Neural Polar Decoders for DNA Data Storage

Henry D. Pfister; Ziv Aharoni

REVIEW 5 major objections 7 minor 37 references

Neural Polar Decoders for DNA Data Storage

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that neural polar decoders, trained only on input-output samples, can decode insertion-deletion-substitution and DNA-storage channels at O(AN log N) complexity, matching or beating trellis and deep-learning baselines in…

desk verdict Solid empirical extension of the NPD framework to synchronization-error channels and DNA storage, but the near-optimal and accurate-MI claims outrun the evidence; needs calibration work and a referee. read the letter →

arxiv 2506.17076 v1 pith:OAM5DNVK submitted 2025-06-20 cs.IT cs.LGmath.IT

Ziv Aharoni , Henry D. Pfister This is my paper

classification cs.ITcs.LGmath.IT MSC 94A1794B35

keywords neuralpolardecodersdeletionchannelinsertion-deletion-substitutionDNAdatastoragesuccessivecancellationdecodingmutualinformationestimationNanoporesequencingsynchronizationerrors

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Insertions and deletions wreck symbol alignment, and exact trellis decoders for those channels cost O($N^{4}$ log N). This paper argues that a neural polar decoder (NPD)—the recursive structure of a polar successive-cancellation decoder with each step replaced by a small neural network—can decode these channels in O(AN log N) time after training on input-output samples alone, with no explicit channel model. On deletion channels the reported frame-error rates are near those of the trellis decoder, while the mutual-information estimates sit between known lower and upper capacity bounds. On real Nanopore sequencing data, the NPD matches a state-of-the-art deep-learning decoder while using roughly thirty times fewer parameters. If true, this makes data-driven decoding a practical route for DNA storage pipelines.

What carries the argument

The machinery is the neural polar decoder: a successive-cancellation decoder whose embedding, check-node, bit-node, and embedding-to-LLR functions are shallow neural networks, so each recursive step costs a fixed network pass and total complexity becomes O(AN log N), with A set by the network size. To handle synchronization errors, the embedding takes the whole received word as input, because a single output symbol no longer corresponds to a single input symbol; the paper tests a CNN-based embedding for deletions and an attention-based embedding for IDS channels, both augmented with positional encodings. The two-pass MI estimator shares the F, G, and H networks between a real-channel pass and a zero-output pass, giving a sample-based estimate of mutual information that is then optimized by a gradient rule.

What would settle it

Take a trained multi-trace NPD and replace Equation (12)'s plain average with a learned or alignment-aware aggregation, such as concatenating trace embeddings or first aligning the reads; if the alignment-aware version gives a clearly lower frame error rate on Nanopore data at the same code rate, the averaging rule is losing information the decoder needs. A simpler test: shift the positions of a fixed set of K traces by a known offset and check whether the frame error rate changes materially, since a sufficient embedding should be largely insensitive to such shifts except through learned positional information.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the NPD's embedding network is what unlocks synchronization-error channels: instead of mapping each received symbol to a belief, the embedding maps the entire received word to per-symbol belief vectors, using positional encodings and either a convolutional network for deletion channels or a two-layer attention mechanism for general IDS channels. It further claims that when K noisy reads of a strand arrive, averaging their embeddings with a 1/sqrt(K) normalization preserves enough information for the successive-cancellation recursion, and that this suffices for Poisson-distributed trace counts. The evidence offered is that NPD mutual-information estimates fall within published capacity bounds on deletion and IDS channels, and that on real Nanopore data the NPD reaches bit error rates comparable to a strong deep-learning baseline at a fraction of the parameter count.

Load-bearing premise

The load-bearing premise is that averaging the per-read embeddings with 1/sqrt(K) normalization keeps enough alignment information for the successive-cancellation recursion; if that averaging discards how noisy reads line up with the original strand, the multi-trace DNA-storage results would degrade even though single-trace decoding works.

