Systematic transcriptome analysis reveals tumor-specific isoforms for ovarian cancer diagnosis and therapy

Significance

Identifying molecules that are specific to tumors for use in early detection, diagnosis, prognosis, and therapy is both a primary goal and a key discovery challenge across diverse areas of oncology. To discover ovarian tumor-specific molecules, we developed custom bioinformatics algorithms to analyze transcriptome sequence data of 296 ovarian cancer and 1,839 normal tissues and validated putative tumor-specific mRNA isoforms by RT–quantitative PCR. The results revealed multiple candidate diagnostic and therapeutic targets with unique sequences that were expressed in most of the cancers examined but not in normal tissues. The process we developed can be readily applied to identify diagnostic and therapeutic targets for any of the 30 or more tumor types for which large amounts of transcriptome data now exist.

Abstract

Tumor-specific molecules are needed across diverse areas of oncology for use in early detection, diagnosis, prognosis and therapy. Large and growing public databases of transcriptome sequencing data (RNA-seq) derived from tumors and normal tissues hold the potential of yielding tumor-specific molecules, but because the data are new they have not been fully explored for this purpose. We have developed custom bioinformatic algorithms and used them with 296 high-grade serous ovarian (HGS-OvCa) tumor and 1,839 normal RNA-seq datasets to identify mRNA isoforms with tumor-specific expression. We rank prioritized isoforms by likelihood of being expressed in HGS-OvCa tumors and not in normal tissues and analyzed 671 top-ranked isoforms by high-throughput RT-qPCR. Six of these isoforms were expressed in a majority of the 12 tumors examined but not in 18 normal tissues. An additional 11 were expressed in most tumors and only one normal tissue, which in most cases was fallopian or colon. Of the 671 isoforms, the topmost 5% (n = 33) ranked based on having tumor-specific or highly restricted normal tissue expression by RT-qPCR analysis are enriched for oncogenic, stem cell/cancer stem cell, and early development loci—including ETV4, FOXM1, LSR, CD9, RAB11FIP4, and FGFRL1. Many of the 33 isoforms are predicted to encode proteins with unique amino acid sequences, which would allow them to be specifically targeted for one or more therapeutic strategies—including monoclonal antibodies and T-cell–based vaccines. The systematic process described herein is readily and rapidly applicable to the more than 30 additional tumor types for which sufficient amounts of RNA-seq already exist.

Identifying molecules that are specific to tumors for use in early detection, diagnosis, prognosis, and therapeutic strategy design is both a primary goal and a key discovery challenge across diverse areas of oncology. Furthermore, the extent of inter- and intratumor heterogeneity indicates that multiple tumor-specific molecules will be needed for any of these applications (1–3). Although DNA alterations constitute the major focus of tumor-specific discovery efforts to date, in many respects mRNA is more attractive for this purpose. This is because RNA can (i) broadly reflect (malignant) cellular phenotypes, (ii) exist in thousands of copies per cell and thereby enable highly sensitive early detection and diagnostic assays, and (iii) sensitively and comprehensively reveal potential candidate antigens for monoclonal antibody targeting, vaccines, and adoptive immunotherapies (4–6). The efficacy of using mRNA for these purposes is highly dependent on the degree of tumor-specific expression.

One of the main themes of microarray-based experiments that have been undertaken during the last decade has been the discovery of tumor-specific “genes.” Aside from the class of cancer-germ-line (aka cancer/testis) genes (7), few have been found. In retrospect, the “gene” concept critically hindered these efforts to discover tumor-specific expression because the word “gene” is a collective term for all mRNA isoforms expressed from a genomic locus. Malignant and normal tissue types can be distinguished by patterns of differential isoform use (8, 9), but when measured in aggregate at the “gene” level the isoform-specific differences are at best recognized as “gene overexpression” or “gene underexpression.” Thus, mRNA expression is not commonly considered to be “tumor-specific”, but “tumor-associated” (via overexpression). The distinction is important, for “tumor-specific” molecules are an ideal that is devoid of detection interpretation ambiguity and off-targeting. So although it has become increasingly clear that there are few, if any, “genes” only expressed in tumors, aside from fusion transcripts (10) the extent to which tumor-specific mRNA isoforms exist is unknown.

Transcriptome sequencing (RNA-seq) is a genomics technology whose principle purpose is to enable genome-wide expression measurements of mRNA isoforms—the level at which distinct tumor-specific mRNA molecules are to be found. To apply RNA-seq for the purpose of identifying mRNA isoforms that tumors express and normal tissues do not express, large databases of RNA-seq data from malignant and normal tissues are required. The Cancer Genome Atlas (TCGA; cancergenome.nih.gov) is an NIH-sponsored effort to study the RNA and DNA in 500 tumors for many cancer types, and the Genotype-Tissue Expression (GTEx) program (11) is an NIH-sponsored effort to study the RNA and DNA in thousands of samples from >50 distinct normal tissue sites. Both of these programs are multicenter efforts that are generating molecular profiling data at a rate, scale, and cost that almost certainly could not be borne by any single entity. The primary intention of these efforts is to generate a public resource to catalyze leaps in progress across all aspects of cancer care, prevention, and therapy. The raw transcriptome data being produced by these efforts has tremendous discovery potential, but to date they have not been rigorously evaluated for tumor-specific molecules for diagnostic and therapeutic applications. Herein we report on our results to date in using these RNA-seq data to identify mRNA isoforms that are only expressed in high-grade serous ovarian adenocarcinomas (HGS-OvCa) and to evaluate the isoforms’ potential for tumor biology insights and oncologic applications.

Results

The overall strategy of our tumor-specific isoform identification process (Fig. 1) is based on (i) computational algorithms we custom-developed for sensitive and accurate isoform identification, (ii) large databases of tumor and normal tissue RNA-seq data produced by TCGA and GTEx, and and (iii) high-throughput RT-qPCR experiments. As reported below, we first used our custom algorithms to efficiently process large amounts of RNA-seq data and applied one prioritization strategy to produce a list of mRNA isoforms rank-prioritized by likelihood of being tumor-specific. We then used our custom-developed software for automated design of isoform-specific PCR primers and performed RT-qPCR using pooled tumor RNA and pooled normal tissue RNA. For isoforms found to only be present in the tumor pool, we measured their expression by RT-qPCR in a larger set of nonpooled tumor and normal samples. The isoforms that were expressed across multiple tumors were then ranked based on whether they were expressed in zero, one, two, three, four, or more normal tissues and evaluated for oncologic applications.

Fig. 1.

Discovery process overview. We applied custom bioinformatics algorithms to large public databases of tumor and normal tissue RNA-seq data to rank-prioritize mRNA isoforms by likelihood of being tumor-specific. We then used RT-qPCR in two phases to confirm tumor-specific expression. First we performed RT-qPCR to analyze the RNA of four tumors pooled together vs. the RNA of four normal tissues pooled together. Then we selected the most likely tumor-specific isoforms based on expression profiles in these two pools. Final validation was RT-qPCR on individual tumor and normal tissues.

Computational Pipeline for RNA-seq.

The standard RNA-seq computational pipeline for organisms with a sequenced genome has three main components (Fig. 2A): (i) alignments of RNA-seq reads to the genome, (ii) an isoform model database, and (iii) an integration algorithm, whose input is the isoform model database and the read pair alignments and whose output is the expression level of the supplied isoforms. We developed a pipeline for isoform identification and expression level estimation that is distinguished by custom methodologies and software algorithms in each of these three components.

