitriles into corresponding acid and ammonia. These enzymes have been well identified and characterized in plants, bacteria and fungi, and are engaged as an industrially important biocatalyst for the production of bulk and fine chemicals. For example, mandelonitrile could be hydrolyzed to optically pure (R)-(-)- mandelic acid, which is widely used for the production of semisynthetic cephalosporins, penicillins, antitumor agents, and anti-obesity agents (Wang et al. 2014). Researchers have revealed that nitrilases play a vital role in various biological processes and plant–microbe interaction, but despite their valuable importance they are relatively less explored for their metabolic functions.

Nitrilases differ variably in substrate specificities and find wide application in the transformation of a range of nitriles to acids (Sharma et al. 2006, 2012; Bhatia et al. 2014). Previous studies have revealed that nitrilases are specific for aromatic nitriles while nitrile hydratase has affinity towards aliphatic nitriles, but in light of rapidly growing information regarding nitrile metabolizing enzymes, various aspects have to be reconsidered (Mylerova and Martinkova 2003). Because of the established fact that amino acids are responsible for protein structure and function (Yeom et al. 2008; Liu et al. 2013), they are found to play a significant role in classifying nitrilases as aliphatic or aromatic.

With the exponential growth in the quantity of biological data in past years, there has been an impressive progress in computational biology. In silico analysis and various machine learning techniques are being applied for knowledge generation from the data. The machine learning approach is one such area of programming computers to optimize the performance criterion using example data or past results. The genome-based discoveries being continually increased, the possibility of finding novel sources of nitrilases has also increased tremendously (Gong et al. 2013; Kaplan et al. 2011). The annotation with functional assignments for their respective classes through various wet lab techniques is time consuming and labor intensive, which makes machine learning to be effectively used to complement them by saving time, money and labor (Pant et al. 2011). ProCoS script version is one such machine learning algorithm that has recently become prominent for in silico analysis, as they have a high dimensionality and accuracy in prediction of results not only for protein–protein complexes but also for enzyme classification (Rishishwar et al. 2010). Amino acid composition is a predictive feature vector for classification of various classes of proteins on the basis of their substrate specificity and position specificity (Kumar et al. 2011; Sharma et al. 2009).

The present article aims to serve for an insightful categorization and classification of nitrilases using script version of the ProCoS. The peptide composition features have been used for making pseudo-amino acid composition (PAAC) and five-factor solution score (5FSS) models in the present study.

Materials and methods

Dataset

The amino acid sequences of the nitrilases were downloaded from the ExPASy (http://www.expasy.org/sprot/) proteomic server and NCBI website. Nitrilases on the basis of their substrate specificity are distributed into two sets, i.e., positive (aliphatic nitrilase) and negative (aromatic nitrilase) dataset. Fifty amino acid sequences were considered in the study for both the datasets (Tables 1 and 2). Test and training sets were designed from a fivefold cross-validation scheme to create a model for the classification of a new sequence of nitrilase. The script used is accessible both as an applet and as a server, which is designed in Java and the server works on Perl-PHP backbone deposited in GitHub (https://github.com/rover2380/Nitrilase.git). The minimum input requirement for the analysis is the protein sequences in fasta format and output can be achieved in the form of tables.

Table 1.

