Resistance undermines targeted therapy. Patients with advanced non-small cell lung cancer (NSCLC) harboring EGFR mutations frequently benefit from third-generation tyrosine kinase inhibitors (TKIs) such as aumolertinib (AUM). However, acquired resistance inevitably develops, severely limiting the long-term clinical benefit of this treatment class.
The T790M secondary mutation in EGFR exon 20 is the most common mechanism of resistance to first- and second-generation EGFR TKIs, which led to the development of third-generation inhibitors like AUM that can target T790M-bearing tumors. AUM has been approved for both second-line T790M-mutant NSCLC and first-line EGFR-activating mutant NSCLC, yet acquired resistance still emerges through mechanisms including additional EGFR mutations, activation of alternative signaling pathways, and aberrant downstream signaling.
Understanding the molecular basis of AUM resistance is an urgent clinical need. Identifying prognostic biomarkers that predict which patients are at highest risk of resistance-driven treatment failure could allow clinicians to intervene earlier with alternative or combination strategies, potentially extending survival.
A resistance model built from cell lines to cohorts. To identify resistance-associated genes, the researchers induced AUM resistance in HCC827 NSCLC cells over six months of stepwise dose escalation followed by maintenance at 10 micromolar drug concentration, achieving a resistance index of 3.35 that reflects clinically relevant acquired resistance. RNA whole transcriptome sequencing compared resistant cells to their parental counterparts.
From this sequencing, 2,987 genes were upregulated and 3,410 were downregulated in resistant cells. These differentially expressed genes were intersected with prognostic genes identified through Cox regression across four independent GEO lung adenocarcinoma cohorts (GSE50081, GSE30219, GSE72094, GSE132313) and the TCGA-LUAD dataset, yielding 20 candidate genes associated with both AUM resistance and overall survival.
To construct the most robust prognostic signature, the authors integrated 10 machine learning algorithms (including Random Survival Forest, Lasso, Ridge, Elastic Net, Stepwise Cox, CoxBoost, plsRcox, SuperPC, GBM, and survival-SVM) across 100 pairwise combinations evaluated in five independent cohorts. The combination with the highest average concordance index across all cohorts was selected as the optimal model.
The TCGA-LUAD cohort served as the training set while the four GEO cohorts served as independent test sets. This multicenter design with 100 model combinations and 5 external validation cohorts represents a particularly rigorous framework for ensuring that the final signature captures biologically meaningful rather than dataset-specific signals.
Lasso plus Random Survival Forest achieves top performance. Among all 100 algorithm combinations, the combined lasso plus RSF model incorporating 12 genes achieved the highest average concordance index and was designated the Aumolertinib Resistance-Related Prognostic Signature (ARRPS). The 12 genes are ANLN, ASPM, BUB1B, CDC20, CDC25C, CDC6, CDCA2, CDCA3, CDKN3, CENPE, CENPF, and FOXM1.
All 12 genes were confirmed to be highly expressed in lung adenocarcinoma cancer tissues compared to non-tumor tissue, and each demonstrated a high discriminative ability for distinguishing LUAD from normal lung. This biological consistency across expression, resistance, and prognosis strengthens confidence that the genes reflect genuine cancer biology rather than statistical noise.
Kaplan-Meier analyses across all five cohorts showed that patients in the high-ARRPS group consistently had significantly worse overall survival than those in the low-ARRPS group. The signature achieved concordance indices of 0.73 in TCGA-LUAD, 0.69 in GSE72094, 0.71 in GSE50081, 0.67 in GSE30219, and 0.66 in GSE13213.
Crucially, ARRPS outperformed 40 previously published prognostic signatures in four of the five cohorts evaluated, demonstrating superior predictive accuracy relative to the existing landscape of LUAD prognostic tools.
Surpassing conventional staging. The prognostic superiority of ARRPS was compared directly against standard clinical and molecular predictors including AJCC stage, TP53 mutation status, KRAS mutation, EGFR mutation, STK11 mutation, smoking pack-years, age, and gender. ARRPS achieved a higher concordance index than all these predictors in every cohort where comparison data were available.
In the TCGA-LUAD cohort, ARRPS reached a concordance index of 0.73 compared to 0.68 for AJCC staging, 0.62 for TP53 mutation status, 0.61 for EGFR mutation, and 0.55 for age. The advantage was consistent across validation cohorts, confirming that ARRPS captures prognostic information not encoded in routine clinical variables.
Univariate and multivariate Cox regression analyses confirmed ARRPS as an independent prognostic factor for overall survival, disease-specific survival (DSS), progression-free interval (PFI), and disease-free interval (DFI) in the TCGA-LUAD cohort. Patients in the high-ARRPS group showed approximately 3.2-fold higher mortality risk on average (hazard ratio = 3.2, P less than 0.001).
ROC curve analysis across all five cohorts demonstrated strong time-dependent predictive accuracy for overall survival, supporting ARRPS as both a robust prognostic tool and a potential guide for risk-stratified clinical decision making beyond what standard staging systems provide.
High-risk patients show activation of growth-promoting pathways. Gene set variation analysis revealed that patients in the high-ARRPS group showed enrichment of signaling pathways associated with cell cycle progression and DNA damage repair, including pathways involving FOXM1 and TOP2A, which are master regulators of mitotic fidelity.
High-ARRPS tumors also exhibited upregulation of metabolic reprogramming pathways including cytochrome P450 drug metabolism, arachidonic acid metabolism, linoleic acid metabolism, ether lipid metabolism, alpha-linolenic acid metabolism, and taurine and hypotaurine metabolism. These metabolic shifts may contribute to the reduced sensitivity to AUM observed in high-risk patients.
