Establishment Of A New Method For Low-frequency DNA Mutation Detection

In the past decades,with the rapid development of next generation sequencing(NGS)technology,NGS-based DNA mutation detection technology has been widely used in disease diagnosis.but because of the high error rate of NGS data,ultra-lowfrequency mutations detection is still challenging.In general,low-frequency mutation detection is mainly applied in circulated cell-free DNA(cfDNA),However,a lot of false positive in the detection because of the very limited amount of cfDNA in blood and the background noise from experimental process including DNA amplification and sequencing.Circulating tumor DNA(ctDNA)is a type of cfDNA produced by tumor cells.Mutation detection based on ctDNA has the potential for cancer screening,noninvasive genotyping of tumor patients,and real-time monitoring of disease.However,ctDNA levels are very low in most early and some advanced solid tumors,and the background noise generated from experimental process including library preparation and NGS,resulting in low detection sensitivity,which makes the detection and analysis of ctDNA difficult.Herein,we have developed a cfDNA low-frequency mutation analysis algorithm,namely Mutseek,based on the Duplex Sequencing strategy and machine learning-based background polishing technology.Combining Mutseek and home-developed molecular barcode technology(dUMI),we were able to reduce the overall base error rate from 0.034% to 0.005%.We analyzed the mutants in cfDNA standards,which containing 8 known mutation sites,and found that the Mutseek can detect mutations as low as 0.1%.In addition,we analyzed cfDNAs from lung cancer patient and found that detection sensitivity for EGFR T790 M mutation is 90%.iDES which is one of the most sensitive mutant detection technologies.Using Mutseek,we found that iDES showed a very obvious G> T / C> A bias,which may be due to oxidative damage during hybrid capture steps.However,this base bias does not exist in dUMI.Both methods have obvious C> T / G> A bias,which is likely because of deamination of methyl-cytosine.When dUMI sequencing data is processed with barcode alone,the base error rate is lower than iDES(0.008% vs.0.01%).Similarly,when the background polishing process is used alone,the base error rate of dUMI is also slightly lower than iDES(0.0052%s vs.0.0068%).Interestingly,when the two techniques are combined,the overall error rate of dUMI after noise reduction is higher than iDES(0.0050% vs.0.0014%),and there is no significant overlay effect.This may be because of the insufficient samples were used for noise reduction model training.In summary,we have developed a NGS bioinformatics analysis program to detect low-frequency mutations at 0.1% in both standards and clinical samples.However,there are still some parts need improvement.First,more clinical healthy samples should be used to train the background polishing model to further improve the detection sensitivity.Second,expedite entire analysis procedure by optimizing the algorism and program,to save the valuable time for patient treatment.