PyBioMed: a python library for various molecular representations of chemicals, proteins and DNAs and their interactions

Background With the increasing development of biotechnology and informatics technology, publicly available data in chemistry and biology are undergoing explosive growth. Such wealthy information in these data needs to be extracted and transformed to useful knowledge by various data mining methods. Considering the amazing rate at which data are accumulated in chemistry and biology fields, new tools that process and interpret large and complex interaction data are increasingly important. So far, there are no suitable toolkits that can effectively link the chemical and biological space in view of molecular representation. To further explore these complex data, an integrated toolkit for various molecular representation is urgently needed which could be easily integrated with data mining algorithms to start a full data analysis pipeline. Results Herein, the python library PyBioMed is presented, which comprises functionalities for online download for various molecular objects by providing different IDs, the pretreatment of molecular structures, the computation of various molecular descriptors for chemicals, proteins, DNAs and their interactions. PyBioMed is a feature-rich and highly customized python library used for the characterization of various complex chemical and biological molecules and interaction samples. The current version of PyBioMed could calculate 775 chemical descriptors and 19 kinds of chemical fingerprints, 9920 protein descriptors based on protein sequences, more than 6000 DNA descriptors from nucleotide sequences, and interaction descriptors from pairwise samples using three different combining strategies. Several examples and five real-life applications were provided to clearly guide the users how to use PyBioMed as an integral part of data analysis projects. By using PyBioMed, users are able to start a full pipelining from getting molecular data, pretreating molecules, molecular representation to constructing machine learning models conveniently. Conclusion PyBioMed provides various user-friendly and highly customized APIs to calculate various features of biological molecules and complex interaction samples conveniently, which aims at building integrated analysis pipelines from data acquisition, data checking, and descriptor calculation to modeling. PyBioMed is freely available at http://projects.scbdd.com/pybiomed.html.


Background
With the increasing development of biotechnology and informatics technology, the past decade has seen an exceptional growth in publicly available data in chemistry and biology, especially in human-specific molecular interaction data. The heterogeneity of data in databases poses a significant challenge to their integration and analysis in practice [1,2]. However, the bioinformatics and the cheminformatics communities have evolved more or less independently, e.g., with an emphasis on macro biomolecules and chemical compounds, respectively. Investigation of interactions is a complex molecular recognition process, which is not only related to the bioinformatics projects that aim at a systematic analysis of the structure and function of proteins and DNAs that scales to the genome level, but also to the chemoinformatics projects that are devoted to the analysis of structure and biological activity of chemicals. More importantly, systematic investigation of generated knowledge in both the chemical and biological knowledge spaces is required, especially in the scenarios of identifying both new targets and their potential ligands, discovering potential biomarkers for complex diseases, understanding the mechanism of interactions, and discovering new regulatory mechanism etc. [3][4][5][6][7][8]. Therefore, it is very necessary to build informatics platforms for unified data or knowledge representation that can integrate the existing efforts from both communities.
Furthermore, wealthy information in these data needs to be extracted and then transformed to useful knowledge by various data mining and artificial intelligent methods. Lots of machine learning methods have been elaborately developed to mine useful biomedicine information [9][10][11][12][13][14][15][16]. However, in order to apply various machine learning approaches on molecular data, it is a common practice to encode molecular information as numerical features. The type of encoding, however, can significantly affect analyses, and choosing a precise and effective encoding is a critical step. Molecular descriptors are one of the most powerful approaches to characterize the biological, physical, and chemical properties of molecules and have long been used in various studies for understanding molecular interactions or drug discovery. These descriptors capture and magnify distinct aspects of molecular topology in order to investigate how molecular structure affects molecular property. Molecular features have frequently been used in the development of machine learning in QSAR/QSPR [17,18], virtual screening [19], similarity search [20], drug absorption, distribution, metabolism, elimination and toxicity (ADMET) eavaluation [21][22][23][24], protein structural and functional classes [25,26], protein-protein interactions [27], compoundprotein interactions [28][29][30][31], subcellular locations and peptides of specific properties [32], meiotic recombination hot spots [33], nucleosome positioning in genomes and other drug discovery processes [34]. In terms of molecular representation importance, some web servers and stand-alone programs, such as RDKit [35], CDK [36], rcdk, PaDEL [37], Cinfony [38], Chemopy [39], ChemDes [40], BioJava [41], BioTriangle [42], bioclipse [43], propy [44], PyDPI [45], Biopython, repDNA [46], CDK-Taverna [47], protr/protrWeb [48], ChemmineR [49], and Rcpi [50] have been established to calculate such descriptors. However, currently available solutions are often limited to certain types of descriptors, lack flexibility, and usually difficult to seamlessly integrate into the predictive modeling pipeline. We still urgently need a comprehensive and flexible toolkit to integrate these separate functionalities into a uniform framework, and further enable us to build a full analysis pipeline.
Here, we developed PyBioMed, a python library, for realizing comprehensive molecular representation of various molecular objects and basic data analysis tasks. PyBioMed mainly focuses on the study of molecular representation techniques for not only single molecular objects, but also interactions between any two individual molecules from pairwise samples. To easily use the PyBi-oMed utilities and functionalities, we provide a uniform interface and highly customized modules in the library to perform data analysis. To introduce and describe PyBioMed utility and application, we selected five wide applications as examples to show that PyBioMed can be used as an integral part of an analytical pipeline. Our computational algorithms are extensively tested and the computed features have also been used in a number of published studies. We recommend PyBioMed to analyze and represent various complex molecular data under investigation. Further, we hope that the library will be incorporated to other research projects such as data analysis pipeline, web server and software applications, in which the molecular representation will play an important role.

