Systematic Prioritization of Druggable Mutations in ∼5000 Genomes Across 16 Cancer Types Using a Structural Genomics-based Approach*

A massive amount of somatic mutations has been cataloged in large-scale projects such as The Cancer Genome Atlas (TCGA) and the International Cancer Genome Consortium projects. The majority of the somatic mutations found in tumor genomes are neutral 'passenger' rather than damaging “driver” mutations. Now, understanding their biological consequences and prioritizing them for druggable targets are urgently needed. Thanks to the rapid advances in structural genomics technologies (e.g. X-ray), large-scale protein structural data has now been made available, providing critical information for deciphering functional roles of mutations in cancer and prioritizing those alterations that may mediate drug binding at the atom resolution and, as such, be druggable targets. We hypothesized that mutations at protein–ligand binding-site residues are likely to be druggable targets. Thus, to prioritize druggable mutations, we developed SGDriver, a structural genomics-based method incorporating the somatic missense mutations into protein–ligand binding-site residues using a Bayes inference statistical framework. We applied SGDriver to 746,631 missense mutations observed in 4997 tumor-normal pairs across 16 cancer types from The Cancer Genome Atlas. SGDriver detected 14,471 potential druggable mutations in 2091 proteins (including 1,516 recurrently mutated proteins) across 3558 cancer genomes (71.2%), and further identified 298 proteins harboring mutations that were significantly enriched at protein–ligand binding-site residues (adjusted p value < 0.05). The identified proteins are significantly enriched in both oncoproteins and tumor suppressors. The follow-up drug-target network analysis suggested 98 known and 126 repurposed druggable anticancer targets (e.g. SPOP and NR3C1). Furthermore, our integrative analysis indicated that 13% of patients might benefit from current targeted therapy, and this –proportion would increase to 31% when considering drug repositioning. This study provides a testable strategy for prioritizing druggable mutations in precision cancer medicine.

A massive amount of somatic mutations has been cataloged in large-scale projects such as The Cancer Genome Atlas (TCGA) and the International Cancer Genome Consortium projects. The majority of the somatic mutations found in tumor genomes are neutral 'passenger' rather than damaging "driver" mutations. Now, understanding their biological consequences and prioritizing them for druggable targets are urgently needed. Thanks to the rapid advances in structural genomics technologies (e.g. X-ray), large-scale protein structural data has now been made available, providing critical information for deciphering functional roles of mutations in cancer and prioritizing those alterations that may mediate drug binding at the atom resolution and, as such, be druggable targets. We hypothesized that mutations at protein-ligand binding-site residues are likely to be druggable targets. Thus, to prioritize druggable mutations, we developed SGDriver, a structural genomics-based method incorporating the somatic missense mutations into protein-ligand bindingsite residues using a Bayes inference statistical framework. We applied SGDriver to 746,631 missense mutations observed in 4997 tumor-normal pairs across 16 cancer types from The Cancer Genome Atlas. SGDriver detected 14,471 potential druggable mutations in 2091 proteins (including 1,516 recurrently mutated proteins) across 3558 cancer genomes (71.2%), and further identified 298 proteins harboring mutations that were significantly enriched at protein-ligand binding-site residues (adjusted p value < 0.05). The identified proteins are significantly enriched in both oncoproteins and tumor suppressors. The follow-up drug-target network analysis suggested 98 known and 126 repurposed druggable anticancer targets (e.g. SPOP and NR3C1). Furthermore, our integrative analysis indicated that 13% of patients might benefit from current targeted therapy, and this -proportion would increase to 31% when considering drug repositioning. This study provides a testable strategy for prioritizing druggable mutations in precision cancer medicine. Several large-scale cancer genome sequencing projects, such as The Cancer Genome Atlas (TCGA) 1 and the International Cancer Genome Consortium (ICGC), have uncovered a large volume of somatic mutations in human cancers. These data provide an unprecedented opportunity to identify somatic mutations that play a pathogenic role in cancer development and progression (e.g. "driver" mutations) that, as such, can be exploited with a therapeutic intent (e.g. druggable targets) (1)(2)(3). However, the majority of these alterations detected in tumor genomes are "passenger" mutations, obscuring those mutations that are actionable in a background of errors, noise, and random alterations (4 -7). The recent advances in structural genomic technologies, like x-ray and nuclear magnetic resonance, have helped investigators generate large-scale three-dimensional (3D) protein structure data. As of February 2015, more than 100,000 protein 3D structures (including ϳ27,000 human 3D protein structures) have been curated in the Protein Data Bank (PDB) database (8). Protein 3D structure and biological function are closely related, especially in the case of protein-ligand binding-site residues, which are local regions that perform critical functions in cells such as binding with small molecules including drugs (9 -12). Thus, development of new computational approaches that efficiently prioritize driver mutations according to their functional impact on protein 3D structures may provide an opportunity for the identification of molecules that can be the focus of targeted therapies.
