Prioritization of driver mutations in pancreatic cancer using cancer-specific high-throughput annotation of somatic mutations (CHASM)

Abstract

Over 20,000 genes were recently sequenced in a series of 24 pancreatic cancers. We applied CHASM (Cancer-specific High-throughput Annotation of Somatic Mutations) to 963 of the missense somatic missense mutations discovered in these 24 cancers. CHASM identified putative driver mutations (false discovery rate ≤0.3) in three known pancreatic cancer driver genes (P53, SMAD4, CDKN2A). An additional 15 genes with putative driver mutations include genes coding for kinases (PIK3CG, DGKA, STK33, TTK and PRKCG), for cell cycle related proteins (NEK8), and for proteins involved in cell adhesion (CMAS, PCDHB2). These and other mutations identified by CHA SM point to potential “driver genes” in pancreatic cancer that should be prioritized for additional follow-up.

Introduction

Pancreatic cancer is the fourth leading cause of cancer death in the United States. It is estimated that in 2009 close to 42,470 Americans were diagnosed with pancreatic cancer, and that 35,240 died from the disease.¹ While the existing therapies for pancreatic cancer are mostly ineffective, the success of gene-specific targeted therapies for lung and colon cancers suggests that a better understanding of the genetic alterations in pancreatic cancer may improve the dismal survival associated with this disease. Indeed, the sequencing of >20,000 coding genes in a series of 24 pancreatic cancers has led to the discovery of a familial pancreatic cancer gene, PalB2, to a better understanding of the DNA alterations associated with cigarette smoking, and to the identification of SMAD4 gene alterations as prognostic markers following surgical resection.²

Genetic alterations that produce bi-allelic inactivation of gene function in a cancer are easily prioritized as likely significant in the development of that neoplasm. Most of the mutations in a cancer, however, are present in only one of the two alleles of the affected gene and/or are not obviously inactivating, and these mutations are much more difficult to classify. The challenge is to determine which of these alterations are “driver” mutations, contributing significantly to tumorigenesis, and which are “passenger” mutations, neutral to cancer fitness.³

We recently developed a high-throughput computational method, called CHASM (Cancer-specific High-throughput Annotation of Somatic Mutations), to prioritize somatic missense mutations found in large-scale tumor sequencing studies. CHASM has been shown to have high sensitivity and specificity for identifying driver mutations, and outperformed other predictors of mutation function.⁴ When applied to a set of 607 somatic missense mutations found in a sequencing study of glioblastoma multiforme tumors, CHASM was able to correctly identify the isocitrate dehydrogenase (IDH1) R132 mutation recently discovered to act as a driver in this cancer type.⁵

Here we applied CHASM to 963 of the missense mutations identified in the sequencing of >20,000 coding genes in 24 pancreatic cancers and found that CHASM prioritized missense mutations in most of the known driver genes in pancreatic cancer (TP53, p16/CDKN2A and SMAD4). (Although CHASM is able to correctly identify known KRAS driver mutations, we have omitted them from this list, since they were included in our training data). In addition, many of the mutations prioritized by CHASM (false discovery rate ≤0.3) are in genes that code for proteins with potentially interesting functions, including tyrosine and lipid kinases, cell cycle regulation and cell-adhesion molecules.

Results

Fifty-six of the 963 somatic missense mutations analyzed had an FDR of ≤0.30 (Table 1, Suppl. Table 3). Of these 56, 41 mutations occur in known pancreatic driver genes (KRAS, TP53, CDKN2A, SMAD4). Thirty-four of these mutations have previously been identified as drivers and were included in the classifier's training set. Thus, it is likely that our scores inflate the statistical significance of these mutations. Only two of the remaining 15 mutations assigned FDR ≤0.30 were in genes that were mutated more than once in the Discovery set of 24 pancreatic cancers from the original sequencing study. Since recurrent mutation was a criteria for classifying a gene as a significant candidate cancer gene in the original study, 13 of these mutations would not have been prioritized. These 13 mutations occur in 13 genes that were not known driver genes nor were prioritized as likely driver genes in the original sequencing study. That is to say that we wouldn't have picked these genes out as significant without CHASM.

