I Introduction
Our society is relying more and more on the digital world, and as a consequence, the demand for digital data storage are exponentially increasing. The duration for which this data should be preserved is also constantly growing, mainly for legal reasons. The current tools are not sufficient for both this reasons, first because the storage hardware cannot follow the storage demand anymore, and secondly, because the hardware replication needed to ensure the durability of the data storage is confronted with the same demand problem. For those reasons, alternatives to the current data storage technologies have to be developed quickly. Synthetic molecules, and particularly synthetic DNA are promising candidates as alternative data storage media [2]. Synthetic DNA molecules have huge encoding abilities (non-synthetic DNA is encoding living beings), are very small, and can be stored for very long periods of time. The DNA molecules are composed of a sequence of nucleotides - adenine (A), cytosine (C), guanine (G) and thymine (T) - that can be used to synthesize DNA and store digital data into it [3], [4]. Information theorists started to develop tools to encode data into DNA and some of them focused on image data storage [1, 5]. Recently, the JPEG DNA^{1}^{1}1https://jpeg.org/jpegdna/index.html research group started an exploration on the opportunity to develop a specific image coding solution adapted to the technologies of DNA synthesis to compress and store images into DNA-like sequences [6]. Some compression algorithm and coders have been developed specifically for this paradigm of data storage [7, 8, 5].
In this work we propose a variable-length quaternary encoder adapted to DNA data storage. This encoder is based on the Shannon-Fano algorithm revisited to work with a quaternary alphabet composed by the four letters (nucleotides) , and constrained by some biochemical restrictions. In this paper we considered mainly the constraint on homopolymers which consists in avoiding codewords constructed by the repetition of the same nucleotide more than 3 times, e.g., . This algorithm showed good performance compared to one of the best state-of-the-art coder proposed in [7]. Furthermore, we integrated the proposed quaternary coder in the image coding standard JPEG. The modified JPEG algorithm has shown better performance when compared to the image codec proposed in [1] which uses the Goldman coder of [7]. We finally established bounds to identify the margin of progress still available in the proposed variable-length coder, to identify the improvement that could benefit the proposed DNA-based JPEG codec.
Ii Context
Ii-a DNA data storage
A workflow has been designed to store data into those DNA molecules. The first step in this workflow is the construction of a coder able to represent data in a sequence of symbols of the alphabet . The encoded information has to be input into a DNA molecule, through the operation of synthesis. The molecules then are stored in hermetic capsules. To access the data, the last operations of the workflow are used: first the DNA molecule is sequenced, which means that we read the sequence of nucleotides from the molecule. Then this nucleotide sequence has to be decoded to retrieve the data. In some of those processes, (synthesis, sequencing and storage), the DNA molecules are either manipulated or stored. During those events, a lot of factors can alter the integrity of the molecule and introduce errors in the nucleotide sequence. Those errors can be substitutions (a nucleotide is changed into another one, for example an becomes a ) and indels (random insertions and deletions of nucleotides in the sequence).
Furthermore, to reduce the noise level of these different operations, the DNA coded information stream must respect some biochemical constraints on the combinations of bases that form a DNA fragment: homopolymers, high/low content and repeated patterns should be avoided. Designing a coding algorithm specific to synthetic DNA data storage appeared as a solution to accurately tackle the issues that come with the DNA storage workflow, particularly the noise produced in the workflow. Designing compression algorithms specific to DNA might also optimize the compression rate at which we store the data, which is still an important topic because of the high cost of the DNA synthesis.
Ii-B Image coding into DNA: A solution based on JPEG
The codec proposed in [1] is a modified version of the classic JPEG standard algorithm, adapted to encode data into DNA. The classic JPEG algorithm compresses and encodes images with the use of two main encoders. Those encoders are used to encode the DC and AC indices obtained through the Discrete Cosine Transform (DCT) computed on the image, block by block. The first one is the Huffman coding that encodes the run/categories of the AC indices, and the categories of the DC indices obtained through the DCT. In [1], this coder has been replaced by the Goldman encoder [7] that generates, instead of binary codes, quaternary DNA-based codes. This work focuses on improving this specific coder to increase the performance of the general algorithm. The second coder used in the JPEG standard is a fixed-length coder that codes the values of the DC and AC indices obtained through the DCT. In [1], this coder has been replaced by the fixed-length quaternary coder proposed in [8]. The figure 1 presents the whole workflow of the JPEG image coder adapted to DNA coding that has been proposed in our previous work [1].