Recently, several computational approaches have been developed for the prediction of putative cancer driver genes (13). These methods are classified into four main classes. The first class utilizes mutation frequency information and other genomic features (e.g. DNA replication time) to detect driver genes under the assumption that these genes have a higher mutation rate in comparison to the expected background mutation model, such as MutSigCV (14) and DrGap (15). The second class utilizes evolutionary conservation of amino acids to predict functional impact of nonsynonymous mutations, such as SIFT (16) and MutationAssessor (17). The third class of approaches examines the distribution of nonsynonymous mutations in specific protein regions or protein 3D structures to estimate their functional impacts, such as Oncodrive-CLUST (18), e-Driver (19), and mutation set enrichment analysis (MSEA) (20). MSEA was implemented by two methods, namely MSEA-clust and MSEA-domain, to predict cancer driver genes using mutation hotspot patterns in coding sequences or protein domain sequences (20). The application of MSEA to somatic mutations in seven cancer types in TCGA found a total of 82 putative cancer genes bearing significant mutation hotspots. The last class is based on the network perturbation hypothesis in cancer, that is, genetic aberrations may cause network architectural changes by affecting or removing a node or its connection within the network, or changing the biochemical properties of a node or protein (21). Representative methods include VarWalker (22) and HotNet2 (23). Although many methods and tools have been developed, so far, how to efficiently integrate somatic mutation data with structural genomics information to identify potentially druggable cancer driver gene products or actionable mutations for precision medicine remains a monumental challenge (24).
In this study, we proposed SGDriver, a novel, structural genomics-based method that incorporates the somatic missense mutations into the protein-ligand binding-site residues to decipher the functional role of somatic mutations and prioritize putative druggable molecular targets using a Bayes inference statistical framework (Fig. 1). We found significantly higher mutation rates at ligand binding-site residues in comparison with the whole protein sequences in all the 16 cancer types that we examined. Additionally, mutations at proteinligand binding-site residues are more likely to be deleterious when compared with those at nonbinding-site residues. We applied SGDriver to over 740,000 missense mutations in ϳ5000 tumor-normal pairs in 16 cancer types in TCGA. We found 298 significantly mutated proteins (SMPs) that have enriched missense mutations at their protein-ligand bindingsite residues (adjusted p value (q) Ͻ 0.05). We further performed a drug-target interaction network analysis by systematically integrating large-scale drug pharmacological data. This analysis found 98 known anticancer targets as well as 126 new mutated proteins that may be repurposed as potential druggable anticancer targets. Our integrative analysis suggested that 13% of cancer patients might benefit from currently approved targeted therapies, with this proportion potentially increasing to 31% by considering drug repositioning. In summary, SGDriver provides a useful tool to identify potential druggable SMPs that can be further studied for precision cancer medicine.  (v69), which is the largest database for somatic mutations. To reduce redundancy and ensure the quality of somatic mutation data, in this study, we only focused on the somatic mutations in TCGA tumor-normal matched samples from aforementioned three datasets. We used ANNOVAR (26) to map these somatic mutations into the protein sequences for identifying the corresponding amino acid changes based on RefSeq ID. The functional impact of nonsynonymous SNVs (single nucleotide variants) was assessed using SIFT (16) and PolyPhen 2 scores (27). For this analysis, we obtained SIFT and PolyPhen 2 scores from ANNOVAR annotation database. Then, we converted RefSeq ID to UniProt ID with the aid of UniProt ID mapping tool (http://www.uniprot.org/uploadlists/).

Collection and Preparation of Somatic
Gene Expression Data-Normalized gene expression data of RNA-Seq (V2) was extracted from TCGA using R package (TCGA-Assembler) (28). All public data files on TCGA Data Coordinating Center (DCC) server were gathered on January 5, 2015.