Editorial extensions

If this is right

Decoding complexity for deletion and IDS channels drops from O(N^4 log N) trellis decoding to O(AN log N), making long-block and list decoding practical.
NPDs can estimate and optimize input distributions for IDS channels without knowing the channel model, producing numerical capacity estimates that sit between published bounds.
On DNA channels with Poisson-distributed trace counts, the NPD inner code plus concatenated outer decoding reaches bit error rates around 0.02 for block lengths up to 2^19 bits.
On real Nanopore data, the NPD matches a state-of-the-art deep decoder in bit error rate while using about 3 million parameters instead of 100 million.
Because the decoder is trained from samples, new channel conditions only require new input-output pairs rather than a re-derived statistical channel model.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Our inference: the same per-read embedding averaging could be tested on other repetition channels, such as multi-pass flash reads or repeated wireless transmissions, where the decoder observes multiple noisy versions of the same codeword.
Our inference: because the NPD only needs samples, it could be retrained directly on reads from a new sequencing chemistry without re-deriving an error profile; the paper stops at fine-tuning a pre-trained model on one Nanopore dataset.
Our inference: the MI estimation procedure, though not a rigorous bound, may serve as a practical code-design oracle for IDS channels at the short block lengths used in real DNA storage, a regime where asymptotic capacity bounds are loose.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

If you skim the abstract, here is the plain version: the paper's headline claims overstate what is actually demonstrated. The genuinely new content is an empirical extension of the authors' earlier neural polar decoder (NPD) framework to channels with synchronization errors, using sequence-level CNN and attention embeddings, plus a multi-trace aggregation rule and experiments on real Nanopore data. That is real work and the engineering looks sensible. But "near-optimal decoding performance" rests on a comparison against the trellis SC decoder from Tal et al. where the NPD achieves slightly better FERs, and the authors themselves say that surprise is probably due to numerical errors in the trellis implementation. That attribution is a conjecture, not a verified fix. If the trellis is actually accurate, the NPD's advantage is unexplained and near-optimality is unsupported; if the trellis is buggy, the comparison should be redone. Either way, the abstract goes beyond what the evidence shows.

The MI estimates have a similar issue. The numbers in Figures 2 and Table I sit inside published bounds, which is encouraging, but they are single-point estimates computed as differences of two numerical upper bounds on entropy at finite block length, with no error bars or repeated-seed statistics. Remark 1 concedes the rigor problem. A single point inside an interval is not a demonstration of estimation accuracy. The clean test would be to calibrate the NPD's learned posteriors against exact trellis posteriors on small block lengths where the trellis is computationally feasible; that experiment is missing.

What the paper does well: the sequence-level embedding design (CNN for deletion-only, attention for IDS) is a sensible response to the variable-length alignment problem; the padding scheme is clean; the multi-trace aggregation rule is explicitly flagged as heuristic, which is honest, and the empirical results on the Nanopore data suggest it does not obviously break. The authors also cite their earlier work for the MI estimator and input-distribution optimizer, so there is no novelty inflation there. The complexity reduction to O(AN log N) relative to the trellis's O(N^4 log N) is real, and the SCL results are a useful extra that the trellis decoder cannot practically reach.

The main missing pieces are code and data, which would let others check the empirical claims, and the absence of any calibration of learned posteriors against exact ones at small N. The central empirical direction is sound, but the near-optimality and accurate-MI wording needs to be toned down or backed by that calibration.

This paper deserves a serious referee: it is a competent empirical contribution to a subfield that has few practical decoders for synchronization-error channels. My own verdict would be conditional acceptance, with the calibration test, error bars, and a revised abstract as conditions.

