Clustered Distribution of Natural Product Leads of Drugs in the Chemical Space as Influenced by the Privileged Target-Sites

Some natural product leads of drugs (NPLDs) have been found to congregate in the chemical space. The extent, detailed patterns, and mechanisms of this congregation phenomenon have not been fully investigated and their usefulness for NPLD discovery needs to be more extensively tested. In this work, we generated and evaluated the distribution patterns of 442 NPLDs of 749 pre-2013 approved and 263 clinical trial small molecule drugs in the chemical space represented by the molecular scaffold and fingerprint trees of 137,836 non-redundant natural products. In the molecular scaffold trees, 62.7% approved and 37.4% clinical trial NPLDs congregate in 62 drug-productive scaffolds/scaffold-branches. In the molecular fingerprint tree, 82.5% approved and 63.0% clinical trial NPLDs are clustered in 60 drug-productive clusters (DCs) partly due to their preferential binding to 45 privileged target-site classes. The distribution patterns of the NPLDs are distinguished from those of the bioactive natural products. 11.7% of the NPLDs in these DCs have remote-similarity relationship with the nearest NPLD in their own DC. The majority of the new NPLDs emerge from preexisting DCs. The usefulness of the derived knowledge for NPLD discovery was demonstrated by the recognition of the new NPLDs of 2013–2014 approved drugs.

T he focus of drug discovery has moved from natural products (NPs) to technology-derived synthetic molecules for about 20 years 1 without the anticipated drug productivity improvement 2 . Although being largely sidelined, NPs are still relevant 3,4 with NP-related small molecule drugs representing 29.5% of the 132 FDA approved drugs in 2008-2012 (Supplementary Table S1), NP-related drugs include NPs and NP semi-synthetic derivatives, mimetics, and pharmacophore-guided synthetic molecules 3 . The NPs from which the NP-related drugs have been derived are named as the NP leads of drugs (NPLDs). There is a renewed interest in discovering drugs 5 from NP privileged structures 6 and derivative libraries 7 . The knowledge of the distribution of the NPLDs in the chemical space provides useful clues for prioritizing the relevant efforts.
Although NPs are in well-defined subspaces of the chemical space 8 , because of their enormous number 9 , structural diversity 9,10 and molecular complexity 11 , only a fraction of NPs can be practically explored in the foreseeable future. Drug discovery efforts need to be prioritized towards the NPs with higher discovery potentials.
The key questions are which NPs to explore and where to find them. Evidences suggest that some NPLDs may congregate in specific drug-productive regions of the chemical space. Certain NP chemical classes (e.g. steroids and nucleosides) are drug prolific 12 . Half of the drugs are made of dozens of molecular frameworks 13 . For many drugs, their individual molecular, physicochemical and topological pharmacophore properties [14][15][16] and the corresponding principal components [17][18][19][20] are constrained in specific ranges. Moreover, the GPCR, kinase and protease targeting agents have been reported to each cluster together in the chemical space 1 . These studies have consistently shown that a substantial percentage of the NPLDs congregate in the chemical space. However, the extent and the detailed distribution patterns of the congregation of NPLDs in the chemical space and the mechanisms leading to such patterns have not been fully investigated.
There is a need to study these questions from different structural and molecular binding perspectives to gain a deeper understanding of the structural characteristics of NPLDs and to find clues for guiding the search of new NPLDs. In this work, we determined the distribution patterns of the 348 and 94 NPLDs of 749 pre-2013 approved and 263 clinical trial small molecule drugs (Supplementary Tables S2, S3) in the chemical space represented by the molecular scaffold trees 21 and the molecular fingerprint based hierarchical clustering tree 22,23 of 137,836 non-redundant NPs 3,24,25 . The number of NPs profiled here is comparable to those of the earlier large-scale NP studies 9,21,26 . Molecular fingerprints were used for representing NPs in the hierarchical clustering tree because of its demonstrated effectiveness in structural similarity searching, and its extensive applications in drug lead discovery 12,23,[27][28][29][30] .
The derived distribution patterns were studied from the perspective of preferential binding of NPLDs to the privileged target-sites for determining whether it contributes to the formation of these patterns. We also evaluated whether these patterns are distinguished from those of the bioactive NPs and how they evolve with time. We further tested whether the derived knowledge can be explored for NPLD discovery by applying it to retrospectively judge the development potential of the new NPLDs of 2013-2014 approved drugs uninvolved in the derivation of the NPLD distribution patterns. New technologies are expected to significantly expand the currently accessible NP chemical space 31,32 and their potential impact is not reflected in this study.