Aliphatic nitrilases with their accession and amino acid number

Aliphatic nitrilases
S. no	Name of the microorganism	Accession number	Length (amino acid)
1	Rhodococcus rhodochrous K22	gi|417382	383
2	Rhodococcus rhodochrous J1	gi|417384	366
3	Nocardia sp. C-14-1	gi|60280369	381
4	Synechococcus sp. ATCC 27144	WP_011243013	334
5	Polaromonas naphthalenivorans	gi|500125486	353
6	Rhizobium leguminosarum bv. viciae 3841	gi|116255137	340
7	Variovorax paradoxus EPS	gi|315596504	344
8	Burkholderia sp. BT03	gi|495013900	356
9	Danaus plexippus F2	gi|357616093	389
10	Comamonas testosterone	gi|1082009	354
11	Sorangium cellulosum So0157-2	gi|521469000	342
12	Rhizoctonia solani 123E	gi|660965364	364
13	Polycyclovorans algicola	gi|659838894	362
14	Rhizobium leguminosarum	gi|659064095	348
15	Methylobacterium sp. L2-4	gi|657247605	358
16	Bosea sp. 117	gi|657241356	350
17	Bradyrhizobium sp. th.b2	gi|656043203	360
18	Azospirillum halopraeferens	gi|655966390	354
19	Bradyrhizobium elkanii	gi|654889008	354
20	Rhizobium sp. JGI 0001019-L19	gi|655350271	348
21	Burkholderia mimosarum	gi|654755069	350
22	Amycolatopsis taiwanensis	gi|654475327	346
23	Variovorax sp.P21	gi|654178860	350
24	Agrobacterium rhizogenes ATCC 15834	gi|653181208	350
25	Saccharomonospora viridis DSM 43017	ACU96985	331
26	Mesorhizobium loti	gi|652688040	348
27	Acidovorax oryzae	gi|651303417	344
28	Achromobacter xylosoxidans	gi|651250268	345
29	Variovorax paradoxus	gi|648592180	350
39	Methylobacterium sp. 88A	gi|648483839	363
31	Burkholderia kururiensis	gi|648430021	359
32	Pseudomonas syringae B728a	WP_011266126	336
33	Methylopila sp. 73B	gi|519032254	350
34	Sphingopyxis alaskensis	WP_011541682	338
35	Bradyrhizobium sp. ORS278	WP_011927383	337
36	Xanthobacter sp. 126	gi|635631313	352
37	Colletotrichum fioriniae PJ7	gi|615443311	362
38	Oligotropha carboxidovorans OM5	gi|209874119	354
39	Methylibium petroleiphilum PM1	gi|124258961	357
40	Marinomonas ushuaiensis DSM 15871	gi|575464044	344
41	Betaproteobacteria bacterium MOLA814	gi|557914537	367
42	Cupriavidus sp. WS	gi|519051014	356
43	Methylopila sp. M107	gi|519021908	352
44	Methyloversatilis universalis	gi|519007573	345
45	Teredinibacter turnerae	gi|518436209	349
46	Shimwellia blattae ATCC 29907	WP_002439083	342
47	Burkholderia gladioli	gi|503455327	373
48	Starkeya novella	gi|502933508	357
49	Serratia sp. M24T3	gi|497320793	342
50	Janthinobacterium sp. Marseille	gi|501028829	355

Table 2.

Aromatic nitrilases with their accession and amino acid number

Aromatic nitrilases
S. no	Name of the microorganism	Accession number	Length (amino acid)
1	Pantoea sp. AS-PWVM4	gi|544758631	328
2	Elizabethkingia	gi|544938496	318
3	Fodinicurvata sediminis	gi|550981872	310
4	Thalassospira lucentensis	gi|550982983	311
5	Rhizobium leguminosarum bv. trifolii WSM1325	gi|240856665	330
6	Cellulophaga algicola DSM 14237	gi|319421185	316
7	Maricaulis maris MCS10	gi|114340126	310
8	Pseudomonas sp. GM41	gi|576708726	324
9	Burkholderia sp. BT03	gi|576730682	328
10	Morganella morganii subsp. morganii KT	gi|455420318	338
11	Rubellimicrobium mesophilum DSM 19309	gi|598658225	319
12	Tomitella biformata	gi|640112707	324
13	Pedobacter jeongneungensis	gi|640722764	318
14	Flexithrix dorotheae	gi|648518461	314
15	Sediminispirochaeta bajacaliforniensis	gi|648603114	316
16	Niabella soli DSM 19437	gi|570745400	321
17	Butyrivibrio sp. MC2021	gi|651408280	310
18	Dyadobacter alkalitolerans	gi|651643084	314
19	Arenibacter latericius	gi|652415782	316
20	Maribacter antarcticus	gi|652759557	316
21	Chryseobacterium sp. UNC8MFCol	gi|653122843	319
22	Meiothermus chliarophilus	gi|654421979	314
23	Sphingobacterium thalpophilum	gi|654603925	318
24	Desulfatibacillum aliphaticivorans	gi|654863925	307
25	Parabacteroides gordonii	gi|655317710	317
26	Pseudonocardia spinosispora	gi|655591302	310
27	Stappia stellulata	gi|656017004	316
28	Rhodococcus aetherivorans	gi|657826219	322
29	Marssonina brunnea sp. MB_m1	gi|597582433	321
30	Pseudomonas pseudoalcaligenes CECT:5344	gi|652791517	324
31	Burkholderia multivorans CGD1	WP_006401663	307
32	Thalassiosira pseudonana	EED91795	320
33	Saccharomyces cerevisiae RM11-1a	EDV09642	322
34	Ajellomyces dermatitidis ER-3	EEquation 85041	297
35	Scheffersomyces stipitis ATCC 58785	XP_001385512	307
36	Methanosarcina mazei BAA-159	WP_011033178	307
37	Arabidopsis thaliana	AEE77890	346
38	Bacillus sp. OxB-1	AB028892	339
39	Synechocystis sp. PCC6803	gi|1001835	346
40	Aeribacillus pallidus	gi|111054396	323
41	Runella slithyformis	WP_013931053	310
42	Pseudomonas entomophila L48	WP_011534641	307
43	Shewanella sediminis HAW-EB3	ABV35137	317
44	Microscilla marina ATCC 23134	WP_002693358	304
45	Janthinobacterium sp. Marseille	WP_012080333	316
46	Burkholderia cepacia J2315	WP_006483427	307
47	Bordetella bronchiseptica	WP_003808910	310
48	Geodermatophilus obscurus ATCC 25078	WP_012946300	260
49	Nocardiopsis dassonvillei ATCC 23218	WP_013156158	280
50	Streptomyces albus J1074	WP_003950974	315

