Base-Calling in DNA Pyrosequencing with Reconfigurable Bayesian Network

Abstract

A reconfigurable computing method based on dynamic Bayesian learning network is proposed for base-calling in pyrosequencing from microarray gene expression data. Due to long memory and stochastic non-idealities in the pyrosequencing process, exact inference on the proposed dynamic Bayesian learning network is computationally prohibitive in both run-time and memory usage for reasonable problem sizes. To circumvent these issues, we design a reconfigurable Bayesian learning network, whereby processing nodes evaluate posterior probabilities of all states in parallel and crossbar switch realizes network topology that interconnects all processing nodes. The success of the proposed method is demonstrated by a prototype system implemented with Berkeley Emulation Engine 3 (BEE3) board, which achieves close to 2 times increase in read length and about 3 orders of reduction in run-time than previously reported for both experimental and simulated pyrosequencing data.