Drug scaffolds have been well analyzed 13 and drug distribution in the chemical space have been extensively studied from the perspectives of specific molecular and physicochemical properties 1,14-20 . To the best of our knowledge, our work is the first large-scale and systematic study of the detailed distribution patterns of the largest set of NPLDs in the chemical space from the perspectives of their molecular scaffolds and structures. The molecular scaffold analysis was intended for determining whether there is a significant change in the congregation patterns of the NPLDs in comparison to the previous studies 13,33 . The molecular structural analysis was intended for further probing the complex structural features of the NPLD congregation phenomenon and the underlying molecular mechanisms that might contribute to the clustering of NPLDs with particular focus on the possible influence of the binding of NPLDs and their derivatives to the privileged target sites.

Methods
We collected 442 NPLDs 1,3,27,28,34,35 and the information about their NP origin 1,28 from the literature. We also collected 169,037 NPs from the ZINC 24 , TCM-ID 29 , TCM@Taiwan 25 , and other literatures 3 . For database entries with multiple non-linked components, only the largest component was selected. Hydrogens were added and small fragments (counter ions, solvent molecules, etc.) were removed by using Corina, The number of NPs were reduced to 137,836 after removing the duplicate entries, small NPs with molecular weight ,50 Daltons (drug leads are .100 Daltons 30 ) and the NPs whose molecular fingerprints could not be computed by using available software tools such as PaDEL 36 . Duplicates were identified and removed by structural comparison based on a set of 98 molecular descriptors we have used for classifying bioactive molecules 37 and implemented in the online server MODEL 38 , open-source software PaDEL 36 , and our own software, which can distinguish different molecules non-distinguishable by the 881-bit Pubchem molecular fingerprints.
In deriving the molecular scaffold trees of the 442 NPLDs and 137,836 NPs, Scaffold Hunter v2.3.0 21 was used to select the NPLDs and NPs with ring structures and to subsequently cluster them into molecular scaffold trees by using default rule set in the Scaffold Tree Generation window. The molecular fingerprint based hierarchical clustering tree of the 442 NPLDs and 137,836 NPs was generated by using the Matlab statistics toolbox with the structures of the NPs represented by 2D molecular fingerprints 23 (specifically, the 881-bit PubChem substructure fingerprints computed by using PaDEL 36 ) and with their similarity levels measured by the Tanimoto coefficient Tc 22,23 and the complete linkage. Tc was used because it is the most popular similarity metric for molecular fingerprint based measurement of compound similarity 23 . Complete linkage was used because of its relatively good performance in clustering bioactive compounds in a recent comparative study 39 . The hierarchical tree graphs were generated by using EMBL automatic tree generator in iTOL version-1.8.1 40 with the distance of the NPs measured by the Tanimoto distance Td 5 1-Tc. In analyzing the physicochemical landscapes of the NPs in specific regions of the chemical space, we used MODEL 38 and Discovery studio 3.1.1 software to compute eight molecular descriptors frequently used for analyzing drug-like 14,41-43 and leadlike 43,44 features. These are molecular weight (MW), lipophilicity AlogP and logD, polarizability (PZ), and the number of O1N (ON), hydrogen bond donor (HD), hydrogen bond acceptor (HA), rotatable bond (RB), and rings (RI).
To determine whether the clustering of the NPLDs in specific sub-regions of the chemical space are statistically more significant than chance, our derived distribution patterns of the NPLDs with respect to those of the randomly shuffled NP communities were analyzed by the method used for determining the statistical significance of the phylogenetic clustering of traditional medicinal plants 45 . In this method, the mean Tanimoto distance MTd of the NPLDs in every NPLD-clustered sub-region was compared to the MTd values of these NPLDs in 60,000 randomly generated NP sub-regions. By using the algorithms implemented in the Phylocom: software 46 , a one-tailed P-value and a net relatedness index (NRI) were calculated for each subregion. The P-value is the number of randomly selected NPs that are more clustered than the NPLDs in each sub-region divided by the number of runs (60,000 in this study). The desired significance level a of the P-value was further adjusted by Bonferroni correction to a9 5 a/N (N is the number of independent statistical significance tests, which is 60 in this study) 47 . The NRI is a standardized effect size measure of the community structure in each sub-regions, which is the difference in average Td between the NPLDs and the 60,000 randomly generated NP sub-regions, and standardized by the standard deviation of the Td values in 60,000 randomly shuffled sets of NP communities. The sign of NRI informs whether the NPLDs are more clustered (NRI . 0) or more dispersed (NRI , 0) than the NPs in each subregion. These quantities were calculated by using Phylocom v4.1 46 with the Td values of the NPLDs and NPs as input data.

