Concurrent genetic and non-genetic resistance mechanisms to KRAS inhibition in colorectal cancer

SUMMARY

KRAS is mutationally activated in 45%–50% of colorectal cancer (CRC) cases, and while KRAS-targeted therapies have shown clinical promise, drug resistance limits their efficacy. To explore the mechanisms underlying KRAS inhibitor resistance, we use targeted exome sequencing and spatial transcriptomics on patient-matched CRC biopsies following combined treatment with KRAS^G12C and EGFR inhibitors. We show that acquired genetic events are identified in most patients at progression but are often subclonal and coexist with transcriptional adaptive states. Mesenchymal, YAP, and fetal-like transcriptional signatures predominate in resistant tumors, while inflammatory programs are induced early on treatment. Single-cell spatial analysis reveals intratumoral heterogeneity, with diverse adaptive states in different zones of individual tumors. Using human and murine organoid models, we show that drug-induced inflammatory programs are, at least in part, cancer-cell autonomous, and precede the emergence of drug resistance. We identify TBK1 as a target to abrogate the inflammatory adaptive phase and enhance responses to KRAS inhibition.

Graphical Abstract

In brief

Combining exome sequencing and spatial transcriptomic profiling across patient-matched CRC biopsies, Alonso et al. demonstrate that genetic and non-genetic resistance co-exist in KRAS/EGFR inhibitor-refractory tumors. KRAS inhibition-induced TBK1 activation represents a treatment-specific vulnerability that can be exploited therapeutically in combination with KRAS/EGFR inhibition.

INTRODUCTION

Colorectal cancer (CRC) is the second leading cause of cancer-related mortality worldwide, accounting for nearly 1 million deaths annually.^1,2 Kirsten rat sarcoma viral oncogene homolog (KRAS) is the most commonly mutated oncogene in CRC, observed in up to 50% of cases.³ The past decade has seen the emergence of multiple selective and non-selective small-molecule inhibitors targeting RAS proteins.^4,5 The first-in-class KRAS inhibitors, adagrasib and sotorasib, have demonstrated clinical benefit in patients with various cancer types harboring KRAS^G12C mutations.^6-9 In addition to these G12C-selective small molecules, there is now a range of additional compounds in pre-clinical and clinical development, including G12D-selective and pan-RAS inhibitors with the potential to target the majority of KRAS variants.^10-13 Thus, over the next decade, it is likely that millions of cancer patients will be treated with KRAS inhibitors.

The pivotal KRYSTAL-1 and CodeBreak 300 trials, which led to the approval of adagrasib and sotorasib, respectively, showed that only a fraction of patients with KRAS^G12C CRC respond to treatment.^6,7 Moreover, objective responses are often shortlived, with a median progression-free survival of less than 7 months.^6,7 Analysis of post-progression cell-free DNA (cfDNA) and tumor samples show that secondary RAS pathway mutations frequently emerge during treatment.^14-17 However, a subset of resistant tumors lack a clear genetic mechanism of escape. Furthermore, even when secondary mutations emerge, they are often present at low allelic frequencies and, at times, disappear under continued therapeutic pressure, indicating low clonal fitness.^14-16 Thus, while subclonal events may enable a subset of cells to escape treatment, they are unlikely to drive the bulk of resistant outgrowth.

Understanding the mechanisms underlying both primary and acquired resistance to KRAS-targeted agents is crucial for developing combination therapies that achieve deeper and more durable responses. To this end, we analyzed patient-matched clinical samples from KRAS-inhibitor trials using Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets (MSK-IMPACT) targeted gene sequencing and single-cell spatial transcriptomics. Additionally, we developed resistant organoid models to functionally interrogate tumor adaptations to KRAS inhibition. Our findings reveal that both genetic and non-genetic resistance mechanisms coexist within individual patients, exhibiting significant intratumoral heterogeneity, with diverse adaptive transcriptional programs predominating in different tumor zones. Furthermore, we show that the initial acute phase of KRAS inhibition is characterized by cell-intrinsic inflammatory signaling and identify a therapeutic strategy combining KRAS plus TANK-binding kinase 1 (TBK1) inhibition to mitigate these early adaptations and enhance treatment responses.

RESULTS

Genetic and non-genetic adaptation to KRAS inhibition in CRC

To identify genetic and non-genetic responses to KRAS inhibition, we prospectively collected patient-matched pre-treatment (n = 11), on-treatment (day 7 to day 21) (n = 11), and post-progression (“resistant”) (n = 8) biopsies from 12 CRC patients treated with combined KRAS^G12C plus EGFR inhibition (Figures 1A and 1B). Eleven patients underwent pre-treatment and on-treatment biopsies, while 7 patients underwent biopsies at three time points; in some instances, genomic profiling was captured through cell-free or circulating tumor DNA (ctDNA) without tissue biopsy (Figure 1A). Patients received treatment in one of two clinical trials combining sotorasib plus panitumumab, or testing adagrasib with or without cetuximab^6,18; all patients had previously received fluoropyrimidine-based chemotherapy, and all patients except one received combined KRAS^G12C and EGFR inhibition. Among these patients, five showed partial responses, while seven had stable disease according to RECIST criteria (Figure 1A). The mean treatment duration was 9.4 months, which closely reflects the larger trial populations.^6,7 Tumor characteristics, treatment response, and exome sequencing data (baseline and resistant) are summarized in Figure 1A.

Figure 1. Remodeling of the tumor microenvironment in response to combined KRAS plus EGFR inhibition.

(A) Percentage change in tumor size (RECIST criteria), clinical characteristics, and oncoprint (pretreatment and post-progression) of patients with CRC treated with combined KRAS plus EGFR inhibition.

(B) Study workflow. In total, 30 pre-treatment, on-treatment, and resistant tissue biopsies were analyzed by single-cell spatial transcriptomics (CosMx).

(C) Uniform manifold approximation and projection (UMAP) of patient samples colored by cell type.

(D) Single-cell heatmap showing normalized expression of selected cell typing markers across tumor compartments.

(E) UMAP analysis of cancer cells grouped by treatment status and colored by patient ID.

(F) Percentage of cell subsets in the pre-treatment (Pre) (n = 11 patients) and on-treatment (On) (n = 11 patients) tissue biopsies. Connected dots represent matched patient samples.

Colored bars represent the mean ± SEM. p values were calculated using a two-tailed Wilcoxon rank-sum test.

(G) Representative pre-treatment and on-treatment patient-matched samples.

Left: immunofluorescence images, middle: cell typing inferred from transcriptomic data.

Dots represent individual cells. Right: spatial localization of cell typing transcripts.

Dots represent individual transcripts. See also Figures S1, S2 and Table S1.

Targeted exome sequencing using MSK-IMPACT showed acquired genetic events in seven out of the eleven patients for whom pre-treatment and resistant samples were available (including solid tumor or liquid biopsies) (Figures 1A and S1). Six patients had one or more acquired mutations in known regulators of RAS signaling, including KRAS^G12C and FLT3 amplifications, and missense mutations in EGFR, FLT3, DUSP4, and ROS1. The majority of putative resistance drivers were subclonal; of the 17 de novo mutations detected in resistant biopsies, only seven were clonal (defined as a variant allele frequency >50% relative to preexisting clonal mutations) (Figure S1). There were no identifiable acquired genetic events in the remaining four patients. Thus, like what was previously reported in CRC and other cancer types following progression on KRAS inhibitors, escape mutations were not universally identified and were often subclonal.^14-16

To identify phenotypic tumor adaptations, we performed spatial transcriptomics of matched pre-treatment, on-treatment, and resistant biopsies (Figure 1B). We employed a cyclic fluorescent in situ hybridization platform (CosMx) that quantifies the relative abundance of 950 transcripts at single-cell resolution directly on formalin-fixed paraffin-embedded (FFPE) tissue.¹⁹ A total of 679,588 cells passed quality control across all patient samples, including 375,985 cancer cells. Unsupervised clustering of single-cell profiles identified 14 cell subsets, which were annotated by known gene expression markers (Figures 1C, 1D, S2A, and S2B). Cancer cells were annotated based on their high expression of epithelial/CRC markers (KRT8, KRT18, KRT19, EPCAM, and CLDN4) that were notably distinct from other epithelial populations in metastatic sites, such as hepatocytes (Figures 1D, S2A, and S2B). Malignant cells showed significant heterogeneity, and generally clustered by patient and treatment status (Figure 1E). Pre-treatment and resistant samples were transcriptionally more diverse than on-treatment samples, which generally clustered together and showed decreased inter-tumoral heterogeneity (Figures 1E and S2C). In contrast, intra-tumoral transcriptional heterogeneity was similar across treatment conditions (Figures 1E and S2C). As expected, analysis of cell type abundance showed a significant decrease in the number of malignant cells early on-treatment relative to the pre-treatment time points (Figures 1F, 1G, and S2D-2G). We also noted an increased proportion of cancer-associated fibroblasts (CAFs), B cells, and endothelial cells in the on-treatment biopsies (Figures 1F, 1G, and S2D), though the difference only remained statistically significant for B cells when cancer cells were excluded from the analysis (Figure S2E).

