Citations

3528 Background: CALGB 80405 is a phase III clinical trial of FOLFOX and FOLFIRI w/ randomly assigned cetuximab or bevacizumab. Novel machine learning approaches to the study dataset provide valuable insights into CRC prognosis and management of CRC progression. Methods: Using a Monte Carlo Bayesian Generalized Linear Model analytical platform, we built an ensemble of models for overall survival (OS). We used 99 baseline and demographic variables, including 1904 patients w/ 1o side and 949 w/ KRAS wild-type status. Building an ensemble of predictive models reduces risk of overfitting, estimates model uncertainty and identifies key variables by model consensus as measured by ensemble frequency (freq). We fit gender and 1o side (L vs R) stratum-specific models to examine differences in drivers of disease in those strata. Results: 1o side (avg Cox hazard ratio = 0.89, R side reference), ECOG performance status (1.30, reference level 0), AST concentration (1.01), peripheral neutrophil percentage (1.01) and local primary and abdominal site of disease indicators (1.22; 1.26) were key variables predictive of OS ( > 75% freq). In 1o side stratum-specific models, urine protein level (1.61), treatment intent (0.75, nonpalliative as reference) and hemoglobin concentration (0.85) were more associated w/ L side progression (freq > 85% in L stratum model, < 20% in R), while liver and lung sites of disease (2.3; 1.09) were more associated w/ R side progression (freq > 65% in R stratum model, < 20% in L). Predictors of 1o left-sidedness included age (avg log odds ratio = 0.02), hemoglobin (0.41), and abdominal (3.79) and liver (0.68) sites of disease. Modest differences in disease prognostic factors existed between genders: women more influenced by metastatic status, age, liver site of disease and creatinine level; men more influenced by urine protein level and prior diabetes. Conclusions: 1o side plays a central role in potentially explaining both variation in OS and differences in drivers of OS. Availability of these measures at baseline enables better sense of disease course at initiation of treatment. Support: U10CA180821, U10CA180882, Eli Lilly & Co., Genentech, Pfizer Clinical trial information: NCT00265850.

3511 Background: CALGB 80405 was a randomized Ph3 trial showing no OS or PFS difference in mCRC pts treated with Bevacizumab (BV) or Cetuximab (Cet) in the first line. A Nanostring platform was used to determine the CMS classification of 392 KRAS wt (codon 12 and 13) primary tumors and correlated it with OS and PFS in patients enrolled in 80405. Methods: CMS for 392 of 431tumors were defined using a custom CRC Nanostring panel (39 CMS classification not possible). Stratified Cox proportional hazard model was used to evaluate the effect of CMS classification stratified by prior radiation, prior chemotherapy, adjusting for age, sex, race, primary in place, liver met only, and sidedness. Results: We found CMS1 (14%), CMS2 (47%), CMS3 (2%), CMS4 (29%), NonConsensus (8%). Results are shown in Table 1. Patients with CMS1 who received BV had significantly longer OS than those who received Cet (HR 0.47, 95% CI [0.24, 0.92]). Patients with CMS2 who received BV tended to have shorter OS than those who received Cet (HR 1.41, CI [0.95, 2.08]). Conclusions: Our data suggest that CMS is associated with OS and PFS in first line therapy in mCRC patients. Preliminary data suggest that certain CMS may be associated with efficacy of Bev and Cet based chemotherapy. CMS classification should be explored as a stratification factor in future trials. Support: U10CA180821, U10CA180830, U10CA180882 Clinical trial information: NCT00265850. [Table: see text]

3504 Background: CALGB 80405 was a randomized phase III trial that found no difference in OS in first-line mCRC pts treated with either bevacizumab (Bev) or cetuximab (Cet). Primary tumor DNA from 361 pts, including KRAS mutant (mut) pts, has been profiled for somatic gene mutations/ML/MSI to discover molecular markers of OS. Methods: Mutations in 11 genes were determined by PCR, MSI by microsatellite analysis, and ML by next-generation sequencing (FoundationOne). Cox proportional hazard models are used, stratified by prior XRT and +/- adjuvant chemotherapy; adjusted by age, race, gender, synchronous vs. metachronous, liver metastases, sidedness, all RAS. Results: BRAF: Mut pts had shorter OS than wild-type (wt) pts (HR 1.92, 95% CI 1.34,2.75; p<0.001); HR 1.65 (1.09,2.50) after adjusting for sidedness (p 0.022). In mut pts longer OS is observed in Bev arm vs. Cet arm (p 0.041); in wt pts no arm difference is observed (p 0.291, table). MSI: OS does not differ between MSI-H and MSI-S pts (HR 0.78 [0.40, 1.52], p 0.450). In MSI-H pts longer OS is observed in Bev arm vs. Cet arm (p 0.002); in MSI-S pts no difference is observed (p 0.305, table). ML: Hypermutated MSI-H pts are excluded. In a subset of 205 pts, pts with ML>5 (N=93) have longer OS than pts with ML≤5 (N=112) (HR 0.65 [0.42,1.00], p 0.048). In Bev arm higher ML confers longer OS than lower ML (HR 0.85 [0.80,0.96], p 0.004); in Cet arm no difference is observed (HR 0.99 [0.90,1.09], p 0.862). Conclusions: BRAF is a strong negative prognostic factor in mCRC, even when sidedness is taken into account. ML is a novel marker for further evaluation. The effect of Bev and Cet in either BRAF mut or MSI-H pts should be tested in larger datasets. Updated results from more screened samples will be presented. [Table: see text]

