A molecular structure matching approach to efficient identification of endogenous mammalian biochemical structures.

Metabolomics is the study of small molecules, called metabolites, of a cell, tissue or organism. It is of particular interest as endogenous metabolites represent the phenotype resulting from gene expression. A major challenge in metabolomics research is the structural identification of unknown biochemical compounds in complex biofluids. In this paper we present an efficient cheminformatics tool, BioSMXpress that uses known endogenous mammalian biochemicals and graph matching methods to identify endogenous mammalian biochemical structures in chemical structure space. The results of a comprehensive set of empirical experiments suggest that BioSMXpress identifies endogenous mammalian biochemical structures with high accuracy. BioSMXpress is 8 times faster than our previous work BioSM without compromising the accuracy of the predictions made. BioSMXpress is freely available at http://engr.uconn.edu/~rajasek/BioSMXpress.zip.

Introduction

Metabolomics is the comprehensive, qualitative, and quantitative study of all the small molecules, called metabolites, in an organism [1]. A major challenge in metabolomics is the interpretation of the vast amount of data produced by the high-throughput techniques used for information extraction and data interpretation [2,3]. The existence of several on-line chemical structure databases has provided a vital support for molecular identification by allowing the search for candidate compounds using experimentally determined features with computationally simulated features. Such searches often result in a large number of false positives, making identification of the compound under investigation extremely difficult. Hence, cheminformatics methods are needed to efficiently search such large chemical databases and potentially identify unknown endogenous biochemical compounds. Several methods have been developed with the objective of discriminating between candidate structures that are synthetic and those that are biochemical. The first attempt to solve this problem was reported by class="Chemical">Nobeli et al. [4] who used two-dimeclass="Chemical">nsioclass="Chemical">nal molecular structures to maclass="Chemical">nually derive a library of 57 structural fragmeclass="Chemical">nts commoclass="Chemical">nly fouclass="Chemical">nd iclass="Chemical">n 745 class="Chemical">n class="Species">E. coli metabolites. Such fragments were used to reveal the main constituents of metabolites and to assist in the classification of the metabolome into biochemically relevant classes. In related work, Gupta and Aires-de-Sousa [5] and Peironcely et al. [6] employed fingerprints and random forest classifiers [7] to classify endogenous biochemical compounds. Molecular fingerprints represent the structure of a molecule as a list of binary values (0 or 1) that indicate the presence or absence of structural features in the molecule [8]. Gupta and Aires-de-Sousa's model correctly annotates 95% of the 1,811 compounds downloaded from the Kyoto Encyclopedia of Genes and Genomes (KEGG) [9] used for training their model. While Peironcely et al. reported that 96% of 457 compounds downloaded from the Human Metabolome Database (HMDB) [10], not used for training the model, were classified as endogenous. In our previous work [11], we developed BioSM, a molecular classifier that can identify endogenous mammalian biochemical structures contained within chemical structure space. BioSM uses the structures of known endogenous mammalian biochemical compounds to aid in the classification process, as opposed to other works that use fragments of known structures. BioSM correctly predicted 95% of 1,388 (KEGG) compounds as endogenous mammalian biochemical in a set of leave-one-out cross validation experiment. Additionally, 89% of 2,330 compounds downloaded from HMDB were identified as endogenous metabolites. One of BioSM's limitations, granting its encouraging results, was its need to exhaustively search all known biochemical structures to be able to make a decision about the candidate compound under investigation which resulted in an undesirably high run time. In this paper we propose an efficient cheminformatics tool, which can be used to identify biological compounds based on their molecular structures, called BioSMXpress. Similar to our previous work, the prediction method applied by BioSMXpress relies on a set of endogenous mammalian biochemical compounds obtained from the KEGG database, hereafter referred to as scaffolds. In a nutshell, BioSMXpress is designed to predict that a given query structure is biological after encountering exactly one scaffold match satisfying a given threshold. This threshold is based on the number of atoms in both the query structure and the scaffold being examined. Knowing this gives us the opportunity to avoid the need to exhaustively search the entire scaffolds list before making a decision about the query compound with confidence. To do this efficiently, BioSMXpress selects the scaffolds that have the potential to promote the candidate in the least possible time. Then, only those scaffolds, with enough atoms to satisfy the given threshold, are checked against the candidate compound for similarity.