Referee Report

5 major / 7 minor

Summary. This paper extends neural polar decoders (NPDs) to insertion-deletion-substitution (IDS) and deletion channels by replacing the embedding function of a successive-cancellation polar decoder with CNN- or attention-based networks that process the entire received sequence. It proposes averaging per-trace embeddings for multiple noisy reads (Eq. 12), estimates mutual information and optimizes input distributions with the NPD, and evaluates the approach on synthetic deletion/IDS channels against trellis decoding and capacity bounds, and on synthetic and real Nanopore DNA storage data against DNAformer and other reconstruction methods. The main claims are that NPDs achieve near-optimal frame-error-rate performance on deletion channels, accurate mutual-information estimates, O(AN log N) decoding complexity, and competitive DNA-storage performance with far fewer parameters than DNAformer.

Significance. If the main claims hold, the paper offers a practical, model-free decoding route for synchronization-error channels and DNA storage, with much better complexity scaling than trellis SC decoding and with list decoding enabled. The paper is honest about several limitations: Remark 1 states that the MI estimates are not rigorous bounds, and Section IV-C labels the multi-trace aggregation rule as heuristic. The external benchmarking against trellis FERs and published capacity bounds is a strength, as is the use of real Nanopore data. However, the central 'near-optimal' and 'accurate MI' claims currently rest on comparisons that are either conjectured to be affected by numerical artifacts or are single-point estimates without error bars, so the significance is conditional on the additional validation described below.

major comments (5)

[V-4, Fig. 3b] The near-optimal FER claim is supported only by a comparison in which the NPD slightly outperforms the trellis-based SC decoder, and the paper attributes this to 'numerical errors in the implementation of the trellis-based decoder' without demonstrating them. If the trellis implementation is accurate, the NPD may share or amplify those errors; if it is inaccurate, the comparison should be repeated with a corrected trellis or against exact posteriors for small N. This point is load-bearing because the abstract's 'near-optimal decoding performance' claim depends on it.
[V-2, Remark 1, Table I] The MI estimates are reported as single-point values computed at finite block length as the difference of two numerical upper bounds on entropy, with no error bars, multiple seeds, or exact-MI reference for small N. A single value falling inside published bounds is weak evidence of accuracy, especially because the input-distribution optimization loop uses the same NPD to estimate MI. Please add seed statistics and validate against exact MI for small block lengths.
[IV-C, Eq. (12)] The multi-trace aggregation rule, averaging per-trace embeddings with 1/sqrt(K) normalization, is explicitly acknowledged to be heuristic, yet it is the mechanism underlying all multi-trace DNA-storage results in Section VI. Without an ablation against alternatives (e.g., learned aggregation or per-trace trellis combination) or a theoretical justification, the DNA-storage claims are not fully supported. Please provide such an ablation or restrict the claims accordingly.
[V-3, Fig. 3a] The decoding-speed comparison runs the NPD on a GPU and the trellis decoder on a CPU, a difference the authors acknowledge is 'not entirely fair.' Because the lower-complexity claim is central, the paper should either provide a same-hardware timing comparison or state the complexity advantage purely in terms of the O(AN log N) architecture and parameter counts.
[VI-B2, Fig. 8b] The text says the NPD 'slightly underperforms DNAformer in terms of decoding error,' while the abstract claims the method 'match[es] or surpass[es] the performance of existing methods.' Please reconcile these statements and report whether the comparison is at uncoded rate only; if coding is applied, give coded BER/FER for all methods.

minor comments (7)

[Section I] In the first paragraph, 'at allOnce a sequencing threshold' is missing a space; please fix typographical spacing throughout the manuscript.
[Section VI-B] There are two subsections with the identical title 'SC Decoding Information Encoded on Multiple Input Strands' and nearly identical text; please consolidate them.
[Figure 5 caption] The phrase 'The column labeled K shows results for K ~ Poisson(5)' is unclear because K is the x-axis; please clarify the caption.
[Table I] Please state the block length N and the number of Monte Carlo samples used for each NPD estimate, and add standard deviations.
[Figures 3 and 4] The figures use 'SCT' while the text refers to the 'trellis-based decoder'; please define SCT in the captions or use consistent terminology.
[Section VI-B1] The role of the 36 random padding bits in the punctured polar code and the way the information set is selected should be stated more precisely; currently the reader must infer it from Figure 7a.
[Eq. (11)] The normalization term in the denominator should be defined explicitly; as written, the notation is ambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: decoding and MI estimates are trained from channel samples and benchmarked against external trellis and capacity bounds; the self-referential MI-optimization loop is a standard variational estimator, not a definitional reduction.

