🧬 biology

Where DNA data storage loses its bits

This paper analyzes the specific sources of information loss in DNA data storage across 17 datasets, revealing that sampling inefficiencies often dominate over sequence errors in low-error pools, and demonstrates that an iterative soft receiver significantly reduces the number of reads required compared to existing methods.

Original authors: Hassan Khodaiemehr, Malihe Aliasgari, Chen Feng

Published 2026-11-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hassan Khodaiemehr, Malihe Aliasgari, Chen Feng

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine a library where every book is written not on paper, but on a microscopic strand of DNA. This is the promise of DNA data storage: a way to pack the world's information into a volume no larger than a shoebox. The process is straightforward in theory. Scientists synthesize short strands of DNA, each carrying a tiny piece of digital data, and store them in a pool. To retrieve the information, they sequence the strands, reading the chemical letters back into a computer. However, the real world is messy. The chemical machines that build these strands are not perfect, and the machines that read them make mistakes. Some strands are lost entirely, while others are read too few times to be trusted. For years, researchers have celebrated how many bits of data they can fit into a single DNA letter, but they have rarely stopped to ask a more fundamental question: where exactly does the data go when it disappears?

A new study from researchers at the University of British Columbia and Kean University tackles this mystery by looking at the data that was lost, not just the data that was saved. Instead of simply reporting a final score, the team examined seventeen different public experiments, analyzing millions of raw DNA reads to build a precise map of where information vanishes. They treated the DNA storage process like a noisy communication channel, fitting mathematical models to the specific errors found in each experiment. By doing this, they could separate the losses into distinct categories: strands that were never read at all, strands that were read but too few times, errors caused by the chemistry of the DNA itself, and the cost of the extra markers needed to organize the data.

The findings reveal that for the most common type of DNA storage, which uses standard laboratory synthesis and high-precision sequencing, the biggest losses do not come from the DNA breaking or the machines misreading letters. Instead, the data is lost simply because the pool of strands is uneven. In a perfectly uniform pool, every strand would be read the same number of times. In reality, some strands are abundant and read dozens of times, while others are rare and never seen. The researchers calculated that at low reading depths, this unevenness and the cost of indexing the strands account for a significant drop in efficiency, costing between 0.2 and 0.35 bits of information for every DNA letter. In contrast, the actual errors made by the sequencing machines—substituting one letter for another or deleting a letter—account for less than 0.01 bits of loss. This suggests that for current systems, the path to better storage lies not in writing more complex error-correcting codes, but in making the synthesis process more uniform so that every strand has an equal chance of being read.

The picture changes dramatically when the DNA is old or the chemistry is more error-prone. When DNA strands are aged or stored under harsh conditions, they break, leading to a situation where the sequencer reads only a fragment of the original message. In these cases, the loss of information is massive, with up to 0.58 bits lost per letter simply because the strands are too short to be useful. Similarly, when using newer, cheaper, but noisier technologies like nanopore sequencing, the machines struggle to keep track of where they are in the sequence. Here, the loss comes from the difficulty of aligning the reads, a problem that costs between 0.04 and 0.26 bits per letter. The study shows that for these difficult conditions, the solution is not better chemistry, but better reading. The researchers developed a new method for decoding the data that uses a few known "pilot" letters as anchors to help the computer align the messy reads. This approach allowed them to recover data with 15 to 42 percent fewer reads than the best existing methods, effectively turning a noisy, unreliable process into a reliable one.

The researchers also tested how much information could theoretically be stored in different types of DNA designs. They found that some designs, which avoid repeating the same letter to prevent chemical errors, actually limit the storage density to about 1.59 bits per letter, sacrificing nearly a quarter of the potential capacity for the sake of robustness. By comparing real-world results against the theoretical limits of their specific conditions, the team showed that recent storage systems are already operating at 90 percent or more of what is possible for their technology. This means that for the most common setups, the gap between current performance and the theoretical maximum is very small. The study concludes that the future of DNA storage depends on matching the right solution to the right problem: improving the uniformity of the DNA pool for standard storage, repairing broken strands for aged archives, and using smarter decoding algorithms for noisy, low-cost technologies. By placing every dataset against its own specific losses, the researchers have provided a clear roadmap for where the next gains in this field will come from.