Materials and methods

BioSMXpress was designed as an enhancement to BioSM with the aim of making the least possible number of structure comparisons to efficiently identify biochemical structures with the aid of a scaffolds list. BioSMXpress decides if a candidate structure is biochemical based upon how similar that structure is to any of the structures in the scaffolds list. In this work, two molecular structures are considered to be a "match", if the smaller structure is an exact substructure (atom and bond types) of the larger structure being compared. The underlying cheminformatics functionality of BioSMXpress is based on an open source Java based toolkit called the Small Molecule Subgraph Detector (SMSD) [12]. This toolkit is used to find the maximum common sub-graph between small molecules using atom type matches and bond n class="Gene">sensitivity iclass="Chemical">nformatioclass="Chemical">n. Iclass="Chemical">n additioclass="Chemical">n to SMSD, BioSMXpress uses Marviclass="Chemical">n, a chemical structure processiclass="Chemical">ng software, to geclass="Chemical">nerate both the caclass="Chemical">noclass="Chemical">nical SMILES aclass="Chemical">nd atom couclass="Chemical">nts from structure data files (.sdf) for all the compouclass="Chemical">nds described iclass="Chemical">n this work.

Computational algorithm

Here, we introduce a tool that can efficiently identify small endogenous n class="Species">mammalian biochemical structures from chemical structure space. First we will start by deficlass="Chemical">niclass="Chemical">ng some class="Chemical">notatioclass="Chemical">ns followed by a detailed explaclass="Chemical">natioclass="Chemical">n of the computatioclass="Chemical">nal model behiclass="Chemical">nd BioSMXpress. Let cbe the molecular structure of a query compouclass="Chemical">nd, S = {s1, s2,...,s} be a set of class="Chemical">n small biological compouclass="Chemical">nds (scaffolds). Let s∼ ciclass="Chemical">ndicate that scaffold sis a substructure of caclass="Chemical">ndidate compouclass="Chemical">nd c, aclass="Chemical">nd AC (s) represeclass="Chemical">nt the class="Chemical">number of atoms iclass="Chemical">n compouclass="Chemical">nd s. Let miclass="Chemical">nAC deficlass="Chemical">ne the miclass="Chemical">nimum class="Chemical">number of atoms required iclass="Chemical">n a scaffold sto ideclass="Chemical">ntify cas biological if c∼ s. If two molecular structures r aclass="Chemical">nd q were fouclass="Chemical">nd to be a match, a similarity score

is computed where r ∼ q. Based on a given substructure threshold (subThr), the minimum atom count is computed as

Similarly, based on a given superstructure threshold (superThr ), BioSMXpress computes the maximum atom count,

Finally, let be the scaffold list assigned to cwhere . A scaffold is assigned to if and only if . Please note that each candidate structure with a different atom count is provided with a different set of scaffolds in . Once is populated with the appropriate scaffolds for c, BioSMXpress examines cagainst each of those scaffolds. As soon as a match (substructure or superstructure) is found, BioSMXpress predicts that cis biological and terminates. Otherwise, it's predicted to be nonbiological. Values for subThr and superThr are determined by cross validation as described in the following section.

In addition to that, BioSMXpress orders the potential scaffolds in such that scaffolds with atom counts closer to the candidate atom count are examined first followed by those with a larger atom count difference. In this case, once a match is found the search terminates and it is guaranteed that this is the best possible match (as a substructure or superstructure). Figure 1 illustrates a visual example of the BioSMXpress scaffolds ordering process.

Figure 1

Scaffold selection and sorting process. In this example, it is assumed that the candidate compound (c) consists of 9 atoms and that subThr = 0.5 and superThr = 0.51. Therefore, minAC = [9 ∗ 0.5] = 4 and . (a) The hashed scaffolds list with minAC and maxAC identified. (b) The sorted scaffolds list consists of all the scaffolds with 9 atoms followed by those with 10 atoms followed by those with 8 atoms and so on.

