NeuralPolish

NeuralPolish polishes nanopore genome assemblies by identifying and correcting base-level errors using an alignment feature matrix and bi-directional GRU networks.

Key Features:

• Alignment Feature Matrix Construction: Represents read-to-assembly alignments as a matrix with each row corresponding to an individual read and each column representing aligned bases at specific contig positions.
• Bi-Directional GRU Networks: Two orthogonal bi-directional Gated Recurrent Unit (GRU) networks constitute the core architecture, with the first network processing the alignment matrix row-by-row to extract per-read sequence information.
• Probability Distribution Calculation: The second bi-directional GRU processes the feature matrix column-by-column to calculate probability distributions across aligned bases for error identification.
• CTC Decoder and Greedy Algorithm: Uses a Connectionist Temporal Classification (CTC) decoder combined with a greedy algorithm to generate the final polished sequence from the computed probabilities.

Scientific Applications:

• Benchmarking and comparison: Evaluated on five real datasets and assemblies produced by Wtdbg2, Flye, and Canu, and benchmarked against Racon, MarginPolish, HELEN, and Medaka.
• Assembly polishing for nanopore data: Demonstrated to produce more accurate assemblies with fewer errors compared to the referenced polishing methods.

Methodology:

Construct an alignment feature matrix from read-to-assembly alignments (rows = reads, columns = aligned bases); process the matrix row-by-row with a bi-directional GRU to extract per-read features; process column-by-column with a second bi-directional GRU to compute probability distributions across aligned bases; decode the polished sequence using a CTC decoder with a greedy algorithm.