Table 1.

Fifty-six somatic missense mutations with CHA SM FDR ≤ 0.30

Hugo	Transcript	Mut	Zygosity	CHASM score	p value	q value<=	In training set?
TP53	CCDS11118.1	Y234C	Homo	0.034	0.0004	0.05	Yes
CDKN2A	CCDS6510.1	H98P	Homo	0.052	0.0004	0.05	Position Only
TP53	CCDS11118.1	I255N	Homo	0.062	0.0004	0.05	Yes
TP53	CCDS11118.1	S241F	Homo	0.066	0.0004	0.05	Yes
CDKN2A	CCDS6510.1	L63V	Het	0.068	0.0004	0.05	Yes
TP53	CCDS11118.1	L257P	Homo	0.072	0.0004	0.05	No
TP53	CCDS11118.1	C275Y	Homo	0.078	0.0004	0.05	Yes
TP53	CCDS11118.1	G266V	Homo	0.078	0.0004	0.05	Yes
TP53	CCDS11118.1	R248W	Homo	0.134	0.0004	0.05	Yes
NEK8	NP_835464	A197P	Het	0.144	0.0004	0.05
PIK3CG	CCDS5739.1	R839C	Homo	0.166	0.0008	0.05
TP53	CCDS11118.1	H179R	Homo	0.180	0.0013	0.05	Yes
SMAD4	CCDS11950.1	C363R	Homo	0.184	0.0013	0.10	Position Only
TP53	CCDS11118.1	R282W	Homo	0.198	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Homo	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Homo	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
KRAS	CCDS8702.1	G12D	Het	0.202	0.0013	0.10	Yes
TP53	CCDS11118.1	R175H	Homo	0.212	0.0017	0.10	Yes
CALB1	CCDS6251.1	D199G	Het	0.220	0.0021	0.10
KRAS	CCDS8702.1	G12R	Het	0.224	0.0021	0.10	Yes
KRAS	CCDS8702.1	G12R	Het	0.224	0.0021	0.10	Yes
KRAS	CCDS8702.1	G12V	Het	0.228	0.0025	0.10	Yes
KRAS	CCDS8702.1	G12V	Homo	0.228	0.0025	0.10	Yes
KRAS	CCDS8702.1	G12V	Het	0.228	0.0025	0.10	Yes
KRAS	CCDS8702.1	G12V	Het	0.228	0.0025	0.10	Yes
KRAS	CCDS8702.1	G12V	Het	0.228	0.0025	0.10	Yes
KRAS	CCDS8702.1	G12V	Het	0.228	0.0025	0.10	Yes
TP53	CCDS11118.1	V217G	Homo	0.232	0.0029	0.10	No
STK33	CCDS7789.1	F323L	Het	0.246	0.0029	0.15
MAPT	CCDS11499.1	G333V	Het	0.268	0.0042	0.15
TP53	CCDS11118.1	L344P	Homo	0.278	0.0046	0.20	No
DGKA	CCDS8896.1	V379I	Het	0.278	0.0046	0.20
TP53	CCDS11118.1	K120R	Homo	0.290	0.0051	0.20	Position Only
TP53	CCDS11118.1	D208V	Homo	0.300	0.0051	0.20	No
FLJ25006	CCDS11237.1	S196L	Het	0.302	0.0051	0.20
CMAS	CCDS8696.1	I275R	Het	0.304	0.0051	0.20
COCH	CCDS9640.1	S400L	Het	0.320	0.0067	0.20
STN2	CCDS9875.1	I590S	Het	0.332	0.0076	0.20
PRKCG	CCDS12867.1	P524R	Het	0.344	0.0088	0.25
PCDHB2	CCDS4244.1	A323V	Het	0.348	0.0093	0.25
TTK	CCDS4993.1	D697Y	Het	0.366	0.0105	0.25
PRDM5	CCDS3716.1	V85I	Het	0.368	0.0114	0.30
UNC13C	ENST00000260323	G1126R	Homo	0.370	0.0114	0.30
TP53	CDS11118.1	T155P	Homo	0.380	0.0122	0.30	Yes