Results and Discussion
Distribution profiles of NPLDs in the chemical space from the perspectives of molecular scaffolds and molecular structures. There are 411 NPLDs and 134,097 NPs with ring structures. These were grouped by Scaffold hunter 21 into molecular scaffold trees of 39,051 scaffolds (114 are drug-productive). The distribution of the NPLD scaffolds in this large-scale analysis is similar to the previous findings 13, 33 , the majority (62.7%) of the NPLDs of the approved drugs and a substantial percentage (37.4%) of the NPLDs of the clinical trial drugs congregate in 62 drug-productive scaffolds or scaffold parent-child sub-branches (DSs) labeled as DS1 to DS62 ( Figure 1, Supplementary Table S4 and Figures S1-S5). A DS is defined as a scaffold with $2 NPLDs that have yielded $1 approved drug or a scaffold parent-child sub-branch with $2 NPLD-producing scaffolds that have yielded $1 approved drug.  These DSs have collectively yielded 69.6% approved and 44.4% clinical trial drugs. The congregation of NPLDs in the DSs coupled with the earlier finding that the GPCR, kinase and protease targeting agents each are clustered together in the chemical space 1 indicates that NPLDs of the same and different scaffolds against the same classes of targets may on a broader scale be clustered together in the chemical space. To facilitate the visualization of our generated scaffold trees by using Scaffold Hunter, the resulting scaffold database was exported as a SQL file that can be downloaded at http://bidd. nus.edu.sg/group/NPLD_Distribution/NP_ScaffoldHunter.zip.
To probe the larger-scale distribution patterns of NPLDs in the chemical space from the perspective of molecular structures, we generated a molecular fingerprint based hierarchical clustering tree of the 442 NPLDs and 137,836 NPs. The derived tree is composed of 33 main branches (Supplementary Figure S6 and Table S5). Most (87.9%) branches are drug-productive, reflecting the fact that NPs primarily co-evolve and interact with proteins 6 and a variety of chemical classes 3,4 and target families 48,49 have been therapeutically explored. Nonetheless, NPLDs within each branch are mostly clustered together, with 341 (77.2%) NPLDs (82.5% approved, 3.0% clinical trial) clustered in 60 drug-lead productive clusters (DCs) labeled as DC1 to DC60 ( Figure 2, Supplementary S7-S10 and Table S6). A DC is defined as a relatively small region of the molecular fingerprint characterized chemical space with moderate to high concentration of NPLDs yielding $1 approved drug.
To facilitate the analysis of the clustered distribution of the NPLDs, we generated the heat map of the proximity matrix of 442 NPLDs against 137,836 non-redundant NPs. The proximity matrix was calculated by using molecular fingerprint Tanimoto distance Td between NPLDs and NPs with the row and column positions representing the NPLDs and NPs in the same order as their respective positions in the hierarchical clustering tree of the NPLDs and NPs. The heat map was created by using the heatmap.2 function of the gplots package in R with the red to yellow colors indicating the stronger to weaker structural similarity between the NPLDs and NPs. The heat map for branch 4 and 9 are shown in Supplementary Figures S11-S12 and those of the other branches can be downloaded from http://bidd.nus.edu.sg/group/NPLD_Distribution/NP_heatmaps. zip.
We found that 11.7% of the NPLDs in the DCs have remotesimilarity relationships (0.57 # Tc , 0.7) with the nearest NPLD in their own DC, and another 24.9% of the NPLDs in these DCs have intermediate-similarity relationship (0.7 # Tc , 0.85) with their nearest NPLD in their own DC. Remote-similarity relationships have been reported in compounds with cross-pharmacology relationships 50 and between a bioactive compound and its scaffold hopping parent bioactive compound 51 . Therefore, the DCs broadly cover the high-similarity to remote-similarity relationships for capturing similar activities, and cross-pharmacology and scaffold hopping types of relationships.