full rationale

The paper's derivation chain is self-contained in the relevant sense. The NPD training objective (Eqs. 5-6) is a standard cross-entropy fit to channel samples, and decoding FERs are compared against the external trellis SC decoder [31]. MI estimates (Eqs. 7-9) are plug-in estimates from the trained NPD; they are not defined as the bounds they are compared with in Fig. 2 and Table I. The alternating MI-optimization loop (Eqs. 10-11) is a score-function/policy-gradient estimator whose consistency the authors cite to their prior work [1], but the resulting estimates are calibrated against independent theoretical bounds, so the central claims do not reduce to a fitted parameter or to a self-citation. The multi-trace aggregation rule (Eq. 12) is explicitly labeled heuristic, which is a limitation rather than a circular step. The conjectured numerical-error explanation for the NPD outperforming the trellis decoder (Sec. V-4) and the absence of error bars on MI estimates (Remark 1) are correctness/validation concerns, not instances of definitional circularity. Overall, the strongest claims are not forced by construction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The framework imports channel-model assumptions and polar-code polarization results from prior literature, and the new components are architecture choices plus a heuristic trace-aggregation rule. No physically new entities are introduced.

free parameters (5)

NPD network weights theta_E, theta_F, theta_G, theta_H = not enumerated; about 3M parameters in the real-data model
Learned by SGD on 5e6 input-output samples; all decoding and MI estimates are functions of these fitted weights.
Embedding dimension d = d = N/2
Chosen architecture hyperparameter in Appendix A, not derived from the channel.
Hidden size h = h = 2N
Chosen architecture hyperparameter for all NN modules in Appendix A.
Attention padding length Lmax = Lmax = ceil(1.1*N/2) in DNA experiments
Chosen so padding covers the expected output length; affects embedding cost and accuracy.
Trace aggregation normalization coefficient = 1/sqrt(K)
Heuristic normalization in Equation (12) to keep the embedding second moment stable; explicitly not derived.

assumptions (6)

domain assumption The IDS channel model of Definition 1, with iid per-position insertion/deletion/substitution events and uniform inserted symbols, captures DNA synthesis and sequencing noise.
Used in all synthetic experiments and to pretrain for real data; deviations from this model would change the trained decoder's behavior.
standard math Polar codes polarize over deletion channels under Markov input distributions (Tal et al. [31]).
Justifies using SC decoding and MI-based code design; imported from [31].
standard math The trellis-based SC decoder of [31] provides exact posterior beliefs for deletion channels and is a fair benchmark.
Used as the optimal benchmark in Section V; the paper conjectures numerical errors in that implementation but does not verify them.
ad hoc to paper SGD training of the NPD converges to a close approximation of the synthetic-channel posteriors.
No convergence or approximation guarantees are given; the method relies on empirical loss minimization.
ad hoc to paper Averaging per-trace embeddings with Equation (12) preserves sufficient statistics for SC decoding.
The paper explicitly labels the rule heuristic in Section IV-C.
domain assumption Real-data frozen bits can be assumed known to the decoder via shared randomness.
Used to reconstruct polar codes from the real Nanopore dataset; assumes a common seed or shared randomness is available.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Neural Polar Decoders for DNA Data Storage." pith.science (2026). https://pith.science/paper/OAM5DNVK