Features

Amino acid composition (AAC)

The amino acid frequency was calculated for both the datasets of proteins (aliphatic and aromatic nitrilases). Calculation of amino acid frequencies gives the value of the occurrence of that amino acid in the particular protein sequence. The fraction of the twenty amino acids was calculated using the following equation:

$$\text{Fraction of amino acids}=\frac{\text{total number of amino acid}\left(i\right)}{\text{total number of amino acids in proteins}}.$$

This gives a significance of a particular amino acid. The script takes an input of 20 vectors corresponding to twenty amino acids. Figure 1 shows that the amino acid frequencies of aromatic and aliphatic nitrilases are different, so they can be easily distinguished.

Fig. 1.

Comparison of amino acid frequencies of aliphatic and aromatic nitrilases using ProCoS

Dipeptide composition (DPC)

Dipeptide composition was calculated for all the 20 × 20 (400) combinations of amino acid. It gives significance to the combination of amino acids. The fraction of each dipeptide was calculated using the following equation:

$$\mathbf{Fraction}\mathbf{of}\mathbf{dipeptide}=\frac{\text{total number of dipeptides}\left(i\right)}{\text{total number of all possible dipeptides}}.$$

Tripeptide composition (TPC)

Tripeptide composition was also calculated like amino acid and dipeptide composition, thus generating all 20 × 20 × 20 (8000) feature vectors for training and testing datasets.

Pseudo-amino acid composition (PAAC)

The use of simple amino acid composition feature misses the important information in order of amino acid present in the peptide. Keeping this in view, the following information is incorporated with the help of PAAC as mentioned by Chou (2001). The feature vectors built according to this concept contains the frequency of 20 amino acids followed by their respective order information. Web server for calculation of PAAC had been proposed which calculates the respective feature (Shen and Chou 2008).

Split amino acid composition (SAAC)

Peptides were split into three parts to compute split amino acid composition of each part of protein separately. In this way, a vector of dimension 60 (3 × 20) was created instead of 20 in case of amino acid composition. In SAAC, each protein was divided into three parts like: (1) 20 amino acids of the N terminus, (2) 20 amino acids of the C–terminus, and (3) remaining protein length after removing 20 amino acids from N– and C– terminus.

Hybrid model 1

First hybrid model was made by combining the feature vectors of amino acid composition and dipeptide composition (AAC + DPC) giving us 420 vectors (20 + 400) for training and testing dataset.

Hybrid model 2

Second hybrid model was made by combining split amino acid feature to the hybrid 1 (AAC + DPC + SAAC) feature resulting in 480 (20 + 400 + 60) feature vectors for SVM.

Machine learning using script version of ProCos (Protein composition server)

The present study uses the script version which has been implemented and is a supervised machine learning algorithm. The idea behind using the script is the classification which attaches the feature vector with each sample (this case its peptide) to represent those points in a high dimensional feature space and then assigning the points into a particular category (positive or negative class) on the basis of an optimal separating hyperplane. The script training most preciously gives a global solution to optimize the hyperplane, thus avoiding the problem of overfitting of the data to one another class.