Given the limited sensitivity of the CosMx platform for resolving specific immune populations, we complemented the spatial transcriptomic analysis with multiplex immunofluorescence (mIF) in a subset of samples with sufficient remaining tissue (Figures S2F and S2G). The relative abundance of epithelial and immune subpopulations was quantified using antibodies against panCK, CD4, FOXP3, CD8, CD68, and CD163. This analysis revealed a significant enrichment of CD4⁺ T cells in on-treatment samples, a population that was not robustly captured by the CosMx assay (Figures S2F-S2G). Other immune and stromal populations did not show statistically significant differences across time points.

To identify adaptive responses to RAS inhibition specifically in tumor cells, we examined differentially expressed genes and programs in the malignant compartment of patient-matched biopsies. Over-representation analysis (ORA) and gene set variation analysis (GSVA) across Hallmark and curated intestinal cell-type specific gene set collections (Table S1) showed activation of intestinal Paneth and stem cell signatures, as well as multiple inflammatory programs in on-treatment and resistant samples (Figures 2A and 2B). Annotation of individual genes that were significantly differentially expressed (log2 fold-change >0.3, adjusted p < 0.05) in at least two-thirds of the samples again showed a marked enrichment of interferon (IFN) and cytokine signaling in on-treatment samples (Figure 2C). In contrast, therapy-resistant samples showed an enrichment of epithelial to mesenchymal transition (EMT), fetal-like and Yes-associated protein 1 (YAP) programs (Figures 2B-2D, 3A, S3, and S4) that we and others have previously linked to CRC metastasis and WNT-targeted therapy resistance.^20-27 There was also a marked downregulation of epithelial and intestinal differentiation markers, and a corresponding increase in intestinal stem cell markers under KRAS inhibitor treatment but in most cases, these signatures returned to baseline in resistant samples (Figure 2B). As expected, the expression of KRAS targets was lower on-treatment compared to baseline and subsequently increased at progression in a subset of patients (Figures 2B-2D, 3A, and S3), likely due to withdrawal from KRAS/EGFR therapy.

Figure 2. Transcriptomic adaptations to KRAS inhibition in the cancer cell compartment.

(A) ORA bubble plot of Hallmark and curated intestinal cell-type specific gene sets showing the ten most significantly enriched pathways in the on-treatment versus pre-treatment, and resistant versus pre-treatment patient samples. Differentially expressed gene sets were identified using a Wilcoxon rank-sum test. Adjusted p values were calculated using Fisher’s exact test with Benjamini-Hochberg correction for multiple comparisons.

(B) Heatmap displaying module score quantiles of Hallmark and intestinal cell type-specific gene sets across individual patient samples.

(C) Cancer cell differential gene expression in the on-treatment and resistant (versus pre-treatment) patient-matched samples. Genes for which log2 FC > 0.3 or < −0.3 in at least two-thirds of samples are displayed. Colored squares represent the signaling pathways associated with each gene.

(D) Representative case (patient 6) showing immunofluorescence and spatial images.

The gradient scale shows the module score quantiles of transcriptional programs in cancer cells (non-malignant cells were removed from the analysis). See also Figures S3, S4, and Table S1.

Figure 3. Co-existing genetic and non-genetic resistance mechanisms to KRAS inhibition.

(A) Oncoprint of the pre-treatment and post-progression biopsies from patient 5.

(B) Immunofluorescence and spatial images of the pre-treatment and resistant biopsies from patient 5. The gradient scale shows module score quantiles of selected gene sets in cancer cells. Far-right show transcripts for RPL21 (endogenous control) and KRAS (each dot represents an individual transcript).

(C) Oncoprint of the pre-treatment and post-progression biopsies from patient 4.

(D) Immunofluorescence and spatial images of patient 4 observed at low magnification. Two areas in the post-progression biopsy with seemingly divergent resistance mechanisms are highlighted.

(E) Area 1 and area 2 from (D) are displayed at higher magnification. The gradient scale shows module score quantiles of selected gene sets in cancer cells.

(F) Volcano plot showing differentially expressed genes in area 1 versus area 2.

Differentially expressed genes corresponding to IFN, YAP, fetal, and RAS programs are highlighted. p values were calculated using a two-tailed Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple comparisons. See also Figures S3-S5 and Table S1.

A notable outlier in our data was patient 4, who, in contrast to other cases, showed robust expression of type 1 and type 2 IFN programs in the pre-treatment biopsy (Figure S3). Importantly, before the on-treatment biopsy, the patient started treatment for a flare of rheumatoid arthritis with the anti-inflammatory drug hydroxychloroquine and following the on-treatment biopsy, received rituximab during the study treatment.

Coexisting genetic and non-genetic resistance to KRAS inhibition

The presence of both genetic and non-genetic transcriptional changes in progression samples raised the question of whether these changes were coincident in individual tumors. In fact, transcriptional reprogramming was identified in each of the eight resistant biopsies analyzed, including those with multiple putative genetic resistance alterations (Figures 3A, 3B, and S3-S5). Resistant tumor exome sequencing of patient 5, for instance, showed amplification of KRAS^G12C (5.9-fold copy number gain), FLT3 and MDM2, as well as acquired missense mutations in DUSP4, ATM, AURKA, SERPINB3, and STAG3 (Figure 3A). In addition to the expected upregulation of KRAS transcripts and downstream reactivation of mitogen-activated protein kinase (MAPK) signaling, likely driven by the KRAS^G12C amplification and other RAS pathway mutations, transcriptomic analysis revealed strong enrichment of YAP and fetal intestinal signatures at progression (Figure 3B), revealing that even in cases with described genetic resistance profiles,^14-17 transcriptional and lineage adaptations can occur.

In some cases, MSK-IMPACT sequencing and transcriptomic data revealed co-existing genetic and non-genetic resistance mechanisms but the spatial analysis revealed significant intratumoral heterogeneity, with enrichment of distinct adaptive programs in different zones of individual tumors (Figures 3C-3F and S3-S5). Patient 4, for instance, progressed after 23.5 months of treatment with sotorasib plus panitumumab, and MSK-IMPACT sequencing showed no genetic events predicted to activate MAPK signaling (Figure 3C). Spatial transcriptomics revealed two distinctive areas with seemingly mutually exclusive transcriptional profiles. Area 1 showed strong enrichment of type 1 IFN, YAP, and fetal intestinal programs but low MAPK signaling. In contrast, area 2 showed robust MAPK reactivation but suppressed type 1 IFN, YAP, and fetal markers (Figures 3D and 3E). We observed a similar phenomenon in progression biopsies from patient 1, showing isolated “YAP-high, MAPK-low” tumor regions juxtaposed to “YAP-low, MAPK-high” areas (Figure S5). These observations highlight the possibility that independent resistance profiles may evolve in parallel within the same tumor and may explain the frequent emergence of multiple subclonal genetic events at progression captured through cfDNA profiling.^14-16

Cancer cell-autonomous activation of inflammatory programs following KRAS inhibition

We next asked whether transcriptional reprogramming is dependent on the tumor microenvironment. Pathway and differential gene expression analyses of the non-malignant compartment revealed broad inflammatory signatures across CAFs, myeloid, and lymphoid cells in both on-treatment and resistant samples (Figures S6 and S7). Overrepresentation analysis demonstrated significant enrichment of cytokine, IFN, JAK-STAT, and immune-related pathways across multiple compartments. Notably, CXCL12, a chemokine previously linked to EMT and therapy resistance,²⁸ was among the most highly upregulated genes in CAFs and monocytes at the on-treatment and resistant time points (Figures S6 and S7).

To investigate potential cell-cell communications, we performed ligand-receptor interaction analysis using CellChat,²⁹ which infers intercellular signaling by estimating the probability of interaction based on the expression levels of ligands in “sender cells” and corresponding receptors in “receiver cells.” To gain insights into temporal and state-specific ligand-receptor interactions, we performed independent CellChat analyses of pre-treatment, on-treatment, and progression samples, as well as IFN-high, YAP-high, and RAS-high biopsy regions (Figure S8). These analyses revealed some statistically significant, albeit weak, ligand-receptor interactions, including from myeloid and lymphocytes to cancer cells such as potential TGFB2-TGFBR2 signaling from T cells to tumors cells specifically in the on-treatment context (Figure S8). Interestingly, we previously identified TGF-β signaling as a driver of lineage change and drug resistance in intestinal cancer cells,²⁰ suggesting such interactions may be functionally important in this context. Still, overall, these data suggest that direct ligand-receptor signaling from immune to cancer cells may be limited. Of course, CellChat analysis is limited to only those ligand-receptor pairs included in the 950-gene CosMx panel; we cannot exclude the involvement of signaling by genes/proteins not represented in this collection.