@misc{pith2026250617076,
  author       = {Pith},
  title        = {Pith review of: Neural Polar Decoders for DNA Data Storage},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OAM5DNVK}},
  note         = {Machine review of arXiv:2506.17076}
}

abstract

Synchronization errors, such as insertions and deletions, present a fundamental challenge in DNA-based data storage systems, arising from both synthesis and sequencing noise. These channels are often modeled as insertion-deletion-substitution (IDS) channels, for which designing maximum-likelihood decoders is computationally expensive. In this work, we propose a data-driven approach based on neural polar decoders (NPDs) to design low-complexity decoders for channels with synchronization errors. The proposed architecture enables decoding over IDS channels with reduced complexity $O(AN log N )$, where $A$ is a tunable parameter independent of the channel. NPDs require only sample access to the channel and can be trained without an explicit channel model. Additionally, NPDs provide mutual information (MI) estimates that can be used to optimize input distributions and code design. We demonstrate the effectiveness of NPDs on both synthetic deletion and IDS channels. For deletion channels, we show that NPDs achieve near-optimal decoding performance and accurate MI estimation, with significantly lower complexity than trellis-based decoders. We also provide numerical estimates of the channel capacity for the deletion channel. We extend our evaluation to realistic DNA storage settings, including channels with multiple noisy reads and real-world Nanopore sequencing data. Our results show that NPDs match or surpass the performance of existing methods while using significantly fewer parameters than the state-of-the-art. These findings highlight the promise of NPDs for robust and efficient decoding in DNA data storage systems.

Figures

Figures reproduced from arXiv: 2506.17076 by the authors.

**Figure 2.** Comparison of NPD-based MI estimates with theoretical bounds for the deletion channel. [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

**Figure 3.** (a) Comparison of decoder speed of the NPD and the trellis-based for deletion channels. (b) [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: FERs attained by SCL decoding of the NPD for [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]

**Figure 5.** Figure 5: Performance of the NPD when decoding blocks of [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]

**Figure 6.** Figure 6: Performance of the NPD when decoding information stored across multiple DNA strands. The [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]

**Figure 7.** Figure 7: Performance when the number of traces follows [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]

**Figure 8.** Figure 8: Experimental results on real data for the case of a single DNA strand is decoded. [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]

**Figure 9.** Figure 9: Performance of the NPD for decoding information stored across multiple concatenated DNA strands [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]

**Figure 10.** Figure 10: Decoding BERs and FERs for a polar code with rate [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

[1]

Code Rate Optimization via Neural Polar Decoders

Ziv Aharoni et al. “Code Rate Optimization via Neural Polar Decoders”. In: 2024 IEEE International Symposium on Information Theory (ISIT) . 2024, pp. 2424–2429. DOI: 10.1109/ISIT57864.2024. 10619429

arXiv 2024
[2]

Data-Driven Neural Polar Decoders for Unknown Channels with and without Memory

Ziv Aharoni et al. “Data-Driven Neural Polar Decoders for Unknown Channels with and without Memory”. In: IEEE Transactions on Information Theory (2024)

work page 2024
[3]

Channel Polarization: A Method for Constructing Capacity-achieving Codes for Symmetric Binary-input Memoryless Channels

E. Arikan. “Channel Polarization: A Method for Constructing Capacity-achieving Codes for Symmetric Binary-input Memoryless Channels”. In: IEEE Trans. Inf. Theory 55.7 (2009), pp. 3051– 3073

work page 2009
[4]

Long-Term Storage of Information in DNA

Carter Bancroft et al. “Long-Term Storage of Information in DNA”. In: Science 293.5536 (2001), pp. 1763–1765

work page 2001
[5]

Scalable and Robust DNA-based Storage via Coding Theory and Deep Learning

Daniella Bar-Lev et al. “Scalable and Robust DNA-based Storage via Coding Theory and Deep Learning”. In: Nat Mach Intell (2025), pp. 1–11. ISSN : 2522-5839. DOI: 10.1038/s42256- 025- 01003-z

doi:10.1038/s42256- 2025
[6]