Collection and Preparation of Protein Structural Genomics Data-We extracted protein-ligand binding site data from BioLiP, which is a semi-manually curated database for high-quality, biologically relevant protein-ligand binding interactions (version October 10, 2014, the most updated version) (29). These annotated ligand molecules include DNA or RNA ligands, peptide ligands, metal ligands and regular ligands (i.e. the common small-molecule ligands). There were a total of 264,692 ligand-protein binding interactions, including 148,054 interactions for regular ligands, 31,797 interactions for DNA or RNA ligands, 72,088 interactions for peptide ligands, and 12,753 interactions for metal ligands. For each UniProt protein, we combined the protein-ligand binding-site residues of all the corresponding PDB structures. In total, there were 17,595 UniProt proteins having protein-ligand binding site information, including 2735 unique human proteins with the number of ligand binding-site residues ranging from 2 to 545 and median number being 19.
Description of SGDriver-SGDriver was developed under the hypothesis that somatic mutations that alter protein-ligand binding-site residues are more likely to induce the initiation and progression of cancer and mediate anticancer drug responses. SGDriver assumes that the observed number of somatic mutations across samples, y, at the ligand-binding-site residues of a protein, follows a Zero Inflated Poisson distribution (ZIP): This means that if a SNV is present, y follows a Poisson distribution with background mutation rate parameter Ͼ 0.
Given one protein sequence having n ligand-binding-site residues, Y ϭ y 1 , . . ., y n is the vector of observed number of mutations across samples per residue. For each ligand-binding-site residue, the number of mutations is counted at the direct protein-ligand binding-site residue and two flanking residues immediately upstream or downstream of the direct binding residue. The likelihood of observation is obtained as f͑ y 1 ,. . ., y n ͉͒ ϭ The posterior distribution is computed for each UniProt protein. The minimal p value of all ligand-binding residues is deemed as the significance for each protein. The previous study revealed that gene expression level affects somatic mutation rates in tumor genomes (14). In this study, we used the specific background mutation rates corrected by gene expression level to predict putative SMPs in each type of cancer. Specifically, we first computed average gene expression for each gene, and assigned these values to the corresponding gene products. For each UniProt protein of interest, we identified the top 100 proteins with the most similar expression levels; their mutational profiles were merged and used to compute the posterior distributions and p values. Finally, p values were adjusted for multiple testing by the Benjamini-Hochberg method (30). The adjusted p value (q) Ͻ 0.05 was used as the cutoff to choose the SMPs that harbor significantly enriched somatic mutations at the protein-ligand binding residues.
Collection of Cancer-associated Genes-We collected a large number of cancer genes from several publicly available sources. First, a total of 547 genes were downloaded from Cancer Gene Census (31) (accessed on November 5, 2014, denoted as CGC genes). CGC genes are well-curated and have been widely used as a reference cancer gene set in many cancer-related projects (32). Second, we collected 693 significantly mutated genes (SMGs) in cancer from TCGA publications and several other publications (details are shown in supplemental Table S3). Finally, we collected 478 oncogenes and 937 tumor suppressor genes as described in our previous studies (21,33).
Construction of Drug-target Interaction Network-We constructed a drug-target interaction (DTI) network by a systematic integration of drug pharmacological data from three databases: DrugBank (v3.0, http://www.DrugBank.ca/) (34), ChEMBL (v18) (35), and BindingDB (36). First, we collected the information of ϳ7700 drugs, including 1600 FDA-approved small molecule drugs, 160 FDA-approved biotech drugs, and over 6000 experimental drugs. In order to improve the quality of DTI data, we only used the DTI pairs with pharmacological profiles (IC 50 , EC 50 or K i ) less than 10 M based on a previous study (37). All drugs were annotated using the detailed therapeutic information based on the Anatomical Therapeutic Chemical (ATC) classification system codes from the DrugBank database as described in previous studies (38,39). All genes encoding drug targets were mapped to their Entrez IDs based on the National Center for Biotechnology Information (NCBI) database (40) as well as their official gene symbols based on HUGO Gene Nomenclature Committee (http://www.genenames.org).
