BlockPolish

BlockPolish improves the base-level accuracy of de novo genome assemblies generated from long-read sequencing technologies such as Nanopore and PacBio by applying block-based segmentation, multiple sequence alignment in complex regions, and multitask bidirectional LSTM consensus prediction.

Key Features:

• Block Division Strategy: Employs a block divide-and-conquer approach that segments contigs into low-complexity and high-complexity blocks based on statistical analysis of aligned nucleotide bases.
• Multiple Sequence Alignment: Uses multiple sequence alignment to realign raw reads within high-complexity blocks to optimize alignments before consensus prediction.
• Multitask Bidirectional LSTM Consensus Prediction: Predicts consensus sequences using two multitask bidirectional LSTM networks trained to handle distinct error profiles in trivial and complex blocks.
• Indel Correction Efficiency: Targets insertions and deletions (indels) to reduce indel-related errors common in long-read assemblies.

Scientific Applications:

• Whole-genome polishing: Applied to whole-genome assemblies including NA12878 assembled with Wtdbg2 and Flye using Nanopore data.
• Comparative performance benchmarking: Demonstrated higher accuracy than Racon, Medaka, MarginPolish, and HELEN on tested assemblies.
• Cross-platform applicability: Applicable to assemblies generated from both Nanopore and PacBio long-read sequencing data.

Methodology:

Contig segmentation into blocks using statistical metrics of aligned nucleotide bases; multiple sequence alignment-based realignment of raw reads in high-complexity blocks; and consensus prediction using two multitask bidirectional LSTM networks trained on trivial and complex block error profiles.