Freely available compound data sets and software tools for chemoinformatics and computational medicinal chemistry applications

We have generated a number of compound data sets and programs for different types of applications in pharmaceutical research. These data sets and programs were originally designed for our research projects and are made publicly available. Without consulting original literature sources, it is difficult to understand specific features of data sets and software tools, basic ideas underlying their design, and applicability domains. Currently, 30 different entries are available for download from our website. In this data article, we provide an overview of the data and tools we make available and designate the areas of research for which they should be useful. For selected data sets and methods/programs, detailed descriptions are given. This article should help interested readers to select data and tools for specific computational investigations.


Introduction
For compound data mining and the development and evaluation of chemoinformatics methods, public domain databases have become indispensable resources. Currently, major public repositories include PubChem 1 , BindingDB 2 , ChEMBL 3 , and ZINC 4 . While the former three databases contain compounds and bioactivity data, the latter collects commercially available compounds that are typically not annotated with activity information. Bioactivity data are usually obtained from original literature or patent sources. From these databases, screening data sets (PubChem) and compound activity classes (BindingDB, ChEMBL) can be extracted. Benchmarking of newly developed computational methods typically depends on the availability of such activity classes. Many compounds and measurements are now also shared between these databases. In addition, there are also a number of smaller public and commercial compound databases, which we do not consider here for the purpose of our discussion (with one exception; see below).
Importantly, depending on the scientific questions under investigation, it is often required to design and assemble data sets with specific features. Such data sets, which are usually reported as a part of a publication describing a computational analysis, a new method, or a benchmark investigation, are only infrequently made available to the public. Herein, we describe data sets originating from our laboratory that can be freely obtained. In addition, we also provide information about software tools developed by us that are available via the same website.

Objectives
The data sets and software tools reported herein have been generated for research activities that essentially fall into four different areas, as reported in Table 1. Area A comprises virtual screening and machine learning applications and is a core area of chemoinformatics. Areas B and C represent molecular selectivity analysis and visualization of structure-activity relationships (SARs), respectively. Furthermore, area D summarizes data mining activities with a focus on structure-activity or -selectivity relationships. Areas B-D are equally relevant for medicinal chemistry (and also chemoinformatics). In addition, especially area B is also relevant for chemical biology. By describing these tools in context, it is hoped that their accessibility to researchers in these areas might be further increased.

Materials and methods
Data sets reported herein were mostly, but not exclusively, assembled from BindingDB and ChEMBL on the basis of defined selection criteria (as specified in the original publications). These sets contain compound structures, provided as SMILES 5 strings or SD files 6 , and -whenever appropriate-associated bioactivity information. Some of the older activity classes that are still available from our website have originated from the license-restricted Molecular Drug Data Report (MDDR) 7 . Therefore, these data sets do not contain compound structures, but only compound identifier information (because a user must obtain a license to access the database). We do no longer license commercial or otherwise restricted databases and will remove corresponding entries from our download section in the near future (to ensure that all compounds are freely available). For the time being, all the information can be accessed. Scripts and programs available from our site generally represent a new computational method or analysis protocol and were implemented in-house in different scripting and programming languages, as specified in the respective entries. Source code is provided. All data sets and programs can be obtained via following URL: http:// www.limes.uni-bonn.de/forschung/abteilungen/Bajorath/labwebsite/downloads. The download section is updated several times per year with materials reported in new publications. Table 1 lists all 30 currently available entries including data sets and/or methods/programs. In each case, research area indices are assigned and publication information is provided. Eleven entries are assigned to area A and one, nine, and five entries to area B, C, and D, respectively. In addition, three entries are assigned to areas A and B and one to B and D. Compound data sets originated from 22 different studies and methods from four. In addition, for four other investigations, both methods and data sets are provided.

Results and discussion
For compound data sets, short descriptions are provided in Table 1. Furthermore, method/program descriptions are given in Table 2. In a number of instances, data entries contain sets of (filtered) compound activity classes (ACs) to ensure reproducibility of results reported in a specific publication. These sets were often directly taken from BindingDB, ChEMBL, the MDDR, or original literature sources and might not be of above-average interest. Nevertheless, for benchmarking of virtual screening methods, these sets are useful. However, other data sets have been especially designed for novel applications. In the following, selected data sets and methods are described in more detail that might be of particular interest for investigators in the designated (or other) research areas. The indices of these entries are highlighted in Table 1 and Table 2.
Selected compound data sets Entry 1: The ACs in this set are designed to have increasing intraclass structural diversity and hence represent test cases of increasing degrees of difficulty for the evaluation of ligand-based virtual screening (LBVS) methods.

Entry 4:
In 26 so-called selectivity sets, compounds are organized on the basis of differential potency against pairs of targets as a measure of selectivity. These sets were originally designed to evaluate an extension of standard similarity searching termed selectivity searching. The sets can be used as test cases for any methods that evaluate or predict molecular selectivity, similar to entries 5 and 6. As such, these data sets are relevant for computational chemical biology.
Entry 5: Selectivity sets focusing on the biogenic amine G protein coupled receptor (GPCR) family.
Entry 6: Eighteen sets with further refined selectivity criteria targeting four different protein families.
Entry 7: Twenty-five different sets are provided that contain compounds with increasing topological complexity and molecular size. These compound sets were designed to evaluate molecular complexity effects in similarity searching. They can be utilized to examine the complexity and/or size dependence of a computational method. Table 1. Publicly available data sets and programs. A list of 30 entries providing data sets and/or methods/programs is shown. For each entry, research area indices are assigned as described in the text, i.e. area 'A' indicates virtual screening (similarity searching), fingerprint engineering and machine learning; area 'B' represents molecular selectivity analysis, area 'C' SAR visualization, and area 'D' structureactivity or -selectivity relationship-oriented data mining. In addition, publication information is given. For compound data sets, short descriptions are provided. Selected compound data sets are highlighted in red and discussed in the text.

