(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



An integrated model for predicting KRAS dependency [1]

['Yihsuan S. Tsai', 'Unc Lineberger Comprehensive Cancer Center', 'University Of North Carolina At Chapel Hill', 'Chapel Hill', 'North Carolina', 'United States Of America', 'Department Of Genetics', 'Yogitha S. Chareddy', 'Brandon A. Price', 'Joel S. Parker']

Date: 2023-05

The clinical approvals of KRAS G12C inhibitors have been a revolutionary advance in precision oncology, but response rates are often modest. To improve patient selection, we developed an integrated model to predict KRAS dependency. By integrating molecular profiles of a large panel of cell lines from the DEMETER2 dataset, we built a binary classifier to predict a tumor’s KRAS dependency. Monte Carlo cross validation via ElasticNet within the training set was used to compare model performance and to tune parameters α and λ. The final model was then applied to the validation set. We validated the model with genetic depletion assays and an external dataset of lung cancer cells treated with a G12C inhibitor. We then applied the model to several Cancer Genome Atlas (TCGA) datasets. The final “K20” model contains 20 features, including expression of 19 genes and KRAS mutation status. In the validation cohort, K20 had an AUC of 0.94 and accurately predicted KRAS dependency in both mutant and KRAS wild-type cell lines following genetic depletion. It was also highly predictive across an external dataset of lung cancer lines treated with KRAS G12C inhibition. When applied to TCGA datasets, specific subpopulations such as the invasive subtype in colorectal cancer and copy number high pancreatic adenocarcinoma were predicted to have higher KRAS dependency. The K20 model has simple yet robust predictive capabilities that may provide a useful tool to select patients with KRAS mutant tumors that are most likely to respond to direct KRAS inhibitors.

Mutant KRAS drives approximately 25% of all cancers and has traditionally been considered “undruggable”. However, the recent clinical approvals of inhibitors targeting KRAS with the specific G12C mutation in lung cancer has shepherded in a new era in precision medicine. Although promising, the responses are often modest and short-lived. Therefore, the ability to predict which tumors are dependent on KRAS will help select patients most likely to derive clinical benefit, and those who will not. We have developed an integrated “K20” model based on features that can improve prediction of KRAS-dependency beyond the presence of an activating KRAS mutation. When applied to lung adenocarcinoma, pancreatic adenocarcinoma, and colorectal cancer patient datasets, the K20 model identified specific subpopulations that correlate with greater dependency on KRAS. These findings present a novel approach for identifying biomarkers that can aid in the selection of patients who are most likely to benefit from KRAS inhibitors.

Competing interests: We have read the journal’s policy and the authors of this manuscript have the following competing interests: C.V.P. is the founder of EnFuego Therapeutics, Inc, which is focused on the development of KRAS therapeutics and holds equity in the company. The remaining authors disclose no potential conflicts of interest.

Funding: Y.C. was supported in part by funding from the National Institute of General Medical Sciences ( https://www.nigms.nih.gov/ ) of the National Institutes of Health (NIH) under the Program in Translational Medicine T32 (award number GM122741). C.V.P. was supported in part by the NIH ( https://www.nih.gov/ ) (award numbers R01CA215075, R01CA258451 and 1R41CA246848), the Lung Cancer Research Foundation, the Free to Breathe Metastasis Research Award and a North Carolina Biotechnology Translation Research Grant (NCBC TRG). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Importantly, previous work has revealed that not all KRAS-mutant cell lines are KRAS-dependent, and that such KRAS-independent cancers exhibit features of an epithelial-mesenchymal transition (EMT) and apoptosis resistance [ 16 ]. De novo resistance to blockade of KRAS-activated MAP kinase signaling has recently been linked with high expression of the EMT regulator zinc finger E-box binding homeobox 1 (ZEB1) [ 17 ]. Additionally, genetic models of KRAS depletion have revealed rapid development of several cell intrinsic and non-cell autonomous mechanisms of KRAS-independence [ 18 – 20 ]. These data imply that direct inhibitors of mutant KRAS may not be as efficacious as is seen with other “oncogene-addicted” cancers, and numerous mechanisms of resistance may emerge. Consistent with this, while response rates of ~60–70% are typically seen in lung adenocarcinomas (LUAD) using inhibitors of aberrant EGFR, ALK, ROS1 or RET, the confirmed response rate in the expansion dose (960 mg) cohort for AMG510 in LUAD patients was about 35%. Even lower, the expansion cohort of colorectal cancer (CORE) patients had a response rate of 12% [ 15 ]. Taken together, with the emergence of well-tolerated and potent KRAS G12C inhibitors, there is an urgent need for a biomarker signature that can select patients most likely to benefit from these treatments beyond the presence of an activating KRAS mutation. While previous groups have developed models to determine RAS activation and dependency in specific cancer types, there is no model that has identified a gene signature across all cancer types with predictive capabilities of KRAS dependency beyond the RAS pathway [ 16 , 21 , 22 ]. By integrating several publicly available datasets that couple KRAS dependencies, genomic features, and tumor-specific transcriptional profiles, we have developed an integrated model (henceforth the K20 model) that improves prediction of KRAS dependency beyond the presence of an activating KRAS mutation.