Datasets

We will briefly describe the source and nature of the datasets selected to train and validate BioSMXpress. Since these datasets will be used to compare between BioSMXpress and BioSM in terms of prediction accuracy, we utilized the same datasets and followed the same curation steps in [11]. In each dataset, compounds with any of the following characteristics were eliminated: (1) compounds with elements other than C, H, class="Chemical">N, O, P aclass="Chemical">nd S; (2) compouclass="Chemical">nds with less thaclass="Chemical">n 4 atoms aclass="Chemical">nd more thaclass="Chemical">n 53 atoms (explaiclass="Chemical">ned below); (3) compouclass="Chemical">nds that were class="Chemical">n class="Chemical">polymers; (4) charged structures except those in which the charge was due to quaternary amines or sulfonium ions; (5) compounds with duplicate structures; and (6) compounds with disjoint structures. We start by defining compounds used to define biological versus non-biological in chemical structure space in this study.

Biological Dataset (Scaffolds list)

The KEGG database was chosen as the source of endogenous n class="Species">mammalian compouclass="Chemical">nds. The list of 1,564 class="Chemical">n class="Species">mammalian scaffolds (KEGGscafs) defined in [11] were used to represent the biochemical structure space in BioSMXpress. Each compound in the scaffolds list comprises of a number of atoms from 4 to 80 atoms per compound.

Non-Biological Dataset (Synthetic compounds list)

The Chembridge http://www.chembridge.com and Chemsynthesis http://www.chemsynthesis.com databases served as the sources of compounds representing the non-biological chemical space. These databases were selected because they comprise synthetic compounds for chemical synthesis and drug screening and design. After curation, a set of 375,930 structures represented the synthetic compounds list. Chemsynthesis and Chembridge databases mainly contain compounds with low molecular weights (a maximum atom count of 53 atoms per compound). Accordingly, 143 of the 1,564 KEGGscafs (with atom count between 54 and 80) were eliminated from any testing set throughout this study and were only used for superstructure scaffold matching. This restriction was enforced to ensure that the sole discrimination between a compound being biological or non-biological is based on the structure of a compound and not on the number of atoms in that compound.

Training Dataset

A total of 2,842 compounds, with at least 4 atoms and at most 53, were used to train and test our predictive model. Half of those compounds were obtained from the scaffolds list (representing the endogenous n class="Species">mammalian chemical space) aclass="Chemical">nd the other half from the syclass="Chemical">nthetic compouclass="Chemical">nds list (represeclass="Chemical">nticlass="Chemical">ng the class="Chemical">noclass="Chemical">n-biological chemical space). The later molecules were raclass="Chemical">ndomly selected from the syclass="Chemical">nthetic dataset to match the atom couclass="Chemical">nt distributioclass="Chemical">n of the 1,421 biological set.

Independent Datasets

To estimate the performance of our predictive model and compare it with that of BioSM, four external validation sets were used: one set of putative class="Species">human metabolites, oclass="Chemical">ne set of placlass="Chemical">nt secoclass="Chemical">ndary metabolites, oclass="Chemical">ne set of drugs, aclass="Chemical">nd oclass="Chemical">ne set of syclass="Chemical">nthetic compouclass="Chemical">nds. For each dataset, aclass="Chemical">ny compouclass="Chemical">nd with a structure ideclass="Chemical">ntical to aclass="Chemical">ny of those iclass="Chemical">n the scaffolds list was removed. Also, structures fouclass="Chemical">nd iclass="Chemical">n more thaclass="Chemical">n oclass="Chemical">ne dataset were removed from all datasets except oclass="Chemical">ne. Molecules iclass="Chemical">n each dataset had to satisfy both mass (50 - 700 Da) aclass="Chemical">nd atom couclass="Chemical">nt (4 - 53 atoms) coclass="Chemical">nstraiclass="Chemical">nts to allow for a fair comparisoclass="Chemical">n betweeclass="Chemical">n BioSMXpress aclass="Chemical">nd BioSM. Additioclass="Chemical">nally, compouclass="Chemical">nds with at least oclass="Chemical">ne class="Chemical">noclass="Chemical">n-biological substructure (class="Chemical">n class="Disease">NBS) were eliminated. NBSs are substructures that are not commonly found in mammalian biochemical compounds. This decision was based on our interest in comparing the core predictive models of BioSM and BioSMXpress since in reality, NBS filters will be applied to both models before any scaffold comparisons are involved. For more details on the curation process followed please refer to [11].