Molecular Docking-We performed molecular docking simulation to repurpose potential drug candidates that are more likely to target mutant protein than wild-type (WT) protein using estrogen receptor (ER) as a case study because ER has both WT and driver mutant protein 3D structures (41,42). We downloaded PDB files of wild-type ER (PDB ID: 1R5K) and p.Tyr537Ser mutant ER (PDB ID: 2QA8) from a previous study (42). We used MGLtools software (v1.5.6) (43) for receptor structure preparation by adding hydrogen and computing Gasteiger charges. The residues within 5Å of ligands were chosen as docking pocket. We collected 40 known ER small-molecule ligands (e.g. ER antagonists and modulators) in SMILES format from the DrugBank database (34). We transferred SMILES to 3D structures for each ligand using SMILES translator and structure file generator from National Cancer Institute (http://cactus.nci.nih.gov/translate/). We performed ligand preparation and optimization using UCSF Chimera software (v 1.10.2) (44) by adding hydrogen and Amber ff14SB charges for standard residues and AM1-BCC for other residues. Structure energy minimum was calculated using the steepest method, with a minimum of 100 steps. Finally, we performed molecular docking experiments by using Autodock_vina software (v.1.1.2) (45) and retained top three docking poses that had the lowest free energy for further analysis.

RESULTS
Overview of SGDriver-We examined the somatic mutation profiles in 4,997 tumor-normal pairs in 16 major cancer types using TCGA data ( Fig. 1A and supplemental Table S1). These 16 major cancer types consist of acute myeloid leukemia (LAML), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), colon and rectal adenocarcinoma (COAD/ READ), glioblastoma multiforme (GBM), head and neck squamous cell carcinoma (HNSC), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), ovarian serous cystadenocarcinoma (OV), prostate adenocarcinoma (PRAD), skin cutaneous melanoma (SKCM), stomach adenocarcinoma (STAD), thyroid carcinoma (THCA), and uterine corpus endometrial carcinoma (UCEC). In this study, we used only missense mutations in our analyses because nonsense mutations typically have different mechanisms, like protein outcome (truncation or termination of translation). The mutational load (the mean number of missense mutations per sample) for each cancer type is shown in Fig. 1B. The missense mutation rate of individual patient ranges from 8 to 497 with an average of 160 across the 16 cancer types. Among the 16 cancer types, LAML had the lowest mean mutation rate (8 missense mutations per sample), whereas UCEC had the highest mutation rate (497 per sample). We then developed SGDriver, a protein structural genomics-based approach to prioritize putative SMPs harboring missense mutations that are significantly enriched at protein-ligand binding-site residues using a Bayes inference statistical framework (see Methods). SGDriver postulates that if a gene product (protein) contains somatic mutations that are significantly enriched in protein-ligand binding-site residues, this protein would be more likely cancer-associated or related to anticancer drug responses (10, 12) (Fig. 1C). By applying SGDriver to 746,631 missense mutations in 4997 cancer genomes in 16 cancer types from TCGA ( Fig. 1A and 1B), and using a statistical significance cutoff value (q) 0.05, we identified 298 putative SMPs (q Ͻ 0.05) in a pan-cancer analysis and a total of 335 putative SMPs in individual cancer analyses ( Fig. 1D and 1E). Finally, we used the drug pharmacological data to explore the potential druggable SMPs or actionable mutations in order to assist anticancer drug development in precision cancer medicine (Fig. 1F).
Pan-cancer Analysis of Protein-Ligand Binding-site Residue Mutations-We mapped missense mutations in 4997 tumor-normal pairs across 16 cancer types from TCGA into 2735 human protein-ligand complexes with unique UniProt ID. The final list was comprised of 7967 unique missense mutations at the protein-ligand binding-site residues covering 1119 unique human proteins (supplemental Table S5). After including the somatic mutations at the two immediate flanking residues of each direct binding residue, we obtained 14,471 mutations on 2091 proteins (including 1516 recurrently mutated proteins).
We first studied the 1119 proteins having mutations directly at the ligand-binding residues. We examined all mutations in the whole protein sequences and the mutations at the direct protein-ligand binding residues, respectively, in 16 cancer types ( Fig. 2A). We found a significantly higher mutation rate at the direct ligand binding residues than that of the whole protein sequence; this is consistent among all the 16 cancer types (p values ranged from 1.2 ϫ 10 Ϫ14 to 7.8 ϫ 10 Ϫ283 , Wilcoxon rank-sum test, supplemental Table S2). For the instance of BRCA, there were 358 somatic missense mutations scattered at 60,946 protein-ligand binding residues; this compared with a total of 3256 mutations on 1,079,303 amino acids at the whole protein level. The mutation rate at the binding residues is nearly twice of that at the whole protein level (p ϭ 8.5 ϫ 10 Ϫ20 ). These observations indicate that protein-ligand binding residues are significantly altered by somatic missense mutations. We further performed the comparison of mutation patterns between nonsmall molecule ligand binding-site residues and regular small-molecule ligand binding-site residues. As shown in supplemental Fig. S1, we did not observe a significant difference in mutation pattern between regular small-molecule ligands and nonsmall molecule ligands.