To directly measure tumor cell responses to KRAS inhibition in the presence or absence of a complete tumor microenvironment, we generated C57Bl/6 colon organoids carrying an Apc truncation (APC^Q1405X), an endogenous Kras mutation (KRAS^G12C), and a Trp53 truncation (p53^Q97X) (hereafter AKP-G12C). Organoids engrafted in both immune competent (C57Bl/6) and immune compromised (NSG) mice showed a marked response to treatment with adagrasib and over 4 weeks showed accelerated outgrowth in the presence of drug. We observed no significant difference in the outgrowth of tumors in C57Bl/6 or NSG mice (Figures S9A and S9B), suggesting that at least in the murine model, adagrasib response and resistance is not dependent on signaling from immune cell types, though we cannot rule out contribution from non-immune populations. To explore resistance in the absence of all other cell types, we treated AKP-G12C organoids with adagrasib alone or the combination of adagrasib and afatinib, mimicking the combined KRAS/EGFRi treatment used clinically. AKP-G12C organoid cultures showed a dose-dependent decrease in ERK phosphorylation (Figure 4A) and growth arrest following KRAS inhibition (adagrasib) or combined KRAS/EGFR inhibition (adagrasib/afatinib) (Figures 4B-4D). Continued treatment gave rise to highly resistant organoids after 10–12 weeks of treatment (Figures 4B-4F), showing an EC50 shift of 25-fold to 100-fold compared to treatment-naive organoids (Figures 4C and 4D). Adagrasib-resistant organoids were cross-resistant to the KRAS^G12C inhibitor sotorasib, and the recently described multi-selective RAS(ON) inhibitor RMC-7977 (Figures S9C and S9D).

Figure 4. Transcriptomic adaptation to KRAS inhibition is cancer-cell autonomous.

(A) Western blot of active and total ERK in parental AKP-G12C organoids treated with adagrasib (0–1,000 nM) for 24 h.

(B) Representative bright field images of parental and KRAS-inhibitor resistant AKP-G12C organoids treated with increasing concentrations of adagrasib. Images were obtained on day 3 of treatment.

(C) Growth of parental and adagrasib-resistant AKP-G12C organoids treated with increasing concentrations of adagrasib. Organoid growth was normalized to untreated conditions (n = 3 experiments).

(D) Growth of parental and adagrasib/afatinib-resistant AKP-G12C organoids treated with increasing concentrations of adagrasib and afatinib. Organoid growth was normalized to untreated conditions (n = 3 experiments).

(E) Growth of parental and adagrasib-resistant AKP-G12C organoids (low KRAS copy number) treated with the EC80 of adagrasib (100 nM) (n = 3 experiments).

(F) Growth of parental and adagrasib/adafatinib-resistant AKP-G12C organoids treated with the EC80 of adagrasib and afatinib (adagrasib 50 nM and afatinib 5 nM) (n = 3 experiments).

(G) KRAS copy number (fold-change) of parental and KRAS-inhibitor resistant AKP-G12C organoids (n = 3 technical replicates).

(H) Over-representation analysis (ORA) bubble plot showing enrichment of Hallmark and curated intestinal gene sets in parental AKP organoids treated with KRAS inhibition (adagrasib 100 nM), or KRAS plus EGFR inhibition (adagrasib 50 nM, afatinib 5 nM). Adjusted p values were calculated using Fisher’s exact test with Benjamini-Hochberg correction for multiple comparisons. (n = 5 independent replicates per condition).

(I) ORA bubble plots showing enrichment of gene sets in resistant versus parental AKP organoids. Adjusted p values were calculated using Fisher’s exact test with Benjamini-Hochberg correction for multiple comparisons. (n = 5 independent replicates per condition).

(J) Western blot of non-phospho and total Yap1 in parental and resistant AKP-G12C organoids treated with DMSO or adagrasib plus afatinib for 24 h.

(K) Volcano plot showing differentially expressed genes in resistant versus parental AKP-G12C organoids. Yap targets are highlighted in blue. p values were calculated using a two-tailed Wilcoxon rank-sum test with Benjamini-Hochberg correction. (n = 5 independent replicates per condition). For organoid experiments (C–G), data represent the mean ± SEM. p values were calculated using a two-tailed t test (C–G). **, p < 0.01. See also Figures S6-S9.

To characterize the response to KRAS inhibition, we performed RNA sequencing of AKP-G12C organoids prior to, during treatment (adagrasib alone (100 nM) or adagrasib/afatinib (50/5 nM) for 72 h), and in resistant cells, mirroring the contexts of patient biopsies. In total, we generated five independent adagrasib-resistant and five adagrasib/afatinib-resistant AKP-G12C lines. A subset of drug-resistant organoid lines exhibited high copy number Kras amplifications (defined as ≥4 copies) (Figure 4G). Interestingly, Kras-amplified organoids were addicted to KRAS inhibition, as drug withdrawal led to a decrease in organoid growth (Figure S9E). These findings are reminiscent of the senescent phenotype observed upon drug withdrawal in KRAS^G12C-amplified human CRC.¹⁶ Transcriptome analysis of adagrasib and adagrasib/afatinib-treated drug-naive organoids revealed enrichment of intestinal stem cell signatures, as recently reported,³⁰ as well as inflammatory programs, including type 1 and type 2 IFN that were reduced in resistant organoids (Figures 4H and 4I; Table S2). These findings mirrored the data from patient biopsies and indicate that inflammatory programs are engaged early following KRAS inhibition^20-25 and not strictly dependent on non-malignant cells within the tumor microenvironment. YAP and fetal-like signatures dominated the transcriptional profile of resistant organoids, independent of KRAS copy number alterations, demonstrating coexistence of genetic and non-genetic changes, as observed in clinical cases (Figures 4H-4K, S9F, and S9G). Indeed, there was a significant positive correlation between gene set enrichment observed in both on-treatment and resistant AKP organoids compared to patient samples (Figure S9H). Consistent with the increase in YAP transcriptional output, resistant organoids showed an elevation in active (non-phosphorylated) YAP compared to parental organoids (Figure 4J); there was no evidence of elevation of non-phospho-YAP following acute adagrasib/afatinib treatment (Figure 4J).

TBK1 inhibition blocks early inflammatory reprogramming and synergizes with KRAS inhibition

Our group and others have shown that inflammatory programs are engaged during the initial response of cancer cells to targeted and cytotoxic therapies.^20-24,31 To test whether blocking of inflammatory signaling would abrogate initial adaptive responses and synergize with KRAS inhibition in controlling tumor cell growth, we conducted a small, focused drug screen in patient-derived, KRAS^G12C-mutant CRC organoids (Patient-derived organoids or PDOs), combining adagrasib with kinase inhibitors targeting inflammatory pathways. We evaluated two concentrations (0.5 and 1 μM) of small molecules targeting eight kinases: PI3K, FAK, KIT, JNK, IKK, TGFβRI, JAK, and TBK1 (Figures 5A and S10A). Of these, only the TANK-binding kinase 1 (TBK1) inhibitor, TBK1/IKKε-in-5, delayed outgrowth in the presence of adagrasib (Figures 5A and S10A), while TBK1 inhibition alone had no significant effect on organoid growth. Similar results were seen with the FDA-approved dual TBK1/JAK inhibitor momelotinib (Figure 5B). Consistent with a role for TBK1 in the cellular response to KRAS inhibition, we observed a dose-dependent increase in TBK1 phosphorylation following treatment of AKP-G12C organoids with adagrasib for 24–72 h (Figure 5C). The same dose-dependent pTBK1 increase was also observed in human KRAS^G13D-mutant HCT116 CRC cells treated with the multi-selective RAS inhibitor RMC-7977 (Figure S10B).

Figure 5. Targeting early inflammatory reprogramming via TBK1 blockage synergizes with KRAS inhibition.

(A) Bar plot showing focused drug screen results of PM1050 PDOs treated with adagrasib alone (50 nM), kinase inhibitors targeting inflammatory programs, or the combination. Two concentrations of each kinase inhibitor were tested. Bar values represent the organoid area fold change during the 13 days experiment (n = 3 experiments). The x axis shows the concentration of the kinase inhibitor (μM).

(B) Growth curve and bright field images of PM1050 organoids treated with KRASi (adagrasib 50 nM), TBK1i (TBK1/IKK-in-5 1 μM, momelotinib 1 μM), or the combination (n = 3 experiments).