The following is a brief description of the five datasets. Please note that the numbers of compounds reported below refer to the datasets after curation. The first dataset consisted of 2,329 compounds and was obtained from HMDB version 2.5 representing putative class="Species">human metabolites. The secoclass="Chemical">nd dataset coclass="Chemical">nsists of 2,416 secoclass="Chemical">ndary placlass="Chemical">nt metabolites, as specified by KEGG, represeclass="Chemical">nticlass="Chemical">ng placlass="Chemical">nt structures. The drug dataset was represeclass="Chemical">nted by 3,282 compouclass="Chemical">nds aclass="Chemical">nd was obtaiclass="Chemical">ned from DrugBaclass="Chemical">nk [13] versioclass="Chemical">n 3.0 aclass="Chemical">nd from the 1989 USAclass="Chemical">n class="Chemical">N and the USP Dictionary of Drug Names. A randomly chosen set of approximately 46,203 molecules was from the National Center for Biotechnology Information's (NCBI) PubChem database [14]. PubChem is the largest freely accessible compound database currently available. Finally, a set of 374,509 compounds from the Chembridge and Chemsynthesis databases, not used in training the model, were used as a synthetic compound test set.

Results and discussion

Classification methods selection

Four classification methods were proposed for BioSMXpress specifically, the SSF method which refers to finding a substructure scaffold match or a superstructure scaffold match in the sorted scaffolds list was utilized. SSSF refers to searching for the best substructure scaffold similarity score (Sc), if existent, and the best superstructure scaffold similarity score (Sc), if existent. It declares the candidate as biological if Sc+ Sc≥ sumThr. From my experience with BioSM, I found that distributing candidates into mass bins, with each bin having its own threshold values, showed an improvement in the prediction quality. Thus, we decided to test if the same concept applies here. I split the set of test compounds into five bins based on the number of atoms per compound and used CV to evaluate the model. This introduced two more methods, class="Gene">SSB aclass="Chemical">nd Sclass="Chemical">n class="Gene">SSB, similar to SSF and SSSF respectively, except that there are independent thresholds assigned to each bin.

Cross Validation (CV) is a method used for estimating how accurately a predictive algorithm will perform in practice while avoiding overfitting as well as tuning meta-parameters [15]. In this study, we used a nested CV framework where parameter tuning was performed by executing 5-fold CV on the training data and the classification accuracy was empirically assessed using 2-fold CV on the testing data. For the results of each training fold, the score where n class="Gene">SENS = SPEC was recorded as the cutoff threshold of that fold. This process was repeated 5 times to iclass="Chemical">nsure that each of the 5 parts was evaluated. Theclass="Chemical">n the average thresholds of all five traiclass="Chemical">niclass="Chemical">ng sets were used as the cutoff values wheclass="Chemical">n evaluaticlass="Chemical">ng the CV testiclass="Chemical">ng data as explaiclass="Chemical">ned iclass="Chemical">n.

We ran 15 CV experiments to evaluate the performance of each method. Several accuracy measures were applied to each experiment such as class="Gene">sensitivity , specificity , aclass="Chemical">nd matthews correlatioclass="Chemical">n coefficieclass="Chemical">nt . Were TP refers to the class="Chemical">number of correctly ideclass="Chemical">ntified compouclass="Chemical">nds while FP refers to the class="Chemical">number of iclass="Chemical">ncorrectly ideclass="Chemical">ntified molecules. Similarly, Tclass="Chemical">n class="Chemical">N refers to the number of correctly rejected compounds and FN refers to the number of those incorrectly rejected. The mean and standard deviation of the 15 experiments are displayed in Table 1. The highest sensitivity of 90% was obtained by the SSF classifier. At the same time, SSF suffered from the lowest specificity of 55% only. Another observation is that the application of sumThr improved the specificity significantly, 82% (SSSB) and 71% (SSSF) versus 62% (SSB) and 55% (SSF) but affected the sensitivity negatively (53% and 73%, respectively). SSB had the best MCC of 51% and hence was selected as the method of choice for BioSMXpress with a sensitivity of 86% and a specificity of 62%.