Ligand-binding-site Residue Mutations Tend to be Deleterious-We hypothesized that somatic missense mutations at protein-ligand binding-site residues have more deleterious impact than other missense mutations. We tested this hypothesis by measuring the functional impact scores of mutations using two popular and complementary tools: SIFT and Poly-Phen2. SIFT (16) score is based on the degree of conservation of amino acid residues in sequence alignments derived from closely related sequences to predict whether an amino acid substitution affects protein function, whereas PolyPhen2 (27) score is based on the integration of eight sequence-based and three structure-based predictive features. We examined the cumulative distribution of SIFT and PolyPhen2 scores for protein-ligand binding-site residue mutations (direct position and two flanking residues) and nonbinding-site residue mutations (amino acid positions excluding the direct binding residues or two flanking residues). The distribution is shown in Fig. 2B and 2C. We defined the mutations with SIFT scores Ͻ 0.95 or PolyPhen2 score Ͼ 0.909 as deleterious mutations based on previous studies (12). We found that protein-ligand binding-site residue mutations are more likely to be deleterious mutations in comparison to nonbinding-site residue mutations when evaluated using both SIFT (p Ͻ 2. Prioritizing Significantly Mutated Proteins in Cancer Using SGDriver-Based on the observation that somatic mutations are enriched at protein-ligand binding-site residues ( Fig. 2A) and that mutations at protein-ligand binding-site residues are more likely to be deleterious than those at other residues ( Fig.  2B and 2C), we used SGDriver to prioritize putative SMPs harboring statistically significant excess number of mutations at protein-ligand binding-site residues among 16 cancer types. As shown in Fig. 3, we identified 298 SMPs harboring mutation-enriched protein-ligand binding-site residues with q Ͻ 0.05 in pan-cancer analysis (supplemental Table S4). All top five SMPs are well-known cancer proteins: BRAF (q Ͻ 10 Ϫ200 ), KRAS (q Ͻ 10 Ϫ200 ), EGFR (q ϭ 4.9 ϫ 10 Ϫ174 ), IDH1 (q ϭ 1.1 ϫ 10 Ϫ147 ), and TP53 (q ϭ 3.1 ϫ 10 Ϫ121 ) (32). We then evaluated our method using the benchmark data. Specifically, we examined the enrichment of these 298 SMPs using two well-curated cancer gene data sets: (1) CGC genes (547 ex-perimentally validated cancer genes), and (2) a curated list of published cancer SMGs (693 genes) from TCGA genome projects and other published literatures (supplemental Table  S3). We found that the genes harboring a significant number of protein-ligand binding-site residue mutations were significantly enriched in CGC (p ϭ 1.6 ϫ 10 Ϫ12 , hypergeometric test) and SMGs (p ϭ 1.2 ϫ 10 Ϫ14 ), respectively.
Identifying Novel Druggable Mutations-One main aim of cancer genomic studies is to identify novel anticancer targets or druggable mutations for targeted therapies and, furthermore, improve patient treatment and clinical care in a fashion of precision cancer medicine. To explore the potential application of SGDriver in precision cancer medicine, we systematically searched the drugs' pharmacological spaces by integrating the predicted SMPs in both pan-cancer (Fig. 3) and individual cancer analysis (Fig. 4). As shown in Fig. 5A, 98 SMPs (such as BRAF, ERBB2, AKT1, and DNMT1) can be targeted by 72 U.S. FDA approved anticancer drugs with binding affinity less than 10 M annotated in DrugBank, Bind-ingDB, and ChEMBL databases. For example, molecules targeting on BRAF Val600Glu mutation (Fig. 5B) have been approved by U.S. FDA in melanoma targeted therapy, and specific drugs include vemurafenib and dabrafenib. In addition, targeting EGFR Leu858Arg mutation is being clinically investigated for lung cancer therapy (57).
