In silico approach for predicting toxicity of peptides and proteins.

Gupta S, Kapoor P, Chaudhary K, Gautam A, Kumar R, Open Source Drug Discovery Consortium, Raghava GP. · 2013

Read In silico approach for predicting toxicity of peptides and proteins. by Gupta S, Kapoor P, Chaudhary K, Gautam A, Kumar R, Open Source Drug Discovery Consortium, Raghava GP. (2013) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc3772798.txt.

Introduction

Sudheer Gupta and co-authors, with the Open Source Drug Discovery Consortium, presented a prediction tool called ToxinPred in PLoS ONE in 2013. Peptides are promising drugs, prized for precision and easy manufacture, but toxicity blocks many candidates. The team built computer models to flag toxic peptides before lab testing.

They trained the models on known toxic peptides of 35 building blocks or fewer, drawn from databases, with harmless peptides for contrast. Certain residues, such as cysteine and proline, appear often and at favored spots in toxic peptides. A model based on pairs of residues reached about 94.5 percent accuracy, and held near 90 percent on fresh data. The free web server also suggests small changes to raise or lower a peptide's toxicity and can scan proteins for toxic regions.

Catalog record

CREATORGupta S, Kapoor P, Chaudhary K, Gautam A, Kumar R, Open Source Drug Discovery Consortium, Raghava GP.
YEAR2013
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC3772798