Fig. 2.

RNA-seq bioinformatics. (A) Our custom RNA-seq computational pipeline broadly conforms to the standard three-component RNA-seq computational pipeline for organisms with a sequenced genome. (B) In our approach, we align read pairs (RP) with maximally sensitive parameterizations and use all known splice junctions, allowing even 1-bp splice junction “overhangs.” Nucleotide-level read-to-isoform consistency analysis identifies and records the read pair-isoform tuples that are exactly concordant and filters out read pairs that are not exactly concordant with some known isoform (I). (C) We minimize isoform nonidentifications (false negatives) with our isoform model database that is a merger of the six major isoform model databases worldwide. Given the read pair-isoform tuples from B, we use a parsimony principle to subsequently minimize false isoform identifications (false positives).

A major distinguishing feature of our approach to RNA-seq read alignment is our use of maximally sensitive alignment parameterizations coupled with nucleotide-resolution read-to-isoform correspondence verification. Such parameterizations enable the thorough detection of all RNA-seq read alignments spanning splice junctions, which are especially informative because they provide exon linkage information that can be crucial for accurate isoform identification. Current practice sets “minimum overhangs” of a read’s alignment over a splice junction into an adjoining exon—often 8 bp or more—to guard against false genomic alignments. To maximally recover the information in RNA-seq reads, we consider alignments with even 1-bp overhangs, but then through nucleotide-resolution read-to-isoform correspondence verification we reject all read pair alignments that do not exactly match the human genome reference sequence. This approach has four consequences (Fig. 2B). First, we maximize the isoform identification information in each set of RNA-seq data. Second, we identify read pairs that do not correspond to any known isoform and prevent their subsequent use for isoform expression estimation. In practice, these rejected read pairs constitute 2–3% of the raw data and are indicative of the presence of isoforms that have not yet been discovered and incorporated into any public database (12). Third, we explicitly associate each read pair with a specific isoform or set of isoforms from which it could have been derived and then use this information in the final expression estimation stage. Owing to the high overlap of isoforms at a genomic locus, read pair alignments often overlap isoforms from which they both could and could not have been physically derived. In some RNA-seq computational protocols, this distinction is not addressed and read pair alignments are erroneously used to estimate the expression of isoforms from which they could not have been physically derived. As shown in Fig. S1 for an exemplar RNA-seq dataset, read-to-isoform correspondence verification markedly reduces the number of isoforms with which read pairs can be associated. Fourth, we explicitly associate read pairs to isoforms to enable our strategy for minimizing both false positives and false negatives in RNA-seq experiments (discussed below).

Fig. S1.

Effect of read-to-isoform verification. RNA-seq read pair alignments often overlap isoforms from which they both could and could not have been physically derived. We perform nucleotide-level correspondence analysis to explicitly associate each read pair with a specific isoform or set of isoforms from which it could have been derived. As shown, this procedure markedly reduces the number of isoforms with which read pairs can be associated.

A major distinguishing feature of our approach to isoform models is the use of a custom isoform model database that we created by merging all of the major isoform model databases (Fig. 2C). Although the use of only one particular isoform model database is standard in current RNA-seq computational protocols, doing so is a source of false negatives (13); if a particular isoform is not in the database, then the integration algorithm (Fig. 2A) cannot know about it and use it for expression estimation. By merging all major isoform model databases, our approach minimizes the possibility of such false negatives. Conversely, isoforms in a supplied isoform model database that are not actually expressed in a sample from which RNA-seq data were generated represent noise for the integration algorithm and can lead to the assignment of nonzero expression for unexpressed isoforms. To minimize the possibility of such false positives, we use the read-to-isoform verification information discussed above and our implementation of a greedy solution to the set cover problem (14) to identify the set of isoforms that most parsimoniously explains the RNA-seq read alignments. In effect, we create an isoform model database that is tailored to each RNA-seq experiment. As shown in Fig. S2, this tailoring reduces the number of isoforms from loci that are used as input to the integration algorithm.

Fig. S2.

Average isoform use at loci. As part of our computational RNA-seq pipeline we apply a parsimony principle to identify the isoforms that can most succinctly account for the RNA-seq reads aligned at a genomic locus. The effect of this procedure is to significantly reduce the number of isoforms models that are supplied to an integration algorithm that estimates isoform expression levels from aligned RNA-seq data. (For visual clarity, only loci with 50 or fewer isoforms are shown here.)

Tumor-Specific Isoform Predictions from 2,135 RNA-seq Experiments.

For our study we sought those mRNA isoforms most pervasively and exclusively expressed in HGS-OvCa. Using 296 curated TCGA RNA-seq datasets for HGS-OvCa, we first identified isoforms expressed in 90–100% of tumors. To capture even very lowly expressed transcripts, we used an expression level cutoff of 10⁻⁶ fragments per kilobase of transcript per million fragments mapped to define whether a transcript was expressed or not. This first filter yielded 117,108 isoforms (Fig. S3A). We next used the 1,839 GTEx RNA-seq datasets to count the number of normal tissues in which the average expression of each of these 117,108 isoforms was equal or higher. As shown in Fig. S3B, most of the isoforms expressed in 90–100% of the TCGA ovarian tumors were also expressed in many normal tissues. For each of the 22,082 isoforms that was equally or more highly expressed in at most one other tissue, we identified the normal tissue with the highest average expression and computed two statistics: (i) the Mann–Whitney P value associated with the two sets of expression values (i.e., tumor vs. normal) and (ii) the fold change of the average tumor expression over the average normal tissue expression. As shown by Fig. S3C, most of the 22,082 isoforms were not appreciably distinguished in their tumor expression from their “closest” normal tissue expression by average expression fold change or the distribution of expression values. Finally, we rank-prioritized the 22,082 isoforms by likelihood of being tumor-specific by sorting them by fold change and P value.

Fig. S3.

Use of RNA-seq to identify the isoforms most likely to be specifically expressed in HGS-OvCa. (A) Using a liberal expression value cutoff to deem an isoform “expressed,” we first identified those isoforms expressed in 90–100% of all 296 HGS-OvCa tumors samples. (B) We then identified those (22,082) isoforms whose average expression was higher in zero or one normal tissue compared with tumor. (C) For each isoform we identified the normal tissue in which its expression was most similar to ovarian tumor expression. We then computed the fold change of average expression and the Mann–Whitney P value. We sorted the isoforms from B by these two statistics to rank prioritize isoforms by likelihood of being tumor-specifically expressed.

High-Throughput mRNA Isoform-Specific PCR Primer Design.

The sequencing technology upon which this study is based has the limitation of only being applicable to ∼200–250 bp fragments of cDNA—restricting its ability to unambiguously identify mRNA isoforms that in the human genome are on average ∼2 kb. For this reason we used RT-qPCR to confirm the tumor-specific expression of mRNA isoforms that we rank prioritized by RNA-seq. To enable a large number of RT-qPCR experiments, we custom-developed software that could exhaustively identify and design primers for all unique amplicons of any target mRNA in the human genome. With this software we attempted to design primers for 671 of the topmost tumor-specific candidate mRNA isoforms. The number 671 was chosen so that we could perform our initial pooled screening (discussed below) with 11 384-well plate PCR experiments. To reach 671 it was necessary to attempt primer designs for the 1,230 topmost tumor-specific candidate mRNA isoforms—corresponding to a 54.6% primer design success rate. Of the unsuccessful attempts, 320 (26.0%) were due to the lack of a unique amplicon sequence in the target isoform and 239 (19.4%) were due to primer design failure. (Primer design failure can occur for reasons related to T_m requirements, forward and reverse primer compatibility, primer or amplicon sequence length constraints, and primer amplification of unintended products.)