Furthermore, we used drug repositioning strategy (38) to expand the scope of drugs targeting SMPs. We found that 126 SMPs were targeted by 596 repurposed drugs (noncancer drugs approved for other disease treatment). Among the 126 SMPs, 36 (e.g. RAC1 and TLR4) can only be targeted by noncancer drugs. To elucidate the potential application of SGDriver for prioritizing new druggable mutations, we briefly highlighted below three proteins (SPOP, ESR1, and NR3C1) harboring the enriched druggable mutations in protein-ligand binding-site residues across different cancer types. SPOP, encoding the cullin-RING ubiquitin ligase adaptor protein speckle-type POZ protein, is mutated in 8 -14% of prostate and endometrial cancers (58). As shown in Fig. 3 and 4, we found that SPOP is significantly mutated in pan-cancer (q ϭ 8.7 ϫ 10 Ϫ12 ) and individual cancer (prostate cancer (q ϭ 1.1 ϫ 10 Ϫ4 ) and uterine cancer (q ϭ 7.4 ϫ 10 Ϫ5 )). Moreover, one driver mutation, p.Phe133Leu (58), is significantly enriched at its ligand binding-site residue (Fig. 5C). This provides potential novel strategy for targeted prostate and uterine cancer therapy.
ESR1, encoding ER, is a member of the nuclear receptor family. 70% of breast cancer samples express ER, and most of these are sensitive to ER inhibitors (41). In this study, we found that ER is significantly mutated in uterine cancer (q ϭ 1.4 ϫ 10 Ϫ59 ). Two mutations: p.Tyr537Ser and p.Asp538Gly, are significantly enriched at its ligand binding-site residues (Fig. 5C). Toy et al. recently reported three highly recurrent mutations (p.Tyr537Ser, p.Tyr537Asn, and p.Asp538Gly) in metastatic breast cancers (41). To further repurpose potential small-molecule drugs that are more likely to target mutant ER than wild-type ER, we docked 40 known ER small-molecule drugs (supplemental Table S6) collected from the DrugBank database into wild-type ER and p.Tyr537Ser mutant ER protein crystal structures, respectively (see Methods). We found that some ligands (e.g. genistein and trilostane) had a higher docking binding affinity on p.Tyr537Ser mutant ER structure compared with wild-type ER structure, whereas other ligands (e.g. tamoxifen, ERA-923, and fluvestrant) showed a lower docking binding affinity on p.Tyr537Ser mutant ER structure compared with wild-type ER structure (supplemental Table  S6). A recent study reported that the resistance of tamoxifen   Table  S3). The horizontal red line denotes the false discovery rate at 0.05. is associated with ESR1 somatic mutations in breast cancer derived xenografts (59). In addition, Robinson et al. found that activation of ESR1 mutations is involved in hormone-resistant (such as tamoxifen and fluvestrant) metastatic breast cancer (60). Thus, tamoxifen, ERA-923, and fluvestrant might have a high resistance risk in ESR1 mutant (e.g. p.Tyr537Ser) breast cancer because of the decreasing binding affinity of the ligand-protein complexes (supplemental Table S6). On the contrary, genistein and trilostane that have the higher predicted binding affinity with p.Tyr537Ser mutant ER protein might be potential candidates to overcome anti-endocrine therapeutic resistance in breast and uterine cancer. Although this computational result has good support in literature, future experimental validation is required. In addition, future work in virtual screening and docking of mutated ER by comparison to biological property-matched decoy molecules is necessary for better interpreting their significance and identifying potential old drugs or new molecules for targeted breast cancer treatment.
NR3C1, encoding glucocorticoid receptor (GR), regulates genes that control the development, metabolism, and immune response (61). Here, we found that GR was significantly mutated in cancers (q ϭ 0.03). The mutation p.Met604Ile is significantly enriched in ligand binding-site residues (Fig. 5C). GR agonists or antagonists had been approved for the treatment of allergic, inflammatory nasal conditions, and hypercortisolism, such as dexamethasone (agonist) and mifepristone (antagonist) (34). Skor et al. showed that GR antagonisms provided a novel therapy for triple-negative breast cancer (62). GR agonist dexamethasone inhibits EGF-induced cell migration by modulating transcriptional activities in MCF10A mammary epithelial cells (63). Collectively, U.S. FDA approved GR agents may provide new indications by targeting ligand binding-site residue mutations such as p.Met604Ile.
