Clinico-pathological associations and concomitant mutations of the RAS/RAF pathway in metastatic colorectal cancer

Somatic hotspots in KRAS, NRAS, BRAF, and PIK3CA, known to be associated with mCRC, were obtained from COSMIC v69, and used as an input for the Ion AmpliSeq Designer™ online tool. Overall, 14 multiplexed primer pairs spanning 1556 base pairs of genomic DNA were designed and used, after titration experiments using two mCRC samples known to be KRAS mutated and wild types, respectively. One hundred pre-stored consecutive samples with known mutations in the aforementioned genes, previously assessed by capillary sequencing as part of our routine diagnostics were resequenced by NGS on Ion Torrent™ 314™ chips, multiplexing an average of 8 samples per chip using AmpliSeq™ barcode kits. The following one hundred nineteen samples were collected consecutively and analyzed by NGS. Sensitivity and specificity of our platform, when sequencing at depth in excess of 1000×, were greater than 99% for FFPE samples with a mutant allelic fraction above 5%.

The analysis pipeline used was as follows: in brief, runs were analyzed on a Dell workstation using the Ion Torrent™ 3.4 alignment and calling pipeline, using hg19 as the reference genome. To increase specificity of our results, as per manufacturer’s instructions, we limited our analyses to target and hotspot regions defined by COSMIC v69 and provided as an output by the Ion AmpliSeq Designer™ tool (files available upon request to the corresponding author). The effective presence and deleteriousness of found mutations, when falling outside the known hotspots identified by COSMIC, was confirmed by manual inspection of the pileups using IGV, and subjected to functional prediction according to a majority consensus of SIFT, PolyPhen2, and MutationAssessor [9‐11]. Mutational variables for downstream statistical analyses were selected using a well-established penalized maximum likelihood regression model implemented in the R package glmnet [12]. For the comparison of mutational and clinical variables, we employed multiple logistic regression corrected with the Firth method, adjusting for stage, age, anatomical site, and gender when appropriate. Pairwise associations between categorical variables were evaluated by Fisher’s exact test, while differences of continuous variables among two or more groups were assessed using Wilcoxon or Kruskal–Wallis tests, respectively.

In survival analyses, we considered as our primary endpoint progression-free survival upon beginning a first line of chemotherapy in the metastatic setting. Survival curves were plotted using the Kaplan–Meier estimators, generated with the package survcomp [13], and p-values were calculated using the log-rank test. For survival analyses including more than one variable, a stepwise backward–forward Cox proportional hazards regression model was employed, starting from all clinical and pathological variables described above, until minimization of the Akaike Information Criterion was achieved (package MASS) [14].

Multiple correspondence analysis (MCA) was performed using the package FactoMineR (http://​factominer.​free.​fr/​contact/​index.​html), after categorizing variables as described, and results were represented with ggplot2 [13]. The mutual exclusivity and lollipop plots in Fig. 4 were generated using the online tools OncoPrinter and MutationMapper [15, 16]. All statistical analyses were conducted within the R environment for statistical computing [17].

We validated using digital PCR (dPCR) assays co-occurring somatic mutations identified in KRAS (G12S, G12 V, G12I, G12C, G12F, G12D, G15S, G13C), NRAS (G13S), BRAF (S467L), PIK3CA (G1050S). The target selection was based on the prevalence of mutational co-occurrences in our set as well as on the commercial availability of pre-validated TaqMan^® SNP Genotyping Assays. These are single-tube assays containing dPCR primers and a probe for both the wild-type and mutant alleles. We prepared a reaction mix for each sample for two technical replicates (2 chips): 4 μL of DNA (20 ng/µL) with 11.7 μL nuclease-free water, 17.4 μL QuantStudio™ 3D Digital PCR Master Mix, 1.7 µL Custom TaqMan^® SNP Genotyping Assays (20×). The dPCR reaction was loaded onto a QuantStudio™ 3D Digital PCR Chip v2, and PCR was performed using the ProFlex 2× Flat PCR System with the program outlined in Additional file 1: Table S1. We analyzed the chips using the QuantStudio™ 3D Digital PCR Instrument and the QuantStudio^® 3D Analysis Suite™ Software.

Over 24 months, 264 specimens from colorectal cancer (CRC) patients were brought to our laboratory for KRAS, NRAS, and BRAF characterization (see Additional file 1: Figure S1). Per protocol, PIK3CA hotspots regions were also assessed in the current study, but not reported to the oncologist in charge of the patient when bearing mutations of potential pathological significance. Of 264 samples that derived all from the primary tumors, 45 were not assessed because there was no indication for RAS testing (non-metastatic patients). Samples from 219 patients were then analyzed with NGS. No sample failed NGS testing, although we had to re-extract tumor DNA in 13 cases due to poor material quality, and we had to generate amplicon libraries twice in 10 cases because of experimental failures along the library preparation pipeline.

Patient characteristics are reported in Additional file 1: Table S2. Median age was 70 years (interquartile range 61–76), 52% patients of the analyzed cohort were male, and 69% of neoplasms with available stage data (n = 156) pertained to the left colon, with most patients presenting with a large (T3 or greater), node-positive tumor at first diagnosis. Microsatellite staining by immunohistochemistry was available for 113 patients of the total population and showed 9 patients (8% of the assessed cases) with loss of MLH1 (8 patients), MSH6 (1 patient), MSH2 (no patients). These demographic and anatomopathological features were well balanced according to sex, and in line with the epidemiologic characteristics of the mCRC patients diagnosed in our region. The relatively high median age is indeed in line with the patients referring to our hospital, which is situated in the region hosting the oldest population in Europe. Outcome data of patients were available for 101 patients of the entire population.

Of 219 assessed samples analyzed, 143 (65%) presented at least one deleterious mutation in any of the four sequenced genes. The number of patients with mutations in a single gene was 109; with two mutations, 31; and with three mutations, 3. Mutations were not detected in any of the KRAS, NRAS, BRAF, and PIK3CA genes (quadruple mutation-negative patients) in 76 (35%) tumors.

We identified 104 (47%) tumors with KRAS mutations, including one tumor with three KRAS p.G12C, p.G12V, and pG13C mutations and six tumors with two KRAS mutations p.G12V and p.G15S; pG12S and p.G12V; p.G12D and p.V9F; p.G12S and p.A59E; p.G12D and p.T58I; p.G12C and p.G13D. The mutations were located in exon 2 in 92 tumors (42%), exon 3 in 7 tumors (3.2%), and exon 4 in 8 tumors (3.6%). p.G12D, p.G12V, and p.G13D were the most common KRAS mutations in our cohort of colorectal cancers (see Fig. 1). Mutations located outside exon 2 were observed in 19 tumors (8%).

Fifteen tumors (7%) harbored a NRAS mutation. Seven mutations (3%) were located in exon 2 and 8 (4%) were located in exon 3. No mutations were found in exon 4. We found four NRAS p.G12C, p.G12D, p.G12V mutations in codon 12, three mutations p.G13D, p.G13R, p.G12S in codon 13, seven mutations p.Q61 K, p.Q61R in codon 61, and one mutation p.G60E in codon 60.

Twenty-eight tumors (13%) harbored a BRAF mutation including one tumor with two BRAF mutation p.S467L and p.V600M. Twenty-three mutations (10%) are located in exon 15 and six mutations (2.7%) were located in exon 11. p.V600E was the most common mutation for BRAF, since it was found in 20 tumors (9%). One tumor harbored an unusual p.V600M mutation with deleterious effect. Eight tumors (3.6%) harbored non-V600 mutations.

Thirty-three tumors (15%) harbored a PIK3CA mutation including one tumor with two PIK3CA mutations p.D725 N and p.H1047Y. Twenty mutations were located in exon 9 (9%), 8 mutations in exon 20 (3.6%), and 6 mutations in exon 13 (2.7%). Three codons (p.E542, p.E545, and p.H1047) account for 26 of 33 PIK3CA mutations (78%). Only three tumors (1.3%) were characterized by a single PIK3CA mutation without any concomitant alterations in the assessed genes.

Sixty-eight (31%) tumors were located in the right colon and 151 (69%) tumors were located in the left colon. Quadruple mutation-negative tumors were observed in 13 right-sided colorectal cancers compared to 49 left-sided colorectal cancers (19 vs 41%, p = 0.001). Right-sided colorectal cancers have had a significantly higher incidence of mutations in BRAF (Additional file 1: Table S3, 8% vs 5%, p = 0.0001). Even though MSI staining was available in 113 of 219 patients, we found 7 patients (6%) with BRAF mutations and MSI status. Tumors with BRAF mutations were more frequently associated with microsatellite instability (6% vs. 2%, p = 0.0009). KRAS mutations were more frequent in females than in males (26% vs. 22%, p = 0.0460), as were BRAF mutations (8% vs 5%, p = 0.0269). In order to explore in an unbiased way the associations of the RAS/RAF pathway with clinical and pathological variables in colorectal cancer, i.e., stage, tumor site, nodal status, sex, and age, we used multiple correspondence analysis (MCA). MCA is a multivariate statistical method akin to principal component analysis but suited for categorical data; as a result, we reduced our variables of interest to two dimensions, which explain the largest fraction of the variance observed in our data set. Clinical and pathological variables together with KRAS, NRAS, BRAF, and PIK3CA status were projected as vectors in a space defined by those two dimensions (see Fig. 2). The position of the variable categories in this two-dimensional space reflects their mutual associations, with no a priori assumption on the underlying structure of the data. We observed that tumors with BRAF mutations fell in proximity of “older age” and “right colon side”. In contrast, RAS mutations clustered with PIK3CA mutations, “younger age”, and “female sex”. In summary, by both MCA analysis and classic statistical tests we could observe two groups of patients enriched in specific mutational and clinico-pathological features (“BRAF mutations”, “right side”, “old age” in group 1 and “RAS mutations”, “PIK3CA mutations”, “younger age”, and “female sex” patients in group 2).

We found 34 tumors (15%) with concomitant mutations within KRAS, NRAS, BRAF, and PIK3CA genes. Thirty-one tumors presented concomitant mutations in two genes; three tumors presented concomitant mutations in three genes including one tumor with BRAF, NRAS and PIK3CA concomitant mutations and two tumors with BRAF, KRAS, and PIK3CA concomitant mutations. Concomitant KRAS mutations were observed in 23 of 33 PIK3CA-mutated tumors (63%), 2 of 15 NRAS-mutated tumors (13%), and 4 of 29 BRAF mutated tumors (13%). Tumors with a PIK3CA mutation showed a significantly higher incidence of concomitant mutations (90%) as compared with tumors with KRAS (26%) or BRAF (28%) mutations. KRAS and PIK3CA tended to be concurrently mutated, in a statistically and biologically significant fashion (OR = 2.97, 95% CI 1.27–7.40, p value = 0.0076, see Table 1).

Concomitant KRAS, BRAF, or NRAS mutations were detected in 18 of 21 tumors (85%) with a PIK3CA exon 9 mutation and in eight of 8 tumors (100%) with a PIK3CA exon 20 mutation. Moreover, we found tumors with multiple mutations in the same gene including six tumors with two KRAS mutations, one tumor with three KRAS mutations, and one tumor with two BRAF mutations. Of these, two tumors harbored two KRAS mutations in the same codon. This combination of multiple mutations in the same gene highlighted by different allelic frequency is reported in Table 2.

We collected clinical data concerning progression-free survival (PFS) and type of chemotherapy for 101 patients. Out of these, 75 underwent first-line chemotherapy. In particular, 24 patients were treated with a combination of fluorouracil-based chemotherapy and an anti-EGFR moAb (either cetuximab or panitumumab), 26 underwent a similar backbone chemotherapy regimen with anti-VEGF moAb (bevacizumab), and 25 received a chemotherapy-only first-line therapy (either single-agent capecitabine or fluorouracil or combination with irinotecan or oxaliplatin). Tumor side (simplified as left for rectal, sigmoid, and descendant; and right for ascendant or transverse) was available for 67 patients. Information concerning tumor side, mutational status, and type of chemotherapy were available for 63 patients.

By univariate analysis, BRAF-mutant patients (N = 7) had a shorter PFS compared with KRAS/NRAS mutant patients (N = 40, hazard ratio—HR = 3.58, 95% confidence interval—CI 1.48–8.66), and wild-type ones (N = 32, HR = 3.44, 95% CI 1.42–8.30) (p-value = 0.0073, logrank test, two-sided)—see Fig. 3a. The worse prognosis associated with mutated BRAF mutation was conserved by multivariable analysis, when taking into account PIK3CA status (adjusted HR = 3.22, 95% CI 1.26–8.20, p-value = 0.0142 by Cox proportional hazards regression) and the interaction of PIK3CA status with KRAS/NRAS or BRAF status (p-value for interaction = 0.4620 and 0.5870 respectively)—see Fig. 3b. Since most PIK3CA mutations co-occurred with mutations in KRAS/NRAS/BRAF (25% vs. 6%), we assessed whether PIK3CA status could stratify the subgroup of mutated patients (N = 47) into different prognostic groups, but this analysis yielded non-significant results (p-value = 0.5686, logrank test, two-sided). Since all PIK3CA exon 20 mutations co-occurred with RAS ones, we could not stratify our analyses for PIK3CA exon 9 vs. 20 mutations. Finally, we aimed at evaluating whether tumor side may have prognostic significance when considered alone or by multivariable analysis together with BRAF status, type of first-line chemotherapy, and the interaction of these two variables. Indeed, right side was associated with worse PFS by univariate statistical assessment (p-value = 0.0494, logrank test, two-sided)—see Fig. 4. Our Cox proportional hazards model was underpowered for analyses with more than 4–5 degrees of freedom, with 44 events and 9 degrees of freedom. As such, our results are of exploratory interest only. Of note, however, tumor side was the only variable independently associated with a PFS, with the direction of the effect in agreement with the univariate log-rank test. Due to the limitation in power size, we do not report formal results concerning this latest analysis.