Iii Huffman Tree Coding
Iii-a Background
Entropy coding methods are variable-length lossless compression methods that represent the different symbols of the source it encodes. The codes are estimated with the help of a frequency table. For every symbol that needs to be encoded, the number of appearances of this symbol in the source is computed. Finally, the length of the codes representing every symbol varies with its probability of appearance: if a symbol is more frequent, the associated code will be short, if it is less frequent, the code will be long.
Symbol | Probability | Code |
---|---|---|
a | 0.40 | 0 |
b | 0.05 | 1010 |
c | 0.18 | 110 |
d | 0.07 | 1011 |
e | 0.20 | 111 |
f | 0.10 | 100 |
The Huffman algorithm is the optimal symbol-by-symbol entropy coding method. Some theoretical limits on its performance have been established with the help of Kraft’s inequality. If the source
is represented by a random variable
with a probability distribution
, the expected length of the Huffman code generated by this source can be defined as:(1) |
being the codeword representing the symbol , the length of this codeword and the set of all possible symbols. Given the base in which the Huffman algorithm encodes the source, the source’s Shannon entropy can be defined as:
(2) |
Shannon’s source coding theorem [9] showed that, like for any other source code, the expected length of the -ary Huffman code representing the source of random variable follows the law:
(3) |
Iii-B The Goldman coder
Nick Goldman introduced in [7] an entropy coding method adapted to DNA. This algorithm satisfies one of the biochemical constraints presented earlier: it doesn’t generate codes with homopolymers. Homopolymers are the repetition of the same nucleotide base several times in a code or in a code stream (for example or ). To avoid those error-prone patterns, the Goldman code is generated by combining a ternary Huffman code and a rotating quaternary transcoder. The ternary Huffman coding generates codes using the alphabet {0,1,2}. The rotating transcoder translates this sequence into a quaternary sequence on the alphabet , without reusing the same element twice in a row. Thanks to this rotating transcoder, no homopolymers will be generated by the Goldman coder.
Previous nucleotide | 0 | 1 | 2 |
A | T | C | G |
T | A | C | G |
C | A | T | G |
G | A | T | C |
The rotating transcoder doesn’t affect the length of the codes generated by Huffman since it transcodes the source ternary symbol per ternary symbol. For that reason, the length of the codes for every symbol, and the expected length of the final code, are only conditioned by the Huffman coder. Hence, we can establish a boundary to the expected length of the entire Goldman coder :
(4) |
Iii-C Limitations for constrained quaternary codes
As previously mentioned, the theory has shown that the Huffman tree coding algorithm was the best entropy coding algorithm when source symbols are independent and identically distributed, i.i.d. source. This means that the quaternary Huffman tree coding algorithm will give the best performance in quaternary entropy coding of a i.i.d. source. The algorithm in itself is the theoretical limit for any quaternary entropy coding, hence the quaternary Huffman code is a bound for any quaternary entropy coder :
(5) |
being the set of all the possible quaternary entropy coders. The problem for DNA data storage is that, as we said previously, an unconstrained code will generate a lot of noise. This was the reasoning behind the development of the Goldman algorithm, and is still a problem that has to be considered. Thus, one cannot simply use the quaternary Huffman coder. This means that to develop a new entropy coder for DNA with improved performance over the Goldman algorithm or any ternary entropy coder , a good objective in terms of performance would be to design a code such as:
(6) |
Iv Proposed variable length quaternary encoder
Iv-a Motivations
The goal of the Goldman coder [7] is to constrain the code and avoid the apparition of homopolymers. The experiments have shown that the biochemical processes of DNA data storage (synthesis, storage and sequencing) can tolerate homopolymers until a certain length (3 for Illumina sequencers and around 5 for Nanopore sequencers), otherwise, the noise starts increasing dramatically [10]. This means that the Goldman coder, that does not allow two consecutive identical nucleotides, is too restrictive when avoiding homopolymers.
In order to meet the homopolymer constraint without losing too much coding potential, we proposed a solution that consists in a quaternary tree-coder constrained at certain positions to avoid too long homopolymers. Since tree structures are used for the code construction, the most simple solution was to constraint the arity (the number of sons of each node) of the tree at certain depths.
The classic Huffman algorithm builds the coding tree from the leaves up to the root, and does not hold any information concerning the final depth of the tree nor the final depth of any of the nodes in the tree. Designing our algorithm from the Huffman algorithm, where there is a local control of the tree structure, did not appear as the most efficient solution, especially for obtaining a balanced tree.
Instead, we based our solution on the classic Shannon-Fano recursive algorithm which builds the coding tree from the root, meaning that the current depth can be propagated through the recursions.
