Abstract 1181P
Background
There is growing evidence suggesting genetic heterogeneity among pts with KRAS-mutated (KRAS mt) NSCLC. The real-world (RW) evidence on prognosis of KRAS and KRAS G12C is limited.
Methods
Adult pts with aNSCLC, without ALK, EGFR, or ROS, were selected from de-identified Flatiron Health-Foundation Medicine Clinico-genomic Database (01/01/2011 - 31/10/2021). The de-identified data originated from ∼280 US cancer clinics (∼800 sites of care). Pts were assigned to 3 cohorts: KRAS mt, KRAS G12C mt, KRAS wild-type (KRAS-wt). KMEANS was used to identify clusters within KRAS mt and KRAS G12C mt. Overall survival was estimated using random forest (RF) and Cox regression. Patient features in the models included demographics, vitals, lab reports, molecular features, first-line treatment, and metastatic sites.
Results
Of 11569 pts included, 7714 (66.7%) were KRAS-wt and 3855 (33.3%) were KRAS mt. Among KRAS mt, 1514 (39.3%) had KRAS G12C mt. 91% of KRAS mt and KRAS G12C mt pts had non-squamous tumors (non-sq), vs 62% in KRAS-wt. 94 features were included in machine learning models. Three-cluster model was identified as the best model based on within-cluster-sum of squared errors. Compared to KRAS-wt, KRAS and KRAS G12C were associated with higher mortality risk in univariable Cox model for pts with non-sq, bone metastasis, or KEAP1 mutations (HRs from 1.11 to 1.42; all p-values <0.05). Using RF and Cox combined model, bone metastasis, ECOG, liver metastases, albumin count and TP53 mutations were top 5 ranked important prognostic factors for KRAS mt and KRAS G12C subgroups. Table: 1181P
| Characteristics | KRAS | KRAS G12C | KRAS-Wt |
| N | 3855 | 1514 | 7714 |
| Median age (y) | 69 | 69 | 69 |
| Female (%) | 58% | 60% | 43% |
| Non-sq (%) | 91% | 91% | 62% |
| Ever smokers (%) | 93% | 97% | 86% |
| Bone metastases (%) | 31% | 33% | 27% |
| Brain metastases (%) | 20% | 20% | 14% |
Conclusions
Machine learning methods demonstrated heterogeneity within KRAS mt and KRAS G12C subtype. KRAS G12C subtype was a negative prognostic factor for pts with non-sq tumors, bone metastases and KEAP1 co-mutations. Machine learning methods can be promising tools to support treatment individualization in pts with aNSCLC in RW.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Novartis Pharmaceuticals Corporation.
Funding
Novartis Pharmaceuticals Corporation.
Disclosure
H.H.F. Loong: Financial Interests, Institutional, Invited Speaker: Boehringer Ingelheim, MSD; Financial Interests, Personal, Invited Speaker: Eli Lilly, Illumina, Bayer, Guardant Health; Financial Interests, Personal, Advisory Board: Novartis, Takeda. T. Pattipaka: Financial Interests, Personal, Full or part-time Employment: Novartis International AG; Financial Interests, Personal, Stocks/Shares: Novartis International AG. T.W. Smith, S.M. Knoll, V. Pretre: Financial Interests, Personal, Full or part-time Employment: Novartis; Financial Interests, Personal, Stocks/Shares: Novartis. R. Caparica Bitton: Financial Interests, Institutional, Full or part-time Employment: Novartis Pharma AG; Financial Interests, Institutional, Stocks/Shares: Novartis. F. Ye: Financial Interests, Personal, Full or part-time Employment: Novartis Pharmaceutical; Financial Interests, Personal, Stocks/Shares: Novartis Pharmaceutical. A. Farago: Financial Interests, Personal, Full or part-time Employment: Novartis; Financial Interests, Personal, Stocks/Shares: Novartis; Non-Financial Interests, Personal, Proprietary Information: Novartis. All other authors have declared no conflicts of interest.