Main

Mutations in the KRAS gene represent the most prevalent oncogenic driver event in non-squamous non-small-cell lung cancer (NSCLC; 25–30%)1,2. KRAS mutations mostly result in a glycine-to-cysteine substitution at position 12 KRAS(G12C), accounting for approximately 13–17% of lung adenocarcinomas1,3,4,5 and approximately 10% of NSCLC6. Discovery of a cryptic allosteric pocket in the switch 2 region of GDP-bound KRAS(G12C) paved the way for the identification, clinical development and regulatory approval of inactive state-selective covalent inhibitors of KRAS(G12C), which has revolutionized the treatment landscape of KRASG12C-mutated NSCLC7,8.

Inhibitors of GDP-bound and/or GTP-bound KRAS(G12C) are in various stages of clinical development7,9,10,11,12. Sotorasib is a first-in-class small molecule that specifically and irreversibly inhibits the KRAS(G12C) protein, exerting selective anti-tumor activity5. Sotorasib has accelerated or full approval in more than 50 countries for adults with KRASG12C-mutated locally advanced or metastatic NSCLC who received one or more prior systemic therapies based on the phase 2 global CodeBreaK 100 (CB100) trial13,14,15. The phase 3 CodeBreaK 200 (CB200) study demonstrated clinical efficacy and tolerability of sotorasib versus docetaxel in patients with previously treated, advanced NSCLC16. The KRAS(G12C) inhibitor adagrasib is also approved for patients with previously treated advanced KRASG12C-mutated NSCLC17,18. Despite remarkable progress, there is considerable interpatient variability in clinical outcomes with KRAS(G12C) inhibitors and a lack of validated biomarkers that can robustly identify patients likely to derive long-term benefit from monotherapy versus patients who exhibit early disease progression and may benefit from an intensified, combination therapeutic strategy.

The efficacy of inactive state-selective KRAS(G12C) inhibitors can be curtailed due to primary insensitivity or the emergence of adaptive or acquired resistance19,20,21. Co-occurring genomic alterations detected at baseline have been reported to affect responses to diverse systemic therapies, including KRAS(G12C) inhibitors22,23. Among patients harboring pathogenic co-mutations in the KEAP1, SMARCA4 and CDKN2A tumor suppressor genes, early disease progression and poor clinical outcomes were previously reported with sotorasib or adagrasib24. Adaptive resistance due to synthesis and GTP loading of new KRAS(G12C) resulting from increased RTK drive and feedback pathway reactivation through wild-type (WT) KRAS, NRAS or HRAS can also impair the activity of KRAS(G12C) inhibitors25,26. Furthermore, tumor cell-extrinsic factors, including innate and adaptive host immune responses and remodeling of the tumor microenvironment (TME), were reported to modulate responses to KRAS extinction or inhibition27,28. Finally, after initial response, diverse and frequently co-incident mechanisms of acquired resistance to KRAS(G12C) inhibition were identified, including emergence of secondary alterations in RAS genes or other components of the RTK–RAS–MAPK pathway, as well as lineage switching19,21,29. In view of the heterogeneity of individual responses to KRAS(G12C) inhibitors, improved understanding of molecular determinants of response and resistance and the development of clinically actionable biomarkers constitute critical unmet needs to inform individual therapeutic decisions and guide the clinical development of combination therapies for patients with KRASG12C-mutated advanced NSCLC.

Here we present a comprehensive, integrative analysis of baseline genomic and transcriptomic biomarker data and on-treatment dynamic changes in circulating tumor DNA (ctDNA) from the CB100 and CB200 datasets, collectively representing one of the largest cohorts to date of patients with NSCLC prospectively treated in clinical trials of any KRAS(G12C) inhibitor.

Results

Patient cohorts and biomarker analysis

Datasets included 317 biomarker-evaluable patients with previously treated advanced KRASG12C-mutated NSCLC who enrolled in CB200 and 112 biomarker-evaluable patients treated in CB100 (Fig. 1). Clinical outcomes in biomarker-evaluable populations were similar to the intention-to-treat (ITT) population across both studies (Extended Data Table 1). Overall, clinical covariates were well balanced between the two studies, although a few small but significant differences were noted (Extended Data Table 2).

Fig. 1: Biomarker analysis methodology.
figure 1

Baseline tissue samples (fresh/archival) were analyzed using the Tempus xT assay, which combined a 648-gene targeted DNA sequencing panel and whole-transcriptome RNA. PD-L1 protein level was assessed by local standard-of-care testing. Pre-treatment and post-treatment samples were assessed using plasma NGS: Resolution ctDx Lung assay (23 genes).

Full size image

Genomic correlates of sotorasib efficacy

Co-occurring genomic alterations introduce biological heterogeneity in KRASG12C-mutated NSCLC and may affect responses to systemic therapies, including KRAS(G12C) inhibitors13,18,22,24,30,31. Hence, we interrogated the interaction of gene mutation status across both studies and treatment arms (n = 312) with clinical outcomes. The most frequently co-mutated genes were TP53 (38.46%), STK11 (34.29%), CDKN2A (18.27%), ATM (16.35%), KEAP1 (15.71%) and LRP1B (10.58%) (Fig. 2a and Extended Data Table 3), which was largely consistent with previous reports from prospective and retrospective studies.

