Thread by @thatdnaguy

#ASHG21 NC: Nicholas Cheng.
Early detection of cancers using plasma cell-free DNA methylomes up to 7 years prior to clinical diagnosis.

#ASHG21 NC: cancer development is a mostly evolutionary process involving sequenctial somatic mutation and methylation / accessibility changes.

#ASHG21 NC: When can you detect changes? The pathogenic transitions can occur earlier than the onset of disease.

#ASHG21 NC: Early detection can improve survivability. Increasing push to find sensitive and specific markers that can detect cancer prior to overt symptoms.

#ASHG21 NC: Cell free DNA can carry the mutational sequences of tumor, as well as the methylation status.

#ASHG21 NC: across cancer genomes there is disruption of methylation patterns. Using 100 CpG sites, can get 0.95 sensitivity and 0.97 specificity for predicting cancer from TCGA panel.

#ASHG21 NC: Does cell free DNA give you enough information to predict the source tissue?

#ASHG21 NC: Often tested for patients that present with symptoms. Population cohorts may do better since it could be presymptomatic. Ontario Health Study data.

#ASHG21 NC: Look for people in the cohort that were healthy at enrollment, but later developed cancer. Linked to the Ontario Cancer Registry to identify people that may be informative.

#ASHG21 NC: pre-cancer blood samples are from up to 7 years prior to diagnosis. Identified cancer-free controls matched for age, sex, smoking status, alcohol consumption, etc.

#ASHG21 NC: 1.8 mL of plasma. immunoprecipitation of methylated cell free DNA sequencing.

#ASHG21 NC: Find differentially methylated regions in the future cancer patient compared to controls. Use general linear models corrected for batch and group.

#ASHG21 NC: Appreciable number of DMRs in future breast cancer and future prostate cancer cell free DNA. Do the DMRs look like the DMRs from breast cancer tissue?
112 regions DMR in cell free DNA are matched to breast cancer tissue.

#ASHG21 NC:
[ editorial note for the UpSet plot, you can increase font size of the numbers on axes and on top of bar plots for readability increase ]

#ASHG21 NC: 10-fold cross valication. Average random forest classifier. Repeated cross-validation 500 times. 200 hypomethylated regions.

#ASHG21 NC: All cases and controls get 0.743 average AUC. In last 3 years get 0.757 AUC. Weighted time dependent AUC, you get 0.731 generally up to 7 years previous to 0 years previous. Jump up to higher AUC in the ~2 years prior to diagnosis.

#ASHG21 NC: Prediagnosis sensitivty is about 45%. Specificity >95%. AUC about 0.76.

#ASHG21 NC: For pancreatic cancer, cell free DNA methlation can be predictive. Post-diagnosis samples have about 0.96 AUC when trained with samples from pre-diagnosis.