All fair enough. The two big immediate challenges in the field are i) that the tumor-derived fraction of total cfDNA can be as low as 1:10000 (stage I) and ii) that it is difficult to make Illumina sequencing more accurate than 1 error in 1000 sequenced bases (in which case the 1:10000 signal is drowned out). This paper uses some clever statistical tricks to reduce Illimina sequencing error; one of these tricks is to leverage population information, i.e. the more samples you sequence the better your understanding of (non-cancer-associated) systematic errors. This follows a long tradition in statistical genetics of using multi-sample panels to improve analysis of individual samples. There are also biochemical approaches like SafeSeq or Duplex Sequencing to reduce sequencing error.<p>Not-so-obvious point #1 is that the presence of cancer-associated mutations in blood != cancer. You find cancer-associated mutations in the skin of older probands, and assumedly many of the sampling sites would never turn into melanomas. A more subtle point is that cfDNA is likely generated by dying cells, i.e. a weak cancer signature in blood might also be indicative of the immune system doing its job.<p>Point #2 is that it's not necessarily about individual mutations, which are, due to the signal-to-noise ratio alluded to above, difficult to pick up. One can also look at the total representation of certain genes in cfDNA (many cancers have gene amplifications or deletions, which are easier to pick up because they affect thousands of bases at the same time), and the positioning of individual sequenced molecules relative to the reference genome. It seems that these positions are correlated with gene activities (transcription) in the cells that the cfDNA comes from, and cancer cells have distinct patterns if gene activity.