Confirmation of Isoform Tumor-Specific Expression by RT-qPCR.

We performed confirmatory RT-qPCR experiments (Fig. S4 and Table S1) using a two-phase approach. In phase 1 we used pooled RNA to efficiently filter out isoforms that were not expressed in tumors and/or were expressed in normal tissues. We formed a pool of four different tumor RNA samples and a pool of four different normal tissue RNA samples and then measured the expression of all 671 isoforms in both pools. As graphed in Fig. 3, we found that 66.2% (n = 445) of isoforms were detected in both pools, 18.2% (n = 122) were detected only in the tumor pool, 1.0% (n = 7) were detected only in the normal pool, and 14.5% (n = 97) were not detected in either pool. Furthermore, our experiments revealed the presence of novel isoforms that are not documented in any of the isoform model databases that we used to construct our custom isoform model database. In the group of isoforms found in both pools, 18.3% of reactions revealed one or two additional products. For the “tumor only” and “normal only” groups, the percentages were 5.7% and 0.4%, respectively.

Fig. S4.

geNorm output. As described in Methods and Materials, qPCR results for 32 candidate reference amplicons in six tumor and normal tissues were used as input to geNorm. For readability we show the geNorm output for the 13 most stable references. The red line in the bottom panel marks geNorm’s standard acceptability cutoff of 0.15.

Table S1.

PCR primers for reference amplicon experiments

Gene symbol	Len	Forward Tm	Reverse Tm	Forward primer	Reverse primer
SLC25A37	223	59.9	60.1	TCGGTGAAGAGACAGGGTCT	GAAGCCTTCGGTCCGCAT
AF011889.4	215	59.8	59.8	GCTGGAGATGAGACCCAGC	CAGGTGTCCTCCCTCCCA
WWC2	212	58.4	59.6	GCTGACTTTGAAGACTATGTGGA	ACCTGGGCGGTCTCTACA
METAP1D	195	60.4	59.1	TGACCGACGCCAACATGG	CTGAAGAAACTGCAGCCGG
ZNF542	237	57	58.7	CAGTAATGGGATGAGTGACATTC	TGAAATATCCTGGCAATGGGC
COX18	173	57.5	59.6	AGAGATGCCAGGCTCACT	TCTGAATGTGCTGCCCCC
SYNRG	203	60.4	60.6	CAGTGGGTGGAGCTGCAG	AGGGGCACTGTTTCCATGC
DHX38	170	58.8	60.0	TCAGCAAGACCCCACAGG	AGGCGCGTTCTCCAGTTC
EIF2B5	127	59.6	59.6	CTGTGGCAGGGTGTTCGA	CCACGACCACCTGGGAAG
SIRT2	179	59.9	60.0	TGCAGGAGGCTCAGGACT	ACAGCGTTCGCTCTGCAT
RBM17	127	59.9	60.1	ACATGGTTGGTGCGGGAG	ACTGCTTCATCATCAGGGGC
ANAPC7	113	59.9	59.66	AAGCCCTGACCCAAAGGC	AGCCAGTGCGTTCCTCAG
MRPS5	118	59.2	60.3	TTGGGGAGGCAGTGTTCC	TGGCGTAGGGATGGGTGT
HSPA9	120	59.6	59.7	TCCGTGCCTCCAATGGTG	TGTGCCCCAAGTAATTTTCTGC
PSMD4	205	59.9	58.6	GCTGACCACACTCACCCC	AGCCAGTTTCACCAGATCCT
TCP1	125	59.5	57.4	TGGTGCAACCATCCTGAAGT	CTGCTGCAATAATAACCACTGAA
TBP	126	58.7	59.7	TTAACAGGTGCTAAAGTCAGAGC	AAAGAAGGGGGTGGGGGA
PGK1	223	57.4	59.2	CGTTATGAGAGTCGACTTCAATG	AACATCCTTGCCCAGCAGA
HSP90AB1	136	60.1	59.24	CTCGTCGGGCTCCCTTTG	ACCACACCACGGATAAAATTGAG
LDHA	114	60.0	59.5	GCCCGAACTGCAAGTTGC	CCAGATTGCAACCGCTTCC
GAPDH	172	60.3	59.9	AAGGTGGTGAAGCAGGCG	CGTTGTCATACCAGGAAATGAGC
GUSB	168	59.0	60.4	TTGCAGGGTTTCACCAGGA	GCACTCTCGTCGGTGACTG
SDHA	263	59.1	60.3	TGGCACTGGGAAGGTCAC	GGTTCCTGGCAAGCTCCC
HPRT1	110	59.4	59.0	TGCTTTCCTTGGTCAGGCA	TTCAAATCCAACAAAGTCTGGCT
ACTB	142	59.7	59.5	CAAGAGATGGCCACGGCT	AGGACTCCATGCCCAGGA
TFRC	173	60.1	59.8	CTGCAGAGGTCGCTGGTC	TCCACGAGCAGAATACAGCC
RPLP0	137	59.9	59.8	TGCCAGTGTCTGTCTGCAG	AGGCCTTGACCTTTTCAGCA
NONO	187	60.2	58.9	GCACAGCCTGGCTCCTTT	GGCGCCTCATCAAATCCTG
PPIA	108	60.0	59.8	AAAGCATACGGGTCCTGGC	TGCTTGCCATCCAACCACT
PPIH	101	59.1	58.9	TGGCCGCATGAAGATCGA	TTGGAACCCCATCTTTCCTGA
RPLP1	100	60.1	59.7	CAATGTAGGGGCCGGTGG	CACTTTCTTCTCCTCAGCTGGA
B2M	272	59.0	60.7	TAGGCTCGTCCCAAAGGC	GGTTCACACGGCAGGCAT

Fig. 3.

Categories of pooled RNA RT-qPCR experiment outcomes. For candidate tumor-specific mRNA isoforms derived from RNA-seq-based analyses, we measured their expression by RT-qPCR in a pool of tumor (T) RNA samples and pool of normal tissue (N) RNA samples. The expression status of the isoforms in aggregate spanned all possible outcomes. By melt curve analysis, we observed instances in which just the target product (single) was amplified and instances in which multiple products (mult) were amplified—indicative of the presence of novel mRNA isoform structures. The number of isoforms in each category is displayed atop each bar.

