Background: The generation of subclonal (low-frequency) mutations is driven by tumor mutations and the relationship between the heterogeneity of tumor mutation abundance and non-small cell lung cancer (NSCLC) remains unknown. We investigate the role of allele frequency heterogeneity (AFH) defined by circulating tumor DNA (ctDNA) profiling in predicting prognosis in advanced NSCLC patients. Methods: Publicly available data set of POPLAR (N=211) and OAK (N=642) trials were used for analyzing. A low ratio of allele frequency (AF) of a mutation to the maximum-somatic-allele-frequency (MSAF) was used to define the presence of AFH. The prognostic value of AF/MSAF ratio that was below a defined cutoff point in overall survival (OS) was evaluated using Cox-proportional hazards regression; and the structural break point was determined by LOESS regression and Chow test. The derived AFH was also explored in an independent cohort (N=259) of advanced NSCLC receiving first-line EGFR-TKIs from the First Affiliated Hospital of Guangzhou Medical University. Results: In the POPLAR and OAK cohort, low AF/MSAF ratio was found to be significantly associated with unfavorable OS in univariate and multivariate analysis. 'Me structural break point analysis demonstrated that AF/MSAF <10% could yield the optimal value to stratify patients with poor OS, which was applied for defining the presence of APH. The presence of AFH significantly correlated with unfavorable OS in advanced NSCLC regardless of treatment arms (overall: HR 1.52, immunotherapy: HR 1.81, chemotherapy: HR 1.32, all P<0.05). In the exploratory FGFR-TKIs cohort, the presence of AFH was also significantly associated with shorter OS (HR 1.72, P=0.039). Conclusions: Our results demonstrate that the presence of AFH predict unfavorable prognosis in advanced NSCLC despite which drug the patients used. The presence of AFH defined by ctDNA might provide an easily-accessible biomarker for risk stratification in the current clinical practice.