Digital data storage typically relies on silicon media, where errors are modeled as clean, symmetric bit-flips. DNA-based data storage operates in a fundamentally different regime. When information is encoded into nucleotide sequences ($A, C, G, T$), it is subjected to complex biochemical processes—such as chemical synthesis, polymerase chain reaction (PCR) amplification, and high-throughput sequencing—that introduce high rates of non-random, asymmetric noise.
As noted by Kiah et al., error rates during synthesis and sequencing can reach up to 15% for long fragments. These errors are not merely substitutions; they frequently manifest as insertions, deletions (indels), and complex structural duplications. If traditional error-correcting codes (ECC) designed for electronic channels are applied blindly to DNA, they require massive data redundancy—often exceeding 30% of the total storage capacity, as highlighted by Pan et al.. Because chemical synthesis is highly expensive, this overhead creates a prohibitive economic barrier.
To overcome this, a major research direction focuses on designing noise-resilient molecular encoding and decoding frameworks that are natively tailored to the specific biochemical biases of the DNA channel.
Coding for Biological Error Profiles
To build efficient codes, researchers must mathematically formalize the unique ways DNA mutates or misbehaves during storage and replication.
Duplication Channels
One prominent biological error pattern is tandem and palindromic duplication, where enzymes accidentally repeat a segment of the sequence or its reverse complement. Standard error-correcting codes fail here because a duplication acts as a variable-length insertion of existing data rather than random noise.
To address this, Lenz et al. developed specialized duplication-correcting codes that exploit the structural constraints of these repetitions, offering far greater efficiency than general-purpose burst-error correction. However, biological duplications are rarely perfect. Tang et al. extended this framework to "noisy duplication" channels, where the duplicated segment contains subsequent substitution errors. As synthesized by Barg et al., designing codes for these channels requires a deep understanding of sequence reconstruction graphs to distinguish intentional data from biological artifacts.
Profile Vectors and Indel Resistance
Rather than forcing the system to fully reconstruct a single, pristine sequence from a sea of noisy reads, alternative approaches alter how the data is represented. Kiah et al. modeled the read-write process as a channel that outputs noisy $l$-gram profile vectors. By combining restricted de Bruijn graphs with rational polytopes, they demonstrated that data can be accurately retrieved from undersampled and noisy sequence pools without the computational overhead of full sequence assembly.
When sequence-level mapping is required, integrating traditional algebraic codes with biological constraints is essential. For instance, Zheng et al. paired Levenshtein codes—which are mathematically optimized to handle insertions and deletions—with deep learning architectures to maintain high data fidelity even at high baseline error rates.
Exploiting Soft Information and Physical Constraints
Another way to mitigate noise without adding physical redundancy is to extract more information from the physical readout process itself.
Soft-Information Decoding
When a sequencing machine reads a DNA strand, it does not just output a string of bases; it generates raw physical signals (like ionic current changes or fluorescence intensities) and associated quality scores. Traditional systems discard this "soft information" and perform hard decoding on the text string.
Lu demonstrated that by scaling the Log-Likelihood Ratios (LLRs) fed into Low-Density Parity-Check (LDPC) decoders based on real-world biochemical error statistics, one can drastically reduce the number of redundant physical reads required to recover the data. Similarly, Park et al. designed a decoding method that leverages the variable lengths of sequencing reads alongside quality scores, lowering the financial cost of data retrieval by making the decoder smarter rather than making the physical pool larger.
Constrained Coding for Nanopores
In nanopore sequencing, a single strand of DNA passes through a microscopic pore, and the machine measures disruptions in an electrical current. A major source of noise here is "segmentation errors," which occur when consecutive nucleotides produce highly similar electrical signals, causing the basecaller to lose track of how many bases passed through.
Standard Sequence: A — G — G — G — T (Current plateaus; causes deletion/segmentation error)
Constrained Code: A — G — C — G — T (Current fluctuates distinctly; error prevented)
To eliminate this issue at the source, Whritenour et al. introduced constrained coding. By algorithmically restricting the encoder to only generate DNA sequences that yield highly distinct, fluctuating electrical currents, they bypassed the limitations of deep learning basecallers and slashed sequencing noise up to 6-fold.
Bypassing Bit-Level Perfection via Perceptual Reconstruction
For certain data types, such as images or video, demanding 100% exact bit-level recovery introduces unnecessary overhead. A parallel strategy within this research direction replaces traditional error-correction codes entirely with domain-specific, lossy reconstruction.
Pan et al. (2019) introduced a "bits-for-brains" approach for image storage. Instead of using physical material to store redundant error-correcting bits, they allowed the DNA sequences to be corrupted by synthesis and sequencing noise. Upon readout, they used deep learning-based image inpainting and specialized quantization to gracefully reconstruct high-quality images from the corrupted sequences.
The authors later extended this philosophy to a hybrid storage medium called "2DDNA" (Pan et al. (2022)), which writes data both into the sequence itself and as physical "nicks" on the sugar-phosphate backbone. By relying on machine learning to resolve the resulting molecular noise, they achieved exceptionally high information density while avoiding the compounding costs of traditional biochemical error correction.
Key Foundational Work
Codes for DNA Sequence Profiles (2015) — contributed 1 concept elements to this direction.
Parallel DNA implementation of fast matrix multiplication techniques based on an $n$-moduli set (2009) — contributed 1 concept elements to this direction.