Reconstructing Strings from Random Traces

Tugkan Batu et al. “Reconstructing Strings from Random Traces”. In: SODA. V ol. 4. 2004, pp. 910– 918

work page 2004
[7]

Trellis Based Lower Bounds on Capacities of Channels with Synchronization Errors

Jason Castiglione and Aleksandar Kavcic. “Trellis Based Lower Bounds on Capacities of Channels with Synchronization Errors”. In: 2015 IEEE Information Theory Workshop-Fall (ITW). IEEE, 2015, pp. 24–28

work page 2015
[8]

Next-Generation Digital Information Storage in DNA

George M Church, Yuan Gao, and Sriram Kosuri. “Next-Generation Digital Information Storage in DNA”. In: Science 337.6102 (2012), pp. 1628–1628

2012

Show all 37 references

[9]

Reliable Communication over Channels with Insertions, Deletions, and Substitutions

Matthew C Davey and David JC MacKay. “Reliable Communication over Channels with Insertions, Deletions, and Substitutions”. In: IEEE Transactions on Information Theory 47.2 (2001), pp. 687– 698

2001
[10]

Shannon’s Theorems for Channels with Synchronization Errors

Roland L’vovich Dobrushin. “Shannon’s Theorems for Channels with Synchronization Errors”. In: Problemy Peredachi Informatsii 3.4 (1967), pp. 18–36

1967
[11]

DNA Fountain Enables a Robust and Efficient Storage Architecture

Yaniv Erlich and Dina Zielinski. DNA Fountain Enables a Robust and Efficient Storage Architecture. https://www.science.org/doi/10.1126/science.aaj2038. 2017. DOI: 10.1126/science.aaj2038

2017 doi
[12]

Bounds on the Capacity of Channels with Insertions, Deletions and Substitutions

Dario Fertonani, Tolga M. Duman, and M. Fatih Erden. “Bounds on the Capacity of Channels with Insertions, Deletions and Substitutions”. In: IEEE Transactions on Communications 59.1 (2011), pp. 2–6. ISSN : 1558-0857. DOI: 10.1109/TCOMM.2010.110310.090039

2011 arXiv
[13]

Sequential Decoding for Binary Channels with Noise and Synchronization Errors

Robert G Gallager. Sequential Decoding for Binary Channels with Noise and Synchronization Errors. British Library, Reports & Microfilms, 1961

1961
[14]

Convolutional Sequence to Sequence Learning

Jonas Gehring et al. Convolutional Sequence to Sequence Learning . 2017. DOI: 10.48550/arXiv. 1705.03122. arXiv: 1705.03122 [cs]

2017 doi arXiv
[15]

Trace Reconstruction from Noisy Polynucleotide Sequencer Reads

Parikshit S. GOPALAN et al. “Trace Reconstruction from Noisy Polynucleotide Sequencer Reads”. WO2017189469A1. 2017

2017
[16]

Robust Chemical Preservation of Digital Information on DNA in Silica with Error-Correcting Codes

Robert N. Grass et al. “Robust Chemical Preservation of Digital Information on DNA in Silica with Error-Correcting Codes”. In: Angewandte Chemie International Edition 54.8 (2015), pp. 2552–2555. ISSN : 1521-3773. DOI: 10.1002/anie.201411378

2015 doi
[17]

Reinhard Heckel, Gediminas Mikutis, and Robert N. Grass. A Characterization of the DNA Data Storage Channel. 2018. DOI: 10.48550/arXiv.1803.03322. arXiv: 1803.03322 [cs]

2018 doi arXiv
[18]

Fundamental Limits of DNA Storage Systems

Reinhard Heckel et al. “Fundamental Limits of DNA Storage Systems”. In: 2017 IEEE International Symposium on Information Theory (ISIT) . 2017, pp. 3130–3134. DOI: 10.1109/ISIT.2017.8007106

2017
[19]