Implementation
The library PyBioMed is implemented in python. It is mainly based on the RDKit and pybel implementation, licensed under BSD 3-Clause License, and depends on the following python modules: scipy, numpy, os and sys. The library is mainly divided into six highly customized modules, namely: (1) PyGetMol, (2) PyPretreat, (3) PyMolecule, (4) PyProtein, (5) PyDNA and (6) PyInteraction. The modules allow the users to get various molecular objects by different ways, pretreat these molecular objects, and finally calculate chemical descriptors, protein descriptors, DNA descriptors and interaction descriptors by different functionalities. The implementation of these six modules is independent, and each module could perform the specific function. Additionally, four additional files include examples and applications, test modules and help documentations, which will greatly facilitate the use of the library. The PyBioMed library is freely available from the home page of the PyBioMed project. URL: http://projects.scbdd.com/pybiomed.html. The formatted documentation of the library is designed based on the sphinx language. The documentation of 149 pages includes the introduction of the library, the installation of the library, how to use each function or module by an example, five real-life applications widely covering different study fields, the detailed documentation of PyBioMed APIs, how to test the library, and the download links for detailed descriptor introduction. The users could learn all knowledge of the library by looking up the detailed documentation. The main functionalities of the library are presented in the following subsections.

Downloading and reading molecular files
Before various studies, it is the first step for the researchers to conveniently obtain various molecular objects from various web sources such as chemical structures, protein sequences, and DNA sequences. The PyBioMed library designed a specific module called PyGetMol to realize the aim. The PyGetMol module is responsible for obtaining and reading various molecular objects, and it contains three functions for downloading molecular objects (Getmol, GetProtein, GetDNA) and various functions for reading molecular objects. The Getmol module provides multiple tools to get molecular structures by the molecular ID from websites including NCBI, EBI, CAS, KEGG and DrugBank. The GetProtein module provides the tool to get protein sequence by the PDB ID and Uni-Prot ID. The GetDNA module provides the tool to get DNA sequence by the Gene ID from websites. Additionally, they are also responsible for reading various molecular objects in different formats, such as SDF, MOL, InChI and SMILES for chemicals, FASTA for proteins and DNAs. The incorporation of these functionalities makes PyBioMed easily accessible in various cheminformatics applications.

Pretreating molecular objects
The check and preprocessing for various molecular objects is of high importance for subsequent descriptor calculation and data analysis, especially those molecular objects from web sources. Some molecules may contain structure defects to a certain extent, and therefore seriously influence or even destroy the subsequent descriptor calculation. The PyPretreat module in the library provides three specific functions which are responsible for pretreating three different molecular objects. The user could check the protein and DNA sequences by PyPre-treatPro and PyPretreatDNA, respectively. The two functions mainly check whether there are additional amino acids types or DNA basic group types or not. If the sequence is right, the result goes back to the number of amino acids and nucleotides or not, otherwise, the result is 0. For chemical compounds, the pretreating step seems to be complex. PyPretreatMol pretreats a chemical structure in the following ways: (1) normalization of functional groups to a consistent format; (2) recombination of separated charges; (3) breaking of bonds to metal atoms; (4) competitive deionization to ensure strongest acids ionize first in partially ionize molecules; (5) tautomer enumeration and canonicalization; (6) neutralization of charges; (7) standardization or removal of stereochemistry information; (8) filtering of salt and solvent fragments; (9) generation of fragment, isotope, charge, tautomer or stereochemistry insensitive parent structures; (10) validations to identify molecules with unusual and potentially troublesome characteristics. The functionalities can be highly customized by setting corresponding parameters according to the job demand. Alternatively, the users could also pretreat the molecular structure using all functions by StandardSmi function.