These DCs have collectively yielded 87.9% approved and 68.8% clinical trial drugs. In particular, 56.0% approved and 67.4% clinical trial NPLDs are clustered in 22 NPLD-prolific DCs ( Table 1) that have collectively yielded 68.4% approved and 39.2% clinical trial drugs, which is consistent with the report that half of the drugs are made of dozens of molecular frameworks 13 . The NPLD-prolific DCs were ranked based on the ratio of the approved NPLDs to the NPs in each DC. Partly because of the inadequate exploration and partly because of the limited availability of the relevant information, these ratios may not fully reflect the reality but nonetheless provide useful indications. We found that 60% of the top-10 NPLD-prolific DCs with .100 searchable NPs in Table 1 are among top-ranked DCs with higher approved NPLD to NP ratios. Thus, drug productivity of these top-ranked DCs seems to arise from higher NPLD yields instead of the higher number of NPs explored. The top-ranked DC38 and DC8 in Table 1 were excluded because they have ,100 searchable NPs. If counted, they are among the DCs with highest approved NPLD to NP ratios.
Statistical significance of the clustering of NPLDs in the DCs. The statistical significance of the clustering of the NPLDs in every DC was evaluated by using the Phylocom software 46 to calculate the P-value and NRI of the NPLDs against the chance clustering of the NPs in the DC from 60,000 sets of randomly selected NPs, as outlined in the Method section. We found that there are statistically more NPLDs in most of the DCs than expected by chance, with 78.3% of the DCs having P # 0.0095 and additional 10% of the DCs having 0.011 # P # 0.0362 respectively (Table 2), which correspond to very strong (P # 0.01) and strong (0.01 , P # 0.05) presumption against null hypothesis respectively 52 . The P-value of the remaining 4 (6.7%) and 3 (5%) DCs are in the range of 0.0525 # P # 0.0736 and 0.1233 # P # 0.1857 respectively, which correspond to low (0.05 , P # 0.1) and no (P . 0.1) presumption against null hypothesis respectively 52 . It is noted that each of these seven DCs has only 2 NPLDs and there is a possibility that the low statistical significance of these DCs are partly due to the few discovered NPLDs in these DCs.
To further provide a more conservative evaluation of the statistical significance of the clustering of NPLDs in the DCs, Bonferroni correction for study-wide hypothesis testing was performed. We found that, under the Bonferroni correction with a9 5 0.05/60, 48.3% DCs still have statistically strong or very strong significance against null hypothesis ( Table 2). On the other hand, 16.7% DCs showed weak and 35% DCs showed no significance. It is noted that the majority (67.7%) of these weak or no significance DCs have 2-3 NPLDs in their respective DCs. The low number of NPLDs in each of these DCs likely leads to a higher tendency of forming a distribution pattern with weaker statistical significance that can be exposed by stricter statistic tests. www.nature.com/scientificreports  53 . We found that the targets of each individual DC are primarily from one to a few target classes (e.g. amine receptors) with their substrates/ ligands from one to a few chemical classes (e.g. amines). This finding is based on the limited target information for 74.9% NPLDs and without considering the additional targets of the non-NPLDs in each DC. While the limited target information may not enable a full investigation of the influence of drug target-sites, it nonetheless provides useful hints about the key factors that promote the clustering of NPLDs. The 203 targets can be classified based on their target-sites into 45 target-site classes (TCs) labeled as TC1 to TC45, which collectively belong to 20 target-site super-classes (TSs) labeled as TS1 to TS20 (Supplementary Table S7). A TS is defined as a group of target-sites bound by substrates/ligands of a specific chemical class (e.g. amine binding sites) irrespective of their targets. A TC represents a sub-group of target-sites of a specific target class (e.g. amine transporters) bound by substrates/ligands of a specific chemical class. The targets in 53 (96.4%) DCs are from 1-3 TCs (27, 19, 7 DCs from 1, 2, 3 TCs) with the majority (65.5%) from either 1 TC (27 DCs) or 2-3 TCs of 1 TS (9 DCs), and the remaining 2 DCs are from 4 TCs (Figure 3). This indicates that the similar target-site structural constraints are likely the key factors in promoting the clustering of NPLDs in individual DCs. The targets of approved and clinical trial drugs are highly selective in their numbers, druggability features, and systems profiles 48,[53][54][55] , and the druggability features have been characterized by the affiliation of the family members of the studied target to the known drug targets 56 and by the existence of a privileged binding site with unique physicochemical properties 57 for enabling favorable binding by drug-like molecules 58 .