Fifteen mutations that occur in genes not previously associated with pancreatic cancer are highlighted. Transcript identifiers from RefSeq, CCDS and E nsembl, zygosity status, CHA SM score, raw p value and q value (multiple testing correction) are shown. Some mutations were included in the CHA SM training set, thus their statistical signficance is likely to be inflated. We have indicated these mutations, as well as mutations whose position (but not amino acid substitution) was in the training set.

Five of the 15 mutations assigned FDR ≤0.30, which weren't in known driver genes, were in genes coding for kinases (PIK3CG, DGKA, STK33, TTK and PRKCG), one was in NEK8, a gene coding for a cell cycle related protein,¹² and two were in genes coding for cell adhesion proteins (CMAS and PCDHB2).

Next, we summarize the relevance of these mutations.

Kinases.

PIK3CG (R839C). This lipid kinase contains the second highest scoring predicted driver mutation among the set of genes not previously identified as drivers in pancreatic cancer. PIK3CG is a member of the phosphatidylinositol 3-kinases (PI3Ks) family and plays a role in mast cell degranulation,¹³ lymphocyte activation¹⁴ and leucocyte chemotaxis.^15,16 Unlike the closely related human oncogene PIK3CA (p110alpha), there is evidence suggesting that PIK3CG functions as a tumor suppressor. Loss of PIK3CG in mice leads to spontaneous, malignant epithelial tumours in the colorectum and PIK3CG can block the growth of human colon cancer cells.¹⁷ PIK3CG is located in a region on chromosome band 7q22 that is frequently deleted in myeloid malignancies.¹⁸ Furthermore, PIK3CG has been found to be downregulated by CpG hypermethylation in Human Colorectal Carcinoma.¹⁹

Several X-ray crystal structures exist for PIK3CG, making it attractive for structural analysis. An investigation of the X-ray crystal structure for PIK3CG bound to ATP (PDB ID: 1E8X),²⁰ may explain how the R839C mutant found in the series of pancreatic cancers could lead to a loss of function. Arg 839 is located within the C-terminal catalytic domain. Specifically, residue 839 resides within a highly conserved stretch of sequence that is part of the ATP-binding pocket. This pocket is highly conserved among all PI3Ks and resembles the canonical protein kinase catalytic site. In the crystal structure, Arg 839 is within hydrogen bonding distance of Glu 702 and Gln 705, and appears to anchor in place several residues in direct contact with ATP, including Asp 836 and Lys 833. Inhibitor binding studies have shown that relative to other PI3Ks, PIK3CG has a higher level of conformational plasticity within the ATP binding pocket.^21,22 The increased plasticity of the PIK3CG active site combined with the lower electrostatic interaction potential of a cysteine compared to arginine (Cys pKa = 8.33; Arg pKa = 12.48) increases the chance that the R839C mutation causes a destabilization of the ATP binding pocket, resulting in decreased affinity for ATP (Fig. 1).

Figure 1.

Structure of the catalytic domain of PIK3CG (PDB ID: 1E8X). ATP is shown bound within the PIK3CG active site. Residues D836 and K831 are within bonding distance of metal catalyst (green sphere) and ATP alpha/beta phosphates respectively. Residue R839 is within hydrogen bonding distance of both E702 and Q705 and appears to serve as an anchor to correctly position ATP contacts along the N-terminal lobe of the catalytic domain.