Nearly all KRAS mutations are concentrated at three codons: glycine-12 (G12), glycine-13 (G13), and glutamine-61 (Q61) [ 4 ]. Recent advances in medicinal chemistry have identified a binding pocket in the glycine-to-cysteine missense mutant KRAS protein at amino acid 12 (G12C), which comprises approximately 12% of all KRAS cancer mutations and accounts for a substantial fraction of KRAS mutations in non-small cell lung cancer (13%), and to a lesser degree in colorectal (4%) and pancreatic cancers (1–3%) [ 4 , 8 – 11 ]. Tool compounds developed by Ostrem and colleagues covalently bind to the mutant cysteine and extend into the binding pocket primarily containing the switch II region (S-IIP), generating a selective response in mutant cells by repressing signaling [ 12 ]. Similar compounds discovered by additional groups have led to the rapid development of direct KRAS G12C inhibitors, the most clinically advanced of which are MRTX849 (adagrasib) and AMG510 (sotorasib) [ 13 – 15 ], both of which are now FDA-approved for use in lung cancer.

Of the three major RAS-family isoforms, mutated KRAS comprises 84% of all RAS-driven diseases and propagates many aggressive tumor types, including lung, colorectal, and pancreatic cancer [ 4 ]. The role of KRAS in cancer progression is well-studied, but until recently the protein has largely been considered “undruggable”. Due to the structure and surface topology of the GTPase, traditional small-molecule inhibitors that can directly antagonize the protein’s function have been widely regarded as untenable. Attempts to block factors involved in MAP kinase signaling (such as MEK and Raf) or binding partners to KRAS have shown some clinical promise in specific cancer types but have presented challenges with toxicity and eventual treatment resistance even in combination with other treatment options [ 5 – 7 ].

Dysregulation of the RAS family of GTPases is responsible for driving nearly 30% of all cancer types (Catalogue of Somatic Mutations in Cancer [COSMIC] v92). Discovered nearly four decades ago as the oncogenes NRAS, KRAS, and HRAS, the RAS family has expanded to include approximately 150 members involved in important cellular processes like cell division, differentiation, migration, and apoptosis [ 1 ]. In healthy cells, membrane-bound RAS-family proteins remain inactive while bound to GDP until stimulated by extracellular signals, which will cause the formation of an intermediate complex with GTP and initiate several downstream signaling cascades. However, missense mutations in RAS proteins render them constitutively active in the GTP-bound state and have been demonstrated to promote nearly all “hallmarks of cancer” [ 2 , 3 ].

Results

An integrated dataset for modeling KRAS dependency We used DEMETER2 as our KRAS cancer dependency dataset as it combined three large-scale RNAi screen datasets and integrated them with model-based normalization [23]. This model system estimates gene dependency on an absolute scale with a score of zero representing no dependency, higher/positive scores representing resistance, and lower/negative scores representing sensitivity (i.e. impaired cell growth upon gene silencing). This dataset allowed us to use RNAi screen datasets from Project Achilles [24], Project DRIVE [25] and a smaller breast cell line cohort [26] and integrate it with publicly-available multi-omic datasets from the Cancer Cell Line Encyclopedia (CCLE) [27]. There were 712 cell lines in total from DEMETER2. As expected, the KRAS dependency scores, which represent absolute KRAS dependency, were significantly lower in KRAS-mutant cells compared to KRAS wild-type (WT) cell lines (Panels A and B in S1 Fig). Using K-means clustering with k = 3, the 712 cell lines were divided into sensitive (n = 29), intermediate (n = 112) and refractory (n = 571) clusters (Panel C in S1 Fig). Within the 126 KRAS-mutant cell lines, there was enrichment of sensitive and intermediate cell lines compared to KRAS wild-type cell lines, which were more refractory (Fisher-exact p-value = 6.2e-55, Panel D in S1 Fig). Among the 126 KRAS-mutant cell lines, lung, pancreas and colorectal were the top three diseases with the highest representation (Panel E in S1 Fig). Because the cell types and expression profiles of the hematologic (e.g. leukemia, multiple myeloma and lymphoma) and non-carcinoma solid tumor cell lines (e.g. CNS tumors) were very different from carcinoma cell lines, we excluded those and cell lines of rare diseases from further analysis. In total, there were 444 carcinoma cell lines with expression, exon mutation and dependency data, of which 106 had KRAS mutations. Fig 1A shows all of the 444 cell lines sorted by KRAS dependency score with color bars to indicate activating mutations in KRAS and other genes in the Ras/Raf pathway. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 1. KRAS dependency in solid cancer cell lines and performance of the K20 model. Cell lines in the training (N = 298) and validation (N = 146) sets had consensus with DEMETER2 KRAS dependency scores. The 20 predictive features in the model were identified from CCLE gene expression and mutation data. (A) Waterfall plot showing KRAS dependency scores of all carcinoma cancer cell lines from DEMETER2 with color bars indicating cell lines with activating mutations in KRAS, NRAS, HRAS, and BRAF. (B) Receiver operating characteristic (ROC) curve comparing the K20 model performance (in red) with the KRAS mutation only model (in blue). (C) Bar plot of K20 model coefficients. (D) Heatmap showing gene expression of 19 classifier genes and mutation status of KRAS in the combined training (in brown) and validation sets (in sky blue) for all 444 cell lines. Entrez ID is indicated in parentheses next to each gene. Red represents higher expression and blue represents lower expression. https://doi.org/10.1371/journal.pcbi.1011095.g001