Cross-validation and evaluation parameter

A fivefold cross-validation for validating pseudo-amino acid composition (PAAC) and five-factor solution score (5FSS) model predictors was used. The performance of all the models was evaluated by the following standard parameter method:

1. Sensitivity or coverage of positive examples: It is the percent of aromatic nitrilase proteins correctly predicted.

   $$\text{Sensitivity}\left(\text{Sn}\right)=\frac{\text{TP}}{\text{TP}+\text{FN}}\times 100.$$

2. Specificity or coverage of negative examples: It is the percent of aliphatic nitrilase proteins correctly predicted aliphatic nitrilase.

   $$\text{Specificity}\left(\text{Sp}\right)=\frac{\text{TN}}{\text{TN}+\text{FP}}\times 100.$$

3. Accuracy: It is the percentage of correctly predicted proteins (aromatic and aliphatic proteins).

   $$\text{Accuracy}\left(\text{Acc}\right)=\frac{\text{TP}+\text{TN}}{\text{TP}+\text{TN}+\text{FP}+\text{FN}}\times 100.$$

4. Mathew’s correlation coefficient (MCC): It is considered to be the most robust parameter of any class prediction method. MCC equal to 1 is regarded as perfect prediction while 0 for completely random prediction.

$$\text{MCC}=\frac{\left(\text{TP}\times \text{TN}\right)-\left(\text{FP}\times \text{FN}\right)}{\sqrt{\left(\text{TP}+\text{FP}\right)\left(\text{TP}+\text{FN}\right)\left(\text{TN}+\text{FP}\right)\left(\text{TN}+\text{FN}\right)}}\times 100$$

where TP and TN are truly or correctly predicted aliphatic and aromatic nitrilases. FP and FN are wrongly predicted aliphatic and aromatic nitrilases.

Results

The script written is a powerful applet and a classification tool that has become increasingly popular in various machine learning applications. Machine learning approach is considered to be one of the vital subfields of artificial intelligence which is more concerned with the development of techniques and methods that enable the computer to learn. The present study classifies nitrilases on the basis of their amino acid composition which is responsible for their substrate specificity, stability and selectivity. The model developed by machine learning technique is used to differentiate between the two groups of nitrilases. The total count of aliphatic amino acids, i.e., alanine (A), glycine (G), leucine (L), isoleucine (I), valine (V), methionine (M) and proline (P), was found to be higher, i.e., 42.7 in case of aliphatic nitrilase as compared to aromatic nitrilases which is 40.1 (Fig. 1). On the other hand, aromatic amino acids, tyrosine (Y), tryptophan (W), histidine (H) and phenylalanine (F) number, were found to be higher, i.e., 12.7 as when compared to aliphatic nitrilases which were 10.7.

For aliphatic and aromatic class of nitrilases, machine was trained using ProCoS, each with a different type of kernel (linear, polynomial, radial basis and sigmoid). The output with the best training results was considered with high sensitivity, specificity, accuracy and Mathew’s correlation coefficient which has been summarized in Table 3 (detailed information provided as supplementary data S1-S7).

Table 3.

Performance of the models based on vectors for amino acid composition (AAC), dipeptide composition (DPC), split amino acid composition (SAAC), pseudo-amino acid composition (PAAC), tripeptide composition (TPC), hybrid 1 (AAC + DPC) and hybrid 2 (AAC + DPC + SAAC), respectively, Matthews correlation coefficient (MCC), rate of false prediction (RFP)

Model	Sensitivity	Specificity	Accuracy	MCC	RFP
AAC	90.00	93.88	91.92	0.84	6.25
DPC	94.00	91.84	92.93	0.86	7.84
SAAC	92.00	81.63	86.87	0.74	16.36
PAAC	100.00	90.00	95.00	0.90	9.09
TPC	94.00	92.00	93.00	0.86	7.84
hyb1	96.00	87.76	91.92	0.84	11.11
hyb2	92.00	93.88	92.93	0.86	6.12

Sensitivity, specificity and accuracy are in percentage (in bold and italics are the maximum accuracy and MCC)

Amino acid composition (AAC)

A sensitivity of 90.00%, specificity of 93.88%, accuracy of 91.92% and MCC of about 0.84 for AAC was achieved which clearly indicates the difference between the two classes of nitrilase, i.e., aliphatic and aromatic nitrilases but with the rate of false prediction (RFP) of 6.25.

Dipeptide composition (DPC)

This model performed better than AAC with sensitivity of 94.00%, specificity of 91.84%, accuracy of 92.93% and MCC of 0.86. RFP was found to be more than AAC, i.e., 7.84, respectively.

Split amino acid composition (SAAC)

This model gave sensitivity of 92.00%, specificity of 81.63%, accuracy of 86.87% and MCC of 0.74, but the RFP was high with the value of 16.36.

Tripeptide composition (TPC)

he model based on TPC feature achieved sensitivity of 94.00%, specificity of 92.00%, accuracy of 93.00% and MCC of 0.86 with the RFP of 7.84.

