Basecalling for DNA Storage

DNA is a promising medium for data storage with its high information density and stability. To retrieve information stored in DNA, sequencing technologies are used to read the encoded bases. The raw signals from sequencing are mapped to a sequence of {A,C,T,G} by machine learning algorithms known as basecallers. Currently, basecallers are optimised mainly on biological DNA instead of focusing on characteristics or constraints unique to data-encoding DNA. Taking advantage of these unique artificial features to fine-tune and adapt the architecture, we demonstrate for the first time that DNA coding scheme constraints can be leveraged to optimise basecallers. Using low-rank adaptation on the basecalling model, we achieve substantial gains with high resource efficiency. Additionally, constraint-aware beam search provides improvements without requiring model retraining.