AIEdit: Alignment-free genome assembly polisher trained on spaced seed match patterns

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genome assembly polishing is crucial for accurate downstream analyses like variant calling and clinical genomics.
• Long-read sequencing improves assembly contiguity but faces challenges with base-level accuracy due to inherent sequencing errors.
• Current polishing tools have limitations, including long run times (alignment-based), poor performance in high-error regions (k-mer-based), or platform specificity (ML-based).

Purpose of the Study:

• To develop AIEdit, a computationally efficient and alignment-free machine learning-based genome polishing tool.
• To create a polisher that generalizes across different sequencing platforms and effectively handles dense error patterns.
• To provide a scalable solution for accurate genome polishing.

Main Methods:

• AIEdit combines spaced seed matching with a neural network for alignment-free error detection and correction.
• The method was trained on simulated and experimental DNA sequencing data.
• Performance was benchmarked against existing tools like ntEdit, POLCA, Medaka, and JASPER.

Main Results:

• On simulated human long-read data, AIEdit reduced error rates by 58%, outperforming ntEdit (21%) and completing faster than POLCA and Medaka.
• AIEdit used significantly less memory (230 GB) compared to JASPER (689 GB).
• For experimental Oxford Nanopore Technologies data, AIEdit increased the Merqury quality score (QV) from 28.7 to 32.9, achieving accuracy comparable to Medaka in a fraction of the time.

Conclusions:

• AIEdit offers a scalable and accurate solution for genome polishing, addressing limitations of existing methods.
• The alignment-free, machine learning approach generalizes across sequencing platforms and excels in correcting dense error patterns.
• AIEdit enables more reliable and efficient downstream genomic analyses by enhancing base-level accuracy in genome assemblies.