Efficient Bidirectional RNNs for Substitution Error Correction in DNA Data Storage

With the exponential growth of digital data, DNA has emerged as a promising high-density storage medium owing to its durability and density. However, sequencing errors can affect the reliability of DNA storage systems. We use the gated recurrent unit (GRU) and long short-term memory (LSTM) to build two Bidirectional Recurrent Neural Networks (BiRNNs) that can fix substitution errors in data that is encoded in DNA. The model is trained and optimized on synthetic DNA strands with controlled errors. Evaluation of engineered sequences and data derived from real sequencing demonstrates reliable error correction across diverse test cases. Comparative analyses highlight the reduced complexity and improved efficiency of the GRU-based model over LSTM-based alternatives. Despite constraints imposed by the codeword dictionaries, the model exceeds the designed error correction capacity by learning sequence dependencies. Our approach improves reliability for high-density archival storage in DNA.