Fig. 2: Analyses of tumor-intrinsic genomic determinants of sotorasib response and resistance.
figure 2

a, Oncoprint of alterations (pathogenic or potentially pathogenic based on OncoKB, SnpEff and FATHMM designation) in patients with available tissue NGS data at baseline (CB100 + CB200 sotorasib and docetaxel arms). The rows indicate genes with reported alterations (short variants, copy number variants (gain or loss), insertions, deletions or fusions) sorted based on prevalence. Of note, 10 patients from CB200 (3.2% of patients across both CodeBreaK studies) were negative for KRASG12C by retrospective tissue NGS. However, 8 of 10 patients were still reported to be KRASG12C positive by the plasma ctDNA assay. Tissue heterogeneity and sample quality likely account for this discordance. All patients, including those with lack of KRASG12C detection by retrospectively performed tissue NGS, had prospective central laboratory confirmation of KRASG12C in tissues using the pre-defined diagnostic companion assay (KRAS RGQ PCR Kit (Qiagen)). Hence, these patients were included in the analyses. b, Forest plot showing hazard of progression (HR (95% CI)) with sotorasib compared to docetaxel treatment in CB200 patients grouped based on specified genomic alterations. FDR-adjusted, two-sided P values for the interaction between arm and gene alteration in Cox proportional hazards models are noted. c,d, Kaplan–Meier curves of PFS according to ATM mutation status in patients treated with sotorasib or docetaxel (CB200). e, Kaplan–Meier curve of PFS according to ATM mutation status in patients treated with sotorasib from the combined dataset. f,g, Kaplan–Meier curves of PFS (f) or OS (g) according to KEAP1 mutation status in patients treated with sotorasib in the combined dataset. CNV, copy number variation; MUT, mutant; NE, not estimable.

Full size image

Sotorasib trended toward increased progression-free survival (PFS) compared to docetaxel across most co-alteration–defined patient subgroups in CB200 (Fig. 2b). In treatment-by-biomarker interaction analyses, testing all reported genomic alterations with sufficient prevalence (Methods), ATM status was the only genomic biomarker associated with differential PFS benefit across study arms (interaction P = 0.009; false discovery rate (FDR)-adjusted P = 0.092; Fig. 2c). Specifically, patients with ATMWT NSCLC had significantly longer PFS with sotorasib versus docetaxel (median PFS (mPFS), 5.72 months versus 4.01 months; P = 0.0026; log-rank test), whereas patients with ATMMUT NSCLC had numerically improved PFS with docetaxel versus sotorasib (mPFS, 8.25 months versus 4.17 months; P = 0.13; log-rank test; Fig. 2d). Patients with ATMWT NSCLC were significantly more likely to exhibit long-term benefit (PFS ≥ 6 months) versus early disease progression (PFS ≤ 3 months) with sotorasib versus docetaxel (P = 0.0197; logistic regression; Extended Data Fig. 1a). Among 205 patients treated with sotorasib in the combined CB100/CB200 dataset, ATM co-mutations were associated with significantly shorter PFS (mPFS; ATMMUT versus ATMWT, P = 0.032; log-rank test; Fig. 2e); consistent trends were observed in individual studies (Fig. 2c, left panel, and Extended Data Fig. 1b). Overall survival (OS) with sotorasib was not affected by pathogenic ATM mutations (Extended Data Fig. 1c).

KEAP1 mutational status was not predictive of differential PFS benefit from sotorasib versus docetaxel in CB200 (interaction P = 0.2; FDR-adjusted P = 0.51). Patients with KEAP1-mutated NSCLC had significantly shorter PFS (P = 0.028, log-rank test; Fig. 2f) and OS (P = 0.00054, log-rank test; Fig. 2g) with sotorasib compared to those bearing KEAP1WT tumors in the combined dataset, with consistent trends across the individual studies (Extended Data Fig. 1d,e). Accordingly, patients with KEAP1WT NSCLC were more likely to exhibit long-term benefit versus early disease progression with sotorasib compared to those bearing KEAP1MUT tumors (P = 0.0125, logistic regression; Extended Data Fig. 1f).

Based on previous reports24, we evaluated the potential effect of co-alterations in SMARCA4 and STK11. In both the combined dataset and CB200, pathogenic alterations in SMARCA4 were not associated with significant differences in PFS or OS with sotorasib or docetaxel, although patients with SMARCA4MUT tumors more frequently displayed early disease progression (Extended Data Fig. 1g). When analyzing studies separately, there was a discordant interaction of STK11 alteration with clinical outcomes (Extended Data Fig. 2a–c). In CB200, pathogenic STK11 alterations were associated with significantly shorter PFS and OS in both sotorasib and docetaxel arms (Extended Data Fig. 2a,b), supporting a poorly prognostic role. In contrast, no significant differences in PFS or OS were detected between sotorasib-treated patients in CB100 with STK11MUT versus STK11WT NSCLC (Extended Data Fig. 2c). These conclusions were upheld when limited to patients bearing STK11MUT/KEAP1WT tumors (Extended Data Fig. 2d). Thus, additional biological modifiers may account for the heterogeneous clinical responses to sotorasib and other KRAS(G12C) inhibitors in patients with KRASG12C-mutated NSCLC with STK11 co-mutations.

KRAS G12C-mutated NSCLC transcriptional subtypes

Because the co-mutational profile alone may not fully capture the diversity of KRAS-mutated NSCLC, the transcriptional landscape of KRASG12C-mutated NSCLC and its impact on clinical outcomes with sotorasib was examined. An unbiased, non-negative matrix factorization-based analysis of KRAS-mutated NSCLC previously identified three robust and reproducible transcriptional subtypes, termed KP, KL and KC, based on distinct co-occurring genomic alteration enrichment patterns30. Key features of KP tumors include enrichment in mesenchymal gene signatures, frequent TP53 co-alterations and a typically ‘hot’ T cell-inflamed TME with high levels of tumor cell programmed death-ligand 1 (PD-L1) expression. KL tumors are characterized by frequent inactivating STK11 genomic alterations and/or functional LKB1 inactivation, a ‘cold’ CD8-positive T cell-depleted TME with low levels of tumor cell PD-L1 expression and increased prevalence of KEAP1 co-alterations with evidence of nuclear factor erythroid 2-related factor 2 (NRF2) pathway activation. Prominent features of KC tumors encompass reduced thyroid transcription factor 1 (TTF-1) expression, frequent mucinous differentiation, evidence of NRF2 pathway activation in a subset of cases and enrichment for bi-allelic CDKN2A/CDKN2B loss events30,32. In the combined dataset of 234 patients with KRASG12C-mutated NSCLC and available baseline tumor RNA sequencing (RNA-seq) data, KP/KL/KC subtypes accounted for 42.3%, 35.0% and 22.6% of tumors, respectively (Fig. 3a and Extended Data Fig. 3a), consistent with their previously reported prevalence across diverse datasets30,33. KP/KL/KC subtypes exhibited the anticipated co-mutation enrichment patterns, and KC tumors expressed substantially lower levels of NKX2.1/TTF1 mRNA (Fig. 3a, Extended Data Fig. 3b,c and Extended Data Table 4).

