Background: Circulating tumor DNA (ctDNA) has been applied and showed potential in cancer early/late-stage detection, tumor genotyping and post-operation recurrence monitoring. The fraction of ctDNA in cell-free DNA (noted as ccf hereby), in addition to standard SNV/INDEL/CNV analysis, has also been showed to associate with the tumor progression and prognosis. In theory, accurate ccf can further be useful in correcting and improving given SNV/INDEL/CNV results. Existing tools capable for calculating ccf (PureCN, FACETS, Sequenza, etc.) use coverage data in targeted regions and SNP allele frequency to calculate the tumor fraction, which fail to give accurate estimation at relatively low ctDNA concentrations. 

Methods: A Maximum Likelihood model was built to estimate ccf. We first select informative SNPs with significantly different VAF in the case and paired-control samples. The mutation type of an informative SNP is determined by the variant allele frequency (VAF) in the paired samples and the copy number of the case sample. Likelihood of each SNP given a specific ccf was then calculated. After clustering SNPs into clones, the ccf of each clone was estimated using a global likelihood. 

Results: Performance of the method was validated by ctDNA dilution series analysis. 6 cfDNA from cancer patient was diluted (concentrations: 1/3 - 1/81). Detection limit of the method is ~2%, and correlation between estimated and expected ccf ranged from 0.93 to 0.98. 

Conclusions: We have developed a novel method to better estimate cancer cell fractions in cell-free DNA. Results showed our method is able to calculate ccf at lower ctDNA concentrations with higher accuracy and stability than benchmarked tools. We describe here a method for target-sequencing data that is more sensible, accurate and stable than currently available tools.