Our revealed links between the clustering of NPLDs in individual DCs and the grouping of their targets in selected TCs are consistent with these findings. NPLDs in these DCs possess structural, physicochemical and/or pharmacophore features complementary to a privileged target-site, are at or near activity peaks against the target, and have good or amendable safety and pharmacokinetic properties. They either have or may be further optimized to gain such additional features as adequate metabolic stability 59 , metabolite safety 60 , absorption 61 and physical forms 62 to reach the drug sweet spots 63 in the chemical space. Therefore, our revealed clustered patterns of NPLDs and their links to the selected TCs provide useful information and enable further study of the distribution profiles of the NPLDs in the chemical space particularly with respect to the relevant target-site classes.
The other drug-prolific DCs are also closely linked to specific TCs (Table 1), with DC19 (steroids) linked to the nuclear receptor ligand binding sites (TC44), DC5 (aminoglycosides) to the DNA metabolism enzyme nucleoside phosphate (TC7) and ribosome 30 s subunit aminoacyl-tRNA (TC18) binding sites, DC17 (acarviosins) to the phosphatase substrate (TC19), ribosome 23S rRNA peptidyl transferase (TC25) and outer membrane lipopolysaccharide (TC35) sites, DC21 (fatty acids, prostanoids) to the retinoid receptor ligand (TC37) and coenzyme A analog metabolism enzyme substrate (TC40) binding sites, DC28 (cardiac glycosides) to the nucleoside Detailed analysis of the physicochemical landscape of the NPLD distribution profile reveals clues for searching the sweet spots in the DCs. While the knowledge of the clustered distribution patterns of the NPLDs in the DCs and the correlation to the TCs is useful for revealing the NPLD-like structural frameworks for targeting specific target classes, more detailed analysis is needed for identifying the NPLDs within each DC. On the other hand, drug-like 14,41-43 and leadlike 43,44 rules have been derived and extensively used for identifying drug leads on the basis of whether their specific physicochemical properties fall into certain drug-likeness or lead-likeness ranges. Optimal hydrophobic and hydrogen bond interactions, and thus the AlogP/logD and ON/HD/HA values, are important for optimizing NPLDs into drugs 64 . Therefore, additional clues for searching the new NPLDs may be obtained by studying the physicochemical landscapes of the known NPLDs and the NPs in the DCs.
We evaluated the physicochemical landscape of the NPLDs and NPs in branch 9 characterized by the eight physicochemical properties MW, AlogP, logD, ON, HD, HA, RB, RI, and PZ frequently used for analyzing drug-like 14,41-43 and lead-like 43,44 features (Supplementary Figure S13). This branch contains four DCs (DC16, DC17, DC18, and DC19). While, the NPs inside and outside these DCs have mixed MW values, there is a significantly higher concentration of NPs with either higher AlogP/logD values or higher ON/HD/HA values. In particular, regardless of their MW values, the NPLDs tend to be located at the peak of either AlogP/logD or ON/HD/HA. For instance, the NPLDs in DC17 (which include macrolides, polyenes, spinosyns and acarviosins) have peak MW values likely due to the added or enlarged hydrophobic groups (peak AlogP/logD values) to optimally interact with, e.g., the outer membrane lipopolysaccharide sites of TC35, or the added hydrogen bonding components (peak ON/HD/HA values) to optimally interact with, e.g., the phosphatase substrate sites of TC19. The NPLDs in DC19 (composed of steroids and derivatives) tend to have peak AlogP or logD values without significantly enlarged MW values over other NPs in the DC, possibly due to enhanced hydrophobic components within the steroid structural framework for achieving optimal hydrophobic interactions with the nuclear receptor ligand sites of TC44. Therefore, the tendency of the NPLDs to be located at either the AlogP/logD or the ON/ HD/HA peaks in the DCs may be potentially used as an indicator for searching new NPLDs.