Calculating descriptors form chemicals, proteins and DNAs
Molecular representation is the core feature of the PyBi-oMed library. 'The molecular descriptor is the final result of a logic and mathematical procedure which transforms chemical or biological information encoded within a symbolic representation of a molecule into a useful number or the result of some standardized experiment' [51]. Molecular representation plays a fundamental role in chemoinforamtics and bioinformatics, and lies at the heart of ligand-based drug design. It is the first even the most key step in the data analysis tasks. A large number of molecular descriptors from chemicals, proteins and DNAs could be easily calculated by the PyBioMed library, which covers wide applications in various cheminformatics and bioinformatics projects.

Molecular descriptors based on chemical structures
The PyMolecule module in PyBioMed is responsible for calculating the commonly used structural and physicochemical descriptors. It computes twelve feature groups composed of fourteen descriptors that include 775 descriptor values. These descriptors capture and magnify distinct aspects of chemical structures, including constitutional descriptors, topological descriptors, kappa shape indices, connectivity indices, Burden descriptors, E-state indices, charge descriptors, Basak information indices, autocorrelation descriptors, molecular properties, MOE-type descriptors, and pharmacophore descriptors.
A detailed list of descriptors for chemicals covered by PyBioMed is summarized in Table 1. The usefulness of molecular descriptors in the representation of molecular information is reflected in their widespread adoption and use across a broad range of applications and methodologies, as reported in a large number of published articles [18,22,52,53]. More detailed description and references can be found in the documentation of PyBioMed. We could import the corresponding module to calculate the molecular descriptors as need. Alternatively, an easier way to compute these descriptors is construct a PyMolecule object, which encapsulates all methods for the calculation of descriptors.
Besides molecular descriptors, PyBioMed also realizes the computation of a number of molecular fingerprints, and a specific fingerprint module is designed to achieve the aim. Molecular fingerprints are string representations of chemical structures, which consist of bins, each bin being a substructure descriptor associated with a specific molecular feature. 19 types of molecular fingerprints and substructure fragments are provided in PyBioMed, including topological fingerprints, E-state fingerprints, MACCS keys, FP4 keys, atom pairs fingerprints, PubChem fingerprints, topological torsion fingerprints, and Morgan/circular fingerprints etc. (see Table 2). The usefulness of these molecular fingerprints covered by PyBioMed have been sufficiently demonstrated by a number of published studies of the development of machine learning classification systems in QSAR/SAR, drug ADME/T prediction, similarity searching, clustering, ranking and classification [54].

Protein or peptide descriptors based on amino acid sequences
The PyProtein module in PyBioMed is responsible for calculating the widely used structural and physicochemical features of proteins and peptides from amino acid sequences. It computes five feature groups composed of fourteen features, including amino acid composition, dipeptide composition, tripeptide composition, normalized Moreau-Broto autocorrelation, Moran autocorrelation, Geary autocorrelation, sequence-order-coupling number, quasi-sequence-order descriptors, composition, transition and distribution of various structural and physicochemical properties, and two types of pseudo amino acid composition (PseAAC) descriptors. These features could be generally regarded as different Chou's PseAAC modes. In addition, it can also easily compute previous descriptors based on user-defined properties, which are automatically available from the AAindex database. A list of features for proteins and peptides covered by PyBi-oMed is summarized in Table 3. These features have been used for predicting protein-and peptide-related problems by using machine learning methods. More detailed description and references can be found in the documentation of PyBioMed.
In fact, the abovementioned features can be regarded as different Chou's PseAAC modes [55]. For example, amino acid, dipeptide, tripeptide, or k-mer peptide (k = 4, 5,…) compositions are just different modes of Chou's PseAAC. Moreover, the higher-level features such as GO (Gene Ontology) information, FunD (Functional Domain) information, and Sequential Evolution information are also skillfully fused into the Chou's PseAAC descriptors to characterize different protein information which is widely used for solving various biological problems. An excellent review by Chou has pointed out their relevancy.