DGKA (V379I) is a diacylglycerol (DAG) kinase that is important for maintaining membrane phospholipid homeostasis. Its various functions include the conversion of DAG to phosphatidic acid (PA), resynthesis of phosphatidylinositols (PIs) and competition for the second messenger DAG with protein kinase C (PKC).²³ DGKA has been implicated in VEGF mediated angiogenesis²⁴ and suppression of TNF-alpha induced apoptosis of human melanoma cells via NF-KB.²⁵ In addition, this kinase is found to be constitutively activated in nucleophosmin/anaplastic lymphoma kinase (NPM/ALK) fusion in malignant lymphomas, where inhibition of DGKA significantly reduced tumor growth.²⁶

Serine/threonine kinase 33 (STK33) (F323L) remains largely uncharacterized but has sequence similarity to calcium/calmodulin dependent kinase (CAMK).²⁷ In addition, STK33 has been identified as a synthetic lethal in KRAS mutant lung cancers.^28,29

TTK (D697Y) (also known as MPS1) is a dual specificity kinase (Tyr and Ser/Thr) associated with proliferation through regulation of chromosome alignment during mitosis and with centromere duplication. It also controls nuclear targeting of c-Abl in response to oxidative stress.³⁰ If the protein fails to be degraded, the subsequent production of multiple centromeres could contribute to tumorigenesis. BRAF signaling was shown to interfere with the mitotic spindle checkpoint during cell division by stabilizing TTK levels in melanoma cells.³¹ This gene has also been implicated in bladder cancer.³²

PRKCG (P524R) is the gamma isoform of protein kinase C (PKC). PKC isoforms play diverse physiological roles through their many downstream targets, including transcriptional regulation and cell growth. The gamma isoform is normally only expressed in the nervous system. PRKCG serves as the receptor for phorbol esters, which are a class of tumor promoter.³³ This gene is also mutated in spinocerebellar ataxia, leading to aberrant MAPK signaling.³⁴

FLJ25006 (S196L) is an uncharacterized serine/threonine kinase.³⁵ Aside from the fact that is has a C-terminal kinase domain, very little is known about this protein.

Cell cycle proteins.

NEK8 (A197P) is a member of the “neverin-mitosis” or NIMA-related kinase (NRK) family of proteins and codes for a protein that plays a role in cell cycle progression from G₂ to M-phase. NEK8 has been shown to localize to cilia and regulate both centrosomal and ciliary stability.^36,37 Interestingly, pancreatic cancer and precursor pancreatic intraepithelial neoplasia lesions have been found to be devoid of primary cilia.³⁸ Furthermore, mutations in NEK8 have been found to cause polycystic kidney disease.^39–41

Cell-adhesion proteins.

CMAS (I275R) catalyzes a reaction that provides substrate for the addition of sialic acid to glycoproteins and glycolipids that are important for the structure and function of tissues.⁴² This gene has not previously been implicated in cancer however, increase of sialylation of tumor associated carbohydrate antigens (TACA) is a well-known feature of transformed cells.⁴³

PCDBH2 (A323V) is thought to code for a calcium dependent cell-adhesion protein and is thought to be critical for neuronal connections in the brain, though it does appear to be expressed in the pancreas. Epigenetic silencing of protocadherin gene clusters are observed in association with Wilm tumor and gastric cancer.⁴⁴ These proteins appear to act as tumor suppressors.

Other putative drivers.

PRDM5 (V85I) encodes a transcription factor and is known to act as a tumor suppressor gene.^45,46

PRDM5 is a target of epigenetic silencing in colorectal and gastric cancer.⁴⁷

CALB1 (D199G) is a calcium-binding protein that is part of troponin C superfamily. Expression of CALB1 in a pancreatic islet β-cell line has been shown to protect against cytokine induced apoptosis and necrosis.⁴⁸

MAPT (G333V) encodes for the microtubule-associated protein tau and is a biomarker for endocrine sensitivity and chemotherapy resistance in estrogen-receptor-positive breast cancer.⁴⁹

COCH (S400L) is highly conserved among human, mouse and chicken. COCH has been found to be mutated in sensorineural deafness and vestibular disorder DFNA9.⁵⁰

STN2 (1590S) is part of the endocytic machinery.⁵¹ STN2 has not been implicated in any cancer types.