Achieving the Capacity of the DNA Storage Channel

Andreas Lenz et al. “Achieving the Capacity of the DNA Storage Channel”. In: ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . 2020, pp. 8846–8850. DOI: 10.1109/ICASSP40776.2020.9053049

2020
[20]

Coding Over Sets for DNA Storage

Andreas Lenz et al. “Coding Over Sets for DNA Storage”. In: IEEE Transactions on Information Theory 66.4 (2020), pp. 2331–2351. ISSN : 1557-9654. DOI: 10.1109/TIT.2019.2961265

2020
[21]

Synthesis of High-Quality Libraries of Long (150mer) Oligonucleotides by a Novel Depurination Controlled Process

Emily M. LeProust et al. “Synthesis of High-Quality Libraries of Long (150mer) Oligonucleotides by a Novel Depurination Controlled Process”. In: Nucleic Acids Research 38.8 (2010), pp. 2522–

2010
[22]

Efficient Reconstruction of Sequences

V .I. Levenshtein. “Efficient Reconstruction of Sequences”. In: IEEE Transactions on Information Theory 47.1 (2001), pp. 2–22. ISSN : 1557-9654. DOI: 10.1109/18.904499

2001 doi
[23]

Synthetic DNA Applications in Information Technology

Linda C. Meiser et al. “Synthetic DNA Applications in Information Technology”. In: Nat Commun 13.1 (2022), p. 352. ISSN : 2041-1723. DOI: 10.1038/s41467-021-27846-9

2022 doi
[24]

Asynchronous Methods for Deep Reinforcement Learning

V olodymyr Mnih et al. Asynchronous Methods for Deep Reinforcement Learning . 2016. DOI: 10. 48550/arXiv.1602.01783. arXiv: 1602.01783 [cs]

2016 doi arXiv
[25]

Random Access in Large-Scale DNA Data Storage

Lee Organick et al. “Random Access in Large-Scale DNA Data Storage”. In: Nat Biotechnol 36.3 (2018), pp. 242–248. ISSN : 1087-0156, 1546-1696. DOI: 10.1038/nbt.4079

2018 doi
[26]

Mutual Information Upper Bounds for Uniform Inputs through the Deletion Channel

Francisco Pernice, Berivan Isik, and Tsachy Weissman. “Mutual Information Upper Bounds for Uniform Inputs through the Deletion Channel”. In: IEEE Transactions on Information Theory (2024)

2024
[27]

Upper Bounds on the Capacity of Deletion Channels Using Channel Fragmentation

Mojtaba Rahmati and Tolga M. Duman. “Upper Bounds on the Capacity of Deletion Channels Using Channel Fragmentation”. In: IEEE Transactions on Information Theory 61.1 (2015), pp. 146–156. ISSN : 1557-9654. DOI: 10.1109/TIT.2014.2368553

2015
[28]

Reconstruction Algorithms for DNA-storage Systems

Omer Sabary et al. “Reconstruction Algorithms for DNA-storage Systems”. In: Sci Rep 14.1 (2024), p. 1951. ISSN : 2045-2322. DOI: 10.1038/s41598-024-51730-3

2024 doi
[29]

Survey for a Decade of Coding for DNA Storage

Omer Sabary et al. “Survey for a Decade of Coding for DNA Storage”. In: IEEE Transactions on Molecular, Biological, and Multi-Scale Communications 10.2 (2024), pp. 253–271. ISSN : 2332-

2024
[30]

Trellis BMA: Coded Trace Reconstruction on IDS Channels for DNA Storage

Sundara Rajan Srinivasavaradhan et al. Trellis BMA: Coded Trace Reconstruction on IDS Channels for DNA Storage . 2024. DOI: 10.48550/arXiv.2107.06440. arXiv: 2107.06440 [cs]

2024 doi arXiv
[31]

Polar Codes for the Deletion Channel: Weak and Strong Polarization

I. Tal et al. “Polar Codes for the Deletion Channel: Weak and Strong Polarization”. In: IEEE Trans. Inf. Theory 68.4 (2021), pp. 2239–2265

