Iterative Learning for Reference-Guided DNA Sequence Assembly From Short Reads: Algorithms and Limits of Performance

Recent emergence of next-generation DNA sequencing technology has enabled acquisition of genetic information at unprecedented scales. In order to determine the genetic blueprint of an organism, sequencing platforms typically employ the shotgun sequencing strategy to oversample the target genome with a library of relatively short overlapping reads. The order of nucleotides in the reads is determined by processing the acquired noisy signals generated by the sequencing instrument. Assembly of a genome from potentially erroneous short reads is a computationally daunting task even when a reference genome exists. Errors and gaps in the reference, and perfect repeat regions in the target, further render the assembly challenging and cause inaccuracies. Here, we formulate the reference-guided sequence assembly problem as the inference of the genome sequence on a bipartite graph and solve it using a message-passing algorithm. The proposed algorithm can be interpreted as the well-known classical belief propagation scheme under a certain prior. Unlike existing state-of-the-art methods, the proposed algorithm combines the information provided by the reads without needing to know reliability of the short reads (so-called quality scores). Relation of the message-passing algorithm to a provably convergent power iteration scheme is discussed. To evaluate and benchmark the performance of the proposed technique, we find an analytical expression for the probability of error of a genie-aided maximum a posteriori (MAP) decision scheme. Results on both simulated and experimental data demonstrate that the proposed message-passing algorithm outperforms commonly used state-of-the-art tools, and it nearly achieves the performance of the aforementioned MAP decision scheme.