UNC13C (G1126R) plays a role in vesicle maturation during exocytosis.⁵²

Discussion

Pancreatic cancer is fundamentally a genetic disease—a disease caused by inherited and acquired mutations in selected genes.⁵³ Recent advances in sequencing technologies now make it possible to sequence virtually the entire pancreatic cancer genome. The challenge is to determine which of the many genetic alterations identified are biologically significant “driver” mutations, and which are neutral “passenger” mutations.³

We applied the CHASM methodology to a large set of somatic missense mutations previously identified in pancreatic cancer and we prioritize missense mutations in 19 genes (assigned an FDR ≤0.30) as possible drivers in pancreatic cancer. Fifteen of these 19 genes are not known driver genes in pancreatic cancer, and would not have been prioritized had we simply looked at the number of times that gene has been shown to be mutated in pancreatic cancer. Included in the set of predicted driver mutations was a mutation in the PIK3CG gene that is likely to lead to a loss of function. Although members of the PI3K family, specifically PIK3CA, have been found to be among the most mutated genes in various cancer types, mutations within PIK3CG are less common. There are only 4 mutations for PIK3CG listed in the COSMIC database and the R839C mutation is not among them. To our knowledge, this is the first time that the PIK3CG gene has been implicated in pancreatic cancer.

Many of the canonical signaling pathways responsible for mediating essential cellular functions such as growth, proliferation and cell-cell adhesion are known to be deregulated in tumorigenesis, however, the extent to which specific pathway protein constituents contribute to disease progression when mutated is still not well understood. Due to the complex nature of proteinprotein interaction networks (cross talk, redundancy, feedback), each neoplasm can accumulate a different set of alterations that lead to a similar phenotype, and many genes not previously implicated in cancer are likely to play a role. By identifying putative driver genes that are mutated at low frequency, it may be possible to understand which specific components of a pathway are responsible for its deregulation in individual patients. Targeting these genes for functional studies could provide insight into the disease process and may assist in the development of more specific drugs and therapies.⁵⁴

Materials and Methods

Data collection.

We assembled two sets of missense mutations to train our computational classifier. First, 3,299 missense mutations in 75 genes were selected from the COSMIC database⁶ as likely drivers (Bert Vogelstein, personal communication). Next, we used the CHASM passenger generation algorithm⁴ to create 4,500 synthetic passenger missense mutations. The passenger mutations were based on di-nucleotide based background mutation rates in pancreatic cancer (Supp. Table 1) and were targeted at genes previously seen to be somatically mutated in either breast, colorectal, glioblastoma multiforme or pancreatic cancers.^2,7,8 Finally, we created a third filtered set of 2,700 passenger mutations in genes that have not been previously associated with cancer to use as an empirical null set for estimating p values.

Algorithm.

We used the CHASM protocol to train a random forest (ensemble of decision trees),^9,10 using 3,299 likely driver mutations as the positive class and 4,500 synthetic passenger mutations as the negative class (Data Collection). Each missense mutation was quantified by 70 predictive features, calculated from properties of genomic and protein sequence, predicted protein structure and multiple sequence alignments (Suppl. Table 2).⁴ Each of the 963 somatic pancreatic missense mutations was then scored by the trained forest. Scores range from 0 to 1 and represent the fraction of trees in the forest that classify the mutation as a passenger.

To transform scores into p values, we used the trained forest to score 2,700 filtered passenger mutations from the empirical null (Data Collection). We assume that very few (if any) of these mutations are drivers that were accidentally created by the passenger generation algorithm. Thus, the fraction of filtered passengers with scores below a threshold of interest can be used as an empiritcal p value estimate of the probability that a mutation in that score range is a false positive.

Finally, we obtained q values (FDR) for each of the 963 somatic mutations using the Benjamini-Hochberg algorithm.¹¹