Halcyon

Halcyon performs basecalling for nanopore sequencing by mapping raw electrical signals to nucleotide sequences using an encoder-decoder neural network with monotonic attention.

Key Features:

• Neural Network Integration: Implements neural-network architectures adapted from machine translation to model complex relationships between nanopore signal and nucleotide sequences.
• Monotonic Attention Mechanism: Uses monotonic-attention to align input signal segments to output nucleotides without requiring bidirectional or global attention.
• No Pre-segmentation Required: Operates directly on continuous raw nanopore signal streams, avoiding explicit signal segmentation steps.
• Encoder–Decoder Model: Employs an encoder-decoder framework to learn direct mappings from raw signals to nucleotide sequences.

Scientific Applications:

• Detection of Structural Variations: Produces basecalls from nanopore long-read data to support identification of structural variants in genomes.
• Haplotype Phasing: Generates precise long-read sequences to assist haplotype phasing and analysis of complex genetic variation.

Methodology:

Uses an encoder-decoder neural model with monotonic attention to learn direct mappings from raw nanopore signals to nucleotide sequences without pre-segmentation; evaluated on a human whole-genome sequencing dataset and compared against third-party basecallers and Oxford Nanopore Technologies basecallers.