Fig. 3: Transcriptomic determinants of sotorasib clinical efficacy.
figure 3

a, Prevalence of transcriptional subtypes (KC, KL and KP) and TTF1 expression status (within KC/KL/KP) for patients with available RNA-seq data (CB100 + CB200 sotorasib and docetaxel arms). Box plot defines maxima (top of the box at the third quartile), minima (bottom of the box at the first quartile), center, whiskers (upper whisker extends to the lowest and highest values within 1.5× interquartile range of the first and third quartile) and percentile. b, Kaplan–Meier curves of PFS in patients with KC, KL and KP tumor types treated with sotorasib or docetaxel (CB200). The FDR-adjusted P values for the post hoc log-rank tests of PFS difference between sotorasib and docetaxel in the KC/KL/KP subtypes were 0.56, 0.03 and 0.50, respectively. c,d, Kaplan–Meier curves of PFS (c) or OS (d) across KC, KL and KP tumor types for patients treated with sotorasib in the combined dataset. e,f, Kaplan–Meier curves of PFS (e) or OS (f) according to TTF1 (high versus low) mRNA expression in patients treated with sotorasib in the combined dataset. g, ORR (% (95% CI)) according to TTF1 (high versus low) mRNA expression in patients treated with sotorasib in the combined dataset. OR and unadjusted, two-sided P value (Fisher’s exact test) are shown. FPKQ, fragments per kilobase per million reads, quantile normalized; NE, not estimable.

Full size image

We assessed the efficacy of sotorasib compared to docetaxel in KP/KL/KC transcriptional subtypes in CB200 and observed subtype-dependent differences in anti-tumor activity. A Cox proportional hazards model of PFS versus the interaction of transcriptional subtype and treatment suggested that the independent effect of subtype (P = 0.008) was more significant than the interaction with treatment (P = 0.28). Patients with KL tumors (representing 27% of CB200 patients receiving sotorasib) had improved mPFS with sotorasib compared to docetaxel (mPFS, 5.85 months versus 2.69 months; FDR-adjusted P = 0.011; log-rank test; hazard ratio (HR), 0.4; P = 0.07; Fig. 3b). In contrast, no significant differences in PFS were observed between sotorasib-treated and docetaxel-treated patients in the KP and KC subgroups. In patients with KP tumors, both treatments resulted in numerically longer but similar PFS (mPFS, 7.75 months versus 7.16 months; P = 0.33), whereas shorter PFS with either sotorasib or docetaxel was observed in the KC subgroup (mPFS, 3.94 months versus 3.02 months; P = 0.56). No significant subgroup-specific differences in OS were detected between treatment arms (Extended Data Fig. 3d).

Next, we compared the efficacy of sotorasib in the KP/KL/KC transcriptional subtypes in the combined dataset. Consistent with CB200, PFS with sotorasib varied widely across KP/KL/KC subtypes and was shorter in patients with KC tumors (mPFS, 2.92 (KC), 8.11 (KL) and 7.75 (KP) months; P = 0.016; log-rank test; Fig. 3c). This patient group was further skewed toward early disease progression versus long-term benefit with sotorasib (Extended Data Fig. 3e). OS was markedly different between subtypes and was significantly curtailed in patients with KC tumors (median OS (mOS), 5.82 (KC), 16.62 (KL) and 16.0 (KP) months; P = 3.5 × 10−5; log-rank test), corresponding to a higher hazard of death compared to those bearing either KL (HR (95% confidence interval (CI), 2.4 (1.4–4.4)), FDR-adjusted P = 2.7 × 10−3) or KP (2.9 (1.7–4.9), FDR-adjusted P = 1.6 × 10−4) NSCLC (Fig. 3d).

Because a substantial fraction of KC NSCLC express low levels of TTF-1, we further assessed whether expression of TTF-1 could more precisely separate clinical outcomes with sotorasib30. We partitioned tumors into TTF1-high (TTF1High) and TTF1-low (TTF1Low) based on NKX2.1/TTF1 mRNA expression. Across the combined dataset, 17.5% of KRASG12C-mutated NSCLC were classified as TTF1Low (Extended Data Fig. 3c), consistent with the previously reported prevalence of TTF-1-negative KRASG12C-mutated NSCLC by immunohistochemistry (IHC) (~15%)34. Patients treated with sotorasib bearing TTF1Low tumors exhibited significantly shorter PFS (mPFS, 2.76 months (TTF1Low) versus 8.11 months (TTF1High); P = 2.9 × 10−9; log-rank test; HR (95% CI), 0.23 (0.14–0.4)) and OS (mOS, 4.47 months (TTF1Low) versus 16 months (TTF1High); P = 2.7 × 10−8; HR (95% CI), 0.26 (0.15–0.43)) and significantly lower objective response rate (ORR) (4.17% versus 42.1%; P = 0.000151; Fisher’s exact test; odds ratio (OR) (95% CI, 16.55 (2.52–701.28))) compared to patients treated with sotorasib harboring TTF1High tumors (Fig. 3e–g). The adverse effect of TTF1Low status was observed when analysis was limited to KC tumors (Extended Data Fig. 3f,g). Although TTF1 status was not predictive of sotorasib benefit versus docetaxel in CB200, patients with TTF1High tumors had improved PFS with sotorasib versus docetaxel (mPFS, 7.26 months versus 4.83 months, respectively; HR (95% CI), 0.62 (0.36–1.07); Tukey adjusted P = 0.11, z-test; Extended Data Fig. 3h,i). These results indicate that TTF-1 status represents a robust and clinically tractable determinant of sotorasib efficacy.