(C) Western blot of active and total ERK and TBK1 in parental AKP-G12C organoids treated with adagrasib (0–1,000 nM) for 24 and 72 h.

(D) Representative case (patient 5) showing images of the pre-treatment and on-treatment samples. The gradient scale shows the module score quantiles of TBK1-IRF3 targets (only cancer cells are displayed).

(E) Volcano plot showing differentially expressed genes in parental AKP-G12C organoids treated with adagrasib (100 nM) versus DMSO. TBK1-IRF3 and IFNγ targets are highlighted. p values were calculated using a two-tailed Wilcoxon rank-sum test with Benjamini-Hochberg correction. (n = 5 independent replicates).

(F) Over-representation analysis (ORA) bubble plot showing enrichment of Hallmark and curated intestinal gene sets in parental AKP organoids treated with adagrasib (100 nM) plus momelotinib (500 nM), versus adagrasib alone. (n = 5 independent replicates).

(G) Volcano plot showing differentially expressed genes in parental AKP-G12C organoids treated with adagrasib (100 nM) and momelotinib (500 nM) versus adagrasib alone. p values were calculated using a two-tailed Wilcoxon rank-sum test with Benjamini-Hochberg correction. (n = 5 independent replicates).

(H) Western blot of AKP-G12C organoids transduced with dox-inducible shRNAs targeting Tbk1 (shTbk1).

(I) Growth curve and bright field images of shTbk1 and shRen (control) AKP-G12C organoids treated with adagrasib (100 nM), doxycycline (1 μg/mL), or the combination (n = 3 experiments).

(J) Growth curve of AKP-G12C organoids treated with adagrasib, afatinib, momelotinib, or the combination (n = 4 experiments).

(K) Day 3 and day 5 Cell Titre Glo (CTG) luminescence of AKP-G12C organoids treated with adagrasib, afatinib, momelotinib, or the combination (n = 3 experiments). p values for organoids and cell growth experiments (A–B) and (I–K) were calculated using a two-tailed t test. Data represent the mean ± SEM (A–B) and (I–K). Bright field images show organoid density on the final day of the experiment. **, p < 0.01, *, p < 0.05. See also Figure S10.

TBK1 and the related kinase IKKε signal through interferon regulatory factor 3 (IRF3), a key mediator of the innate antiviral immune response.^21,32 Indeed, IRF3 phosphorylation was elevated in RMC-7977-treated HCT116 cells, coincident with increases in pTBK1 (Figure S10B). Analysis of patient-matched samples revealed that IRF3 target genes included in the CosMx panel were upregulated following combined KRAS and EGFR inhibition (Figures 5D and S10C). Similarly, differential gene expression of AKP-G12C organoids treated with adagrasib showed robust enrichment of IRF3 targets (Figures 5E and S10D), while co-treatment with momelotinib showed dramatic downregulation of inflammatory signatures with IRF3 targets among the most significantly suppressed (Figures 5F and 5G). Together, these data support the notion that TBK1-IRF3 signaling is triggered in response to KRAS inhibition. To ensure the combined effect of adagrasib and TBK1 inhibition was not specific to a single PDO line, we treated AKP murine organoids carrying either KRAS^G12C (Figure S10E) or KRAS^G12D (Figure S10F) mutations as well as a second KRAS^G12C-mutant CRC PDO (Figure S10G) and HCT116 CRC cells (Figure S10H). In each case, TBK1 blockade had minimal impact as a monotherapy but stunted cell proliferation in combination with KRAS inhibition. Both TBK1/IKKε-in-5 and momelotinib inhibit related kinases in addition to TBK1; however, silencing of only TBK1 using two independent doxycycline-inducible short hairpin RNAs (shRNAs) mimicked the effect of drug treatment (Figures 5H and 5I), implying that TBK1 inhibition is sufficient to enhance the activity of adagrasib treatment on drug-naive organoids. Importantly, momelotinib also improved the response to the current CRC standard of care combination therapy, KRAS/EGFR inhibition (Figures 5J and 5K). Together, these data highlight that the acute induction of inflammatory programs following KRAS inhibition, at least in part through TBK1, may play a role in the emergence of drug-resistant tumor outgrowth.

DISCUSSION

Adagrasib and sotorasib have demonstrated significant clinical benefit in treating advanced epithelial cancers, and several next-generation RAS inhibitors are now in late-stage clinical development, offering the potential to significantly improve outcomes for millions of cancer patients in the coming years. However, as with other targeted therapies, primary or acquired resistance is likely to occur in virtually all patients. Understanding the mechanisms underlying drug escape is essential for developing strategies to delay, or even prevent, treatment failure. Recent studies show that KRAS inhibitor-resistant tumors frequently acquire de novo mutations predicted to reactivate RAS signaling.^14-17 In agreement with these studies, our analysis identified acquired RAS pathway mutations in the majority of progression samples; however, as in previous examples,^14-17 most mutations were subclonal, raising the question of whether they are bona fide drivers of resistant outgrowth.

By integrating genomic and transcriptomic data from serially collected, patient-matched biopsies and cancer models, our study reveals that acquired genetic alterations often co-exist with non-genetic resistance mechanisms. We show that resistant tumors engage YAP-associated transcriptional programs reminiscent of regenerative or fetal-like intestine, which we and others have reported to be associated with metastasis and multidrug resistance.^20-27,33-38 In some tumors, these adaptive responses are engaged in parallel to MAPK pathway reactivation, likely driven by genetic events, such as KRAS^G12C amplifications.¹⁶ Notably, transcriptional programs in the resistant biopsies were more diverse compared to the pre-treatment and on-treatment samples. Indeed, we observed significant intratumoral heterogeneity at progression, with inflammatory responses, EMT, YAP signaling, fetal-like programs, and MAPK pathway reactivation predominating in different zones of individual tumors, implying that discrete regions may evolve independently via the engagement of diverse putative escape programs. In other cases, genetic and non-genetic programs co-existed within the same tumor zones and individual cells. These findings may explain the detection of multiple subclonal events in post-progression cfDNA samples, which, as opposed to tumor biopsies, capture genetic drivers across a patient’s metastatic tumors.^14-17

The analysis of on-treatment samples revealed transcriptional rewiring as early as 7 days following treatment initiation, or as early as 3 days in ex vivo organoid culture. Notably, on-treatment programs were significantly less diverse compared to resistant phenotypes and often involved inflammatory signatures previously shown to play a role in therapy resistance,^20-25,33,34 indicating they may serve as early mediators of tumor adaptation to KRAS inhibition. Focused profiling of drugs targeting kinases involved in inflammatory programs showed that combined KRAS plus TBK1 inhibition or KRAS/EGFR/TBK1 inhibition restrained the growth of patient-derived and murine organoids. This is not the first indication of a role for TBK1 inhibition in KRAS mutant cancers^39-41; however, in contrast to previous work by Barbie et al., which suggested a specific TBK1-dependency in KRAS mutant cells, our work implies that TBK1 is engaged following KRAS inhibition and that targeting both proteins enhances the anti-tumor efficacy. These results align with recent studies in EGFR-mutant non-small cell lung cancer, where EGFR inhibition activates an innate antiviral response through the TBK1-IRF3 axis, and combined TBK1 and EGFR inhibition yields synergistic effects in preclinical models.²¹

In all, our study reveals the complexity of the adaptive response to KRAS inhibition in CRC. The inter- and intra-tumoral heterogeneity in progression samples, and the coexistence of genetic and non-genetic resistance mechanisms underscores the challenge of reversing therapeutic resistance. Targeting early inflammatory adaptations, which we show to be more conserved and generalizable within and across tumors, may offer a path for deeper and more durable responses to RAS-targeted therapies in CRC.

Limitations of the study

While our study sheds light on critical adaptive processes and the significant heterogeneity of drug resistance mechanisms within individual tumors, there are limitations. Though patient-matched samples provide an excellent resource to profile disease progression under treatment, the relatively small cohort size limits our ability to correlate specific clinical features such as metastatic site or RECIST response with transcriptional adaptations. Similarly, we are unable to deconvolute the impact of specific genomic profiles on phenotypic responses. Further, due to limited sample volume from individual biopsies, the scope of transcriptional and epigenetic profiling is restricted, constraining our ability to dissect rare malignant or immune populations, and the analysis of tumor-stroma interactions. Using multiplexed immunolabeling, we did detect a significant increase in T cell infiltration in on-treatment samples, which has been observed in various pre-clinical contexts.^4,42,43 In these settings, KRAS inhibition can relieve an otherwise immunosuppressive environment and synergize with immune checkpoint blockade. In our murine model, the treatment of engrafted AKP-G12C tumors revealed no significant differences in response in the presence or absence of a functional immune system. However, it is important to note that these organoid models carry relatively low mutational burden and thus, are unlikely to trigger CD8-mediated responses as reported in pancreatic cancer models with high baseline T cell infiltration.⁴² Nonetheless, the reproducibility of key inflammatory signatures across samples and organoid models underscores the relevance of these adaptive mechanisms that are, at least in part, cell intrinsic.