Development of a KRAS dependency classifier To model “clinical benefit”, which we deemed reflective of tumors either shrinking or becoming cytostatic upon direct KRAS targeting, we grouped cell lines in the “sensitive” (n = 24) and “intermediate” (n = 83) clusters into a “non-refractory” class. We aimed to build a prediction model that would classify cell lines into binary groups (non-refractory vs. refractory). The 444 solid cancer cell lines were randomly divided into training (67%, 72 KRAS-mutant and 226 KRAS WT) and validation (33%, 34 KRAS-mutant and 112 KRAS WT) sets, balancing for tissue types and KRAS dependency clusters. Monte Carlo cross validation via ElasticNet [28] was used to develop and compare the prediction accuracy of KRAS dependency among different features in the training set and to pick the best tuning parameters. We compared five different feature sets to see which one had better KRAS dependency prediction: [1] gene expression only, [2] gene expression and gene level mutation status (binary), [3] gene expression and KRAS mutation status (binary), [4] gene expression and specific KRAS mutation type and [5] gene expression, KRAS mutation status and the interaction between them (Panel A in S2 Fig). For gene expression, in order to select genes that are differentially expressed in tumors and avoid genes that are cell-line specific, we only included genes whose expression had variations (standard deviation >0.5 in log2 normalized gene expression) in the lung adenocarcinoma (LUAD), pancreatic adenocarcinoma (PDAC), and colorectal cancer (CORE) mRNA datasets from The Cancer Genome Atlas (TCGA) [22,29–31]. The feature set 5 with “KRAS mutation interaction” had the highest AUC at 0.94 (shown as a black dotted line in S2A Fig) when setting the α at 0.3 and the number of features at 80, while the feature set 3 with “gene expression and KRAS mutation status” had the second highest AUC at 0.93 when using 20 features at α = 0.9 (Panel A in S2 Fig). We selected the feature set with “gene expression and KRAS mutation status” as our final model because of the minimal performance difference (i.e., AUC = 0.94 vs. 0.93) and the smaller number of features required. We named this final model as “K20”, which includes the gene expression of 19 genes and the mutation status of KRAS. Panel B in S2 Fig shows the AUC comparison among all feature sets with a small number of features. Feature set 4 (“gene expression with specific KRAS mutation type”) performed similarly to feature set 3 and had only subtle feature differences. When applying the K20 model to the whole dataset, it had an AUC at 0.96, which is an improvement upon the KRAS mutation status only model (AUC = 0.85) (Fig 1B). The performance improvement was seen in the validation set as well (Panel C in S2 Fig). Fig 1C shows the weight of all 20 features included in the final model. While KRAS mutation status is a binary variable (KRAS wild-type = 0, KRAS-mutant = 1), the rest of the features are based on gene expression (RPKM on log2 scale). Importantly, aside from KRAS mutational status, we found that KRAS expression itself played the strongest role in predicting KRAS dependency (Fig 1C). Indeed, for KRAS WT cell lines, higher KRAS expression was significantly associated with increased KRAS dependency, and we see a similar trend for KRAS-mutant cell lines (Panel C in S3 Fig). High KRAS expression is oncogenic by nature [32] and generally correlates with poor prognosis, and some clinical studies have assessed the differential impact of KRAS expression over KRAS mutation in certain cancers. Increased KRAS amplification, as opposed to KRAS mutations, can lead to increased metastatic endometrial disease [33]. In colorectal cancer, independent of the presence of KRAS mutations, KRAS copy number gain and amplification can also be a negative predictor to anti-EGFR treatment [34,35]. Previous modeling of KRAS dependency found that KRAS gene copy number and protein expression were correlated with an increased dependency on the oncogene [16]. Fig 1D shows a heatmap of all 20 features in the K20 prediction model. Cell lines were sorted by DEMETER2 KRAS dependency scores. We found that higher expression in CCRL2, NR1l2, CCDC170, ANO1, SDR16C5 and KRAS occur more frequently in KRAS-dependent cell lines, while expression in genes such as SIRPA, MRAS, LRP12, EVC and APBA2 are higher in KRAS-independent cell lines. Several genes, such as STK38L and SDR16C5, were previously shown to correlate with oncogenic KRAS [36,37], and CCRL2 expression on cells within the TME is known to suppress KRAS-mediated tumor progression [38].

[END]
---
[1] Url: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011095

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.