In phase 2 we measured the expression of a subset of the isoforms in an expanded set of individual, nonpooled RNA samples. For the subset we selected isoforms that were detected in only the tumor pool, that were associated with a single peak melt curve, and that were at least moderately expressed [i.e., quantification cycle (C_q) <31–32]. (Low expression of an isoform in a pool could mean moderate expression in only one sample of the pool or low expression across all samples in the pool.) These selection criteria resulted in a subset constituting 86 of the 122 isoforms only detected in the tumor pool. To expand the set of RNA samples we added an additional 8 tumor samples and an additional 14 normal tissue samples—for a total of 12 tumor samples and 18 normal tissue samples. We then measured by RT-qPCR the expression of the 86 isoforms in the 30 individual samples and then ranked the isoforms by the number of normal tissues in which they were expressed. The top-ranked 33 isoforms, shown in Fig. 4 and Table S2, constitute 5% of the original 671 isoforms investigated and either have tumor-specific or restricted normal tissue (normal-restricted) expression. The top six isoforms, or 0.9% of the original 671, were expressed in 6–12 of the 12 tumors and were undetectable in all 18 normal tissues examined. An additional 11 isoforms (1.6% of 671) were only observed in one normal tissue, which in most cases was either fallopian tube or colon. In the remaining 16 cases (2.4% of 671) in which the isoforms were present in two, three, or four normal tissues, fallopian tube and/or ovary were most consistently among the normal tissues.

Fig. 4.

RT-qPCR expression measurements for mRNA isoforms. In total, we selected 671 mRNA isoforms for tumor-specific confirmation RT-qPCR experiments. Using pooled RNA, we found a subset of them to be only expressed in the tumor RNA pool (Fig. 3). We selected 86 of these for a second set of RT-qPCR experiments with 12 tumor and 18 normal tissue RNA samples, which were not pooled. The 33 mRNA isoforms with tumor-specific and the most normal-restricted expression from the set of 86 are shown, constituting 5% of the original 671.

Table S2.

PCR primers for isoforms in Fig. 4 of the main text

Gene symbol	Isoform	Len	Forward Tm	Reverse Tm	Forward primer	Reverse primer
WFDC2	ENST00000462062	146	60	59.3	TTGCTGGAGCTGCAGTCT	GGAACCCTCCTTATCTTGGTTC
RP11-3J1.1	ENST00000505347	99	58.4	60.0	ACACAATGTGTCCTAGAAGAAGA	AGCTGCCCATTCGACTGT
TMPRSS3	ENST00000398405	77	59.8	59.4	GCCCCCTTCTCATTCCGA	GCAGCAACAGCATCTGGT
CTD-2616J11	ENST00000574814	169	60	59.8	CTGGGCCTGAAGGGAACA	AGGGTGTCCAGGCGTATG
ETV4	lAug10	200	59.9	60.7	GGCGAGCAGTGCCTTTAC	CGCACCCGGTGACATCTAT
PRAME	gAug10	171	59.9	60.2	GGCGTGAATGCGTGGATT	TGCCACGCACGTGTTTTT
huhare	fAug10	269	60.2	59.9	GCCGTGGTGGTGTATTGC	CACATCACTGGGCGTTCG
SPC24	ENST00000429831	106	59.8	60.1	GCTGCTGGAAACGCAAGA	CTGTCGCTCCTGCTCCTT
PTH2R	ENST00000413482	174	60.0	60	AGGTTCCTTGAACAGCTGGA	CACTGTTCCTCTGGGCCA
VTCN1	ENST00000359008	284	59.1	60.0	GCTGACCTCGCGCATAAT	CTGTCCGGCCTCTGAACA
SLC44A4	ENST00000414427	114	60.1	59.9	TCACTGTCGCCCAGAAGG	CGCTGATCCCCTGCTGTA
ESR1	ENST00000456483	332	60.6	58.2	CCGGCATTCTACAGGCCA	TCCACAAAGCCACCTTTCA
TNFRSF8	ENST00000263932	121	59.8	59.9	GCCTTCCCACAGGATCGA	ACTGCTGTGTCGGGAACA
VASN	HIT000304576	420	60.6	60	CCTATCGGGCCCTGTTGG	TGTAGGGCTTTGCGTGGA
MYLPF	ENST00000563728	110	60	60.1	ACCAGGAACCCAATCGCA	GCCCACATGTTCTTGATCTCC
FOXM1	lAug10	150	60.1	60	CGGCCTCAAACCCAAACC	GGCTCCTCAACCACAGGT
FOXM1	ENST00000536066	234	59.9	60.0	GTCCCCCTGCTCCTGATC	TCCCCTCCTCAGCTAGCA
C19orf53	ENST00000588841	158	59.9	58.4	CCCAAGAAAAGGCGGTAAGG	ACTTCTAGGTTCTTCTTGAGCTT
CD9	iAug10	146	59.8	59.9	GGGGTCAGCGGGACTTTA	GCAAGGACAGCAATCCCG
RAB11FIP4	ENST00000578694	94	59.8	59.8	GCCTGGGAGGTCGTGTTA	CATCAGCAAAGGTGGGGC
CHODL	ENST00000465099	96	59.0	60.2	TGAGCCAATTCCCTGGAGA	GAATCAACGTGCTGGCCC
AURKA	sAug10	257	60	59.8	ACTTGGGTCCTTGGGTCG	TGCACTCCAGCCTCTAGC
CDCA5	ENST00000529290	218	58.9	60.0	GAACCTGCCCACCTTATTGT	GGTCACTGCAGGCAGAGA
CDH24	uc001wil	167	59.2	60	TACAGAGCTCGGCTGGAG	TCTGGATGGCCACTTGCA
FGFRL1	fAug10	428	60.2	60	GCTCCTCTGGGGGTCAAG	GCGGTTTTGGGTCTTGCA
LSR	uc002nyp	160	60	59.8	CCTCAGGTGTTCCCAGCA	CCACTGCGGACTGAGCTA
SLC22A18	ENST00000312221	345	59.9	60	GGCTGGAACTCAGACCCA	TAGAGCGCTCATCCTGCC
STON2	aAug10	222	60.8	59.7	AACTCAGCTTCCGGTCACC	TGTTTCTGTTGTCTGGTAGCTG
SLC44A4	pHIT000078073	112	59.9	59.5	TACAGCAGGGGATCAGCG	TGTCCCACAGCCACAAGA
OPN3	hAug10	135	60.0	58.9	TGCTGGTGTCCCTCTTCG	TGCATTTGTGACTGGAACTCT
AC019117	ENST00000419463	69	59.3	60.1	GGAACATCTACACACAGAGGAAA	GGGGACTGTTGGGAATGGA
LINC00284	ENST00000439707	323	59.8	60.0	TCAGAAGGCAAAGATTGACCAG	TCCTGCTGAGCCAGGAAC
MUC16	HIT000048730	108	59.8	59.9	ACCCATCGGAGCTCTGTG	GGAACAGTTACTTGTGGGGC
TFRC	Multiple (ref.)	173	60.1	59.8	CTGCAGAGGTCGCTGGTC	TCCACGAGCAGAATACAGCC
RBM17	Multiple (ref.)	127	59.9	60.1	ACATGGTTGGTGCGGGAG	ACTGCTTCATCATCAGGGGC
PPIA	Multiple (ref.)	108	60.0	59.8	AAAGCATACGGGTCCTGGC	TGCTTGCCATCCAACCACT

Biologic Basis and Applications of Candidate Tumor-Specific Molecules.

Because the 33 mRNA isoforms in Fig. 4 are expressed in 6–12 of the 12 different tumors and have highly restricted or undetected normal tissue expression, they are of immediate and high interest for both understanding tumor biology and for oncologic applications. A complication that arises when interpreting isoform-level findings is that most isoforms of most genes have not been explicitly studied, and even small differences in mRNA or protein isoform primary sequence from a well-studied canonical isoform can alter the molecule’s function, localization, lifetime, structure, and/or interaction network (15). With this caveat in mind, we highlight below isoforms that are likely to play a causative functional role in the malignant state and that have potential use for diagnosis and therapy.