Star Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Patients

All patients received treatment as part of KRAS inhibitor clinical trials approved by the MSKCC Institutional Review Board/Privacy Board under protocols 19–408 (NCT03785249) and 20–183 (NCT04185883). Pathology, radiology reports, medical oncology, and surgery notes were reviewed in the electronic medical record to annotate clinical characteristics for patients. Patient sample collection followed the appropriate Institutional Review Board/Privacy Board protocols and waivers (protocols 06–107, 12–245, 14–019). All participants provided written informed consent for both the clinical trials and biospecimen collection. This study adhered to the ethical principles outlined in the Declaration of Helsinki.

Cell lines

HEK293T cells were purchased from ATCC (CRL-3216) and maintained in Dulbecco’s Modified Eagles Medium (DMEM, Corning, 10-013-CV) containing 1% Pen/Strep (Fisher Scientific 15140163) and 10% FBS (Gibco A5670801) at 37°C with 5% CO₂. HCT116 cells were purchased from ATCC (CCL-247) and maintained in Roswell Park Memorial Institute (RPMI) 1640 with L-glutamine (Corning, 10-040-CV) supplemented with 10% FBS (Gibco A5670801), 1% Pen/Strep (Fisher Scientific 15140163), and 1% GlutaMAX (Gibco 35050061) at 37°C with 5% CO₂. Cell lines were tested for mycoplasma routinely every six months.

Animal models and mouse tumor treatments

All mouse experiments described were approved by the Institutional Animal Care and Use Committee at Weill Cornell Medicine and Memorial Sloan Kettering Cancer Center, under protocol number 2014-0038. C57Bl/6n mice were purchased from Charles River and NOD.Cg-Prkdc^scid Il2rg^tm1Wjl/SzJ (NSG) mice were purchased from the Research Animal Resources and Compliance core facility at Memorial Sloan Kettering Cancer Center. Mice were allowed to acclimatize to the facility for 1 week prior to subcutaneous transplant of AKP-G12C organoids (~200,000 cells) in sterile saline with 50% Matrigel. Tumor growth was monitored by caliper measurements and once the average size of tumors reached 100mm³, mice were randomized to vehicle and adagrasib treatment groups. Adagrasib (50mg/kg) was administered by oral gavage daily (5 days on, 2 days off) for 30 days.

Murine colon cancer organoids

Female mice (4–8 weeks old) were euthanized by inhaled carbon dioxide and sprayed down with 70% ethanol. The abdominal cavity was opened, the colon was removed and opened longitudinally with sterile scissors. Feces and excess mucus were scraped off with a slide. The colon was transferred to a 50 mL tube and washed with 20 mL of ice-cold PBS. The colon was then cut into 5–10 mm pieces, placed in 10 mL of ice-cold EDTA/PBS, and incubated on a shaker at 4°C for 60 min. Following incubation, the tube was shaken vigorously 10 times, and the supernatant was removed. The tissue was resuspended in 20 mL of ice-cold sterile PBS containing 200 U DNase I and shaken approximately 30 times to release colonic crypts. The suspension was filtered twice using 100 μm cell strainers. The filtrate was centrifuged for 5 min at 1100 RPM, and the pellet was resuspended in 400–500 μL of Matrigel and plated in 6 independent wells of a 12-well plate. Wild type organoids were cultured in advanced DMEM/F12 (Gibco, 12634028), 10 mM HEPES (Corning, 25-060-Cl), 1× GlutaMAX (Gibco, 35050061), 0.5 nM WNT Surrogate-Fc fusion protein (Immunoreactive Antibodies, N-001), 50 ng/mL EGF (Peprotech AF-100-15-500UG), 40ng/mL LDN193189 (MedChemExpress, HY-12071) and RSPO3 conditioned media (2% of final volume). Fungin (InvivoGen, ant-fn-1) and Primocin (InvivoGen, ant-pm-05) were added to the culture for the first seven days.

To generate colon cancer organoids, colon crypts derived from LSL-Kras^G12C or LSL-Kras^G12D mice were transfected with a construct containing a Trp53 sgRNA and Cre recombinase, and selected by withdrawing EGF from the media and treating with Nutlin-3 for 10 days.⁴⁴ To engineer missense Apc mutations, organoids were nucleofected with a construct containing the base editing enzyme FNLS and a synthetic gRNA targeting codon Q1405, and selected by withdrawing RSPO3 from the media.^48,49 AKP organoids were cultured in murine organoid (MO) medium containing advanced DMEM/F12, 10 mM HEPES, 1× and GlutaMAX (Gibco, 35050061). Organoid passage and cryopreservation were done as previously described.^50,51 Organoids were passaged in 300 μL of Matrigel plated in one well each of a 6-well dish (Corning 353224). DNA, protein, and RNA isolation were done as previously described.^50,51 Resistant models were generated by treating organoids with adagrasib alone (100 nM), or adagrasib plus afatinib (50 and 5 nM, respectively), until resistant organoids emerged (4–12 weeks). Organoids were passaged approximately every 3–7 days as necessary, and drugs were refreshed with each passage.

Patient-derived organoids

Patient-derived organoids were established as previously described.⁵² Patient tumor samples were washed, cut into 5 mm fragments, and incubated in a dissociation buffer containing advanced DMEM/F12 (Gibco, 12634028), 4 mM L-glutamine, 100 U/mL penicillin-streptomycin, 250 U/mL type III collagenase (Worthington, 9001-12-1), and 1 U/mL dispase (Sigma-Aldrich, D4818). The samples were shaken at 37°C for 30 min, filtered through a 70 μm cell strainer, centrifuged at 1100 RPM for 5 min, and washed with advanced DMEM/F12. Cells were counted and resuspended in Matrigel at a concentration of approximately 10,000 to 50,000 cells per 200 μL of Matrigel in 6-well plates. PDOs were maintained in human organoid (HO) medium, containing advanced DMEM/F12, 10 mM HEPES, 4 mM L-glutamine, 1mM N-Acetylcysteine (Sigma-Aldrich A9165), 10 mM nicotinamide (Sigma-Aldrich N0636), 50 ng/mL EGF (Peprotech AF-100-15), 20 ng/mL FGF-10 (Peprotech, 100-26), 10 μM Y-27632 (MedChemExpress, HY-10071), 500 nM A-83-01 (MedChemExpress, HY-10432), 10 μM SB202190 (MedChemExpress, HY-10295), 200 nM LDN193189 (MedChemExpress, HY-12071), 10 nM Gastrin-I (MedChemExpress, HY-P2671), 0.5 nM WNT Surrogate-Fc fusion protein (ImmunoPrecise Antibodies, N-001), RSPO3 conditioned media (2% of final volume), and B27 supplement (Gibco 17504044).

METHOD DETAILS

Cloning

PCR-amplified 97mer oligos were inserted into EcoRI and XhoI-digested pCF806⁵³ (Addgene 186711) by standard restriction cloning. Vectors and inserts were ligated using T4 DNA ligase (NEB M0202M) at 16°C overnight and transformed into Stbl3 chemically competent cells. Oligos for shRNA cloning are listed in Table S3. For lentiviral production, HEK293T cells were transfected with the appropriate lentiviral transfer vector along with the packaging plasmids psPax2 (Addgene, catalog no. 12260) and pMD2.G (Addgene, catalog no. 12259) using polyethylenimine. Viral supernatants were harvested at 48, 72, and 96 h post-transfection, centrifuged at 25,000 RPM and 37°C for 2 h, and stored at −80°C.

Organoid growth experiments

Mouse and human colorectal organoids were cultured in Matrigel and maintained in organoid growth media without EGF. Organoids were seeded in 48-well plates. Organoid growth was monitored in real-time using the Incucyte S3 Live-Cell Analysis System (Sartorius). Representative brightfield images were acquired every 24 to 72 h over a period of 7 days for mouse organoids and 14 days for human organoids. Organoid area was quantified using the organoid module, with automated segmentation and analysis parameters optimized for colorectal organoid morphology. For each time point, organoid area was normalized to the initial area measured on day 0. The following drugs were used: adagrasib, (MedChemExpress HY-130149), afatinib (MedChemExpress HY-10261), RMC-7977 (MedChemExpress HY-148439), TBK1/IKK-IN-5 (MedChemExpress HY-128679), momelotinib (MedChemExpress HY-10961), GSK8612 (MedChemExpress HY-111941), alpelisib (MedChemExpress, HY-15244), defactinib (MedChemExpress HY-12289), TPCA-1 (MedChemExpress HY-10074), pexidartinib (MedChemExpress HY-16749), SP600125 (MedChemExpress HY-12041), LY2157299 (MedChemExpress HY-13226). Concentrations used for each drug and treatment durations are described in detail in the results section and figures. For shRNA validation experiments, AKP-G12C organoids were transduced with lentiviral vectors encoding dox-inducible shRNAs targeting Tbk1 or Renilla. Two independent shRNAs targeting Tbk1 were tested. Organoids were cultured in the presence of doxycycline (1 μg/mL) for 7 days, with or without adagrasib (50 nM). Organoid growth was monitored in the incucyte.

