Analysis of a large food chemical database: chemical space, diversity, and complexity

Databases and data curation

Four chemical databases were homogeneously curated and analyzed, namely: FooDB version 1.0 (accessed November, 2017) (The Metabolomics Innovation Centre, 2017), drugs approved for clinical use available in DrugBank 5.0.2. (Law et al., 2014), GRAS (Burdock & Carabin, 2004), and a random subset of drug-like natural products from ZINC 12 (Irwin & Shoichet, 2005), of a size comparable to FooDB. The GRAS and DrugBank sets used in this work also have been used as reference in other comparative studies (Medina-Franco et al., 2012). The random set from ZINC was employed just as reference and other random sets from ZINC could be used. Compounds from all databases were washed and prepared using Wash MOE 2017 node in KNIME version 3.5.3 (Berthold et al., 2008). Briefly, the washing protocol implemented in MOE included removing salts and neutralizing the charges in the molecules. The largest fragments were kept and duplicates in each dataset deleted. Table 1 summarizes the databases and sizes after data preprocessing.

Chemical space visualization

The visual representation was generated with ChemMaps, a novel method for large chemical space visualizations (Naveja & Medina-Franco, 2017). Briefly, ChemMaps is able to generate two- and three-dimensional representations of the chemical space based. It uses as input the pairwise chemical similarity computed using fingerprints data. This approach exploits the ‘chemical satellites’ concept (Oprea & Gottfries, 2001), i.e., molecules whose similarity to the rest of the molecules in the database yield sufficient information for generating a visualization of the chemical space. Further details of ChemMaps are described elsewhere (Naveja & Medina-Franco, 2017).

Physicochemical properties

Six physicochemical properties (PCP) were calculated with RDKit KNIME nodes version 3.4, namely: SlogP (partition coefficient), TPSA (topological polar surface area), AMW (atomic mass weight), RB (rotatable bonds), HBD (hydrogen bond donors) and HBA (hydrogen bond acceptors). For the analysis reported in this short communication, these properties were selected based on their broadly extended use for cross-comparison of compound databases of biological relevance. However, additional properties can be calculated.

Molecular complexity

Fraction of sp³ carbons and number of stereocenters were computed for FooDB as measures of structural complexity. Despite the fact that there are several other measures, these two are straightforward to interpret, easy to calculate and are becoming standard to make cross comparisons among databases (Méndez-Lucio & Medina-Franco, 2017). As described in the Results and Discussion section, the computed values for FooDB were compared to literature data already reported for the reference data sets.

Scaffold content

The term “molecular scaffold” is employed to describe the core structure of a molecule (Brown & Jacoby, 2006). Different approaches have been proposed to consistently obtain a molecule’s scaffold in silico. In this work, scaffolds were generated under the Bemis-Murcko definition using the RDKit nodes available in KNIME (Bemis & Murcko, 1996). Bemis and Murcko define a scaffold as “the union of ring systems and linkers in a molecule”, i.e., all side chains of a molecule are removed.

Global diversity

The so-called “global diversity” (or total diversity) of FooDB was assessed and compared to other reference collections using a consensus diversity plot (González-Medina et al., 2016). As described recently, a consensus diversity plot simultaneously represents, in two-dimensions, four diversity criteria: structural (based on pairwise molecular fingerprint similarity values), scaffolds (using Murcko scaffolds computed as described in the Scaffold content section), physicochemical properties (based on the six properties described in Physicochemical properties section), and database size (the number of compounds) (González-Medina et al., 2016). The structural diversity of each data set is represented on the X-axis and was defined as the median Tanimoto coefficient of MACCS keys fingerprints. The scaffold diversity of each database is represented on the Y-axis and was defined as the area under the corresponding scaffold recovery curve, a well-established metric to measure scaffold diversity (Medina-Franco et al., 2009). The diversity based on PCP was defined as the Euclidean distance of six auto-scaled properties (SlogP, TPSA, AMW, RB, HBD, and HBA - vide supra) and is shown as the filling of the data points using a continuous color scale. The relative number of compounds in the data set is represented with a different size of the data points (smaller data sets are represented with smaller data points).