Iv-B Proposed constrained tree coder
The coder is implemented with two algorithms. The algorithm 1 is a simple initialization: it starts the whole construction of the tree, at the root (at depth 1). The algorithm 2 manages the recursive calls in function of two parameters: a global one, the maximum length of homopolymer that is allowed , and a local one, the depth at which the call was made. If the depth is in the class 0 of congruence modulo , the tree will be constrained to ternary, otherwise, it will not be constrained. The figure 2 shows an example of such a coding tree.
In the algorithms 1 and 2, we assume that the symbols in the set and their frequency table have both already been sorted by decreasing frequency. In the algorithm 2, the function slices into divisions the set of symbols and frequencies , with the cumulated frequencies of every slice being as close as possible to each other. The function takes the list containing 3 or 4 trees and generates a tree with a new root, and each son of this root being one of the trees in the input sequence . The first son is labeled with a 0, the second with a 1, the third with 2 and the last one with a 3.
Iv-C Transcoding into DNA
The algorithm described earlier generates a code containing for every symbol a sequence of bases: 0, 1, 2, and 3. These sequences still need to be translated into DNA sequences respecting the homopolymer constraint described earlier.
To transcode the sequences into DNA, two different kinds of transcoding tables are used. According to the tree structure, the symbol in the sequence is a ternary symbol if and a quaternary one otherwise, being the maximum homopolymer length. When the symbol is ternary, it is translated into a nucleotide with the Goldman transcoding table (Table II), otherwise, the quaternary trancoding table (Table III) is used.
Symbol | 0 | 1 | 2 | 3 |
---|---|---|---|---|
Nucleotide | A | T | C | G |
Iv-D Property of the proposed tree coder
If is the maximum homopolymer length that can be accepted, constraining the coding tree to ternary at every depth that is a multiple of , would prevent the apparition of homopolymers of length more than (see Property 1).
Proposition 1.
Let be a code obtained thanks to the proposed coder (algorithm 1). Let be the set of homopolymer patterns found in the codeword of the symbol . Let be a homopolymer subsection of . Let , and respectively be the depth at wich the section starts, respectively ends, in the coding tree.
Then, with,
(7) |
Proof.
By contradiction: If a code has a codeword containing a homopolymer pattern of length bigger than , the homopolymer can be located in the coding tree between depths and with . By construction of the coding tree, there is a constrained depth in the middle of the homopolymer pattern, where the Goldman transcoding table II is used, hence two consecutive nucleotides are not identical, so is not a homopolymer pattern. ∎
V Experimental results
V-a Experiments on synthetic sources
Before integrating the coder inside a complete image codec, we estimate the raw gain that we can obtain in comparison to the Goldman coder (). We also compare our results to a Huffman-based constrained quaternary code () that we implemented, even if as mentionned in the motivations subsection (IV-A) it was not considered as a good candidate from the beginning.
The metric used to compare the performance of variable-length coders is the expected length, that has been previously introduced in equation (1).
We used i.i.d
Gaussian sources for those experiments. It was made sure to generate enough samples in the sources to approach the Gaussian distribution: 100 relisations, each one containing 10000 samples. Each sample is quantized using a dictionnary composed of 162 symbols. Since the Goldman coder is based on a Huffman ternary tree coder, comparing the expected length of the two coders to the ternary entropy (equation (
2) with ) is appropriate: if the new coder beats the ternary entropy, then it is better than any tenary-based entropy coder for DNA (see equation (6)).3.48 | 3.56 | 3.81 | 4.31 | 4.39 | 4.45 |
The length expectancy presented in table IV shows a gain of 0.64 nucleotide per symbol for the proposed algorithm () over Goldman (), and a difference of 0.25 nucleotide per symbol with the unconstrained quaternary Huffman coder (). As expected in the motivations subsection (IV-A), in comparison, the constrained Huffman-based coder clearly underperforms. Not having a control over the depth of the coding tree means that it is harder to balance it and optimize the expected length.
As a last experiment, the performance of the coder was evaluated on a source following a distribution that is similar to the distribution of the AC indices produced by a JPEG codec in most natural images. The results presented in the table V have been generated using a source that follows a frequency table of the AC coefficients computed on the Kodak image dataset^{2}^{2}2http://r0k.us/graphics/kodak/. Here again, the proposed constrained Shannon-Fano quaternary encoder is the best performing.