We next found that 152 SMPs were targeted by 790 clinical or preclinical agents. For example, KRAS Gly12Asp/Ser/Val mutations (Fig. 5B) have been widely investigated for molecular targeting in nonsmall cell lung cancer (64), including a covalent KRAS inhibitor SML-8 -73-1 in lung cancer (65). Fig.  6 shows a bipartite drug-target interaction network connecting 180 SMPs (squares) and 673 U.S. FDA approved drugs (circles). The original Cytoscape file for preparing Fig. 6 is provided in supplementary file 1. All drugs are grouped by ATC classification system codes. As shown in Fig. 6, some FDA approved agents (i.e. sorafenib, afatinib, pazopanib, and nilotinib) can target multiple SMGs. Notably, some new SMGs, such as ALB, ADRB2, CHRM2, and NR3C1, may be potential candidates for drug repurposing in targeted cancer therapy. Put together, Fig. 6 offers a rich resource for identi- -log (q) 10 -log (q) 10 -log (q) fying new repurposing approved drugs to target cancer SMPs for future cancer therapeutic development.
How Many Patients May Benefit From the Approved Therapeutic Agents?-We next investigated the tumor sample coverage for U.S. FDA approved drugs by targeting SMPs detected in aforementioned pan-cancer and individual analyses. Among 16 cancer types, there were 13 cancer types whose patients could benefit from the approved therapeutic agents by targeting SMPs; the three cancer types that did not benefit are: KICH, LAML and PRAD (Fig. 7). In total, of the 4997 TCGA tumor samples used in this study, 659 (13%) samples might be treated with 56 FDA approved anticancer drugs. If we expanded the search space to include repurposed drugs from all 596 FDA approved drugs, this coverage would increase to 31%. Specifically, there were 1556 samples harboring ligand binding-site residue mutations in SMPs that might be targeted by 470 U.S. FDA approved drugs. When focusing on individual cancer type, we found up to 60% of SKCM patients might benefit from 214 approved therapeutic agents (29 anticancer drugs and 185 repurposing drugs). This high proportion is largely attributed to BRAF Val600Glu mutation that is a well-studied target for targeted cancer therapy (66), occurring in 89 samples. The similar feature appeared in THCA where BRAF Val600 was altered in 189 samples and 56% of patients could benefit from 9 approved therapeutic agents. However, in our analysis, few available FDA approved targeted agents could be used for KIRC and BLCA targeted therapeutics (Fig. 7B). Drug repositioning may provide great opportunities for such kind of cancer treatments (e.g. KIRC and BLCA). For example, RARG, encoding retinoic acid receptor gamma, was significantly mutated in BLCA (q ϭ 3.0 ϫ 10 Ϫ7 ). Several dermatological drugs such as acitretin (34), tretinon (67), alitretinoin (68), and adapalene (69) may have new indications for BLCA treatment by targeting RARG (Fig.  6). Notably, alitretinoin showed preclinical and clinical antitumor activities in bladder (70), breast (71), and other cancer treatments (68). Although the results above are promising, they are still exploratory and future wet-lab experimental validation is needed. DISCUSSION Massive amounts of cancer genomics data generated from next-generation sequencing have motivated investigators to develop novel computational approach for prioritizing new druggable genes or mutations for the development of precision cancer medicine. Although the results we reported here are exploratory, the prioritized list can help cancer biologists and physician scientists to find the top candidates for experimental and clinical validation. In this study, we proposed SGDriver, a structural genomics-based approach under the Bayes inference statistical framework, to identify significantly mutated proteins and druggable mutations in cancer. To ver-ify the hypothesis of SGDriver, we examined the mutational load distribution of missense mutations at protein-ligand binding-site residues and whole protein level. We found that a higher mutation rate at protein-ligand binding-site residues than that of whole protein level in all the 16 cancer types we examined. Furthermore, we found that missense mutations located at protein-ligand binding-site residues were more likely to be deleterious when compared with other residues. In this network, nodes include 180 targets (SMPs, squares) and 668 U.S. FDA approved drugs (circles), and edges denote the interactions with pharmacological profiles (IC 50 , EC 50 or K i ) less than 10 M. All drugs were grouped using the Anatomical Therapeutic Chemical (ATC) classification system codes. The original Cytoscape file is provided in supplementary file.

Drug target node
Here, "deleterious" mutations are based on prediction that might be disruptive to its protein structure or function, rather than to the organism. SGDriver is not designed for detecting the effect of gain-of-function of somatic mutations either. To expand our study, we performed gene set enrichment analysis of SMPs identified by SGDriver in two important types of cancer genes: tumor suppressor genes, whose somatic mutations tend to be loss-of-function, and oncogenes, whose somatic mutations tend to be gain-of-function in cancer. Using the tumor suppressor genes and oncogenes collected in our recent studies (21,33), we found that the identified SMPs in individual cancer type and pan-cancer analysis were significantly enriched in both tumor suppressor genes (p ϭ 3.1 ϫ 10 Ϫ33 , Fisher's exact test) and oncogenes (p ϭ 1.5 ϫ 10 Ϫ26 ) (supplemental Fig. S3). These results further indicated that the SMPs identified by SGDriver could be either oncoproteins or tumor suppressors, which is consistent with a recent study (72).
