Alleviating cell-free DNA sequencing biases with optimal transport

Cell-free DNA (cfDNA) is a rich source of biomarkers for various (patho)physiological conditions. Recent developments have used Machine Learning on large cfDNA data sets to enhance the detection of cancers and immunological diseases. Preanalytical variables, such as the library preparation protocol or sequencing platform, are major confounders that influence such data sets and lead to domain shifts (i.e., shifts in data distribution as those confounders vary across time or space). Here, we present a domain adaptation method that builds on the concept of optimal transport, and explicitly corrects for the effect of such preanalytical variables. Our approach can be used to merge cohorts representative of the same population but separated by technical biases. Moreover, we also demonstrate that it improves cancer detection via Machine Learning by alleviating the sources of variation that are not of biological origin. Our method also improves over the widely used GC-content bias correction, both in terms of bias removal and cancer signal isolation. These results open perspectives for the downstream analysis of larger data sets through the integration of cohorts produced by different sequencing pipelines or collected in different centers. Notably, the approach is rather general with the potential for application to many other genomic data analysis problems.