Table 1

Mean and standard deviation of accuracy measures obtained for 15 cross validation experiments using 4 different scoring.

		SSF	SSSF	SSB	SSSB
SENS	Mean	0.90	0.73	0.86	0.58
	StdDev	0.02	0.04	0.02	0.03
SPEC	Mean	0.55	0.71	0.62	0.82
	StdDev	0.04	0.05	0.04	0.03
MCC	Mean	0.41	0.45	0.51	0.48
	StdDev	0.03	0.02	0.05	0.05

Leave-One-Out cross validation analysis

As an additional method to evaluate how well BioSMXpress can identify endogenous class="Species">mammalian biochemical structures usiclass="Chemical">ng the class="Chemical">n class="Gene">SSB classifier, we carried out a set of LOOCV experiments using the SSB method with the averaged subThr, superThr and bin boundaries determined by CV as explained in the Methods section. Here, 1,421 experiments representing KEGG structures (with atom count between 4 and 53 atoms/compound) were carried out. Please note that in every experiment, the scaffolds list was composed of 1,420 compounds plus 143 compounds (those with atom count between 54 and 80 atoms/compound) as the scaffolds list. As a result, BioSMXpress annotated 94% of the 1,421 compounds as being biological structures.

Using 1,387 scaffolds in a set of LOOCV experiments implemented by BioSMXpress and BioSM independently were implemented and compared. These 1,387 compounds were the scaffolds that satisfied the constraints of both BioSM and BioSMXpress (mass in the range of 50 - 700 Da and number of atoms in the range of 4 - 53). BioSM was capable of identifying 94.5% of the 1,387 scaffolds as biochemical structures while BioSMXpress identified 94.2%. Figure 2 shows the breakdown of the results of this comparison with compounds binned by atom count.

Figure 2

Biological predictions resulting from a set of LOOCV experiments by BioSM. Compounds were binned by atom count.

BioSMXpress performed best when identifying compounds in the first bin (99% positive identification of biochemical compounds) while BioSM was able to predict only 92% of those compounds. In general, Figure 2 indicates that BioSMXpress is better at identifying biochemical compounds in the first and second bins (99% and 97%, respectively), while BioSM is better at identifying biochemical compounds in the fourth and fifth bins (96% and 94%, respectively).

A note worth mentioning is that in addition to the 1,387 scaffolds annotated, BioSMXpress was able to positively identify 34 compounds in the scaffolds list that were rejected by BioSM without classification due to mass restrictions (masses were greater than 700 Daltons). This indicates that BioSMXpress has broadened the range of compounds examined just by restricting the number of atoms in a candidate compound versus its molecular mass.

Prospective validation

Independent datasets containing plant secondary metabolites, drugs, class="Species">human metabolites, syclass="Chemical">nthetic molecules, aclass="Chemical">nd PubChem compouclass="Chemical">nds were classified by BioSMXpress. Subsequeclass="Chemical">ntly, the same datasets were also aclass="Chemical">nclass="Chemical">notated by BioSM. Table 2 preseclass="Chemical">nts a comparisoclass="Chemical">n of BioSM's predictioclass="Chemical">ns versus those of BioSMXpress. The results iclass="Chemical">ndicate that 91% of the 2,329 HMDB molecules were correctly classified as eclass="Chemical">ndogeclass="Chemical">nous class="Chemical">n class="Species">mammalian metabolites while 88% of them were identified by BioSM. Predictions for the 2,416 plant metabolites by BioSMXpress and BioSM were comparable with 72% and 73%, respectively. As for the 3,282 drug compounds, 58% were predicted to be biological by BioSMXpress versus 62% by BioSM. In contrast, only 25% of the randomly selected 46,203 PubChem compounds were predicted as biological by BioSMXpress as opposed to 35% by BioSM.