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

In Silico Approach for Predicting Toxicity of Peptides and Proteins By Sudheer Gupta, Pallavi Kapoor, Kumardeep Chaudhary, Ankur Gautam, Rahul Kumar, Gajendra P S Raghava, Randen Lee Patterson Published in: PLoS ONE DOI: 10.1371/journal.pone.0073957 ABSTRACT Abstract Background Over the past few decades, scientific research has been focused on developing peptide/protein-based therapies to treat various diseases. With the several advantages over small molecules, including high specificity, high penetration, ease of manufacturing, peptides have emerged as promising therapeutic molecules against many diseases. However, one of the bottlenecks in peptide/protein-based therapy is their toxicity. Therefore, in the present study, we developed in silico models for predicting toxicity of peptides and proteins. Description We obtained toxic peptides having 35 or fewer residues from various databases for developing prediction models. Non-toxic or random peptides were obtained from SwissProt and TrEMBL. It was observed that certain residues like Cys, His, Asn, and Pro are abundant as well as preferred at various positions in toxic peptides. We developed models based on machine learning technique and quantitative matrix using various properties of peptides for predicting toxicity of peptides. The performance of dipeptide-based model in terms of accuracy was 94.50% with MCC 0.88. In addition, various motifs were extracted from the toxic peptides and this information was combined with dipeptide-based model for developing a hybrid model. In order to evaluate the over-optimization of the best model based on dipeptide composition, we evaluated its performance on independent datasets and achieved accuracy around 90%. Based on above study, a web server, ToxinPred has been developed, which would be helpful in predicting (i) toxicity or non-toxicity of peptides, (ii) minimum mutations in peptides for increasing or decreasing their toxicity, and (iii) toxic regions in proteins. Conclusion ToxinPred is a unique in silico method of its kind, which will be useful in predicting toxicity of peptides/proteins. In addition, it will be useful in designing least toxic peptides and discovering toxic regions in proteins. We hope that the development of ToxinPred will provide momentum to peptide/protein-based drug discovery (http://crdd.osdd.net/raghava/toxinpred/). INTRODUCTION The last decade has seen an unprecedented revival of interest in therapeutic peptides as potential drug candidates [1]. A plethora of research articles has been published every year demonstrating discovery of variety of novel therapeutic peptides e.g. tumor homing peptides [2], cell penetrating peptides [3], anti-microbial peptides [4], anticancer peptides [5], [6], [7], etc. and applications of these peptides in various diseases like cancer, diabetes, cardiovascular diseases, etc. As a result of these efforts, rate of therapeutic peptides entering into clinical trials has improved significantly over the decade [1]. However, despite the discovery of hundreds of such therapeutic peptides, only few peptide-based drugs have made it to the market. Peptides have numerous advantages over small molecules that include high biological activity, high specificity, low production cost, and high penetration [1], [5]. However, toxicity, immunogenicity and stability remain the main concerns in the development of peptide-based drugs. Stability of peptides can be enhanced by various ways [8], including incorporation of D-amino acids (making peptides protease resistant), changing the backbone chemistry, cyclization, and incorporation of α-aminoxy amino acids [9]. Similarly, there are numerous in silico tools, which can predict the immunogenicity of the peptides [10], [11], [12], [13], [14], but there is hardly any way/method to predict the toxicity of peptides. Computational methods for predicting toxicity of peptides not only save time and money, but also facilitate the designing of better therapeutic peptides with low toxicity while retaining the functionalities. Keeping these facts in mind, in this study, for the first time, an attempt has been made to develop an in silico method for predicting toxicity of peptides. Toxic peptides have been collected from various databases/studies including ATDB [15], DBETH [16], BTXpred [17], NTXpred [18], Arachno-Server [19], Conoserver [20]. In silico models have been developed using the machine-learning technique support vector machine (SVM), for discriminating toxic peptides from non-toxic peptides. In addition, various motifs from toxic peptides were discovered and used for toxicity prediction. MATERIALS AND METHODS Dataset Creation We extracted small toxins (proteins/peptides) from different databases and studies that include ATDB [15], Arachno-Server [19], Conoserver [20], DBETH [16], BTXpred [17], NTXpred [18], and SwissProt [21]. We removed all proteins/peptides having more than 35 residues or any non-natural amino acid. As a result, 1805 unique toxic proteins/peptides were obtained. By employing the similar criteria, toxic proteins/peptides were also searched in SwissProt database using keyword KW800 (keyword 800 stands for toxin as molecular functions). A total of 803 toxic proteins, having length less than 35 amino acids were obtained. It is possible that many toxic peptides obtained from various databases could also be present in SwissProt. Therefore, identical toxic proteins/peptides were removed and finally we got 303 unique toxic proteins/peptides from SwissProt. These proteins/peptides were considered as toxic peptides or positive examples. Though it is possible to extract well-annotated or experimentally validated toxic peptides, but it is difficult to obtained non-toxic peptides. Therefore, to create a negative dataset, we have searched protein/peptide sequences in UniProt using keywords NOT KW800 NOT KW20 (keyword 800 and 20 stand for toxin and allergen as molecular functions). Proteins/peptide sequences having length less than 35 amino acids were extracted. After removing sequences with non-natural amino acids, two types of negative datasets were created; first dataset consists of 3893 sequences from SwissProt (NOT KW800 NOT KW20) and second dataset consists of 13541 sequences from TrEMBL (keyword NOT KW800 AND KW33090) [21]. While searching non-toxins in TrEMBL, additional keyword plant proteins were applied as search criteria as most of the plants are edible and therefore, the probability of plant proteins/peptides to be toxic is very low. Above toxic and non-toxic peptides/proteins were used to generate various datasets for training, testing and evaluating our models developed for predicting toxicity of peptides (Figure 1). Following is the brief description of these datasets: [Figure: Overview of datasets’ creation.] Main and alternative datasets The main dataset used for training and testing of SVM models was generated from experimentally validated toxic peptides (obtained from various databases) and well-annotated non-toxin peptides/proteins obtained from SwissProt. It includes 1805 toxic peptides as positive examples and 3593 non-toxic peptides as negative examples. In addition, we also generated an alternate dataset, which is similar to the main dataset except negative examples. It consists of 1805 toxic-peptides/proteins as positive examples and 12541 non-toxin peptides/proteins obtained from TrEMBL (instead of SwissProt). Independent datasets In order to evaluate biases in the performance of developed models, we created different independent datasets. First independent dataset comprises of 303 toxic proteins/peptides (called positive examples) and 300 non-toxic peptides/proteins or negative examples extracted from SwissProt. None of the negative or positive examples was included in the main dataset. This dataset is referred as main independent dataset and used for evaluating models developed on the main dataset. In order to evaluate the performance of models developed on alternate dataset, we developed another independent dataset referred as alternate independent dataset. This dataset consists of 303 positive examples extracted from SwissProt and 1000 negative examples extracted from TrEMBL, which were not included in the alternate dataset. Models Developed using Support Vector Machine There are a number of machine learning techniques used for building prediction models. In this study, support vector machine (SVM) has been

Read the complete text