NRF2 activation status

Because KEAP1 alterations were associated with inferior clinical outcomes, we examined whether activation of a NRF2-driven transcriptional program that may capture both KEAP1/NFE2L2/CUL3-mutated and a subset of WT tumors could affect the clinical efficacy of sotorasib. For this, we used a previously validated gene expression signature35. In the combined dataset with available RNA-seq data, 21.8% of tumors were classified as NRF2High and 78.2% as NRF2Low (Fig. 4a). Notable differences in the fraction of NRF2High tumors were observed across distinct co-mutation-defined cohorts (Extended Data Fig. 4a). Concordant with the role of KEAP1 as a regulator of NRF2 stability, 78.8% (26/33) of KRASG12C-mutated NSCLC with KEAP1 co-alterations were NRF2High. NRF2High status was further documented in 22.9% (11/48) and 9.1% (12/132) of STK11MUT/KEAP1WT and STK11WT/KEAP1WT tumors, respectively. The prevalence of NRF2High tumors differed considerably among the KP/KL/KC subgroups (1.0%, 32.9% and 43.4%, respectively; P = 3.3 × 10−13; Fisher’s exact test; Extended Data Fig. 4b), in agreement with earlier reports30, as well as between TTF1High and TTF1Low tumors (16.1% versus 48.8%, respectively; P = 3.212 × 10−5; Fisher’s exact test).

Consistent with previous reports linking NRF2 activation and aggressive NSCLC35, patients with NRF2High tumors exhibited significantly shorter PFS with sotorasib compared to those bearing NRF2Low tumors in the combined dataset (mPFS, 2.73 months versus 7.75 months; P = 0.0016; log-rank test; Fig. 4b), with trends observed in individual studies (Extended Data Fig. 4c). OS was significantly shorter in patients harboring NRF2High tumors (mOS, NRF2High 6.05 months versus NRF2Low 16.0 months; P = 3 × 10−4; Fig. 4c). These results were in agreement with a recent study that reported inferior clinical outcomes with adagrasib in patients with NRF2High tumors in the KRYSTAL-1 trial36. Among patients with NRF2Low KRASG12C-mutated NSCLC, sotorasib resulted in improved PFS compared to docetaxel in CB200 (mPFS, 6.24 months versus 4.47 months; P = 0.015; log-rank test; Extended Data Fig. 4d), although NRF2Low was not predictive of differential PFS benefit (treatment-by-biomarker interaction analyses; FDR-adjusted P = 0.7). No difference in PFS was observed between the two treatment arms in patients with NRF2High tumors (Extended Data Fig. 4d). OS was not different between treatment arms in either NRF2Low or NRF2High tumors.

Fig. 4: Impact of NRF2 activation status on clinical outcomes.
figure 4

a, Prevalence of NRF2 activation status for patients with available RNA-seq data (CB100 + CB200 sotorasib and docetaxel arms). b,c, Kaplan–Meier curves for PFS (b) or OS (c) based on NRF2 (high versus low) activation status for patients treated with sotorasib in the combined dataset. d,e, Kaplan–Meier curves of PFS (d) or OS (e) based on TTF1 (high versus low) mRNA expression and NRF2 (high versus low) activation status for patients treated with sotorasib in the combined dataset. NE, not estimable.

Full size image

Finally, we examined whether integration of TTF-1 and NRF2 status would further refine stratification of clinical outcomes with sotorasib in the combined dataset. Remarkably, the integrated analysis yielded three groups with markedly distinct clinical trajectories. Low TTF1 expression was the hallmark of the poor prognosis group, regardless of NRF2 status (mPFS (95% CI), 2.76 (1.41–3.12) months; TTF1Low; Fig. 4d; mOS (95% CI), 4.47 (3.15–7.92) months; TTF1Low; Fig. 4e). Among patients with TTF1High tumors, NRF2Low status identified a group (70.3% of the entire population) with prolonged PFS (mPFS (95% CI), 8.31 (5.72–11.04) months) and OS (mOS, (95% CI), 16.62 (12.48–not estimable) months) with sotorasib. Patients with TTF1High; NRF2High tumors (13.1%) exhibited intermediate PFS (mPFS (95% CI), 4.07 (1.41–9.99) months) and OS (mOS (95% CI), 9.49 (4.07–13.96) months). Thus, both TTF-1 expression and NRF2 activation status represent major determinants of sotorasib efficacy that can stratify patients with advanced KRASG12C-mutated NSCLC into groups with distinct clinical outcomes. In an assessment of the concordance of TTF-1 status based on RNA-seq (high/low) and IHC (positive/negative) in a subgroup of 27 patients with available data from both modalities, six of eight tumors that were TTF1Low by RNA-seq were confirmed to be TTF-1-negative by IHC, whereas 16 of 19 tumors deemed to be TTF1High by RNA-seq were TTF-1-positive by IHC (OR (95% CI), 13.81 (1.58–207.07); P = 0.006; Fisher’s exact test). There was overall concordance between TTF-1 status as determined based on RNA-seq and IHC, in line with a previous study reporting a strong correlation between TTF-1 expression per IHC and RNA expression37.