Table 2

Predictive results using the SSB classifier for 6 different datasets.

Dataset	Number of Compounds	BioSM	BioSMXpress
HMDB Compounds	2,329	88%	91%
Plant Metabolites	2,416	73%	72%
Drug Compounds	3,282	62%	58%
PubChem Compounds	46,203	35%	25%
Synthetic Compounds	374,509	33%	36%

Predictive results using the n class="Gene">SSB classifier for 6 differeclass="Chemical">nt datasets.

In addition to these four prospective datasets, a set of 374,509 synthetic compounds (represented by Chembridge and Chemsynthesis compounds) were evaluated by BioSMXpress and BioSM with 36% and 33% of these being predicted to be biochemical, respectively. Overall, the comparison in Table 2 shows that the biochemical prediction percentages made by BioSM and BioSMXpress are practically comparable except that BioSMXpress predicted 10% lesser compounds of the PubChem compounds as biological.

Execution and CPU time comparison

n class="Chemical">Now that we have showclass="Chemical">n that the predictive performaclass="Chemical">nce of BioSMXpress is aclass="Chemical">nalogous to BioSM, iclass="Chemical">n this sectioclass="Chemical">n we discuss their time performaclass="Chemical">nce. We used a high-eclass="Chemical">nd cluster http://becat.ucoclass="Chemical">nclass="Chemical">n.edu/hpc/ hosted by the School of Eclass="Chemical">ngiclass="Chemical">neericlass="Chemical">ng aclass="Chemical">nd the Taylor L. Booth Eclass="Chemical">ngiclass="Chemical">neericlass="Chemical">ng Ceclass="Chemical">nter for Advaclass="Chemical">nced Techclass="Chemical">nology (BECAT) at the Uclass="Chemical">niversity of Coclass="Chemical">nclass="Chemical">necticut to ruclass="Chemical">n aclass="Chemical">nd compare the performaclass="Chemical">nce of both BioSM aclass="Chemical">nd BioSMXpress. We raclass="Chemical">n both classifiers with a set of raclass="Chemical">ndomly geclass="Chemical">nerated datasets as caclass="Chemical">ndidate datasets for predictioclass="Chemical">n. Each dataset was evaluated by both predictive models uclass="Chemical">nder the same circumstaclass="Chemical">nces (same class="Chemical">number of cluster class="Chemical">nodes, threads, same scaffolds list, etc.) aclass="Chemical">nd the time for each model was recorded. We were also iclass="Chemical">nterested iclass="Chemical">n compariclass="Chemical">ng the CPU time utilized by both BioSM aclass="Chemical">nd BioSMXpress. CPU time is the amouclass="Chemical">nt of time for which a ceclass="Chemical">ntral processiclass="Chemical">ng uclass="Chemical">nit (CPU) was used for processiclass="Chemical">ng iclass="Chemical">nstructioclass="Chemical">ns of a computer program or operaticlass="Chemical">ng system, as opposed to, for example, waiticlass="Chemical">ng for iclass="Chemical">nput/output operatioclass="Chemical">ns. Figure 3 shows the average CPU time utilized by each of the classifiers wheclass="Chemical">n makiclass="Chemical">ng predictioclass="Chemical">ns for each data set size (50 - 50,000 compouclass="Chemical">nds). Similar to respoclass="Chemical">nse time, BioSMXpress has outperformed BioSM by utiliziclass="Chemical">ng aclass="Chemical">n average of 7 times less CPU time. We geclass="Chemical">nerated multiple caclass="Chemical">ndidate datasets with 50, 100, 500, 1,000, 5,000, 10,000, aclass="Chemical">nd 50,000 compouclass="Chemical">nds. Each dataset was composed of raclass="Chemical">ndomly selected compouclass="Chemical">nds from a pool of all the iclass="Chemical">ndepeclass="Chemical">ndeclass="Chemical">nt datasets used iclass="Chemical">n this study as described iclass="Chemical">n the Methods sectioclass="Chemical">n. To eclass="Chemical">nsure that the oclass="Chemical">nly factor we are measuriclass="Chemical">ng is the class="Chemical">number of compouclass="Chemical">nds iclass="Chemical">n a set class="Chemical">no matter what the class="Chemical">nature of the compouclass="Chemical">nds iclass="Chemical">ncluded is, we geclass="Chemical">nerated multiple raclass="Chemical">ndom datasets (specifically 3) with the same class="Chemical">number of compouclass="Chemical">nds for each size required. So we raclass="Chemical">n 3 groups each coclass="Chemical">ntaiclass="Chemical">niclass="Chemical">ng 50 raclass="Chemical">ndomly selected compouclass="Chemical">nds, 3 groups of 100 compouclass="Chemical">nds aclass="Chemical">nd so oclass="Chemical">n, aclass="Chemical">nd theclass="Chemical">n reported the average respoclass="Chemical">nse time of each group size. Figure 3 displays the average ruclass="Chemical">n time of BioSM versus that of BioSMXpress wheclass="Chemical">n aclass="Chemical">nclass="Chemical">notaticlass="Chemical">ng datasets of sizes 50 to 50,000 compouclass="Chemical">nds as explaiclass="Chemical">ned above.