2021
[32]

Codes Which Correct Single Asymmetric Errors

R. R. Varshamov and G. M. Tenengolts. “Codes Which Correct Single Asymmetric Errors”. In: Automation and Remote Control 26.2 (1965), pp. 286–290

1965
[33]

Attention Is All You Need

Ashish Vaswani et al. “Attention Is All You Need”. In: Advances in Neural Information Processing Systems. Ed. by I. Guyon et al. V ol. 30. Curran Associates, Inc., 2017

2017
[34]

Improved String Reconstruction over Insertion-Deletion Channels

Krishnamurthy Viswanathan and Ram Swaminathan. “Improved String Reconstruction over Insertion-Deletion Channels”. In: Proceedings of the Nineteenth Annual ACM-SIAM Symposium on Discrete Algorithms . SODA ’08. USA: Society for Industrial and Applied Mathematics, 2008, pp. 399–408

2008
[35]

Construction of Polar Codes for Channels with Memory

R. Wang et al. “Construction of Polar Codes for Channels with Memory”. In: 2015 IEEE Information Theory Workshop-Fall (ITW). IEEE, 2015, pp. 187–191

2015
[2540]

DOI: 10.1093/nar/gkq163

ISSN : 0305-1048, 1362-4962. DOI: 10.1093/nar/gkq163

doi
[7804]

DOI: 10.1109/TMBMC.2024.3403488

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Code Rate Optimization via Neural Polar Decoders

Ziv Aharoni et al. “Code Rate Optimization via Neural Polar Decoders”. In: 2024 IEEE International Symposium on Information Theory (ISIT) . 2024, pp. 2424–2429. DOI: 10.1109/ISIT57864.2024. 10619429

arXiv 2024

[2] [2]

Data-Driven Neural Polar Decoders for Unknown Channels with and without Memory

Ziv Aharoni et al. “Data-Driven Neural Polar Decoders for Unknown Channels with and without Memory”. In: IEEE Transactions on Information Theory (2024)

work page 2024

[3] [3]

Channel Polarization: A Method for Constructing Capacity-achieving Codes for Symmetric Binary-input Memoryless Channels

E. Arikan. “Channel Polarization: A Method for Constructing Capacity-achieving Codes for Symmetric Binary-input Memoryless Channels”. In: IEEE Trans. Inf. Theory 55.7 (2009), pp. 3051– 3073

work page 2009

[4] [4]

Long-Term Storage of Information in DNA

Carter Bancroft et al. “Long-Term Storage of Information in DNA”. In: Science 293.5536 (2001), pp. 1763–1765

work page 2001

[5] [5]

Scalable and Robust DNA-based Storage via Coding Theory and Deep Learning

Daniella Bar-Lev et al. “Scalable and Robust DNA-based Storage via Coding Theory and Deep Learning”. In: Nat Mach Intell (2025), pp. 1–11. ISSN : 2522-5839. DOI: 10.1038/s42256- 025- 01003-z

doi:10.1038/s42256- 2025

[6] [6]

Reconstructing Strings from Random Traces

Tugkan Batu et al. “Reconstructing Strings from Random Traces”. In: SODA. V ol. 4. 2004, pp. 910– 918

work page 2004

[7] [7]

Trellis Based Lower Bounds on Capacities of Channels with Synchronization Errors

Jason Castiglione and Aleksandar Kavcic. “Trellis Based Lower Bounds on Capacities of Channels with Synchronization Errors”. In: 2015 IEEE Information Theory Workshop-Fall (ITW). IEEE, 2015, pp. 24–28

work page 2015

[8] [8]

Next-Generation Digital Information Storage in DNA

George M Church, Yuan Gao, and Sriram Kosuri. “Next-Generation Digital Information Storage in DNA”. In: Science 337.6102 (2012), pp. 1628–1628

2012