Pseudo-amino acid composition (PAAC)

Model based on PAAC feature vector achieved the highest sensitivity of 100.00%, specificity of 90.00%, accuracy of 95.00% and MCC of 0.90 and the RFP of 9.09, respectively (Tables 3 and 4). Among all the models, this model has the maximum accuracy and MCC so we considered this feature model as the best out of all models built yet in this study for nitrilase classification.

Table 4.

Performance of ProCos model using pseudo-amino acid calculation (PAAC) and five-factor solution score (5FFSS) features

Threshold	PAAC				5FFSS
	Sn	Sp	Acc	Mcc	Sn	Sp	Acc	Mcc
− 0.1	100.00	90.00	95.00	0.90	96.00	84.00	90.00	0.81
0.0	96.00	90.00	93.00	0.86	92.00	86.00	89.00	0.78
0.1	94.00	92.00	93.00	0.86	90.00	88.00	89.00	0.78

Sn sensitivity, Sp specificity, Acc accuracy, Mcc Matthews correlation coefficient

Discussion

As the next generation DNA sequencing (NGS) techniques have become cheaper and more efficient in yielding sequence data in a short time, the number of sequences in the public domain has increased significantly but still important annotations are missing (Chakravorty and Hegde 2017). Experimental validation of every uncharacterized, putative and hypothetical sequence may not be possible with the same pace (Rottig et al. 2010) and assigning functions to all the predicted genes/proteins would be time and cost ineffective (Kim et al. 2013). The characterized set of sequences deposited in the gene/protein databases for nitrilases is fewer in number; therefore, automated computational methods are needed to assign a putative function to uncharacterized sequences reliably (Mills et al. 2015). To the best of our knowledge, no study has been carried out for reliable classification of nitrilases as aliphatic or aromatic.

Previous analysis has confirmed that functional annotation between a test sequence and annotated sequence is above 60%, below which the probability of predicting the function of the test to the query sequence is rather low (Tian et al. 2003; Arakaki et al. 2009; Rottig et al. 2010). It has been inferred in the past that low sequence similarities (below 30%) have resulted in more of paralogs with the query sequence instead of orthologs (Chen and Jeong 2000). Nitrilases with sequence identity as low as 27% with that of characterized nitrilase retained true nitrilase activity if the catalytic triad was found to be conserved (Kaushik et al. 2012). Overall data in the present study share average value of more than 30% identity and conserved catalytic triad. This has led us to infer that sequences retain true nitrilase activity with identity as low as 27% and catalytic triad is conserved throughout. This information will be helpful for the analysis and to predict the models to gain insights into the mechanism of enzyme–substrate specificity as reported in the past (Stachelhaus et al. 1999; Challis et al. 2000; Sharma et al. 2017). Substrate range for nitrilases is rather broad including aliphatic, aromatic and arylnitriles which depends on the groups attached to the side chain (Gong et al. 2012). Characteristics of residues surrounding the active site and the presence of specific amino acids increase the probability for predicting the substrate affinity of nitrilases.

In the present analysis, the script is used to classify the amino acid composition and their dominance in aliphatic and aromatic nitrilases which is responsible for differences in substrate affinity. Cysteine acts as a nucleophile for substrate attack and is activated due to the deprotonation of sulfhydryl group of cysteine by glutamic acid (Zang et al. 2014). Glutamic acid acts as a general base, whereas lysine as general acid (Martinkova and Kren 2010). The aliphatic amino acid alanine (A) also plays a significant role in overall activity of nitrilases (Sharma et al. 2009; Kaushik et al. 2012). Glycine (G), leucine (L), isoleucine (I), valine (V), methionine (M) and proline (P) are other important amino acids which support the aliphaticity of nitrilases. On the other hand, aromatic substrate affinity for some nitrilases is due to tyrosine (Y), tryptophan (W), histidine (H) and phenylalanine (F) which are found to be higher in aromatic nitrilases. These amino acids create aromatic-rich environment near the catalytic centre of nitrilases which prefer aromatic substrates (Liu et al. 2013; Zang et al. 2014). The present data clearly define the role of amino acids for the substrate specificity determination which will further play a significant role in mutational studies of nitrilases to achieve better stability, specificity and reactivity.

Conclusion

The article focuses on the use of the script based method for classification of aliphatic and aromatic group of nitrilases. The results clearly exhibited that the algorithm can be used as a tool to classify nitrilases as aliphatic and aromatic class. The overall accuracy achieved by writing the following script is 95.00%. These machine learning techniques can be used to predict different features of the gene/protein and selection of these algorithms for the prediction of gene/protein function.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Compliance with ethical standards

Conflict of interest

The authors declare that they have no conflict of interests.