Genomic alteration analysis of TCGA-LUAD data revealed that high-ARRPS patients had significantly higher mutation frequencies in genes including NPAP1, TSHZ3, PCDHB8, PTPRT, GRM8, and ANK1, as well as copy number amplifications at chromosomal regions including 3q26.2 and 12q15 and deletions at multiple loci. These genomic differences provide mechanistic context for the survival differences observed between risk groups.
The 12 ARRPS genes, particularly FOXM1, CDC20, and CENPF, are master regulators of mitotic checkpoint control and DNA repair pathways that are recurrently altered in osimertinib-resistant tumors. This mechanistic overlap suggests that ARRPS may capture resistance biology common to third-generation EGFR TKIs beyond AUM alone.
Drug screening identifies compounds targeting resistant cells. Using drug sensitivity data from the CTRP and PRISM databases, the researchers performed differential drug response analysis between patients in the top 10% and bottom 10% of ARRPS scores. Spearman correlation analysis filtered for compounds whose sensitivity (as measured by AUC) was negatively correlated with ARRPS score at a threshold of r less than -0.3.
Fifteen compounds passed initial screening, six from CTRP and nine from PRISM. Further filtering using the CMap database identified CD-437 and TPCA-1 as the most promising candidates based on high CMap scores and limited prior study in LUAD. Both compounds showed 4 to 6-fold selectivity for AUM-resistant cells over parental cells, confirming that they target resistance-specific dependencies rather than exerting generic cytotoxicity.
When combined with AUM in CCK8 proliferation assays, both CD-437 and TPCA-1 exerted synergistic anti-proliferative effects in resistant cell lines, with HSA synergy scores exceeding 20 for the CD-437 combination and exceeding 10 for TPCA-1. Colony formation assays confirmed the antiproliferative synergy under combination treatment conditions.
In vivo validation using subcutaneous xenograft tumors in BALB/c nude mice showed that both combinations significantly inhibited tumor growth compared to AUM alone, as measured by tumor volume and tumor tissue weight. These in vivo results extend the in vitro synergy findings and support the potential clinical relevance of these combination strategies for high-ARRPS patients.
From biopsy to personalized treatment decisions. The authors propose a practical clinical workflow in which RNA sequencing of biopsy tissue from newly diagnosed NSCLC patients is used to calculate an ARRPS score. Patients are then stratified into low-risk and high-risk groups using the median score as a threshold.
Patients in the low-ARRPS group are expected to respond adequately to conventional AUM monotherapy and can be managed with standard protocols. Patients in the high-ARRPS group, who face a substantially elevated risk of disease progression and resistance-driven treatment failure, would be candidates for combination regimens incorporating CD-437 or TPCA-1 alongside AUM.
The 12-gene signature was validated across 1,412 patients from five independent cohorts, demonstrating generalizability to the clinical heterogeneity of real-world LUAD populations. The ARRPS also performed as an independent prognostic factor across multiple survival endpoints in multivariate analyses, supporting its use alongside rather than instead of standard clinical variables.
Practical implementation barriers include the cost of RNA sequencing and the limited clinical-grade availability of CD-437 and TPCA-1. The authors are developing a targeted 12-gene NanoString assay as a lower-cost alternative suitable for CLIA-accredited laboratories, and are exploring FDA-approved retinoids as CD-437 analogs that could be more readily deployed.
Resistance biology beyond a single drug. The 12 ARRPS genes encode proteins involved in mitotic checkpoint regulation and DNA damage response, which are biological processes recurrently disrupted in tumors that acquire resistance to osimertinib and other third-generation TKIs. This mechanistic overlap raises the possibility that ARRPS captures a conserved resistance biology shared across EGFR TKI generations.
The therapeutic relevance of CD-437 and TPCA-1 may extend beyond AUM-specific resistance. CD-437, a retinoic acid receptor agonist, and TPCA-1, an IKK-2 inhibitor with anti-inflammatory and anti-tumor properties, appear to target survival pathways that are upregulated in TKI-resistant cells rather than simply suppressing primary EGFR signaling, which TKIs already accomplish.
Validation in osimertinib-resistant patient-derived xenograft models is ongoing and represents the next step toward confirming whether ARRPS and the associated therapeutic candidates apply across the third-generation EGFR TKI class. If validated, this would substantially broaden the potential clinical impact of the findings beyond patients specifically treated with AUM.
The study's multicenter design with five independent cohorts totaling over 1,400 patients, evaluated using 100 machine learning combinations, represents a methodological standard that reduces the risk of overfitting or cohort-specific bias that has limited the clinical translation of many previously published prognostic signatures.
A dual-purpose biomarker and treatment guide. ARRPS represents an advance in the management of EGFR-mutant NSCLC by serving simultaneously as a prognostic stratification tool and a guide for selecting patients who may benefit from resistance-targeted combination therapies, addressing a gap not filled by standard clinical staging or mutation profiling.
The signature's consistent performance across five independent validation cohorts and its superiority over 40 published alternatives suggest it has achieved a level of generalizability rarely demonstrated in single-center prognostic studies, making it a credible candidate for prospective clinical validation.
The identification of CD-437 and TPCA-1 as synergistic partners for AUM in resistant disease provides a specific therapeutic hypothesis that can be tested in prospective clinical trials, with both drugs characterized by mechanisms complementary to EGFR inhibition rather than overlapping with it.
Future work includes developing a cost-accessible NanoString-based implementation of the 12-gene assay, completing validation in additional TKI-resistant preclinical models, and initiating clinical trials of AUM-based combination regimens in high-ARRPS patient subgroups identified at diagnosis.