Immune phenotype and sotorasib efficacy

Because both innate and adaptive host immune responses may contribute to the anti-tumor efficacy of KRAS(G12C) inhibitors5, we investigated potential tumor cell-extrinsic determinants of sotorasib efficacy. Across patients with low, intermediate or high PD-L1-expressing tumors, there was consistent sotorasib activity and PFS benefit compared to docetaxel (Fig. 5a). In agreement with the previously reported role of LKB1 inactivation as a driver of the cold TME in KRAS-mutated NSCLC22, a higher fraction of KL or STK11-mutated tumors had PD-L1 tumor proportion score (TPS) < 1%, whereas KP or TP53-mutated tumors had PD-L1 ≥ 50% (Extended Data Fig. 5a). We then explored the effect of PD-L1 expression on PFS within the transcriptional subtypes. Although PD-L1 status did not seem to influence PFS in patients with KC/KP tumors, there was a numerical improvement in PFS in patients with PD-L1 TPS < 1% versus PD-L1 TPS ≥ 1% KL tumors (mPFS, 8.34 months versus 3.14 months; Extended Data Fig. 5b). Finally, we examined whether sotorasib outcomes diverged across previously identified tumor immune subsets from an integrated pan-cancer large-scale immunogenomic analysis38. Across the combined dataset with available RNA-seq data, the most prevalent immune subtypes were ‘inflammatory’ (23.4%) and ‘interferon-γ (IFNγ)-dominant’ (21.4%), followed by ‘transforming growth factor-β (TGFβ)-dominant’ (19.3%), ‘lymphocyte-depleted’ (18.6%), ‘wound healing’ (12.4%) and ‘immunologically quiet’ (4.8%) (Fig. 5b). We observed enrichment of inflammatory and immunologically quiet subtypes in tumors with PD-L1 TPS < 1%, whereas the IFNγ-dominant tumor type was overrepresented among PD-L1-expressing tumors (P = 0.0004; Fisher’s exact test; Extended Data Fig. 5c), consistent with a role for IFNγ in the upregulation of tumor PD-L1 (ref. 39). PFS and OS with sotorasib differed significantly between patients with distinct tumor immune subtypes in the combined dataset and individual studies, with the inflammatory subtype consistently exhibiting the best clinical outcomes (mPFS, 9.99 months; mOS, 17.54 months) and the IFNγ-dominant and wound healing subtypes exhibiting the worst clinical outcomes (mPFS, 4.14 months and 3.61 months; mOS, 6.64 months and 7.43 months, respectively) (Fig. 5c,d and Extended Data Fig. 5d). Patients with the IFNγ-dominant subtype were skewed toward early disease progression versus long-term benefit, whereas the inflammatory subtype was skewed toward long-term benefit with sotorasib treatment (Extended Data Fig. 5e). Among patients with inflammatory tumors, we observed a higher PFS in patients with PD-L1 TPS < 1% (mPFS, 16.3 months) versus PD-L1 TPS ≥ 1% tumors (mPFS, 5.29 months) (Extended Data Fig. 5f). Taken together, results highlight diverse outcomes across immune phenotypes and further identify a PD-L1 low but inflamed set of tumors that may benefit from sotorasib treatment.

Fig. 5: Tumor immune contexture and sotorasib efficacy in KRASG12C-mutated NSCLC.
figure 5

a, Forest plot showing hazard of progression with sotorasib versus docetaxel treatment in CB200 patients grouped based on tumor cell PD-L1 expression. Two-sided P values (Cox proportional hazards model) for PD-L1 level HRs were FDR adjusted. b, Pie chart indicating the prevalence of specific tumor immune subtypes (Thorsson immune phenotyping) in patients treated with sotorasib from the combined dataset. c,d, Kaplan–Meier curves of PFS (c) or OS (d) based on specific tumor immune subtypes for patients treated with sotorasib in the combined dataset. NE, not estimable.

Full size image

ctDNA dynamics in response and resistance

Having explored the association between key baseline molecular features and sotorasib efficacy, we assessed the potential utility of dynamic, on-treatment biomarkers of anti-tumor activity. Across multiple studies of targeted therapies in advanced NSCLC and other tumor types, longitudinal monitoring of ctDNA may provide an early readout of treatment response, predict the emergence of adaptive or acquired resistance before radiologic progression and yield prognostic information40,41,42,43. In CB200 at baseline, KRASG12C detection with ctDNA was 82.5% for the sotorasib arm versus 73.7% for the docetaxel arm, consistent with what was previously observed with liquid biopsies from KRASG12C-mutated NSCLC42,43. In patients with detectable KRASG12C mutations at baseline, treatment with sotorasib resulted in more rapid reduction in KRASG12C variant allele frequency (VAF) relative to docetaxel, consistent with shorter time to radiologic response, and was evident as early as 1 week after treatment initiation (Fig. 6a)16. Clearance of KRASG12C ctDNA occurred more rapidly in the sotorasib arm and was achieved in 43% of sotorasib-treated patients at the cycle 2/day 1 timepoint compared to 14% of patients treated with docetaxel (Fig. 6b). Across both treatment arms, patients with detectable KRASG12C ctDNA during treatment exhibited increased hazard of disease progression or death and worse PFS compared to those with KRASG12C ctDNA clearance across cycles 1–4 (Fig. 6c–f and Extended Data Fig. 6a,b). Of note, KRASG12C ctDNA negativity observed at cycle 1/day 8 with sotorasib treatment (Fig. 6d) was associated with improved PFS (mPFS, 7.26 months versus 4.01 months; P = 0.0067; log-rank test). Overall, there was a trend toward numerically higher hazard of progression or death in patients with detectable KRASG12C ctDNA at later timepoints, consistent with tumor cell persistence and the possible emergence of adaptive/acquired resistance. Conversely, undetectable KRASG12C ctDNA at later timepoints likely reflects a sustained molecular response and may indicate prolonged clinical benefit. Notably, clearance of KRASG12C ctDNA at cycle 2/day 1 was associated with significantly longer PFS with sotorasib even among patients with TTF1High tumors, highlighting the potential additional predictive utility of on-treatment biomarkers when assessed in tandem with baseline molecular determinants of clinical outcome (Fig. 6g).

Fig. 6: Longitudinal monitoring of ctDNA to capture sotorasib response and resistance.
figure 6

a, Longitudinal modeling to depict changes in KRASG12C VAF (LSM estimate (95% CI)) in patients treated with sotorasib versus patients treated with docetaxel at various timepoints indicated on the x-axis. FDR-adjusted, two-sided P values are shown for the difference between sotorasib and docetaxel at each timepoint from linear mixed models. b, Longitudinal modeling to depict changes in KRASG12C detection status (% (95% CI)) in patients with detectable ctDNA treated with sotorasib versus patients treated with docetaxel at various timepoints. FDR-adjusted, two-sided P values (McNemar test) are shown for the difference between post-baseline visit and baseline visit for each treatment group. c, PFS dependence on KRASG12C detection status (HR for KRASG12C status (95% CI)) from Cox proportional hazards models stratified by treatment arm. Center of the error bar is the HR. Two-sided P values were FDR adjusted. d–g, Kaplan–Meier curves of PFS in patients with detectable versus undetectable KRASG12C at C1D8 (d), C2D1 (e and f) or TTF-1 high (+) sotorasib patients at C2D1 (g). h, Oncoprint depicting pathogenic acquired variants at progression from sotorasib-treated patients from the combined dataset and pie chart showing the variant distribution among pathways of interest. The oncoprint rows indicate genes with reported alterations (short variants, copy number variants (gain or loss), insertions, deletions or fusions), sorted based on prevalence. C, cycle; CNV, copy number variation; D, day; LSM, least squares mean; NE, not estimable; SFU, safety follow-up.

Full size image

Finally, we characterized candidate genomic mechanisms of acquired resistance to sotorasib in ctDNA obtained at the time of radiologic disease progression. Plasma samples at progression were collected ≥1 month before the earliest progression date among independent and investigator assessments per Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1 in the combined dataset. Among 134 patients with detectable ctDNA at the time of radiologic progression, emergent genomic variants absent in paired baseline ctDNA samples were detected in 44 patients (32.8%), with pathogenic alterations identified in 30 patients (22.4%). The most prevalent acquired pathogenic genomic alterations involved TP53 (40%), MET (30%, mostly amplifications) and EGFR (20%, mostly amplifications) (Fig. 6h, Extended Data Fig. 6c and Supplementary Table 1). Interestingly, acquired pathogenic alterations in KEAP1 and STK11 were identified in three patients (10%). Overall, 43% of acquired pathogenic variants involved RTK pathways (Fig. 6h, right panel).

Discussion

This study dissected molecular determinants of response and resistance to sotorasib, leveraging in-depth genomic and transcriptomic analyses of tumor tissue and serial analysis of ctDNA from two unique and complementary clinical cohorts: the single-arm phase 2 CB100 study and the global randomized phase 3 CB200 study that compared the efficacy of sotorasib with docetaxel. Collectively, our datasets constitute one of the largest aggregate populations of patients prospectively treated in clinical trials of KRAS(G12C) inhibitors assembled. Furthermore, availability of randomized training datasets and a distinct validation cohort facilitate assessment of potential prognostic and predictive effects. Thus, our analysis represents, to our knowledge, the most comprehensive and detailed exploration of sotorasib efficacy biomarkers to date.

Our results indicate that the transcriptional diversity of KRASG12C-mutated NSCLC can affect clinical outcomes with sotorasib. Previous work identified three major transcriptional subtypes of KRAS-mutant NSCLC, termed KC, KL and KP, with distinct co-mutation enrichment patterns, differentiation states, histologic features and immune profiles. In our analysis, the KC transcriptional subtype, marked by reduced TTF-1 expression in tumor cells, frequently mucinous histologic features and enrichment of CDKN2A/CDKN2B bi-allelic loss events, was associated with significantly shorter PFS and OS with either sotorasib or docetaxel, indicative of a poor prognostic effect. In support of this finding, a mucinous transcriptional state akin to the classical subtype of pancreatic adenocarcinoma was recently reported to support long-term tumor cell persistence in response to diverse RAS inhibitors in NSCLC44 and pancreatic ductal adenocarcinoma preclinical models45,46. Notably, our work identified TTF-1 status as a major determinant of sotorasib clinical efficacy in KRASG12C-mutated NSCLC, resulting in divergent outcomes in patients with TTF1Low (ORR, 4%; mPFS, 2.8 months; mOS, 4.5 months) compared to TTF1High (ORR, 45%; mPFS, 8.1 months; mOS, 16 months) tumors. Because TTF-1 expression by IHC is routinely assessed as part of the diagnostic workup of thoracic tumors and metastatic carcinomas, this finding may provide a feasible biomarker strategy to identify approximately 15–20% of patients with KRASG12C-mutated NSCLC who exhibit poor prognosis and may benefit from treatment intensification with RAS inhibitor–anchored combination therapeutic regimens. In addition, this marker may further emerge as a clinically relevant stratification factor in future clinical trials of RAS inhibitors.

Our study raises the possibility that the composition of the TME may influence clinical outcomes with sotorasib, beyond tumor cell-intrinsic biomarkers of sotorasib efficacy. Strikingly, patients with the KL transcriptional subgroup, characterized by enrichment of STK11 co-occurring alterations and a ‘cold’ TME with frequent lack of PD-L1 expression on tumor cells, appeared to derive improved benefit from sotorasib versus docetaxel in CB200 compared to other subtypes. This calls into question whether the effect of KRAS(G12C) inhibitors can be further improved by combining with PD-(L)1 antibodies. This finding was consistent with the increased prevalence of immune subtypes associated with improved sotorasib efficacy in patients with PD-L1-negative tumors. Because patients with PD-L1-negative KRAS-mutated NSCLC exhibit inferior outcomes with first-line PD-(L)1 inhibitor-based chemo-immunotherapy compared to those harboring PD-L1-positive tumors, these findings may support efforts to combine sotorasib with platinum-doublet chemotherapy for the first-line treatment of patients with advanced KRASG12C-mutated PD-L1-negative NSCLC, such as the ongoing randomized phase 3 CodeBreaK 202 trial47. Additional work using direct visualization and enumeration of distinct immune cell subsets and characterization of immune cell functional states in tumor tissue is required to validate these hypothesis-generating findings.

Concordant with findings from large retrospective studies and post hoc analyses of prospective KRAS(G12C) inhibitor clinical trials1,13,18,24,36, co-occurring genomic alterations in key tumor suppressor genes were associated with heterogeneous clinical outcomes with sotorasib in our cohorts. Improved PFS with sotorasib compared to docetaxel was observed in patients bearing KRASG12C-mutated NSCLC with ATMWT genomic status. Further work is required to validate the role of ATM co-mutations and evaluate their significance in the context of chemotherapy-naive patients. In agreement with previous reports24,36, pathogenic alterations in KEAP1 were associated with inferior clinical outcomes with sotorasib. Similar trends were observed in patients bearing NRF2High tumors that captured a subset of KEAP1WT NSCLC. Integration of NRF2 activation status and TTF-1 expression enabled robust partitioning of patients into favorable (~70%), intermediate (~13%) and poor (~17%) prognostic groups. Future application of unbiased, machine learning-supported approaches to derive predictive signatures of sotorasib clinical benefit may enable refinement of clinical outcome prediction algorithms in patients with TTF1High KRASG12C-mutated NSCLC.

Finally, our analysis revealed that on-treatment clearance of KRASG12C ctDNA with sotorasib occurred rapidly (as early as cycle 1/day 8 in 31% of patients) and was associated with significantly improved PFS at early (cycle 1/day 8) and later (cycle 4/day 1) timepoints, with progressively increasing hazard of progression in KRASG12C ctDNA-positive patients at later timepoints. Thus, KRASG12C ctDNA clearance may serve as an early on-treatment surrogate of long-term outcomes with sotorasib and may refine clinical response prediction algorithms and baseline biomarkers.

Our study has several limitations. Despite the large cohort size, the subset of biomarker-evaluable patients is more limited, reducing the statistical power of molecular analyses. Although we do not see major differences in outcomes between biomarker-evaluable patients and the ITT population, we cannot rule out potential bias that may result from analyzing a subset of the enrolled population and/or instances of informative censoring that could confound interpretation of biomarker data. Additionally, most analyses reported here are post hoc in nature and were not prespecified. Because only genomic alterations considered pathogenic or likely pathogenic were included in the analysis, we cannot rule out the possibility that certain biologically relevant variants may have been erroneously considered WT. Furthermore, inference of immune cell abundance based on whole tumor RNA-seq profiles has caveats that reduce the accuracy and limit the resolution of the analysis. Lastly, given the limited ctDNA panel size, analysis of acquired resistance variants may not reflect the full spectrum of genomic alterations contributing to clinical resistance or non-genomic resistance mechanisms that cannot be captured by ctDNA assays. Thus, these analyses are exploratory and signal seeking, requiring further validation in additional datasets.

Our work provides critical new insights into tumor cell-intrinsic and tumor cell-extrinsic molecular determinants of sotorasib efficacy in advanced KRASG12C-mutated NSCLC. These results may pave the way for the development of clinically applicable biomarker predictors for personalization of treatment strategies, further inform patient stratification in future clinical studies and aid selection of optimized, rational and tailored combination partners of KRAS(G12C) inhibitors to prolong disease control.

Methods

Study design and patient population

The objective of this analysis was to characterize patients with advanced KRASG12C-mutated NSCLC, focusing on molecular and clinical features associated with response and resistance. Patients were pooled from the CB200 and CB100 sotorasib groups for power reasons, to assess prognostic factors. Patients in the CB200 sotorasib and docetaxel groups were analyzed separately to assess predictive factors. The predictive factors were then evaluated in patients in the CB100 sotorasib group to confirm consistency with the CB200 sotorasib group.

The multicenter, single-group, open-label, phase 1/2 CB100 study (NCT03600883) enrolled patients ≥18 years of age with KRASG12C-mutated locally advanced or metastatic NSCLC after progression on prior therapies13. Key eligibility criteria included locally advanced or metastatic NSCLC, KRASG12C mutation confirmed through molecular testing, an Eastern Cooperative Oncology Group (ECOG) performance score of ≤2 (phase 1) and ≤1 (phase 2), measurable disease according to RECIST version 1.1 and treatment with one or more prior systemic therapies unless intolerant of or ineligible for available therapies known to provide clinical benefit48. Exclusion criteria included previous treatment with a KRAS(G12C) inhibitor and active brain metastases. Patients with asymptomatic, treated and stable brain metastases were eligible. As of 22 February 2022 (data cutoff), 174 patients received sotorasib 960 mg orally once daily. The primary endpoint was OR (complete or partial response) as assessed by blinded independent central radiologic review (BICR). Secondary endpoints included duration of response, disease control, time to response, PFS, OS and safety.

The randomized, multicenter, open-label, phase 3 CB200 study (NCT04303780) compared the efficacy of sotorasib versus docetaxel in patients with KRASG12C-mutated advanced NSCLC. Enrolled patients were ≥18 years of age who had disease progression after previous platinum-based chemotherapy and PD-1 or PD-L1 inhibitor16. Key eligibility criteria included histologically or cytologically documented, locally advanced and unresectable or metastatic NSCLC, KRASG12C mutation confirmed by central laboratory testing, tumor progression after receiving at least one previous systemic therapy for advanced disease, ECOG performance status of 0 or 1 and measurable disease according to RECIST version 1.1 (ref. 16). Patients with treated, stable brain metastases were eligible. Exclusion criteria included new or progressing untreated brain lesions or symptomatic brain lesions, previously identified oncogenic driver mutation other than KRASG12C for which an approved therapy is available, previous treatment with docetaxel (with conditions), previous treatment with a direct KRAS(G12C) inhibitor, systemic anti-cancer therapy within 28 days of study day 1 and therapeutic or palliative radiation therapy within 2 weeks of treatment initiation. As of 2 August 2022 (data cutoff), 171 patients were assigned to once-daily sotorasib 960 mg orally, and 174 patients were assigned to docetaxel 75 mg m−2 intravenously every 3 weeks; 46 patients crossed over from docetaxel to sotorasib. The primary endpoint was PFS by BICR per RECIST version 1.1, and secondary endpoints included OS, ORR and patient-reported outcomes.

Inclusion criteria for biomarker-eligible patients were prospectively defined. Patients who provided appropriate consent were biomarker evaluable if they met protocol inclusion criteria and had either genomic (baseline and/or treatment) or transcriptomic (baseline) data for baseline associations, ctDNA dynamics or acquired resistance (Fig. 1). Attribution of predictive versus prognostic effects was based on treatment arm by biomarker interaction analysis in CB200, whereas assessment of molecular determinants of sotorasib efficacy was performed in the pooled population of sotorasib-treated patients across both CB200 and CB100, with subsequent confirmation of the consistency of observed effects in the individual trials. Sex and/or gender were not considered in this biomarker analysis. Baseline characteristics for sex are reported in CB100 (ref. 13) and CB200 (ref. 16).

Institutional review board approval before study initiation and participating country regulatory authority approval were received, and all patients provided written informed consent. Patients were not compensated for their participation in the study.

