Adaptive bandwidth kernel density estimation for next-generation sequencing data

High-throughput sequencing experiments can be viewed as measuring some sort of a "genomic signal" that may represent a biological event such as the binding of a transcription factor to the genome, locations of chromatin modifications, or even a background or control condition. Numerous algorithms have been developed to extract different kinds of information from such data. However, there has been very little focus on the reconstruction of the genomic signal itself. Such reconstructions may be useful for a variety of purposes ranging from simple visualization of the signals to sophisticated comparison of different datasets.

Here, we propose that adaptive-bandwidth kernel density estimators are well-suited for genomic signal reconstructions. This class of estimators is a natural extension of the fixed-bandwidth estimators that have been employed in several existing ChIP-Seq analysis programs.

Using a set of ChIP-Seq datasets from the ENCODE project, we show that adaptive-bandwidth estimators have greater accuracy at signal reconstruction compared to fixed-bandwidth estimators, and that they have significant advantages in terms of visualization as well. For both fixed and adaptive-bandwidth schemes, we demonstrate that smoothing parameters can be set automatically using a held-out set of tuning data. We also carry out a computational complexity analysis of the different schemes and confirm through experimentation that the necessary computations can be readily carried out on a modern workstation without any significant issues.