Entry
Year Area Index Provided Data set description

Selected methods and programs
Entry 10: A graphical data structure termed combinatorial analog graph (CAG) is introduced to systematically organize analog series on the basis of substitution patterns and identify subsets of analogs having high in SAR information content.

Entry 14:
A further extended and refined CAG implementation for the study of SARs across multiple targets.
Entry 15: SARANEA (a semantic construct of SAR and "Araneae", i.e., the scientific order of spiders) is a collection of different tools for graphical and numerical SAR analysis. It contains the networklike similarity graph (NSG), an SAR network (reminiscent of "spider webs") in which compounds are nodes and edges structural similarity relationships. In addition, nodes are annotated with different levels of SAR information. Several NSG variants have been introduced for different aspects of SAR exploration. The SARANEA tool collection was designed for large-scale SAR data mining and analysis, comparison of global and local SAR features, and the study of structure-selectivity relationships.
Entry 16: A program to calculate and display three-dimensional activity landscapes of compound data sets. An activity landscape is defined as any graphical representation that integrates molecular similarity and activity relationships. A 3D activity landscape can be conceptualized as a 2D projection of a chemical reference space (in which compound dissimilarity increases with inter-compound distance) with an interpolated potency surface added as the third dimension.

Entry 18:
The similarity-potency tree (SPT) is a graph representation that organizes compound neighborhoods in large data sets on the basis of structural nearest neighbor relationships and reveals chemically interpretable SAR information. This data structure can be understood as a compound-centric activity landscape view. A basic SPT implementation is also available as a part of SARANEA.
Entry 17: Sets of matched molecular pairs (MMPs) are given that were systematically extracted from BindingDB and ChEMBL. An MMP is defined as a pair of compounds that only differ by the exchange of a single fragment (substructure).

Entry 24:
On the basis of systematic similarity search profiling of ChEMBL, 50 ACs were selected. These sets represent meaningful test cases for benchmarking of LBVS methods. The ACs were assembled because they were neither too "easy" nor too "difficult" for standard similarity searching using different molecular fingerprints.
Entry 25: This database contains a collection of known active reference compounds, newly identified actives (hits), and screening database information extracted from original literature sources reporting prospective LBVS applications. Only studies were considered that provided sufficiently detailed information to reproduce the search calculations. These studies were identified in a systematic survey of published LBVS applications. The database provides an alternative benchmark system for LBVS. For example, on the basis of these compound sets, it can be determined whether a new methodology is capable of reproducing the results of successful prospective virtual screens using other approaches (i.e., screens that have identified structurally novel and experimentally confirmed hits).
Entry 30: Sets of activity cliffs are provided that belong to five newly introduced structural categories. These cliffs were systematically extracted from ChEMBL (latest release). An activity cliff is defined as a pair of structurally similar or analogous compounds with a large difference in potency. Accordingly, activity cliffs typically represent a rich source of SAR information.
Limitations of data sets Entries 1, 4-7, 9, 11, and 12 (assembled until 2010) only contain MDDR compound identifiers, but no structures, due to license restrictions, as commented on above. Entry 27: "Scaffold hopping", i.e., the detection of active compounds having different structural frameworks (core structures), is the ultimate goal of LBVS and its primary measure of success. However, the evaluation of the scaffold hopping potential of different LBVS methods is complicated by the fact that scaffold hops can involve similar or different core structures, which is generally not taken into account in the statistical assessment of benchmark investigations. An algorithm is presented that calculates the structural distance between any two scaffolds, regardless of their chemical composition or size. Application of this method makes it possible to quantify the degree of difficulty involved in computational scaffold hopping exercises.

Conclusions
Herein we have given an overview of specialized compound data sets and methods/programs that have originated from different research projects in our laboratory and that are made freely available to others with interests in chemoinformatics, computational medicinal chemistry, and chemical biology. These tools were presented and described in context. We hope that this report will further alert investigators in our and other scientific fields to available resources for specific computational applications and help to select data sets and tools that are relevant for given research topics. It is also hoped that the introduced methodological concepts will further evolve through wide use by others.
Author contributions JB conceived the study, YH collected and organized the data and information, YH and JB wrote the manuscript.

Competing interests
No competing interests declared.

Grant information
The author(s) declared that no grants were involved in supporting this work.
In addition to releasing a large number of datasets, the authors have also provided software that allows those reading their papers to critically evaluate their group's published methods on their own datasets. The release of this software also enables the reader to compare new methods with methods they may already be using. I applaud Hu and Bajorath for their efforts to promote open science and to provide datasets that will hopefully provide the basis for a large body of future work.
I have read this submission. I believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.
No competing interests were disclosed. Competing Interests: Considerable effort is spent in bio-and chemo-informatics groups in the development of curated datasets to validate methods or draw conclusions regarding specific questions. Even if such data appears in the literature as part of particular papers, it may not be easily found when looking for data for related but different questions. As the data may be useful for such related questions, Considering that literature searches are the primary method to search such data, I find it worthwhile to make datasets more widely available though a publication such as this. Two critiques follow though:

Article Comments
To the authors: their website is not providing access to the data in any way other than individual 1. downloads of the data from specific studies. A more dedicated interface, with some of the discussion of the datasets as it appears in the paper would be worthwhile considering that in the future more datasets will be added to the site but the paper will not be updated.

To
: The abstract section of a paper like this should have a section on availability where 2.
F1000Research the URL giving access to the data should be included. A similar abstract section appears in papers in the journal Bioinformatics.
No competing interests were disclosed. Competing Interests: