DNA is an ancient and efficient information carrier in living organisms. At present, it is thought to have great potential as an alternative storage medium because standard storage media can no longer meet the exponentially increasing data archiving demands. Compared with common information carriers, the DNA molecule exhibits multiple advantages, including extremely high storage density (estimated physical density of 455 EB per gram of DNA1), extraordinary durability (half-life >500 years (refs. 2,3)) and the capacity for cost-efficient information amplification.
Many strategies have been proposed for digital information storage using organic molecules, including DNA, oligopeptides and metabolomes4,5,6,7,8. Since current DNA sequencing technology has advantages in terms of both cost and throughput, storing digital information using DNA molecules remains the most well-accepted strategy. In this approach, the binary information from each file is transcoded directly into DNA sequences, which are synthesized and stored in the form of oligonucleotides or double-stranded DNA fragments in vitro or in vivo. Then, sequencing technology is used to retrieve the stored digital information. In addition, several different molecular strategies have been proposed to implement selective access to portions of the stored data, to improve the practicality and scalability of DNA data storage9,10,11.
However, the use of basic transcoding rules (that is, converting [00, 01, 10, 11] to [A, C, G, T]) generates some specific patterns in DNA sequences that result in challenges regarding synthesis and sequencing9,12,13. For example, single-nucleotide repeats (homopolymers) longer than 5 nt might introduce a higher error rate during synthesis or sequencing14,15. Meanwhile, because of the nature of complementary base pairing (with A pairing to T and G to C), DNA molecules may form structures such as hairpins or topological pseudoknots (i.e., secondary structure), which can be predicted by calculating the free energy from its sequence. It is reported that DNA sequences with stable secondary structure can be disadvantageous for sequencing or when using PCR for random access to and backup of stored information16,17,18,19. Additionally, DNA sequences with GC content <40% or >60% are often difficult to synthesize. Therefore, the length of homopolymers (in nt), the secondary structure (represented by the calculated free energy in kJ mol−1) and the GC content (in %) are three primary parameters for evaluating the compatibility of coding schemes.
Previous studies on transcoding algorithm development have attempted to improve the compatibility of the generated DNA sequences. Early efforts, including those of Church et al. and Grass et al., introduced additional restrictions in the transcoding schemes to eliminate homopolymers, but this came at the expense of reduced information density1,20,21. Later studies pioneered other base conversion rules without compromising the information density. For example, the DNA Fountain algorithm adopted Luby transform codes to improve the information fidelity by introducing low redundancy as well as screening constraints on the length of homopolymers and the GC content while maintaining an information density of 1.57 bits nt−1 (refs. 6,22). However, the major drawback is the risk of unsuccessful decoding when dealing with particular binary features due to fundamental issues with Luby transform codes. This approach relies on the introduction of sufficient logical redundancy, that is, at the coding level, for error tolerance to ensure successful decoding. This is different from physical redundancy, which refers to the synthesis of excess DNA molecules, that is, increasing the copy number of DNA molecules for each coding sequence23,24. Reducing the logical redundancy could lead to a high probability of decoding failure, but excessive logical redundancy will decrease the information density and significantly increase the cost of synthesis25. Furthermore, specific binary patterns using these early algorithms may also create unsuitable DNA sequences, with either extreme GC content or long homopolymers (Supplementary Table 1). Therefore, developing a coding algorithm that can achieve high information density but, more importantly, perform robust and reliable transcoding for a wide variety of data types in a cost-effective manner is necessary for the development of DNA-based information storage in practical applications25,26,27.
To achieve this goal, we propose herein the yin–yang codec (YYC) coding algorithm, inspired from the traditional Chinese concept of yin and yang, representing two different but complementary and interdependent rules, and we demonstrate its performance by simulation and experimental validation. The advantage of the YYC is that the incorporation of the yin and yang rules finally leads to 1,536 coding schemes that can suit diverse data types. We demonstrate that YYC can effectively eliminate the generation of long homopolymer sequences while keeping the GC content of the generated DNA sequences within acceptable levels. Two representative file formats (.jpg and .txt) were chosen for storage as oligo pools in vitro and a 54 kbps DNA fragment in vivo in yeast cells to evaluate the robustness of data recovery. The results show that YYC exhibits good performance for reliable data storage as well as physical density reaching the scale of EB per gram.