DNA descriptors based on nucleotide sequences
The PyDNA module in PyBioMed is responsible for calculating the widely used structural and physicochemical features of DNAs from nucleotide sequences. Generally, three groups of features from nucleotide sequences are calculated to represent DNA in PyBioMed. (1) three nucleic acid composition features describing the local sequence information by means of k-mers (subsequences of DNA sequences); (2) six autocorrelation features     Amphiphilic pseudo amino acid composition describing the level of correlation between two oligonucleotides along a DNA sequence in terms of their specific physicochemical properties; (3) six pseudo nucleotide composition features, which can be used to represent a DNA sequence with a discrete model or vector yet still keep considerable sequence order information, particularly the global or long-range sequence order information, via the physicochemical properties of its constituent oligonucleotides. A detailed list of descriptors for DNAs covered by PyBioMed is summarized in Table 4. The usefulness of these features covered by PyBioMed for representing DNA sequence information have been sufficiently demonstrated by a number of published studies in computational genomics and genome sequence analysis. More detailed description and references can be found in the documentation of the PyBioMed library. There are two ways to calculate DNA descriptors in the PyDNA module. One is to directly use the corresponding methods, the other one is firstly to construct a PyDNA class and then run their methods to obtain the descriptors. The users could select one or more groups to represent DNAs under investigation. It should be noted that the output is a dictionary form, whose keys and values represent the descriptor name and the descriptor value, respectively. The user could clearly understand the meaning of each descriptor.

Interaction descriptors based on pairwise samples Descriptors from the interaction between two molecules with the same type
The interaction between two molecules with the same type includes drug-drug interaction, protein-protein interaction, etc. However, the construction process of the interaction descriptors of them is similar to each other. We will show how to construct an interaction feature by the protein-protein interaction example. Let F a = {F a (i), i = 1, 2,…, p} and F b = {F b (i), i = 1, 2,…, p} are the two descriptor vectors for interaction protein A and protein B, respectively. There are three methods to construct the interaction descriptor vector F for A and B: 1. Two vectors F ab and F ba with dimension of 2p are constructed: F ab = (F a , F b ) for interaction between protein A and protein B and F ba = (F b , F a ) for interaction between protein B and protein A. 2. One vector F with dimension of 2p is constructed:

Descriptors from the interaction between two molecules with different types
The interaction between the molecules with different types includes chemical-protein interaction, protein-DNA interaction, and chemical-DNA interaction. However, the calculation of these interaction descriptors is similar to each other. Likewise, we will show how to construct an interaction feature by the chemical-protein interaction example. There are two methods for construction of descriptor vector F for chemical-protein interaction from the protein descriptor vector F t (F t (i), i = 1, 2,…, p t ) and chemical descriptor vector F d (F d (i), i = 1, 2,…, p d ): Dinucleotide-based auto-cross covariance 2888 Trinucleotide-based auto covariance 24

Trinucleotide-based cross covariance 264
Trinucleotide-based auto-cross covariance 288 Pseudo nucleic acid composition Pseudo dinucleotide composition 18 Pseudo k-tuple nucleotide composition 18 Parallel correlation pseudo dinucleotide composition 18 Parallel correlation pseudo trinucleotide composition 66 Series correlation pseudo dinucleotide composition 90 Series correlation pseudo trinucleotide composition 1. One vector V with dimension of p t + p d is constructed: F = (F t , F d ) for interaction between protein t and chemical d. 2. One vector V with dimension of p t × p d is constructed by the tensor product:

How to use PyBioMed functions
In order to familiarize the users with PyBioMed, It is recommended that the user works through the tutorial examples provided. The tutorial will go through the process of installing and running an example in some detail.
The PyBioMed library would be applied to solve various research tasks in the field of cheminformatics, bioinformatics and systems biology. We introduced five examples of its applications in the documentation including Caco-2 cell permeability, aqueous solubility, drug-target interaction data, protein subcellular location, and nucleosome positioning in genomes. Next, we will briefly introduce the installation of PyBioMed, and how to calculate molecular descriptors by writing few lines of codes.
PyBioMed has been successfully tested on Linux and Windows systems. The installation process of PyBioMed is very easy. However, the user first needs to install RDKit and pybel successfully. The detailed installation steps are described in the README page of the GitHub repository of PyBioMed.
There are two means to compute these molecular descriptors from small molecules. One is to use the builtin modules. We could import related functions to compute these features. For example, the topology module includes a number of functionalities used for calculating various topological descriptors. The user could conveniently use them as need.
The other is to call the related class by importing the class module, which encapsulates commonly used descriptor calculation methods. Pymolecule, Pyprotein and Pydna are responsible for the calculation of chemical descriptors, protein descriptors, and DNA descriptors, respectively. We could construct a corresponding object with a molecule input, and then call corresponding methods to calculate these features.
The interaction descriptors for pairwise samples are calculated as follows:

Discussion
Considering the amazing rate at which data are accumulated in chemistry and biology fields, new tools that process and interpret large and complex interaction data are increasingly important. However, to our knowledge, no open source or freely available tool exists to perform all these functions above in a uniform framework. PyBi-oMed is a powerful python library for the extraction of features of complex interaction data. After representation, different statistical learning tools can be applied for further analysis and visualization of the data. Several case studies from wide applications show how PyBioMed was used to describe various molecular features and establish a model in a routing way (See documentation). The application domain of PyBioMed is not limited to the specific data type. It can, as Fig. 1 shows, be applied to a broad range of scientific fields such as QSAR/SAR, similarity search, virtual screening, ADMET prediction, ligandbased drug discovery, protein function/substructure/ family classification, subcellular locations, post-translational modification (PTM), DNA structure/function/site prediction, and various interaction data analysis such as drug-target/drug interaction and protein-protein interaction studies. We expect that PyBioMed will better assist chemists, pharmacologists and biologists in characterizing, analyzing, and comparing complex molecular objects.
As mentioned in the background, there are several tools that have been developed to calculate chemical, protein or DNA/RNA description. However, PyBioMed is unique because it can be used to extract features of complex interaction data. Commonly, these tools can be divided into three types: software production, software package and web server. Software production tends to be commercial; web server tries to provide user-friendly way to beginners or users without programming skills; software package provides most flexible way to calculate descriptors in different application scenarios. The three types of tools have their own advantages. Here, PyBi-oMed belongs to software package of Python community. In order to give a more clear insight into the different software packages, we made a comparison between them (see Table 5).
From the Table 5, we can see: (1) most of these tools can only realize one or two functionalities while PyBi-oMed made all these functionalities into a uniform framework. (2) For some tools, the implications are actually not the same though the same functionality is ticked. Both RDKit and PyBioMed have the function of 'Pretreating molecules' , while PyBioMed reorganized and wrapped the basic functions of RDKit into 10 new functions. By combining these functions in different orders and setting the corresponding parameters, users can realize the customized pretreating process. The specific functions for pretreating molecules and calculating protein descriptors of BioJava and PyBioMed are completely different. ChemminR can retrieve molecules from PubChem database while PyBioMed can retrieve molecules from NCBI, CAS, KEGG databases. ChemminR can calculate some basic molecular properties while PyBioMed can calculate various and diverse molecular descriptors, especially a series of featured fingerprints. PyDPI is able to calculate interaction descriptors but only limited to chemical-protein interaction feature. PyBioMed realized to represent nine types of interaction features. (3) The application scenarios are not the same though some tools can calculate the same kinds of descriptors. For example, BioJava is for Java environment, and ChemminR is for R environment, while PyBioMed is designed for Python environment. In summary, PyBioMed is a different tool from them and Fig. 1 The main modules of the PyBioMed library and their corresponding wide applications in chemoinformatics, bioinformatics and drug discovery process provides lots of unique functionalities, especially when extracting features of complex interaction data and realizing a full data analysis pipeline.

Conclusion
PyBioMed provides a freely available and ease-to-use python library to calculate various features of biological molecules and complex interaction samples conveniently. It makes a step in this direction providing a way to fully integrate information from chemical space and biology space into an interaction space, which cannot be performed by other existing tools. As far as we know, PyBioMed is the first python library that calculates both individual descriptors of three types of molecular objects and any interactions between two of them in a uniform framework. It provides not only the detailed information about all descriptors and how to calculate them but also several tutorials and corresponding model scripts for different applications. In addition, the functions and modules related in PyBioMed and the stability of the library was extensively tested. We hope that the library will be helpful when exploring questions concerning structures, functions and interactions of various molecular data in the context of chemoinformatics. We also expect that our/other groups may use the free code of PyBioMed and the new machine learning models to implement public web servers. The increasingly diversified and further applications of molecular descriptors urge new descriptors and new tools to be developed, and researches have achieved some results [56][57][58][59]. In future work, we plan to apply the integrated features on various biological research questions, and to extend the range of functions with new promising descriptors for the coming versions of the library.

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