Visual representation of the chemical space

Chemical space of FooDB in comparison with the compounds of the three reference databases is visualized in Figure 1. The figure also shows the individual comparisons of FooDB with GRAS, DrugBank and natural products subset from ZINC, respectively. As shown in Figure 1a, the coverage of chemical space of FooDB is quite large as compared to other datasets. Most GRAS compounds lie within the chemical space framed by FooDB (Figure 1b): indeed, 1,193 compounds (53% of GRAS) are structurally identical between the two databases. Hence, FooDB largely contains and upgrades structural information from GRAS. There is significant overlap with approved drugs (Figure 1c) and natural products from ZINC with FooDB (Figure 1d).

Figure 1. Representation of the chemical space of FooDB.

The visual representation was generated with ChemMaps (Naveja & Medina-Franco, 2017). a) Comparison of FooDB with three reference collections. Panels b–d) show comparisons of FooDB with individual data sets.

Distribution of physicochemical properties

Figure 2 shows the boxplots for the distribution of PCP in all the four databases. For better visualization, the outliers above or below the median +/- 1.5 interquartile range are omitted. As expected, due to the large structural diversity, distribution of PCP in FooDB is broad, in many cases overcoming even approved drugs. For most properties, except RB, several compounds in FooDB share the properties of drugs, and drug-like natural products in ZINC. The comparable physicochemical properties between compounds from FooDB and DrugBank encourages additional systematic investigations for bioactivity of food components. Of course, during this search one needs to consider that compounds with similar properties may have different activity profile. In turn, GRAS consists mostly of small-sized compounds. Table S1 (Supplementary File 1) summarizes the statistics for FooDB and other reference collections.

Figure 2. Distribution of physicochemical properties.

Box plots of the distribution of six physicochemical properties of FooDB and reference data sets. SlogP (partition coefficient), TPSA (topological polar surface area), AMW (atomic mass weight), RB (rotatable bonds), HBD (hydrogen bond donors) and HBA (hydrogen bond acceptors).

Molecular complexity

For FooDB, the fraction of sp³ carbons (mean: 0.62; standard deviation: 0.28) and the number of stereocenters (mean: 4.7; standard deviation: 7.1) indicated a high structural complexity. For comparison, it has reported that the mean of the fraction of sp³ carbons for approved drugs, compounds in the clinic and a general screening collections of organic compounds is 0.47, 0.41 and 0.32, respectively (González-Medina et al., 2016; Lovering et al., 2009). Moreover, the reported mean of the fraction of sp3 carbons for natural products collections ranges between 0.41 and 0.58 (for natural products in ZINC and Traditional Chinese Medicine (López-Vallejo et al., 2012). The complexity of compounds in FooDB is comparable to molecules in GRAS (mean: 0.63; standard deviation: 0.28) (González-Medina et al., 2016).

Scaffold content

Figure 3 shows the frequency of the most common scaffolds in FooDB. Many compounds are acyclic (32%), followed by monocyclic compounds with a benzene (6%), cyclohexene (2%) and tetrahydropyran (1%) as a core structure. The benzene ring is the most common core scaffold in chemical databases used in drug discovery (Bemis & Murcko, 1996; Singh et al., 2009; Yongye et al., 2012). Many of the most frequent scaffolds in FooDB are also common in other compound databases of natural products (González-Medina et al., 2017). In a follow-up work, it will be interesting to explore the type of functional groups commonly present in the acyclic structures of FooDB.

Figure 3. Frequency of the ten most common scaffolds in FooDB.

Recently, Schneider et al. published an analysis on the selectivity of Bemis-Murcko scaffolds based on public bioactivity data available in ChEMBL (Schneider & Schneider, 2017). 78 of the 585 scaffolds reported therein were present in FooDB. The list of the 78 matching scaffolds, along with the original statistics calculated by Schneider et al., is made available as Dataset 1 (Naveja et al., 2018a). Of note, the three most frequent scaffolds in FooDB (benzene, cyclohexane and tetrahydropyran, with more than 300 compounds - Figure 3) are matching scaffolds. Interestingly, the mean Information content (I) value of all 585 Schneider’s scaffolds is 2.8 (sd= 0.6), while the subset of the 78 scaffolds also present in FooDB has a mean I value of only 2.1 (sd = 0.7). Lower I values point towards more promiscuous scaffolds (Schneider & Schneider, 2017), an expected finding given the nature of the database. As example, Table S2 (Supplementary File 1) shows and discusses briefly the statistics for the three most frequent matching scaffolds.