Isoforms of genes related to oncogenesis, stem cells, and stem cell-like cancer cells.

A structurally distinct mRNA isoform lAug10 of ETV4/PEA3 (Fig. 4) was expressed in all studied tumors and was detectable only in normal heart. ETV4 is a transcription factor that is active in developing embryos and adult tissues and that has a demonstrated transforming role in Ewing’s tumors and prostate, ovarian, breast, and other solid tumors (16). The lAug10 isoform is incompletely known at the 3′ end, but enough of the transcript has been sequenced to reveal that lAug10 is the only ETV4 isoform with a truncated N-terminal amino acid sequence and a skipped exon 5. The functional implications of this distinguishing structure are unknown.

FOXM1 is a transcription factor that is both a potent oncogene and an important molecule for maintaining stem cell renewal (17). The gene is highly expressed across a broad range of different solid tumor types, including ovarian cancer. Integrated genomic analyses of ovarian cancer performed by TCGA found the FOXM1 regulatory network to be the most significantly altered in expression level across 87% of the 489 tumors studied. FOXM1 has multiple isoforms, two of which have been studied for their transforming potential (18). The study found that isoforms FOXM1b and FOXM1c both had transforming potential, and that FOXM1c was likely to be constitutively active because it was proteolytically processed to yield a short isoform without the N-terminal inhibitory domain. The lAug10 and gAug10/ENST00000536066 isoforms that we identify in Fig. 4 were neither of the isoforms studied, but interestingly both are short isoforms that are missing the N-terminal inhibitory domain. Thus, it may be that one or both of the FOXM1 isoforms that we identified are constitutively active transforming isoforms of FOXM1.

Tetraspanin proteins are increasingly viewed as therapeutic targets because of their emerging key roles in tumor initiation, progression, metastasis, and sometimes angiogenesis (19). We identified isoform iAug10 of CD9/tetraspanin-29 that was expressed in 10 of 12 tumors and absent from all but one normal nongynecological tissue sample. CD9 is a cell surface marker for normal human embryonic stem cells and for cancer stem cells in non-small-cell lung carcinoma (20). It has various anti- and protumorigenic roles, with the latter including that of an oncogene in an ovarian cancer line (21). The varied and opposing roles of CD9 have been suggested to be a consequence of its different interaction partners in the plasma membrane (19). An additional and compatible reason, though, may be its multiple protein isoforms.

The lipolysis-stimulated lipoprotein receptor (LSR) is a gene that in basal-like triple-negative breast cancer cell lines is a biomarker of cells with cancer stem cell features and with a direct role in driving aggressive tumor-initiating cell behavior (22, 23). These observations are relevant to our study because of the discovery that basal-like breast cancers and ovarian serous cancers exhibit very similar mRNA expression programs and share critical genomic alterations (24)—indicating related etiology and therapeutic opportunities. At the gene level LSR is transcribed in multiple normal tissues, but our investigation revealed LSR isoform uc002nyp.3 to be expressed across all 12 tumors studied and undetectable in all 12 normal tissues studied. Intriguingly, because of this isoform’s structure (Fig. 5D) it has dual therapeutic potential; its splice junction forms a unique amino acid sequence that is a predicted extracellular epitope and is computed to have a high binding affinity for three different MHC I alleles. Thus, this isoform has the potential of encoding a protein with one tumor-specific polypeptide that is both an antibody and T-cell target on ovarian cancer stem cells and that, if found to be expressed in breast basal-like tumors, could be relevant for multiple difficult tumor types.

Fig. 5.

Candidate protein therapeutic targets. (A) The candidate isoform PTH2R.bAug10 is distinguished from the canonical PTH2R isoform by its alternative first exon, which alters the N-terminal amino acid sequence. Both protein isoforms are predicted to contain signal peptides (that are likely cleaved). After signal peptide cleavage, the first exon of PTH2R.bAug10 would still retain a unique 12-aa sequence, which, because the protein is a class B GPCR, is expected to be extracellular and thus amenable for antibody targeting. (B) The candidate isoform CD9.iAug10 is distinguished by a unique exon, which is expected to add 41 uniquely distinguishing amino acids—some of which project into the extracellular environment and constitute a protein-specific antibody target. (C) The candidate isoform ETV4.lAug10 has a unique exon structure that creates a unique splice junction spanning amino acid sequence with high computed binding affinity to two common MHC 1 alleles. (D) The LSR mRNA isoform ucnyp002.3 contains a unique splice junction spanning amino acid sequence that is expected to reside in the extracellular domain of this plasma membrane protein and that also contains subsequences that are computed to have moderate to high binding affinity to multiple common MHC 1 alleles. Thus, the single amino acid sequence is amenable to two therapeutic modalities.

Isoforms for early detection and monitoring of HGS-OvCa.

The Papanicolaou test has recently been demonstrated to be a viable source of ovarian tumor cells (25). This observation allows for the possibility of an early ovarian cancer diagnostic test based on the detection of ovarian tumor-specific mRNA isoforms that are expressed in tumor cells that have disseminated to the cervix. For such an early detection strategy to work, one would need to identify mRNA isoforms that are only expressed in ovarian tumors and not in normal gynecologic tissues. Extensive experimental evidence (26–29) indicates that fallopian tube, and to a lesser extent the ovary, are the tissue(s) of origin of HGS-OvCa. Additionally, many studies (30–32) have demonstrated that expression profiles of tumors are more similar to those of their tissue of origin than to any other normal tissue, so for HGS-OvCa fallopian tube and ovary are the most stringent tissues against which to judge the tumor specificity of an mRNA isoform. As shown in Fig. 4, we found that 15 (2.2%) of our original starting set of 671 isoforms were not expressed in the ovary or fallopian tube, and so constitute an initial candidate set of mRNA isoforms upon which a new and innovative strategy for the early detection of ovarian can be developed.

Isoforms predicted to encode cell surface targets.

The parathyroid hormone receptor 2 gene PTH2R encodes a class B (type II) G protein-coupled receptor (GPCR) that is predominantly expressed in endocrine and limbic regions of the forebrain and to a lesser extent in restricted cell types of peripheral tissues (33). Its function in nonbrain tissues and in cancer has not been studied. The mRNA isoform that we identified is highly expressed in 10 of the 12 tumors used herein (Fig. 4). The isoform is distinguished by its alternative first exon, which is predicted to retain a (likely cleaved) signal peptide (Fig. 5A). In addition to the signal peptide, the first exon would confer on the protein isoform a unique 12-aa sequence. Because the protein is a class B GPCR, its N-terminal sequence is expected to be extracellular and thus amenable to antibody targeting.

The CD9 isoform identified herein, which was expressed in 100% of the late-stage 296 TCGA tumors and in 10 of the 12 tumors (Fig. 4), contains a unique exon (Fig. 5B) that imparts upon the protein a unique, in-frame 41-aa sequence that encompasses the first two transmembrane regions of the protein and the extracellular domain between them—making it amenable to specific antibody targeting if expressed.

Isoforms predicted to encode epitopes for tumor vaccines.

Although the C-terminal portion of the normal-restricted ETV4 isoform identified herein is incompletely known, the portion that is known reveals the isoform to have an exon-skipping event that is unique among all ETV4 isoforms—conferring on the resulting protein at least 14 unique amino acids (Fig. 5C). We analyzed the epitope potential of this region using a computational method (34) that has been recently validated by retrospective prediction against a large set of bona fide T-cell antigens that induced immune responses and were associated with tumor regression and long-term disease stability (35). We identified a 10-mer epitope centered directly over the unique splice junction and calculated it to have a very strong affinity (12.9 nM) for the HLA allele A*02:01 and a moderate affinity (363 nM) for the B*08:01 allele. Because the A*02:01 and B*08:01 alleles are among the most common HLA alleles in the Caucasian population of the United States (36), the ETV4 isoform is a strong candidate for immunotherapeutic application for ovarian cancer.

Discussion

We have developed a highly customized RNA-seq bioinformatics pipeline that is designed for isoform identification and that is distinct from standard approaches because of (i) its use of an isoform model database that is a merger of all isoform model databases available worldwide, (ii) its capability for maximally sensitive genome-wide read alignment, and (iii) the nucleotide resolution consistency analysis that is performed for every sequencing read–isoform combination. Furthermore, we developed a workflow for high-throughput, isoform-level RT-qPCR experiments that is distinguished by custom software for automated design of PCR primers that are specific to individual mRNA isoforms at complex genomic loci (i.e., loci in which no isoform may even have a uniquely distinguishing splice junction or exon). Both the RNA-seq pipeline and RT-qPCR infrastructure have been developed to the point of being highly automated, requiring for a new cancer type modest manual involvement and timeframes (∼6–7 wk) to generate the level of analysis reported herein. We used our combined computational/experimental pipeline to generate detailed molecular hypotheses in the form of specific molecules (i.e., mRNA isoforms and/or the protein isoforms that they encode) with ovarian tumor-specific expression and with particular oncologic application(s). Importantly, the hypotheses we generated would not have been possible with gene-level analyses that by definition encompass numerous mRNA and protein isoforms in aggregate.

Analogous to the challenge of distinguishing driver from passenger mutations in cancer genomics (37), cancer transcriptomics must contend with the challenge of distinguishing those mRNA molecules that are important for the malignant phenotype from those that are not. We addressed this challenge by requiring the mRNA isoforms interrogated in this study to be expressed in 90–100% of the TCGA ovarian tumors, with the rationale being that a tumor-specific isoform that is present in most tumors is likely to be functionally important rather than due to a deregulation side effect. In support of this rationale, among the topmost 5% (n = 33) tumor-specific or normal-restricted isoforms are variants of genes that are demonstrated oncogenes, known to maintain the malignant state, have a direct role in driving aggressive tumor initiating cell behavior, or are necessary for maintaining a stem-cell phenotype. In addition to the cancer genomics goal of identifying driver mutations is the goal of identifying driver mutations that are “actionable.” The RT-qPCR experiments revealed 15 mRNA isoforms that have the tumor specificity required for an early detection diagnostic of ovarian cancer. Additionally, at least 5 of the 33 mRNA isoforms confirmed by RT-qPCR to have tumor-specific or normal-restricted expression encode protein targets that have unique primary structures that would allow them to be specifically targeted by one or more therapeutic strategies, including monoclonal antibody therapy/chimeric T-cell generation, and peptide- or T-cell–based vaccines.

Beyond protein, mRNA itself has the potential to be a therapeutic target (38, 39). If proven to be so, mRNA has a great advantage over protein as a class of target molecule because MHC epitope and cell surface restrictions would not apply. However, like protein therapeutics, mRNA would need to be targeted isoform-specifically because of the high degree of identical nucleotide sequence among the isoforms from a genomic locus. Our study is pertinent to mRNA therapeutics because we demonstrate a feasible strategy for finding tumor-specific mRNA targets.

Herein we have proposed the idea—inspired by a DNA-based approach (25)—of an ovarian cancer detection test based on the detection of tumor-specific mRNA isoforms from malignant cells that have disseminated to the cervix and been collected during a Papanicolaou test. A strategy based on RNA and not DNA could have distinct advantages. Tumor types have characteristic expression profiles that are distinctive from both those of other tumor types and normal tissues. An approach based on RNAs that are broadly indicative of characteristic expression programs could be more robust because it would not rely on particular mutations but on a characteristic cancer cell expression phenotype. Furthermore, because somatic DNA mutations occur in one or a few copies per tumor cell and RNA isoforms can occur in hundreds to thousands of copies per cell, an assay based on mRNA is potentially much more sensitive. The first requirement for such a test is the enumeration of mRNA molecules that indicate the presence of an ovarian tumor. In our experiments, we identified isoforms that were expressed in most or all tumors and were not detected in any normal tissues. Furthermore, we identified additional isoforms that were expressed in most or all tumors and in only one normal tissue that, importantly, was not ovary or fallopian tube. These additional isoforms are also candidates for a detection test because, not being found in the gynecologic tissues tested, they would be indicative of tumor cells if detected in a Papanicolaou test.

There are a number of hard limitations to the approach for tumor-specific isoform identification and validation. These hard limitations are due to the “short read” nature of RNA-seq data and to the great extent to which mRNA isoforms at a genomic locus share exons and splice junctions. RNA-seq reads represent, essentially, 200–250 contiguous base pairs of processed mRNA. Because most mRNAs are much longer than 250 bp, RNA-seq reads cannot provide the information that links distant exons and that is often necessary for unambiguous identification of the source mRNA isoform. Our RNA-seq computational procedure was designed for maximum accuracy in identifying those isoforms that were, and were not, represented in an RNA-seq dataset. To achieve this goal, we minimized false negatives by merging all of the major isoform model databases and then developed nucleotide-level correspondence and parsimony algorithms to minimize false positives. Nonetheless, determining which isoforms generated a set of RNA-seq reads is an inference problem that will always be error-prone and because of this no isoform identification procedure will be completely accurate. However, even if one were able to identify the mRNA isoforms underlying an RNA-seq dataset with complete accuracy, there is a severe limitation on the rate at which their expression can be confirmed by PCR. To confirm an mRNA isoform one must design PCR primers that amplify a uniquely distinguishing nucleotide sequence. At complex genomic loci this is a very challenging task because of the extent to which exons and splice junctions are shared among isoforms. A major component of our study is the algorithms that we developed for automated design of isoform-specific PCR primers. Even with our specialty software we found that we could only design primers for ∼55% of isoforms, meaning that almost half of the isoforms that we predicted by RNA-seq to be tumor-specific could not be investigated by RT-qPCR. Furthermore, for ∼25% of the isoforms for which we could design primers, melt curve analysis revealed the presence of multiple PCR products (often two or three)—indicating the presence of new isoforms. These observations are compatible with recent transcriptome sequencing experiments that have reported on new isoform discovery rates (12, 40, 41). That RT-qPCR discovers isoforms at a higher rate attests to its higher sensitivity and lack of library preparation procedures.

As opposed to the hard limitations that exist for our approach, there are three “soft” limitations that could be readily addressed to potentially improve our tumor-specific isoform identification rate. First, we used only two metrics to rank-prioritize isoforms by likelihood of being tumor-specific. The output of our RNA-seq computational procedures has six metrics. Additionally, our procedures have three threshold values that have not been optimized. We expect that the use of more or other metrics for rank prioritization and of optimized threshold values will yield additional results of the same qualitative nature as reported herein. Second, ovary and fallopian tube were the most common normal tissues in which isoforms were expressed (Fig. 4). As the tissue of origin and primary tumor site, these are exactly the normal tissues in which a tumor-expressed isoform is most likely to be expressed. Unfortunately, these are also exactly the normal tissues for which we had the fewest normal control RNA-seq datasets (three ovary and one fallopian tube). Thus, our ability to negatively filter tumor-expressed isoforms was limited. The GTEx project is actively sequencing ovary and fallopian tube, so this soft limitation will diminish in the near future. Third, we did not account for the known expression subtypes of HGS-OvCa (42–44), but instead sought mRNA isoforms that were expressed in all tumor subtypes (i.e., 90–100% of the 296 TCGA tumors). Incorporating subtype classification into our procedures could yield tumor subtype-specific mRNA isoforms.

We note that additional experiments will be required for the proposed applications of the tumor-specific isoforms that we identified. Tumor cells that disseminate to the cervix or into the bloodstream may down-regulate the isoforms that are expressed in primary tumors, so for utility in a Papanicolaou test-based early detection diagnostic or in identifying circulating tumor cells the continued expression of isoforms in these nonprimary tumor sites will need to be confirmed. Additionally, mRNA expression does not always equate to protein expression, so for the protein isoforms with therapeutic target potential their expression and cellular localization in tumor cells will need to be experimentally confirmed.

In summary, we have developed and rigorously evaluated a systematic process for identifying tumor-specific mRNA isoforms that leverages the large and growing public databases of tumor and normal tissue RNA-seq data. We have quantified the rate at which tumor-specific isoforms can be identified for HGS-OvCa and have demonstrated that they have the potential to provide the specificity needed for extremely specific diagnostics and therapeutics. Our findings are relevant in a larger context because the procedures we developed can be readily and rapidly applied to any of the 30 or more tumor types for which large amounts of RNA-seq data now exist.

Methods and Materials

RNA-seq Bioinformatics.

We created a custom isoform model database by merging the six major isoform model databases available worldwide. We used the set of all isoform splice junctions from our custom database and lenient parameterizations to perform highly sensitive genome-wide alignment of RNA-seq paired-end reads. We then performed an alignment-filtering step to remove spurious alignments that can be generated by using lenient parameterization. To filter, we analyzed each read pair alignment to determine whether or not its implied cDNA fragment was a contiguous subsequence of any mRNA isoform(s). We then use the filtered read alignments to compute the subset of our custom isoform model database that most parsimoniously accounted for the filtered alignments. In effect, we created a tailored isoform model database for each RNA-seq dataset. Finally, we converted read pair genome alignments to transcriptome alignments and explicitly used the strict correspondence between read pairs and isoforms to compute isoform-level expression.

RT-qPCR.

We performed RT-qPCR experiments according to Minimum Information for Publication of Quantitative Real-Time PCR Experiments (MIQE) guidelines (45), which among other criteria include the use of multiple references for intersample comparison and the calculation of PCR efficiencies for quantification. Tumor RNA was obtained from the University of California, San Diego Moores Cancer Center Biorepository and commercially (Origene). Normal tissue RNA was obtained commercially (Biochain and Origene).

Further details are provided in SI Methods and Materials.

SI Methods and Materials

Primary RNA-seq Data.

The primary data for our study consisted of the RNA-seq data generated by TCGA (cancergenome.nih.gov) for high-grade ovarian serous cystadenocarcinoma (OV) and by the GTEx project (11) for 43 different, nondiseased normal tissues. TCGA has generated RNA-seq datasets for 420 OV samples, but many of them have been redacted or are from replicate aliquots. We used table 1 from the October 10, 2013 GDAC Summary Report (gdac.broadinstitute.org/runs/stddata__2013_10_10/samples_report/OV.html) to identify the nonredacted samples and the best single sample from replicate aliquots. This curation resulted in 296 samples, for which we then downloaded the raw RNA-seq paired-end read data for our study. We used all of the 1,839 RNA-seq datasets available from GTEx as of June 1, 2013. We downloaded all of the available paired-end read data specified in the file “SraRunTable_4-15-2013.txt” that was obtained from the SRA Run Selector page on the dbGap (46) website for the GTEx project.

RNA-seq Bioinformatics Pipeline.

Stage one: Consolidated isoform model database.

We created a merged, nonredundant set of gene isoform models by first combining the isoform models from the “ncbi_37_Aug10” version of Aceview (47), the version of RefSEq. (48) available on December 7, 2012 for GRCh37.p10, the version of UCSC Known Genes (49) for hg19 available on December 8, 2012, version 14 of Gencode (50), the Human lincRNA Catalog (www.broadinstitute.org/genome_bio/human_lincrnas/), and version 8.3 of the H-invitational database (51). We then used the “cuffcompare” program from version 0.9.3 of the Cufflinks software package (52) to make the set nonredundant.

Stage two: Paired end read duplicate removal and genome alignment.

For each RNA-seq dataset, we first removed all but one read pair in each group of read pairs that were identical in both the left and right read. We then aligned the resulting set of read pairs to version hg19 of the human genome reference sequence using STAR (53). We supplied STAR with the set of all splice junctions in our custom isoform model database from stage one and used the following nondefault parameter settings: –outStd SAM –outSAMstrandFieldintronMotif –alignSJDBoverhangMin 1 –outFilterMismatchNmax5 –readFilesCommand zcat –seedSearchStartLmax 12 –alignSplicedMateMapLminOverLmate0.08 –outFilterScoreMinOverLread0.08 –outFilterMatchNminOverLread 0.08 –outFilterMultimapNmax100 –outFilterIntronMotifs RemoveNoncanonicalUnannotated –outSJfilterOverhangMin 12 6 6 6.

Stage three: Read pair consistency analysis and isoform selection.

We developed custom software to evaluate each read pair alignment together with each mRNA isoform to which it aligns and to determine at nucleotide resolution whether the RNA fragment implied by the read pair was a strict subsequence of the mRNA isoform nucleotide sequence. From this consistency analysis we constructed a bipartite graph linking isoforms to consistent read pairs. Read pairs not consistent with any isoform were not included in the bipartite graph and were excluded from further use. To identify the isoforms expressed at a genomic locus, we used the bipartite graph and a custom implementation of a greedy solution to the set covering problem (54) to determine the set of isoforms that most parsimoniously accounted for all of the filtered read alignments. The results of this stage were (i) a set of paired end read alignments that were base pair-level consistent with one or more isoforms and (ii) the subset of isoforms that could completely and most parsimoniously account for them.

Stage four: Calculation of isoform expression levels.

We used the eXpress software package (55) to estimate isoform expression levels. The eXpress software requires two input files: a fasta file with mRNA isoform nucleotide sequences and a BAM file with paired end read alignments in transcriptome (i.e., isoform-specific) coordinates. Because read alignments generated in stage three were in genomic coordinates, we downloaded and custom modified the UBU software (https://github.com/mozack/ubu) to convert the genomic alignment coordinates of each filtered read pair from stage three to isoform coordinates for each isoform to which the read pair was found to be base pair-level consistent. We generated the input fasta file by including only nucleotide sequences for those isoforms constituting the parsimonious set from stage three. The only nondefault parameter setting was “–max-indel-size 20.”

RT-qPCR.

Automated design of PCR primers for mRNA isoforms.

Using Primer3 (56) at its core, we developed custom automated software to design PCR primers that would only amplify a product for a target isoform. For a target isoform, the software first extracted all isoforms in our consolidated isoform model database at the same genomic locus. This set of isoforms was then used to identify (i) all single splice junctions, (ii) all pairs of (not necessarily adjacent) splice junctions, and (iii) all splice junction-unique exonic region combinations that were unique to the target isoform. The software then constructed parameterizations that instructed Primer3 how to search for primers for each of these three cases. For single splice junctions, Primer3 attempts to find (i) primer pairs that enclose but do not overlap the splice junction and (ii) primer pairs in which one of the primers overlaps the splice junction. For pairs of splice junctions, Primer3 attempts to find (i) primer pairs that surround but do not overlap either splice junction, (ii) primer pairs in which only one primer overlaps a splice junction, and (iii) primer pairs in which each primer overlaps one of the splice junctions. For splice junction–exonic region pairs, Primer3 attempts to find (i) primer pairs that surround but do not overlap either the splice junction or the exonic region, (ii) primer pairs in which only one primer overlaps either the splice junction or the unique exonic region, and (iii) primer pairs in which one primer overlaps the splice junction and the other overlaps the exonic region. These parameterizations were set on a case-by-case basis through the Primer3 arguments SEQUENCE_PRIMER_PAIR_OK_REGION_LIST, SEQUENCE_OVERLAP_JUNCTION_LIST, PRIMER_MIN_LEFT_THREE_PRIME_DISTANCE, PRIMER_MIN_RIGHT_THREE_PRIME_DISTANCE. The following Primer3 parameter settings were constant for every case: PRIMER_TASK=”generic”, PRIMER_EXPLAIN_FLAG = 1, PRIMER_OPT_SIZE = 18, PRIMER_MIN_SIZE = 18, PRIMER_MAX_SIZE = 23, PRIMER_PRODUCT_OPT_SIZE = 100, PRIMER_PRODUCT_SIZE_RANGE = 60–450, PRIMER_PAIR_MAX_DIFF_TM = 3, PRIMER_MIN_TM = 58, PRIMER_MAX_TM = 62, PRIMER_OPT_TM = 60, PRIMER_SALT_DIVALENT = 2.5, PRIMER_DNTP_CONC = 0.8.

Until a suitable primer pair is found, the software evaluated the primer pairs returned by Primer3 above in rank order of smallest Primer3 penalty. For the evaluation, it first used the nearest-neighbor thermodynamics based PCR primer specificity checking program MFEPrimer-2.0 (57) to verify that only the one intended product was amplified when using the human genome reference sequence and our consolidated transcriptome database as the template. For primer pairs that passed this specificity evaluation step, the software then queried the uMelt web server (58) to verify that the PCR product would produce only one peak in a melt curve analysis. The first primer pair that passed the amplification specificity and melt curve evaluations was used to define the product that was specific to the target mRNA isoform.

High-throughput qPCR.

All qPCR experiments were performed in 384-well plates and with a total reaction volume of 10 μL. PCR primer oligo (IDTDNA) molarity was 300 nM and the template cDNA concentration was 10 ng/μL. Experiments were performed on Roche LightCycler 480 for 35 cycles. We used the KAPA SYBR FAST qPCR kit optimized for the LightCycler 480 and programmed the instrument according to KAPA recommendations. The primer annealing temperature was 54 °C. Upon completion of a qPCR experiment, we exported the raw amplification and melting data to a text file.

We developed custom qPCR software for analysis, quality control, and expression quantification. To calculate the efficiency of a PCR, we first baseline-adjusted the amplification curve to zero fluorescence intensity units at cycle 2. Next, we simultaneously smoothed the amplification curve and calculated its second derivative using a Savitzky–Golay filter (59) with order 5 and window size 7. We then computed the cycle corresponding to the maximum of the second derivative and used it and the three preceding cycles (for a total of four cycles) to define the exponential region of the amplification curve. Finally, we use our own implementation of the taking-difference linear regression method (60) to compute reaction efficiency. To determine the quantification cycle, C_q, for each curve in a 384-well plate experiment, we determined the value of fluorescence intensity that was most commonly included in the exponential regions of the wells that were not no-template-control wells. We defined this fluorescence intensity as the threshold intensity, N_q. The C_q value for each reaction was then set as the fractional cycle value at which the well’s amplification curve equals N_q. We set C_q = 37 for amplification curves that did not reach N_q.

Genome-wide search, evaluation, and selection of reference amplicons.

Stably expressed reference amplicons are a critical component of a qPCR experiment. To identify the most stably expressed reference amplicons for our study we first used our consolidated isoform model database to identify 2,201,622 splice junctions and splice junction pairs that would give rise to a single, unique amplicon <450 bp from any number of underlying isoforms. For each splice junction/splice junction pair, we then computed the sum of the underlying isoforms’ expression values in each of the 295 tumor and 1,839 normal tissue RNA-seq datasets used in this study. Next, we computed the mean expression and the coefficient of variation (CoV) corresponding to each splice junction/splice junction pair for each of the 44 tissue types (1 tumor plus 43 normal). Finally, we summed the CoV values for each splice junction/splice junction pair across the 44 tissue types and ranked the sums from smallest to largest. From this final ranking of the most stably expressed reference amplicons, we selected the 16 top-ranked reference amplicons that did not originate from standard “reference genes” and the 16 top-ranked that did. (The symbols of the standard reference genes that we considered are ACTB, B2M, GAPDH, GUSB, HPRT1, HSP90AB1, LDHA, NONO, PGK1, PPIA, PPIH, RPLP0, RPLP1, SDHA, TBP, and TFRC.) After using our primer design software to design primers for the 32 candidate reference amplicons (Table S1), we performed qPCR with three ovarian tumor samples (University of California, San Diego Moores Cancer Center Biorepository) and three normal tissues (heart, liver, and kidney; Biochain) and used the resulting expression values as input into our custom implementation of the geNorm algorithm (61). From the output of geNorm (Fig. S4), we selected the three most stably expressed amplicons (annotated as references in Table S2).

Relative quantification.

We wrote a custom software implementation of the qBase relative quantification framework (61) to calculate all normalized relative quantities in this study. In accordance with MIQE guidelines (45), we included computed reaction efficiencies and three reference amplicons (discussed above) in the calculations.

Total RNA and cDNA.

All normal tissue total RNA was purchased from Biochain. Tumor total RNA was either purchased from Origene or derived from frozen tumor samples obtained from University of California, San Diego Moores Cancer Center Biorepository. RNA was extracted manually from frozen tumor tissue samples (approximately 25mg) using Qiagen RNeasy (74104) kit as described by the manufacturer. One microgram of RNA as determined by a Nanodrop 1000 (Thermo Scientific) was converted to cDNA using the SuperScript III Reverse Transcriptase Kit (180800051; Life Technologies) with random hexamers priming as described by the manufacturer. Final cDNA was diluted to the equivalent of 10 ng/μL starting RNA concentration. For normal tissue, cDNA from each tissue type was pooled at equal concentrations to minimize reaction efficiency variation.