There are some limitations in current study. First, this study focused on missense mutations in protein-ligand binding-site residues and excluded other types of mutations, such as nonsense mutations, insertions/deletions (indels), synonymous mutations, copy number variation, or gene fusion. All these mutations may play important roles in tumor initiation and development. For example, some approved drugs target genetic events such as indels, copy number variant and gene fusion further reflects this limitation (29). Second, mutational heterogeneity of tumor samples may lead to false positive findings in prioritizing SMGs. In this study, we attempted to incorporate gene expression information to weaken the influence of mutational heterogeneity based on previous study (14). However, this practice may cause other kind of unknown biases in our analysis. Third, we did not distinguish different functional roles for driver mutations, such as loss-of-function and gain-of-function. For traditional cancer drug discovery and development, agonists are often designed to target lossof-function drivers (e.g. p53 activators) (73), whereas inhibitors or antagonists often designed to target oncogenes (e.g. EGFR and BRAF inhibitors) (74). Fourth, we limited our analyses only to the significantly mutated protein-ligand bindingsite residues, which are directly targeted by approved or repurposed therapeutic agents. However, targeting the interactive partner proteins of the significantly mutated proteins, even these partner proteins are wild type, may be also an important strategy for anticancer drug design and development (24). Finally, structural genomics-based approaches are limited in the number of proteins with high-resolution 3D structures available. Until now, only ϳ15% proteins have known 3D structures in the whole human genome. In this study, only 11,119 proteins with high-resolution 3D structures were investigated in pan-cancer and individual studies. Finally, the prioritized SMPs are based on large-scale analyses. Their clinical implications need experimental validation, especially pharmacological work, although such work is usually labor-extensive and time-consuming.
Although there exist some limitations in current study, SGDriver provides useful features and an alternative tool for speeding up the discovery of potential druggable targets for targeted cancer therapy. Furthermore, SGDriver employs protein-ligand binding-site information for identifying SMGs and druggable mutations. Thus, SMGs and mutations identified by SGDriver should be more druggable than traditional computational approaches. In total, we identified 14,471 druggable somatic mutations, including well-known driver mutations and new mutations. These mutations serve a rich resource for future experimental or computational validation.
We further employed drug repositioning strategies to identify more druggable targets or mutations for cancer targeted therapeutics (38). For instances, integrated analysis indicated that the percentage of patients benefitting from current targeted therapy would increase from 13% to 31% by drug repositioning strategy. Considering the rapid accumulation of cancer sequencing data and protein structural genomics data, we plan to improve SGDriver in the near future in the following directions: (1) to predict SMGs with higher quality through the integration of more cancer mutations (including missense mutations, small indels, and nonsense mutations) and structural genomics data; (2) to reduce the false positive rates of SGDriver using the corrected background mutation model based on gene expression and DNA replication time information (14); (3) to characterize driver mutations into lossof-function which may be targeted by agonists or activators and gain-of-function targeted by antagonists or inhibitors; (4) to utilize molecular network and gene expression data to integrate the interacting partners of significantly mutated proteins to expand the candidate pool for drug repurposing study; (5) to identify druggable mutations related to tumor cell colony formation using improved SGDriver to reduce the anticancer drug resistance (75); and (6) to identify old drugs or new small molecules to target specific mutations or mutated proteins for precision cancer therapies by performing molecularly docking-based virtual screening on the significantly mutated proteins identified by SGDriver. In summary, this study provides a novel computational framework by uniquely incorporating large-scale cancer genomics profiles (ϳ5000 tumornormal matched pairs) into protein structure genomics data and identified thousands of druggable mutations for therapeutic development in cancer. * This work was partially supported by National Institutes of Health grants (R01LM011177, P50CA095103, P50CA098131, and P30CA068485), The Robert J. Kleberg, Jr. and Helen C. Kleberg Foundation (to Z.Z.), a Vanderbilt-Ingram Cancer Center's Breast SPORE pilot project (to Z.Z.), and Ingram Professorship Funds (to Z.Z.). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. § § These authors contributed equally to this work. Competing interests: The authors declare that they have no competing interests.