The distribution profiles of NPLDs with respect to bioactive NPs. The more clustered distribution of NPLDs (Figure 2 and Supple-mentary S7-10) are in contrast to the much less clustered distribution of the 48,216 bioactive NPs from the TCM@Taiwan database 25 and the literatures 3 ( Supplementary Figures S14-17). Although the number and diversity of our collected bioactive NPs are limited in representing bioactive NPs, useful indications may be revealed. These bioactive NPs are more diversely distributed in 32 of the 33 branches with 78.8% of the bioactive NPs located outside the DCs. Therefore, NPLDs are distinguished from bioactive NPs in their tendency to more closely cluster together in the chemical space, which is consistent with the distribution pattern of drug-productive species families in the phylogenetic tree (drug-productive species families are more closely clustered than the species families of bioactive NPs) 3 . To investigate whether the more clustered distribution of NPLDs in the DCs is due to the more extensive exploration efforts towards these DCs, the exploration times of the 442 NPLDs, crudely estimated by the time since the first literature report, were compared to those of the 11,816 bioactive NPs inside and outside the DCs, which are largely comparable to each other (Supplementary Figures S18-21). Hence, there is no clear indication to link drugproductivity of the DCs to the biased exploration efforts.
The distribution profiles of NPLDs with respect to time and disease classes. Since 1988, the number of DCs has been gradually increased at an average rate of 3.2 new DCs per 5 years, and the majority (60.0%-69.0%) of the 15-32 new NPLDs approved in every five-year period from 1560 to 2012 are from preexisting DCs (Table 3). Drug discovery focus has been shifting in terms of targets, chemotypes, diseases and therapeutic strategies 48,65 . To study if novel drugs derived from shifted focuses are outside pre-existing DCs, we analyzed 27 new NPLDs approved in 1990-2012 each targeting a novel target previously unaddressed by an approved drug (Supplementary Table S8) and thus are novel NPLDs of the time. At the time of their first drug approval, 18 (66.7%) of these novel NPLDs were from preexisting DCs, suggesting that existing DCs remain good sources of novel NPLDs and drugs.
The usefulness of the knowledge of NPLD distribution profile for facilitating new NPLD discovery. The tendencies of NPLDs to cluster together in the chemical space and to preferentially bind to the privileged target-sites in the target-space may be explored for assessing the development potential of new NP leads. Based on the insights derived from our analysis, one can postulate that, apart from the ability of an NPLD or its derivatives to modulate a validated target, an NPLD may have a higher probability to be developed into a drug if it is inside a DC, near a DC (to form an expanded DC) or near an NPLD outside existing DCs (to form a new DC) in the chemical space, and if its target belongs to an existing TC or a new TC in an existing TS. This postulation was tested by the retrospective analysis of the new NPLDs of FDA approved drugs in 2013-June 2014 66,67 that were not used in the derivation of the NPLD distribution patterns and the target-site linkages. Our literature search led to the finding of 4 new small molecule NPLDs, 3 of which were retrospectively recognized as developable based on our postulation (Table 4). Specifically, the NPLD uridine monophosphate of sofosbuvir is inside DC5 and target TC7, the NPLD phlorizin of canagliflozin is near DC57 (Tc 5 0.91 to the nearest NPLD) and target monosacharide transporter substrate sites as a new TC in TS12 (saccharide binding sites), and the imidazole-based NPLD (e.g. mizoribine) of luliconazole is inside DC13 and target a steroid metabolism enzyme substrate site in TS19 (steroid binding sites). Therefore, the insights derived from this and other studies of NPLD distribution profiles may be explored for facilitating the assessment of the development potential of NP leads.

Concluding Remarks
This study systematically exposed the clustered distribution profiles of NPLDs and revealed useful insights into the mechanisms that partly contribute to the formation of these profiles, i.e., the tendency of NPLDs to preferentially bind to the privileged target-sites. The insights from this and other studies of NPLD distribution profiles provide useful clues to and enable further studies of the lead sweet spots in the chemical space with respect to the corresponding targetsites. The distribution of NPLDs and the lead sweet spots in the chemical space is collectively influenced by potent binding to the target-sites and such additional factors as the optimization potential to reach the drug sweet spots in the chemical space 63 with more adequate metabolic stability 59 , metabolite safety 60 , absorption 61 and physical forms 62 ). Further studies are needed for a deeper understanding of the collective influence of these multiple factors on the distribution of NPLDs in the chemical space. These advances coupled with expanded knowledge of lead-like and drug-like structures and physicochemical properties 13 may enable more prioritized and rational exploration of the NP-subspaces for drug discovery.