Molecular analyses

Baseline tissue samples (fresh/archival) were analyzed using a targeted 648-gene DNA sequencing panel (Tempus xT; Tempus Labs) and a whole-transcriptome RNA assay (Tempus RS.v2). Reported genomic alterations were filtered to retain only somatic, non-synonymous or splice alterations, copy number loss of tumor suppressor genes and copy number gain of oncogenes; the latter two (oncogene versus tumor suppressor gene) were based on the COSMIC Cancer Gene Census classification49. Fusions included in analyses involved oncogenic tumor-suppressing genes or oncogenes or one of these genes: MET, EGFR, ERBB3, FGFR1, FGFR2, FGFR3, NTRK1, NTRK2, NTRK3, ALK, ROS1 and RET. Further filtering for likely pathogenic alterations was based on the alteration having an OncoKB therapeutic level and being oncogenic or likely oncogenic50, being predicted as pathogenic or likely pathogenic by the functional analysis through hidden Markov model (FATHMM) algorithm51 or being predicted to be of high impact by SnpEff software52. For statistical analysis of prognostic or predictive effect in the CB200 study, the filtered alterations were grouped at the gene level, and only genes with prevalence of five or more per treatment arm were included. Effect validation in the CB100 study also required meeting the same criteria. PD-L1 protein level was assessed by local standard-of-care testing. In a subset of biomarker-evaluable patients, TTF-1 status (IHC) was available through information reported by clinical investigators. Post-treatment samples were assessed using plasma next-generation DNA sequencing (NGS): Resolution ctDx Lung assay (23 genes; Exact Sciences Corporation).

Analyses of immune subtype and TME were performed by Tempus Labs based on methodology previously described38. Immune subtypes were determined by calculating single-sample gene set enrichment analysis (ssGSEA) scores of five representative signatures using transcripts per million (TPM) gene expression values. The data were standardized by median centering and scaled by median absolute deviation (MAD), and each study sample was assigned to the nearest cluster centroid based on Euclidean distance. Immune subtypes assessed were IFNγ-dominant, immunologically quiet, inflammatory, lymphocyte-depleted, TGFβ-dominant and wound healing.

Statistical analysis

Statistical analysis was performed using the R ‘stats’ package (version 4.2.3; https://rpkgs.datanovia.com/). Time-to-event analysis (PFS and OS) was performed with Cox proportional hazards regression modeling using the ‘survminer’ package (version 0.4.9). Where noted in the text and figures, comparison of survival between groups was also performed using a log-rank test. To evaluate whether a biomarker was prognostic or predictive in the CB200 study, a Cox proportional hazards model with a treatment-by-biomarker interaction term was fitted. Analysis of whether each biomarker was predictive or prognostic of CB200 categorical outcomes (that is, long-term benefit versus early progression) was performed with logistic regression of treatment-by-biomarker interaction model. For the clustering algorithm, k set to 2 was applied to log10-transformed mRNA expression of NKX2.1/TTF1 across all samples. For continuous and binary variables, statistical significance was assessed through the significance of the interaction term, whereas, for categorical variables with more than two levels (that is, Thorsson et al.38 cluster comparison), a likelihood ratio test was performed (a model with all covariates but without the interaction term versus a model with the interaction term), with P < 0.05 considered statistically significant. Stratified models evaluating the association of each biomarker with clinical response within each treatment arm of the CB200 study (sotorasib, docetaxel) and the CB100 study followed the same statistical modeling approach as outlined above but without any interaction terms included. When analyses of the combined dataset were performed, study was included as a covariate in the model. If prevalence of a categorical variable was compared across two or more groups, a Fisher’s exact test was employed.

Longitudinal ctDNA analyses

KRASG12C status was modeled as a binary variable (detected/not detected) and as a continuous variable (VAF). In the case of the latter, patients who had ctDNA data but no KRASG12C detected had their VAF imputed to a constant value (1−5). For changes in KRASG12C detection status from baseline, a McNemar test was performed at each timepoint within each treatment arm comparing prevalence of detection to day 1 (D1) (baseline). KRASG12C VAF was modeled using linear mixed-effects models, with treatment arm, timepoint (coded as a categorical variable) and treatment arm–by–timepoint interaction treated as fixed effects, whereas an individual patient was treated as a random effect using the R packages ‘lmerTest’ (version 3.1) and ‘emmeans’ (version 1.8.9). Comparisons between docetaxel and sotorasib at individual timepoints were performed by specifying contrasts from the full model.

Adjustment for multiple testing

FDR was controlled using the Benjamini–Hochberg method53 for the hypothesis-generating tests of gene mutation status association with outcomes and post hoc contrasts between factor levels (NSCLC transcriptional subtype, PD-L1 and longitudinal ctDNA). P values for the remaining descriptive ad hoc tests were unadjusted for multiple testing.

Data-sharing statement

The data discussed in this paper have been deposited in the National Center for Biotechnology Informationʼs Gene Expression Omnibus (GEO)54 and are accessible through GEO Series accession number GSE295032 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE295032).

Data-sharing requests relating to data in this paper will be considered after the publication date and after (1) this product and indication (or other new use) have been granted marketing authorization in both the United States and Europe or (2) clinical development discontinues and the data will not be submitted to regulatory authorities. There is no end date for eligibility to submit a data-sharing request for these data.

Qualified researchers may submit a request containing the research objectives, the Amgen product(s) and Amgen study/studies in scope, endpoints/outcomes of interest, statistical analysis plan, data requirements, publication plan and qualifications of the researcher(s).

In general, Amgen does not grant external requests for individual patient data for the purpose of re-evaluating safety and efficacy issues already addressed in the product labeling. A committee of internal advisors reviews requests. If not approved, requests may be further arbitrated by a data-sharing independent review panel. Requests that pose a potential conflict of interest or an actual or potential competitive risk may be declined at Amgen’s sole discretion and without further arbitration.

Upon approval, information necessary to address the research question will be provided under the terms of a data-sharing agreement. This may include anonymized individual patient data and/or available supporting documents, containing fragments of analysis code where provided in analysis specifications.

Further details are available at http://www.amgen.com/datasharing.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.