1.21 | 1.39 | 1.43 | 1.53 | 1.61 | 1.64 |
V-B Experiments on image coding
As explained in the introduction, the variable length coder is only one of the coders used in the JPEG DNA experimentation software’s algorithm [6]. It is used to encode the category of a DC coefficient or the run/category (a combination of the AC category and of the run-length-encoded sequence of zeros preceding this AC coefficient) of an AC coefficient. In the classic JPEG DNA experimentation software’s algorithm, the Goldman encoder encodes those values. To study if and how the proposed variable length coding algorithm improves the compression, we need to make it so that during the comparison, the only variant is the choice of the variable length coder for the categories and run/categories.
To evaluate the performance of the two codecs we computed the compression ratio, expressed in bits per nucleotide. It corresponds in our case to the ratio of the size of the input image (in bits) over the total size of the DNA sequences (i.e., the number of nucleotides) generated to compress the image. This compression ratio is evaluated for the same quality parameters for the two different codecs. The PSNR in function of the compression ratio is presented in the figure 3. The kodak dataset was used to evaluate the performance of the new image compression algorithm.
Through all the images in the test image dataset, our new variable-length coder has consistently improved the performance of the compression algorithm over the one using the Goldman entropy coder. For example with the image kodim23 from the kodak dataset, the compression rate gain varies between 0.5 and more than 2 bits per nucleotide as shown in the figure 3.
The relative gain is less important than when using the entropy coder as a standalone coder because in the case of the image codec, other coders are used, and those have not been optimized.
Vi Conclusion
We have developed a new variable-length tree coder adapted to DNA data storage. This new coder has shown improvements over the Goldman entropy coder that is usually chosen in variable length DNA coding problems. We have even shown that this new coder would do better than any ternary entropy coder by having an expected lenth inferior to the ternary entropy of the source. We then have integrated this new entropy coder inside an image codec to analyze its influence on the performance of the codec. We have seen consistent improvements on the compression rate with the dataset that we used for testing.
In the future, it would be interesting to study the resilience of this new encoder to the noise produced by the different processes of the DNA storage workflow, namely the synthesis, storage and sequencing. Developing new consensus algorithm, or maybe even error correction codes could also bring a lot to this new coder. Finally, it could be interesting to mathematically establish theoretical bounds to the performance of the proposed coder.
References
- [1] M. Dimopoulou, E. G. San Antonio, and M. Antonini, “A JPEG-based image coding solution for data storage on DNA,” EUSIPCO, 2021.
- [2] S. M. H. T. Yazdi, R. Gabrys, and O. Milenkovic, “Portable and error-free DNA-based data storage,” Scientific Reports, vol. 7, no. 1, pp. 1–6, 2017.
- [3] R. N. Grass, R. Heckel, M. Puddu, D. Paunescu, and W. J. Stark, “Robust chemical preservation of digital information on dna in silica with error-correcting codes,” Angewandte Chemie International Edition, 2015.
- [4] Y. Erlich and D. Zielinski, “Capacity-approaching DNA storage,” bioRxiv, p. 074237, 2016.
- [5] C. Pan, S. M. Hossein Tabatabaei Yazdi, S. Kasra Tabatabaei, A. G. Hernandez, C. Schroeder, and O. Milenkovic, “Image processing in dna,” in ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 8831–8835.
- [6] M. Antonini, L. Cruz, E. Da Silva, D. Melpomeni, T. Ebrahimi, S. Foessel, E. Gil, S. Antonio, G. Menegaz, F. Pereira, X. Pic, A. Pinheiro, and M. Raad, “DNA-based Media Storage: State-of-the-Art, Challenges, Use Cases and Requirements,” Oct. 2021, working document of the JPEG DNA AHG. [Online]. Available: https://hal.archives-ouvertes.fr/hal-03447138
- [7] N. Goldman, P. Bertone, S. Chen, C. Dessimoz, E. M. LeProust, B. Sipos, and E. Birney, “Towards practical, high-capacity, low-maintenance information storage in synthesized DNA,” Nature, vol. 494, no. 7435, p. 77, 2013.
- [8] M. Dimopoulou, M. Antonini, P. Barbry, and R. Appuswamy, “A biologically constrained encoding solution for long-term storage of images onto synthetic DNA,” EUSIPCO, 2019.
- [9] C. E. Shannon, “A mathematical theory of communication,” Bell System Technical Journal, vol. 27, pp. 379–423, 1948.
- [10] M. Ross, C. Russ, M. Costello, A. Hollinger, N. J. Lennon, R. Hegarty, C. Nusbaum, and D. B. Jaffe, “Characterizing and measuring bias in sequence data,” Genome Biology, vol. 14, no. 5, p. R51.