XHap: haplotype assembly using long-distance read correlations learned by transformers.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Haplotype assembly from sequencing reads is computationally complex, especially for polyploid organisms.
• Existing methods struggle with read length limitations and sequencing errors, particularly for non-overlapping reads.

Purpose of the Study:

• To develop a novel method, XHap, for accurate haplotype assembly.
• To leverage deep learning, specifically transformers, to identify correlations between distant sequencing reads.

Main Methods:

• XHap utilizes transformers and their attention mechanism to learn dependencies between sequencing reads, even those that do not overlap.
• The method focuses on discovering correlations across large genomic distances.

Main Results:

• XHap demonstrates superior performance compared to state-of-the-art techniques in both diploid and polyploid haplotype assembly.
• The method is effective for both short and long sequencing reads, as shown in experiments on real and semi-experimental data.

Conclusions:

• XHap offers a significant advancement in haplotype assembly, addressing key challenges posed by ploidy and sequencing data limitations.
• The deep learning approach provides a powerful new tool for genomic research.