Polyphenols. Since polyphenols are an important class of compounds in food chemistry (Rasouli et al., 2017), we investigated and quantified the amount of polyphenols in FooDB. Polyphenols are well-known antioxidants, which may play a role in the prevention of several diseases including type 2 diabetes, cardiovascular diseases, and some types of cancer (Neveu et al., 2010). In this line, it is known that oxidative/nitrosative stress has a pivotal role in pathophysiology of neurodegenerative disorders and other kinds of disease (Ebrahimi & Schluesener, 2012). Polyphenols have been demonstrated to elicit several biological effects in in vitro and ex vivo tests (Del Rio et al., 2010; Scalbert et al., 2005).

The molecular structure of polyphenols includes at least two phenolic groups, or one biphenol, and up to any additional number of OH substitutions in aryl rings. They may be classified by their structure in two major groups: flavonoids and non-flavonoids (phenolic acid derivatives) (Del Rio et al., 2013). Some polyphenols, such as quercetin, are found in all plant products, whereas others are specific to particular foods. In many cases, food contain complex mixtures of polyphenols, which are often poorly characterized (Manach et al., 2004).

Polyphenols are also a common chemical motif among natural products, and they are often associated to promiscuity (Tang, 2016). In this work it was found that 3,228 (13.5%) compounds in FooDB are polyphenolic. The list of all 3,228 polyphenolic compounds is made available as Dataset 2 (Naveja et al., 2018b). This set of polyphenols is larger than the 502 polyphenols from food indexed in Phenol-Explorer (Neveu et al., 2010). For comparison, all the reference databases used in this work contained less polyphenols than FooDB. GRAS, ZINC and DrugBank contained 15 (0.6%), 24 (0.1%) and 325 (3.7%) polyphenols, respectively. The large list of polyphenols identified from FooDB is larger than the list of 1,395 polyphenols identified and used in the recent work of Lacroix et al. (Lacroix et al., 2018) that was retrieved from Phenol-Explorer and the Dictionary of Natural Products. Indeed, the list of 3,228 polyphenolic compound made available in this work can be used to augment the already extensive polyphenol-protein interactome work of Lacroix et al. (Lacroix et al., 2018).

Global diversity

Since the diversity of compound data sets depend on the molecular representation (Sheridan & Kearsley, 2002), a global assessment of the diversity of FooDB was analyzed using different criteria: molecular fingerprints, scaffolds, physicochemical properties and number of compounds. The four criteria were analyzed in an integrated manner through a Consensus Diversity Plot generated as described in the Global diversity section of the Methods. The Consensus Diversity Plot in Figure 4 shows that FooDB has about average diversity both by fingerprints and relatively low diversity by scaffolds. Although PCP (represented with the color of the data points) are extremely diverse, structural motifs seem to reappear with slight variations. Figure 4 shows the overall large fingerprint and scaffold diversity of approved drugs (e.g., data points towards the lower left region of the plot). Similarly, the relative global diversity of GRAS i.e., high fingerprint diversity but low scaffold diversity (e.g., upper left region of the plot), is consistent with previous comparisons of these compounds with other reference data sets (González-Medina et al., 2016; Medina-Franco et al., 2012).

Figure 4. Consensus Diversity Plot of FooDB and reference data sets.

The structural diversity of each data set is represented on the X-axis and was defined as the median Tanimoto coefficient of MACCS keys fingerprints. The scaffold diversity of each database is represented on the Y-axis and was defined as the area under the corresponding scaffold recovery curve. The diversity based on physicochemical properties (PCP) was defined as the Euclidean distance of six auto-scaled properties (SlogP, TPSA, AMW, RB, HBD, and HBA) and is shown as the filling of the data points using a continuous color scale. The relative number of compounds is represented with a different size of the data points (smaller data sets are represented with smaller data points).