3503 Background: 80405 found no OS or Progression Free Survival (PFS) difference when bevacizumab (BV) or cetuximab (Cet) was added to 1st-line FOLFOX or FOLFIRI in All RAS wild type (wt) mCRC pts. There was a significant 1° side by biologic interaction (P int: OS = 0.008, PFS = 0.001) favoring pts with left-sided (L) 1°. Analyses of 1° tumors beyond All RAS includes Consensus Molecular Subtype (CMS), BRAF and MSI. (CMS results - see Lenz et al; BRAF -see Innocenti et al) We asked whether 1° tumor location - L vs right (R) - is an independent prognostic marker when these other molecular features are considered. Methods: We used a Cox proportional hazard model stratified by prior XRT and +/- adjuvant chemo; adjusted for age, gender, synchronous vs metachronous, CMS, MSI and BRAF status. Pts with transverse (T) tumors were excluded in this analysis. Results: Sidedness was determined in 782 pts (L - 472; R - 256; T -54). Molecular data from 728 pts (with L - and R-sided 1°s) was available as follows: KRAS -- 291, NRAS -393, BRAF - 393, MSI - 378, CMS - 533. L vs R mOS: 32.9 v 19.6 months (mo) (p < 0.0001). See Table for OS results in All RAS / BRAF wt and BRAF mutant (mut) pts. Sidedness (R vs L) is an independent prognostic marker even after adjusting for all these molecular features: HR = 1.392 (1.032, 1.878), p = 0.031. Conclusions: Primary tumor location is an independent prognostic factor when adjusted for age, gender, synchronous/metachronous, CMS, MSI and BRAF status. We are exploring clinical variables such as tumor burden, metastatic sites and measurability of disease in an attempt to explain the impact of sidedness. Support: U10CA188021, U10CA180882. Eli Lilly and Co, Genentech/Roche, Pfizer, Sanofi. Clinical trial information: NCT002655850. [Table: see text]

Abstract Background Previous genome-wide association studies (GWAS) have identified 42 loci (P < 5 × 10−8) associated with risk of colorectal cancer (CRC). Expanded consortium efforts facilitating the discovery of additional susceptibility loci may capture unexplained familial risk. Methods We conducted a GWAS in European descent CRC cases and control subjects using a discovery–replication design, followed by examination of novel findings in a multiethnic sample (cumulative n = 163 315). In the discovery stage (36 948 case subjects/30 864 control subjects), we identified genetic variants with a minor allele frequency of 1% or greater associated with risk of CRC using logistic regression followed by a fixed-effects inverse variance weighted meta-analysis. All novel independent variants reaching genome-wide statistical significance (two-sided P < 5 × 10−8) were tested for replication in separate European ancestry samples (12 952 case subjects/48 383 control subjects). Next, we examined the generalizability of discovered variants in East Asians, African Americans, and Hispanics (12 085 case subjects/22 083 control subjects). Finally, we examined the contributions of novel risk variants to familial relative risk and examined the prediction capabilities of a polygenic risk score. All statistical tests were two-sided. Results The discovery GWAS identified 11 variants associated with CRC at P < 5 × 10−8, of which nine (at 4q22.2/5p15.33/5p13.1/6p21.31/6p12.1/10q11.23/12q24.21/16q24.1/20q13.13) independently replicated at a P value of less than .05. Multiethnic follow-up supported the generalizability of discovery findings. These results demonstrated a 14.7% increase in familial relative risk explained by common risk alleles from 10.3% (95% confidence interval [CI] = 7.9% to 13.7%; known variants) to 11.9% (95% CI = 9.2% to 15.5%; known and novel variants). A polygenic risk score identified 4.3% of the population at an odds ratio for developing CRC of at least 2.0. Conclusions This study provides insight into the architecture of common genetic variation contributing to CRC etiology and improves risk prediction for individualized screening.

Abstract The efficacy of immunotherapy varies widely among different gastrointestinal cancers. Response to immune checkpoint inhibitors is shown to correlate with tumor mutation load (TML), mismatch repair deficiency (dMMR) status, and programmed cell death-ligand 1 (PD-L1) expression. Herein, we quantify TML, dMMR, and PD-L1 expression and determine their interrelationship in gastrointestinal cancers. Here, a total of 4,125 tumors from 14 different gastrointestinal cancer sites were studied using validated assays. Next-generation sequencing was performed on genomic DNA isolated from formalin-fixed paraffin-embedded tumor specimens using the NextSeq platform. TML was calculated using only somatic nonsynonymous missense mutations sequenced with a 592-gene panel. Microsatellite instability (MSI) was assessed using direct analysis of altered known MSI loci in the target regions of the sequenced genes. PD-L1 expression was analyzed by IHC. Interestingly, right-sided colon and small-bowel adenocarcinomas had the highest prevalence of TML-high tumors (14.6% and 10.2%, respectively). Pancreatic neuroendocrine tumors and gastrointestinal stromal tumors had the lowest rates of TML-high (1.3% and 0%, respectively). TML-high was strongly associated with MSI-H (P &lt; 0.0001). However, all TML-high anal cancers (8.3%) were microsatellite stable (MSS). Higher PD-L1 expression was more likely to be seen in MSI compared with MSS tumors (20.6% vs. 7.8%, P &lt; 0.0001). Implications: TML-high rate varied widely among gastrointestinal cancers. Although MSI is conceivably the main driver for TML-high, other factors may be involved. Future clinical trials are needed to evaluate whether the integration of TML, MSI, and PD-L1 could better identify potential responders to immunotherapy. Mol Cancer Res; 16(5); 805–12. ©2018 AACR.