Fig 1.
Overview of the CSNAP approach for high-throughput compound target identification using Chemical Similarity Networks (CSNs).
(A) Discovery of diverse ligands from cell-based screens with unknown cellular targets. Note that structurally distinct compound classes are represented by different shapes, while structurally-related analogs within each class are labeled with different colors. (B) Target identification using CSNAP. Bioactivity database searches to identify structurally similar reference compounds with known target annotations. The grey nodes represent target annotated compounds. (C) A pair-wise similarity matrix was computed by considering both intra and inter-ligand similarity between query and reference compounds using Tanimoto coefficient (Tc) with cutoff > 0.7. (D) Structurally diverse ligands are clustered into chemical similarity subnetworks based on representative chemotypes (consensus chemical patterns). (E) The network topology was used to guide and quantify the protein-ligand interactions for drug target prediction. Two neighbor counting functions, S-score and H-score were applied to identify and rank the most common targets among the first-order neighbors of the query compounds within the CSN. In this example, compound α has a consensus Target A score = 3 and a Target C score = 1, whereas compound β has a consensus score = 1 for Target A, B and C. (F) Experimental target validation. The predicted targets were validated by comparing RNAi with compound-induced cellular phenotypes and by testing direct protein-ligand interactions in in-vitro assays.
Fig 2.
CSNAP validation using benchmark compounds.
(A) 206 compounds from six major drug classes (ACE, CDK2, HIVRT, HMGA, HSP90 and PARP) were analyzed using CSNAP with a Z-score cutoff of 2.5 and a Tanimoto coefficient (Tc) cutoff of 1. With the exception of 7 molecules, all compounds were ordered into chemical similarity subnetworks specific to each drug target. (B) Outcome of applying the neighbor counting function, S-score to predict the top 5 most common targets shared by the annotated-neighbor nodes of all input ligands within the CSN. The prediction accuracy (percentage of correctly predicted ligands) was determined by comparing the predicted target to ligand target annotations. CSNAP target prediction assessment for each drug class ranked by different S-score cutoffs (S-cutoff = 0, 5 and 10) gave an overall prediction accuracy of 89%, 73% and 60% respectively. (C) Comparison of the total percentage of target pool reduction (percentage of the total number of predicted targets with S-score cutoff over total number of predicted targets with S-score cutoff) against the overall prediction accuracy indicated that an S-score cutoff of 4 is optimal for hit enrichment and target virtual screening. (D) CSNAP target and off-target prediction for benchmark compounds. Predicted targets for the validation compounds were plotted against each drug class to identify targets and off-targets using Ligand-Target Interaction Fingerprints (LTIFs) analyzed on heat maps. The color intensity was scaled according to the S-score (0–1). Note that ACE and CDK2 inhibitors have predicted off-targets based on the additional coloring patterns, indicating drug poly-pharmacology. See S1 Fig for LTIF analysis of the combined drug classes.
Fig 3.
Target prediction accuracy comparison of network-based and ligand-based approaches.
(A) Comparison of the overall target prediction accuracy based on the top hit, top five hits and top ten hits analyzed by CSNAP or the SEA approach using 206 benchmark compounds comprised of six major drug classes (ACE, CDK2, HIVRT, HMGA, HSP90 and PARP). The result shows that CSNAP provides a substantial improvement in target prediction accuracy over the traditional ligand-based approach by pair-wise chemical similarity comparison. (B and C) Detailed target prediction accuracy comparison breakdown of each of the six drug classes predicted by (B) CSNAP and (C) SEA approach respectively. The comparison showed that CSNAP provided a greater success rate at identifying the major targets of promiscuous ligands such as CDK2 and ACE inhibitors, which resulted in low prediction accuracies by the traditional ligand-based method.
Fig 4.
Integration of CSNAP with knowledge databases for mitotic target prediction and phenotypic target validation.
(A) Mitotic compound chemical similarity network. CSNAP analysis of 212 mitotic compounds yielded 85 chemical similarity clusters representing diverse chemotypes, only 21 compounds were not clustered into annotated similarity graphs. (B) LTIF analysis of CSNAP mitotic target predictions. The target spectrum identified four major classes of targets from the top peaks including fatty acid desaturase (SCD), ABL kinase (ABL1), phosphatase (PTPN) and tubulin (TUBB). An independent LTIF analysis of each target class is presented in S2 Fig. (C) Mitotic compound deconvolution. Target associated chemical similarity sub-networks of four predicted targets (SCD, ABL1, PTPN and TUBB) were “pulled-down” from the mitotic CSN. For each cluster, at least one mitotic compound connected to one or more reference nodes with Tc threshold> 0.7. Note that the predicted SCD and ABL1 compounds display over-lapping neighbors, indicating that the predicted targets may be modulated by both compound sets. (D) Phenotypic validation of predicted mitotic targets. Asynchronous HeLa cells were treated with indicated compounds for 24 hours, fixed and stained for DNA and Tubulin. The observed compound-induced cell division defects were compared to target gene expression knockdown defects within the MitoCheck database. All compounds matched the previously characterized phenotypes associated with knockdown of target protein expression. See S6 Fig for complete compound-induced phenotypes.
Fig 5.
Network-based elucidation of a novel tubulin-targeting chemotype.
(A) In-vitro tubulin polymerization assays were used to test the effect of the 212 mitotic compounds on microtubule assembly at 50μM concentration. The end-point absorbance, based on change in OD (dOD), was used to quantify the degree of microtubule polymerization and was converted to percentage fold change relative to DMSO (0%). Among the tested compounds, 134 compounds (63%) had an effect (>20% fold change) on tubulin polymerization. (B) Chemical similarity sub-network consisting of 7 novel anti-tubulin ligands based on a phenyl-sulfanyl-thiazol-acetamide privilege scaffold. The connected analogues within the network showed a consensus tubulin destabilization effect where each step in the path (red) of the sub-network corresponded to a minimum structural change correlating with the observed structure-activity-relationship (SAR). (C) Docking of compound 6 into the β-tubulin colchicine-binding site based on the crystal structure (PDB: 1AS0) exhibited a similar predicted binding mode to colchicine. (D) Ligand alignment between compound 6 and colchicine identified a conserved pharmacophore critical for ligand binding, including the 2 and 10-methoxy groups and a 9-keto group that interacts with Cys-241 of beta tubulin and Val-181 (not shown) of alpha tubulin respectively. (E) Hydrophobicity map of docked compound 6 within the colchicine-binding site revealed a hydrophobic sub-pocket enclosed by Leu-248 and Lys-352. The model showed that compounds 7 and 8 enhance binding affinity by fitting the N-propyl and N-phenyl group in the hydrophobic cavity, consistent with the SAR analysis. See S11 Fig for molecular modeling of compounds 6–12. (F) The most potent compound 8 was tested for direct colchicine site binding using mass spectrometry competitive binding assays. Compound 8 competed strongly with colchicine for the colchicine-binding site, similar to the colchicine-site binder podophyllotoxin. Note that the negative control vincristine did not compete. (G) Immunofluorescence microscopy images of HeLa cells treated with DMSO, taxol, colchicine or compounds 6–8 for 20 hours. Cells were fixed and stained for DNA (Hoechst 33342) and tubulin (primary rat anti-tubulin antibodies and secondary anti-rat Cy3 antibodies). Scale = 5 μm. Note that colchicine, and compounds 6–8 depolymerize microtubules. See S10 Fig for compound 6–12 induced phenotypes.