DNA Sequence Trace Reconstruction Using Deep Learning

Deciphering DNA sequences is fundamental to unlocking the mysteries of life, but the high dimensionality and complexity of biological sequence data significantly hinder knowledge discovery. In particular, the challenges of sequence length, repetitive regions, and structural complexity make it difficult to directly reconstruct complete DNA sequences from raw data. Therefore, this paper proposes a DNA sequence trace reconstruction model, DNARetrace, which performs preprocessing and dataset construction, and then employs a Bidirectional Fourier-Kolmogorov-Arnold Network (Bi-FKGAT), using an extremely unbalanced loss function for link prediction, so as to reconstruct the original DNA sequence. In multi-angle experiments using both simulated and real data, DNARetrace successfully reconstructs DNA sequence traces across large-scale datasets derived from various DNA sequencing methods, overcoming the bias of current approaches toward specific sequencing platforms, and achieves competitive outcomes in DNA storage and genomics downstream tasks. We further validated the expandability of the proposed methods in DNA sequence classification and metagenomic binning tasks. In summary, DNARetrace is compatible with various sequencing scenarios; it reduces the difficulty of discovering novelty knowledge directly from high-complexity raw data, and it provides a reusable tool to accelerate DNA sequence processing and applications.