Western Blot

Organoids in Matrigel were washed with ice-cold PBS and incubated in 1 mL of cold Cell Recovery Solution (Corning 354253) for ~60 min to soften Matrigel. After complete dissolution of the Matrigel, the organoid suspension was transferred to a 15 mL conical tube and 5mL of ice-cold PBS was added. Organoids were centrifuged at 800g for 5 min at 4°C. Organoid pellets were washed with 1 mL PBS and centrifuged at 800g for 5 min at 4°C. Pellets were resuspended in 30 μL of RIPA buffer, shaken at 4°C for 25 min, and centrifuged at 20,000g at 4°C for 20 min. Protein supernatant was collected and quantified using the DC Protein Assay Kit (BioRad 5000116). 30 μg of protein was loaded per Western Blot. Antibodies used for Western blot analysis were: anti-phospho-p44/42 MAPK (Erk1/2) (Thr202/Tyr204) (Cell Signaling Technology 4370), anti-p44/42 MAPK (Erk1/2) (Cell Signaling Technology 9107), anti-phospho TBK1/NAK1 (Ser172) (Cell Signaling Technology 5483), anti-TBK1 (Cell Signaling Technology 3504), anti-non-phospho (Active) Yap (Ser127) (Cell Signaling Technology 29495), anti-Yap (Cell Signaling Technology 12395), anti-phospho-IRF-3 (Ser396) (Cell Signaling Technology 4947), (anti-beta-Actin HRP conjugate (Cell Signaling Technology 12262), anti-Vinculin HRP conjugate (Cell Signaling Technology 18799).

Cell TiterGlo viability assay

Mouse organoids were cultured in Matrigel and maintained in organoid growth media without EGF in 48-well plates under appropriate treatment. After 3 or 5 days, cell viability was measured using CellTiter-Glo 3D Cell Viability Assay (Promega G9681) using manufacturer’s instructions. The following drugs were used: adagrasib, (MedChemExpress HY-130149), afatinib (MedChemExpress HY-10261), and momelotinib (MedChemExpress HY-10961). Concentrations used for each drug and treatment durations are described in detail in the results section and figures.

KRAS copy number assay

Organoids were dissociated and pelleted after three days in culture. Following gDNA isolation, copy number assays were performed using the TaqMan Copy Number Assay (Thermo Fisher Scientific, 4400291) according to the manufacturer’s instructions.

Colony formation assay

1,500 HCT116 cells/well were seeded in a 12-well plate. 24 h after seeding, media was removed and replaced with fresh media containing appropriate treatment. Treatments were refreshed every 2–3 days until controls reached confluency. Cells were then fixed using 4% PFA for 15 min at room temperature, shaking. Cells were stained using 10% Giemsa stain in PBS, shaking overnight at room temperature. Fixed and stained cells were then rinsed in water three times and left to dry overnight. After imaging, cells were destained using 10% acetic acid for 20 min at room temperature, and absorbance measured using Omega Plate Reader.

RNA isolation and RNA-seq

Organoids were collected in 5 mL of ice-cold PBS and centrifuged at 2500 RPM for 5 min at 4°C. The pellet was resuspended in 800 μL of TRIzol (Invitrogen, 15596-026). RNA extraction was carried out following the manufacturer’s protocol. To eliminate DNA contamination, the RNA was treated with recombinant DNaseI (Roche Diagnostics, 04716728001) for 15 min at room temperature, followed by column purification using the Qiagen RNeasy Mini Kit (74106). cDNA synthesis was performed using 1 μg of RNA and quantified with a NanoDrop spectrophotometer (Thermo Fisher Scientific). The Genomics Core Laboratory at Weill Cornell Medicine assessed RNA quality with an Agilent 2100 Bioanalyzer, prepared the RNA library using the TruSeq Stranded mRNA Sample Library Preparation Kit (Illumina), and conducted RNA sequencing (single-end, 75 cycles) on an Illumina NextSeq 500. Raw FASTQ files have been submitted to the SRA under accession number SRA: PRJNA1244364.

RNA-seq analysis

Raw FASTQ files were pseudoaligned to the mouse genome (GRCm39) reference using Kallisto (version 0.51.1).⁴⁶ Differential gene expression was estimated using the R package DEseq2.⁵⁴ Gene set enrichment analysis was performed using R packages fgsea⁵⁵ and GSVA.^45,47 Over-representation analysis (ORA) was performed on a list of up- or down-regulated genes for specific treatment conditions with a log2-fold change of >1.0 (bulk RNAseq) or >0.3 (CosMx) and an adjusted p-value <0.05. Z-scores and qvalues were calculated using the R package “clusterProfiler”⁵⁶ and custom genesets that incorporate the Hallmarks gene collections with curated intestinal and colon specific genesets that identify specific cell types or signaling states (Table S2) (https://github.com/lukedow/Genesets). R (version 4.5.1) and R Studio (version 2025.05.1 + 513) were used to create all visualizations and principal component analysis. Volcano plots, heatmaps, and other visualizations were produced using ggplot2, Prism (version 10.0.3), and Illustrator (version 28.7.4). Statistical details can be found in figure legends (Figures 4 and 5, S8 and S8). Non-parametric statistical tests (Wilcoxon rank-sum test) were used to compare groups; therefore, no assumption of normality was required.

Multiplex immunofluorescence

Multiplex immunofluorescence (mIF) was performed on 5 μm FFPE tumor sections by the Molecular Cytology Core at Memorial Sloan Kettering Cancer Center. Slides were baked at 58°C for 1 h, then loaded onto the Leica Bond RX automated staining system. Deparaffinization and heat-induced antigen retrieval were performed using EDTA-based epitope retrieval solution (ER2; Leica, AR9640) at 100°C for 20 min. Five markers were stained sequentially. The following primary antibodies were applied: CD4 (rabbit, 1:20, Ventana, 790–4423), CD8 (rabbit, 1:20, Ventana, 790–4460), FOXP3 (mouse, 1.25 μg/mL, Abcam, ab20034), CD68 (mouse, 0.04 μg/mL, DAKO, M0814), and CD163 (mouse, 1:3000, Invitrogen, MA1-82342), each incubated for 1 h at room temperature. For rabbit primary antibodies, detection was performed using Leica Bond Polymer anti-rabbit HRP (Leica Polymer Refine Detection Kit, DS9800). For mouse primary antibodies, a rabbit anti-mouse secondary antibody (Abcam, ab133469) was applied as a linker prior to the Leica Bond Polymer anti-rabbit HRP. Fluorescent detection was carried out using Alexa Fluor tyramide signal amplification reagents (Life Technologies, B40953, B40958) or CF dye tyramide conjugates (Biotium, 92172, 96053, 92174). Between each sequential staining round, epitope retrieval was repeated to denature the preceding primary and secondary antibodies before application of the next primary antibody. After completion of all staining rounds, slides were washed in PBS and incubated with 5 μg/mL DAPI (Sigma Aldrich) in PBS for 5 min, rinsed, and mounted in Mowiol 4–88 (Calbiochem). Slides were stored overnight at −20°C prior to imaging. Following fluorescence imaging, coverslips were removed and an additional round of chromogenic immunohistochemistry (IHC) was performed on the Leica Bond RX using an antibody against pan-cytokeratin (panCK; mouse, 0.25 μg/mL, Abcam, ab8068) to delineate epithelial tumor regions, as previously described. Slides were then scanned, and the resulting brightfield images were computationally co-registered with the corresponding fluorescence images to generate final 6-plex composite images (DAPI, CD4, CD8, FOXP3, CD68, CD163, and panCK). Slides were imaged at 5× and 20× magnification. Cell phenotyping and quantification were performed by the Molecular Cytology Core, with cells classified as CD4⁺ T cells, CD8⁺ cytotoxic T cells, FOXP3⁺ regulatory T cells, CD68⁺ macrophages, and CD163⁺ M2-polarized macrophages based on fluorescence intensity thresholds established during panel optimization. Cell densities were reported as cells per mm² of tumor area.

cfDNA analysis

Circulating free DNA (cfDNA) analysis was conducted using the MSK-ACCESS assay, as previously described.¹⁶ The MSK-ACCESS assay is a high-depth, custom-designed test that analyzes key exons and domains of 129 genes, along with introns of 10 genes containing recurrent breakpoints. For each patient, cfDNA and white blood cell (WBC) DNA were extracted from plasma and buffy coat, respectively. A minimum of 3 ng and up to 20 ng of plasma cfDNA was used for library construction, during which unique molecular identifiers (UMIs) and dual-index barcodes were incorporated to suppress sequencing errors. Libraries were captured using custom xGen Lockdown probes targeting key exons of 129 genes and recurrent intronic breakpoints in 10 genes, then sequenced as paired-end reads to ~1500× raw coverage. Raw reads were demultiplexed, trimmed, and aligned to GRCh37, after which UMI-based consensus read collapsing was performed to generate duplex, simplex, and all-unique BAM files. Variant allele frequencies (VAFs) reflect combined simplex and duplex read counts. Copy number alterations were derived from all-unique BAMs, and structural variants were called using Manta (v1.5.0), requiring ≥3 fusion-spanning reads for de novo events.

