Benchmarking digital PCR partition classification methods with empirical and simulated duplex data

Digital PCR (dPCR) is a highly accurate technique for the quantification of target nucleic acid(s). It has shown great potential in clinical applications, like tumor liquid biopsy and validation of biomarkers. Accurate classification of partitions based on end-point fluorescence intensities is crucial to avoid biased estimators of the concentration of the target molecules. We have evaluated many clustering methods, from general-purpose methods to specific methods for dPCR and flowcytometry, on both simulated and real-life data. Clustering method performance was evaluated by simulating various scenarios. Based on our extensive comparison of clustering methods, we describe the limits of these methods, and formulate guidelines for choosing an appropriate method. In addition, we have developed a novel method for simulating realistic dPCR data. The method is based on a mixture distribution of a Poisson point process and a skew-$t$ distribution, which enables the generation of irregularities of cluster shapes and randomness of partitions between clusters ('rain') as commonly observed in dPCR data. Users can fine-tune the model parameters and generate labeled datasets, using their own data as a template. Besides, the database of experimental dPCR data augmented with the labeled simulated data can serve as training and testing data for new clustering methods. The simulation method is available as an R Shiny app.