Figure 3

(a) Average runtime (in hh:mm:ss) needed to make predictions using BioSM versus BioSM. (b) Average CPU time (in seconds) for BioSM and BioSMXpress when annotating sets of compounds of different sizes.

Obviously, BioSMXpress impressively outperformed BioSM across all datasets. BioSMXpress was 10 times faster than BioSM when analyzing 10,000 compounds. Across all datasets examined, BioSMXpress provided an 8 times average speed up over BioSM. Another interesting observation is that it takes BioSM an average of 6 minutes and 51 seconds to evaluate 1,000 compounds while it takes BioSMXpress an average of 5 minutes and 8 seconds to evaluate 10,000 compounds (10 times more compounds in less time).

This drastic difference in run time and CPU time can be explained by observing the number of scaffold comparisons required by each predictive model to make a prediction about any given candidate compound. BioSM needs to compare the candidate structure with each and every structure in the scaffolds list accumulating scores and then finally comparing that score with a threshold to make a prediction. BioSMXpress intelligently selects and sorts the scaffolds that would produce the highest match scores, based on the thresholds, if they were to match the candidate. Only a portion of the scaffolds are added to the list that is actually used by BioSMXpress as the scaffolds list and once a match is found the candidate is predicted to be biological with no other computational steps further needed.

Conclusions

In this paper, we describe the development and validation of BioSMXpress, an efficient supervised cheminformatics tool that uses endogenous n class="Species">mammalian biochemical scaffolds to predict whether a caclass="Chemical">ndidate chemical structure is biochemical or syclass="Chemical">nthetic. BioSMXpress is at average 8 times faster thaclass="Chemical">n BioSM without compromisiclass="Chemical">ng the accuracy of the predictioclass="Chemical">ns. BioSMXpress was able to correctly classify 94% of 1,421 biochemical compouclass="Chemical">nds iclass="Chemical">n a set of leave-oclass="Chemical">ne-out cross validatioclass="Chemical">n experimeclass="Chemical">nt. Thus BioSMXpress may be useful for searchiclass="Chemical">ng large chemical databases iclass="Chemical">n metabolomics applicatioclass="Chemical">ns where the class="Chemical">number caclass="Chemical">ndidates is extremely large as well as the class="Chemical">number of poteclass="Chemical">ntial false positives iclass="Chemical">n aclass="Chemical">n efficieclass="Chemical">nt maclass="Chemical">nclass="Chemical">ner.

Competing interests

The authors declare that they have no competing interests.

Authors' contributions

M.A.H. was responsible for designing the algorithm, software development, and manuscript preparation. She was also responsible for testing and benchmarking BioSMXpress. S.R. was involved in the overall supervision of the project, manuscript preparation, intellectual input, and guidance. All authors have given approval to the final version of the manuscript.