Patient tissue exome sequencing

MSK-IMPACT, a hybridization capture-based next-generation sequencing assay, was used for sequencing as previously described.^16,57 Genomic DNA was extracted from formalin-fixed paraffin-embedded (FFPE) tumors and matched normal blood samples using the Qiagen DNeasy Tissue kit and the EZ1 Advanced XL system (Qiagen 69504 and 9018702), respectively. Targeted sequencing was conducted on all exons and selected introns of 468 genes. Sequencing libraries were generated using the KAPA HTP protocol (Kapa Biosystems KK2611) and the Biomek FX system. Captured DNA fragments were pooled into libraries and sequenced on the Illumina HiSeq 2500, achieving high, uniform coverage (>500× median coverage). All genomic alteration types, including substitutions, indels, copy number variations, and rearrangements, were identified by comparing tumor samples to their matched normal counterparts. All sequencing and analyses were performed in a CLIA-certified clinical laboratory. Somatic alterations, including nonsynonymous substitutions, small insertions/deletions, focal gene-level amplifications, homozygous deletions, and select gene fusions, were identified through a clinically validated bioinformatic pipeline.⁵⁸ Each alteration was subsequently annotated for oncogenicity and clinical relevance using OncoKB (v3.9, December 30, 2021).⁵⁹ Alterations across all categories (mutations, copy number alterations, and fusions) classified as “oncogenic” or “likely oncogenic” by OncoKB, as well as those corresponding to newly identified mutational hotspots, were designated as driver alterations.

Spatial transcriptomics

Spatial transcriptomics was performed by the Integrated Genomics Operation at Memorial Sloan Kettering using the NanoString CosMx platform. Formalin-fixed, paraffin-embedded (FFPE) tissue sections (5 μm thickness) from 30 patient samples were mounted on glass slides, deparaffinized, and rehydrated according to Nanostring’s protocols. Targeted RNA detection was carried out using the NanoString CosMx Human RNA panel (950 genes). Following hybridization, slides were processed and imaged on the CosMx instrument, generating IF images and spatially resolved single-molecule RNA transcript coordinates. Immunofluorescence staining included DAPI for nuclear identification, pan-cytokeratin (panCK) for epithelial cells, CD45 for immune cells, and CD298/B2M as membrane markers (NanoString). Cell segmentation was performed by NanoString’s proprietary pipeline using combined nuclear and membrane signals, producing per-cell transcript counts and spatial coordinates.

Data import, quality control, and preprocessing

Raw CosMx outputs were imported into Rstudio and converted into Seurat objects (Seurat v5) using custom scripts based on NanoString’s recommended workflow. Individual slides were initially processed independently and subsequently merged into a single Seurat object after confirming the absence of duplicated cell identifiers across slides. Cells with fewer than 100 detected transcripts were excluded to remove low-quality or poorly segmented cells. Only cells passing this threshold were retained for downstream analysis. Putative doublets were removed by excluding cells that co-expressed canonical markers from two distinct cell lineages.

Expression values were normalized using Seurat’s NormalizeData function (log-normalization with default scaling factor).⁶⁰ Highly variable genes were identified using FindVariableFeatures, followed by scaling of expression values with ScaleData. Cell clustering was performed using the Louvain algorithm (FindClusters) with a resolution parameter of 3, selected to allow separation of major epithelial, stromal, endothelial. Two-dimensional embeddings were generated using Uniform Manifold Approximation and Projection (UMAP) with min.dist = 0.1 and spread = 4. For visualization purposes, UMAP plots were generated on random subsamples of cells. The cancer cell subset was independently re-processed with normalization, variable feature selection, scaling, PCA, neighbor graph construction (dims 1:30), clustering (resolution 0.5), and UMAP embedding (min.dist = 0.2, spread = 1). For visualization, up to 200,000 cancer cells were randomly subsampled prior to UMAP plotting.

Cell type annotation

Cluster identities were manually annotated based on canonical marker gene expression using feature plots and heatmaps. Marker genes used for cell type annotation included COL1A1, COL1A2, COL3A1, and THBS2 (CAFs); CD74, CD68, SPP1, and MSR1 (myeloid cells); IGHG1, IGHM, IGKC, and JCHAIN (B cells); CD3D, CD3G, CTLA4, and TCF7 (T cells); KRT8, KRT18, KRT19, and KRT20 (epithelial cancer cells); APOA1, TTR, SERPINA1, and FGG (hepatocytes); and PECAM1, VWF, ENG, and SPARCL1 (endothelial cells). Final cell identities were consolidated into the following cell types: CRC epithelial cells, cancer-associated fibroblasts (CAFs), endothelial cells, monocytes/macrophages, B cells, T cells, and hepatocytes. Only cells belonging to these categories were retained for downstream analyses. Statistical details can be found in Figures 1 and S2. Non-parametric statistical tests (Wilcoxon rank-sum test) were used to compare groups; therefore, no assumption of normality was required.

Differential expression analysis

To assess transcriptional changes at the individual patient level (Figure 2C), matched within-patient comparisons were performed for all patients with paired pre-treatment and on-treatment biopsies (n = 11 pairs) and pre-treatment and resistant biopsies (n = 8 pairs) using FindMarkers with the Wilcoxon test. Significance was calculated using Benjamini–Hochberg–adjusted p-values. Statistical details can be found in Figures 2, 3, 5, S3, S5 and S6

Pathway-level activity was quantified using UCell scoring and spatial smoothing. UCell scores were computed on the cancer cell subset (cancercells) and spatially smoothed using SmoothKNN (k = 100, reduction = PCA). Scores were then rank-transformed to generate per-cell quantile scores (0–1 scale), enabling cross-sample comparisons. These quantile scores were subsequently transferred from the cancer cell object (which included all cancer cells from all slides) to individual spatial slide objects via cell barcode matching. Curated gene sets from the Hallmark collection, CRC, and intestinal-relevant literature were filtered to retain genes detected in the CosMx dataset. UCell scores were computed for each gene set on a per-cell basis. Statistical details can be found in Figures 2, S5, S6, and S8. Non-parametric statistical tests (Wilcoxon rank-sum test) were used to compare groups; therefore, no assumption of normality was required.

Spatial visualization

Representative regions of interest (ROIs) were defined by manual cropping of individual tissue slides using the Crop function in Seurat, with crop coordinates selected based on tissue morphology, cell type composition, and signal quality as assessed by whole-slide ImageDimPlot. Cell-type images were generated using ImageDimPlot with cell segmentation boundaries. Pathway activity was visualized spatially on CRC cells alone using ImageFeaturePlot with expression values clipped to the 5th–95th percentile range (min.cutoff = 0.05, max. cutoff = 0.95) to reduce the influence of outliers. Single-molecule transcript overlays were generated using ImageDimPlot with the molecules argument. For region-specific comparisons, CRC cells were stratified into RAS-high and RAS-low regions based on manually curated field-of-view (FOV) assignments, and differential expression between regions was performed using the Wilcoxon test via FindMarkers. Non-parametric statistical tests (Wilcoxon rank-sum test) were used to compare groups; therefore, no assumption of normality was required.

Cell–cell communication analysis

Cell-cell communications were analyzed using CellChat (v1.6)²⁹ on the combined dataset, grouping cells by final annotated cell types. Only secreted signaling interactions from the human CellChat database were considered. Genes and ligand–receptor pairs were filtered based on overexpression criteria, and communication probabilities were computed with a minimum of 10 cells per interacting group. Aggregated communication networks were displayed using circle plots, heatmaps, and chord diagrams. Statistical details can be found in Figure S7. Non-parametric statistical tests (Wilcoxon rank-sum test) were used to compare groups; therefore, no assumption of normality was required.

