Remap: Assembly-Mediated Long-Read Sequence Re-mapping

Remap is a projection-based re-mapping framework designed to improve the alignment consistency of long-read sequencing data (PacBio HiFi and ONT).

Standard read-to-reference alignment often suffers from ambiguity in repetitive regions (e.g., segmental duplications) and breakpoint jitter in structural variants (SVs). Remap addresses this by using assembly contigs as a bridge: reads are first mapped to their local assembly, and then projected onto the reference genome. This approach significantly improves the precision and recall of downstream SV and SNP/Indel calling.

Key Features

• Assembly-Mediated Projection: Resolves multi-mapping ambiguity by utilizing the longer context of assembled contigs.
• Exact Anchor Consistency: Uses strict one-to-one mapping anchors to prevent breakpoint drifting.
• High Precision in Complex Regions: Specifically optimized for low-mappability regions, segmental duplications, and extreme GC regions.
• Versatile Support: Works with PacBio HiFi, ONT(need existing assembly), and R10+ (HQLR) data.
• Compatible Output: Generates standard BAM files compatible with existing callers like Sniffles2, SVIM, and DeepVariant.

Method Overview

Instead of aligning reads directly to the reference ($R \to G$), Remap decomposes the problem into two easier steps:

1. Read-to-Assembly ($R \to A$): Reads are aligned to de novo assembled contigs.
2. Assembly-to-Reference ($A \to G$): Contigs are aligned to the reference genome.
3. Coordinate Projection: Valid anchors are projected from $R$ to $G$ via $A$, and gaps are filled using local alignment.

Installation

Prerequisites

• Python >= 3.10
• minimap2
• samtools
• hifiasm (if raw reads are provided for assembly)

Option 1: Using Conda (Recommended)

We provide a environment file.

git clone https://github.com/micahvista/Remap.git
cd Remap
conda env create -f remap_environment.yml
conda activate remap_env
pip install .

Usage

Remap requires an input read set, a reference genome, and a working directory. It can perform de novo assembly internally or use a pre-computed assembly.

Quick Start

Run full pipeline (Assembly + Remap) for PacBio HiFi data:

remap -i "reads/*.fastq.gz" \
      -r reference.fasta \
      -w ./workdir \
      -o ./output/sample_name \
      -d hifi \
      -t 16

Run full pipeline (Assembly + Remap) for ONT R10.4 data:

remap -i "reads/*.fastq.gz" \
      -r reference.fasta \
      -w ./workdir \
      -o ./output/sample_name \
      -d hqlr \
      -t 16

Argument	Required	Description
-i, --inputreadpath	Yes	Path to input reads (supports wildcards, e.g., *.fq.gz).
-r, --refpath	Yes	Path to the reference genome (FASTA).
-w, --workdir	Yes	Working directory for intermediate files.
-o, --outputprefix	Yes	Prefix for the final output BAM file.
-d, --datatype	Yes	Data type: ont (R9.x), hqlr (R10.x), or hifi (PacBio HiFi).
-t, --threads	No	Number of threads (default: 8).
-a, --asmdata	No (yes if ont)	Path to existing assembly (FASTA or GFA). If omitted, hifiasm is run.
--localasm	No	Enable local assembly mode for reads containing SVs.
--asm2ref	No	Path to existing Assembly-to-Reference BAM.
--read2asm	No	Path to existing Read-to-Assembly BAM.
--rutg	No	Use raw unitig graph for assembly (default: False).

Output

The pipeline generates a standard BAM file containing the re-mapped reads.

Merged Output:

{outputprefix}all.bam (Contains both re-mapped and rescued reads).

Tags:

Includes standard SAM tags plus custom tags for alignment consistency.

Method Overview

Read-to-Assembly ($R \to A$):

Reads are aligned to high-quality contigs to resolve local ambiguities.

Assembly-to-Reference ($A \to G$):

Contigs are aligned to the reference, spanning repetitive regions with uniqueness.

Projection:

Coordinates are projected mathematically from $R$ to $G$ via $A$ using exact anchors.

Refinement:

Gaps between anchors are filled using local alignment (Smith-Waterman) to ensure base-level precision.

Performance Benchmarks

Based on the GIAB HG002 T2T benchmark: Structural Variants: Remap increases F1 scores by +0.04 ~ +0.05 over direct mapping. Segmental Duplications: Significant recall improvement ($\Delta R \approx +0.15$ for ONT) in these difficult regions. Small Indels: Achieves an F1 score of 0.82 (vs 0.77 for direct mapping) for HiFi data22.

Citation

If you use Remap in your research, please cite:

This project is licensed under the MIT License.