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical analyses were performed in GraphPad Prism (version 10.0.3) and R (version 4.5.1, R Studio version 2025.05.1 + 513). Statistical details for each experiment, including the specific test used, the exact value of n and what n represents and measures of center and dispersion, are provided in the corresponding figure legends (Figures 1, 2, 3, 4, 5, and S1-S10). Unless otherwise noted, data are presented as mean ± standard error of the mean (SEM), with connected dots representing matched patient samples where applicable. Unless otherwise stated, each datapoint represents the average of ≥3 independent experiments. Formal tests for normality or equal variance were not performed on data analyzed by Student’s t test; for all spatial transcriptomics and bulk RNA-seq differential expression comparisons, the non-parametric Wilcoxon rank-sum test was used and therefore no assumption of normality was required. For mouse experiments, investigators were not blinded to treatment conditions.

Supplementary Material

Supplemental information can be found online at https://doi.org/10.1016/j.ccell.2026.04.009.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
anti-phospho-p44/42 MAPK (Erk1/2) (Thr202/Tyr204)	Cell Signaling Technology	Cat# 4370; RRID: AB_2315112
anti-p44/42 MAPK (Erk1/2)	Cell Signaling Technology	Cat# 9107; RRID: AB_10695739
anti-phospho TBK1/NAK1 (Ser172)	Cell Signaling Technology	Cat# 5483; RRID: AB_10693472
anti-TBK1	Cell Signaling Technology	Cat# 3504; RRID: AB_2255663
anti-non-phospho (Active) Yap (Ser127)	Cell Signaling Technology	Cat# 29495; RRID: AB_2798974
anti-Yap	Cell Signaling Technology	Cat# 12395; RRID: AB_2797897
anti-phospho IRF3 (Ser396)	Cell Signaling Technology	Cat# 4947; RRID: AB_823547
anti-beta-Actin HRP conjugate	Cell Signaling Technology	Cat# 12262; RRID: AB_2566811
anti-Vinculin HRP conjugate	Cell Signaling Technology	Cat# 18799; RRID: AB_2714181
Bacterial and virus strains
One Shot Stbl3 Chemically Competent E. Coli	Invitrogen	C737303
Biological samples
P1_On	Sloan Kettering	–
P1_Pre	Sloan Kettering	–
P1_Res	Sloan Kettering	–
P2_On	Sloan Kettering	–
P2_Pre	Sloan Kettering	–
P3_On	Sloan Kettering	–
P3_Pre	Sloan Kettering	–
P3_Pre	Sloan Kettering	–
P4_On	Sloan Kettering	–
P4_Pre	Sloan Kettering	–
P4_Res	Sloan Kettering	–
P4_Res	Sloan Kettering	–
P5_Pre	Sloan Kettering	–
P5_Res	Sloan Kettering	–
P6_On	Sloan Kettering	–
P6_Pre	Sloan Kettering	–
P6_Res	Sloan Kettering	–
P7_On	Sloan Kettering	–
P7_Pre	Sloan Kettering	–
P8_Res	Sloan Kettering	–
P9_On	Sloan Kettering	–
P9_Pre	Sloan Kettering	–
P9_Res	Sloan Kettering	–
P10_On	Sloan Kettering	–
P10_Pre	Sloan Kettering	–
P11_On	Sloan Kettering	–
P11_Pre	Sloan Kettering	–
P12_On	Sloan Kettering	–
P12_Pre	Sloan Kettering	–
P12_Res	Sloan Kettering	–
Chemicals, peptides, and recombinant proteins
Dulbecco’s Modified Eagle’s Medium (DMEM)	Corning	10-013-CV
DMEM-F12	Gibco	12634028
Roswell Park Memorial Institute (RPMI) 1640	Corning	10-040-CV
Fetal Bovine Serum	Gibco	A5670801
Penicillin/Streptomycin	Fisher Scientific	15140163
HEPES	Corning	25-060-Cl
GlutaMAX	Gibco	35050061
N-Acetylcysteine	Sigma-Aldrich	A9165
Nicotinamide	Sigma-Aldrich	N0636
EGF	PeproTech	AF-100-15
FGF-10	PeproTech	100–26
WNT Surrogate-Fc fusion protein	ImmunoPrecise Antibodies	N-001
Y-27632	MedChemExpress	HY-10071
A-83-01	MedChemExpress	HY-10432
SB202190	MedChemExpress	HY-10295
LDN193189	MedChemExpress	HY-12071
Gastrin-I	MedChemExpress	HY-P2671
B27 supplement	Gibco	17504044
Fungin	InvivoGen	ant-fn-1
Primocin	InvivoGen	ant-pm-05
Matrigel	Corning	353224
type III collagenase	Worthington	9001-12-1
dispase	Sigma-Aldrich	D4818
Adagrasib	MedChemExpress	HY-130149
Afatinib	MedChemExpress	HY-10261
RMC-7977	MedChemExpress	HY-148439
TBK1/IKKε-IN-5	MedChemExpress	HY-128679
Momelotinib	MedChemExpress	HY-10961
GSK8612	MedChemExpress	HY-111941
Alpelisib	MedChemExpress	HY-15244
Defactinib	MedChemExpress	HY-12289
TPCA-1	MedChemExpress	HY-10074
Pexidartinib	MedChemExpress	HY-16749
SP600125	MedChemExpress	HY-12041
LY2157299	MedChemExpress	HY-13226
Cell Recovery Solution	Corning	354253
Paraformaldehyde (PFA)	Sigma-Aldrich	P6148
Giemsa stain	Sigma-Aldrich	48900
TRIzol	Invitrogen	15596–026
DNaseI	Roche Diagnostics	04716728001
EcoRI-HF	New England Biolabs	R3101L
XhoI	New England Biolabs	R0146L
Polyethylenimine	Sigma-Aldrich	408727
Critical commercial assays
Qiagen DNeasy Tissue kit	Qiagen	69504
EZ1 Advanced XL system	Qiagen	9018702
KAPA HTP Library Preparation Kit	Kapa Biosystems	KK2611
DC Protein Assay Kit	BioRad	5000116
CellTiter-Glo 3D Cell Viability Assay	Promega	G9681
TaqMan Copy Number Assay	Thermo Fisher Scientific	4400291
Qiagen RNeasy Mini Kit	Qiagen	74106
MycoAlert Mycoplasma Detection Kit	Lonza	LT07-418
Deposited data
RNA sequencing	This paper	SRA: PRJNA1244364
CosMx data	This paper	GEO: GSE293124
Experimental models: Cell lines
Human: HEK293T	ATCC	CRL-3216
Human: HCT116	ATCC	CCL-247
Mouse: MZK190 AKP	This study	–
Mouse: MZJ786 AKP	This study	–
Mouse: AO876 AKP (G12D)	This study	–
Human: WCM1050 PDO	This study	–
Human: WCM2176 PDO	This study	–
Experimental models: Organisms/strains
C57Bl/6n	Charles River	–
NOD.Cg-PrkdcscidIl2rgtm1Wjl/SzJ (NSG)	MSKCC RARC	–
Oligonucleotides
TGAACTCGAGAAGGTATATTGCTGTTGACAGTGAGCG	Integrated DNA Technologies	shRNA cloning FWD
AGAAGGCTCGAGAAGGTATATTGC	Integrated DNA Technologies	shRNA cloning REV
TGCTGTTGGCAGTGAGCGACCAGTGGATGTT CAAATGAGATAGTGAAGCCACAGATGTATCTCATTT GAACATCCACTGGGTGCCTACTGCCTCGGA	Invitrogen	shTbk1.179
TGCTGTTGGCAGTGAGCGCCGCAGACTAGCTTATAA TGAATAGTGAAGCCACAGATGTATTCATTATAAGCTAG TCTGCGTTGCCTACTGCCTCGGA	Invitrogen	shTbk1.1754
Recombinant DNA
psPax2	Addgene	12260
pMD2.G	Addgene	12259
pCF80644	Addgene	186711
pLenti-FNLS-P2A-Puro	Addgene	110841
LRT2B	Addgene	110854
Software and algorithms
Incucyte S3 Live-Cell Analysis System	Sartorius	–
Seurat v5	Satija et al.45	–
Fgsea	Korotkevich et al.46	–
clusterProfiler	Yu et al.47	–
Prism 10	GraphPad	Version 10.6.1

Highlights.

• Genetic and non-genetic resistance profiles co-exist in KRASi-treated tumors

• Resistance-linked transcriptional profiles are heterogeneous within refractory tumors

• Early tumor responses are dominated by cell intrinsic inflammatory programs

• TBK1 blockade synergizes with KRAS inhibition