Technical Summary: "Where DNA data storage loses its bits"

Problem Statement
DNA data storage is typically evaluated by the net information density (bits stored per synthesized nucleotide), which has risen from 0.33 to over 1.8 bits per nucleotide. However, these metrics lack a standardized reference for optimality. Current evaluations fail to account for where specific experiments lose information. Without a granular accounting of losses—distinguishing between sampling inefficiencies, sequencing errors, synchronization issues (insertions/deletions), and design constraints—it is impossible to determine if a new coding scheme is near-optimal, whether gains must come from chemistry or coding, or how much sequencing is required for a target density. Existing capacity bounds are often based on idealized independent-error models or substitution-only assumptions, which do not reflect the position-dependent, correlated, and truncation-prone nature of real DNA storage channels.

Methodology
The authors analyzed 17 public datasets comprising 6.2 million raw reads from eight studies, covering various synthesis technologies (array, photolithographic) and sequencing platforms (Illumina, Nanopore). Their approach involves three main technical pillars:

Channel Fitting: Using an Expectation-Maximization (EM) algorithm over alignment lattices, they fitted position-dependent insertion-deletion-substitution (IDS) channels to the raw reads. The model captures deletion and substitution probabilities at each payload position, geometric insertion distributions, and specific substitution kernels.
Information-Theoretic Bounds: They derived theoretical bounds for these fitted models.
- Theorem 1: Provides a "sandwich" bound for the capacity of the positional model with $k$ reads. The upper bound assumes a "genie" reveals the alignment of each read; the lower bound subtracts a synchronization-loss term ( $\Delta$ ) accounting for the entropy of alignment not revealed by read length.
- Theorem 3 & 4: Introduce block-based lower and upper bounds (tiling bounds) to tighten the capacity estimates, particularly for channels with high synchronization loss.
- Propositions: Establish conditions for input uniqueness and methods to convert simulations into confidence intervals.
Achievable Information Rates (AIRs): They measured AIRs directly on the reads using two decoding strategies:
- Consensus Decoding: Aligns reads and votes before decoding.
- Joint Soft Decoding: A new receiver that combines reads softly using the fitted channel model, employing sparse pilot bases for error-prone chemistries and iterative belief propagation.
Model-Based Accounting: They decomposed the theoretical maximum of 2 bits per nucleotide into specific loss categories: unread strands (Poisson sampling), uneven coverage, genie error loss (errors with known alignment), full-read gaps (truncated reads), genie-to-AIR residuals (synchronization and receiver mismatch), and index costs.

Key Contributions

Granular Loss Accounting: The paper provides the first model-based decomposition of information loss in DNA storage, quantifying exactly how many bits are lost to sampling, coverage bias, truncation, and synchronization for 17 distinct datasets.
Capacity Bounds for Real Channels: Unlike previous work relying on substitution-only bounds, this study proves capacity bounds for fitted position-dependent IDS channels, bridging the gap between theoretical limits and experimental reality.
Novel Receiver Design: The authors developed an iterative soft receiver that utilizes sparse pilot bases. In a prespecified, paired benchmark against the best published method (Trellis BMA), this receiver required 7–42% fewer reads to achieve the same file error rate across eight of nine datasets.
Open-Source Framework: The release of the dnacap software package allows any new dataset to be evaluated against its own fitted limits and loss profile.

Key Results

Loss Dominance: For low-error array pools read by Illumina at low sequencing depth, losses are dominated by the pool level (unread strands, uneven coverage, and indexing), costing 0.2–0.35 bits per nucleotide. Sequence errors (substitutions/deletions) contribute less than 0.01 bits.
Error-Prone Chemistries: For photolithographic and Nanopore data, the "genie-to-AIR residual" (information hidden by unknown alignments and receiver mismatch) is significant, ranging from 0.04 to 0.26 bits per nucleotide. This indicates substantial room for improvement via better receivers.
Design Constraints: Designs that forbid repeated bases (to avoid homopolymer errors) cap the theoretical density at approximately 1.59 bits per nucleotide, representing a loss of ~0.41 bits compared to the unconstrained 2-bit maximum.
Decoding Performance: The proposed iterative soft receiver significantly outperformed Trellis BMA on error-prone data. For Nanopore and photolithographic data, it achieved file recovery at coverages where Trellis BMA failed on 83–100% of files. On array data, the gains were smaller but still present.
Cost Analysis: Under current pricing (sequencing is cheap relative to synthesis), the optimal strategy for single-read archives is to maximize density at high coverage. Soft decoding becomes a cost-saving lever primarily for frequent reading or when using cheap, error-prone synthesis chemistries.

Significance and Claims
The paper claims its primary contribution is a cross-technology accounting framework that integrates fitted channels, theoretical bounds, and measured rates to place any DNA storage experiment against its own specific limits. The authors assert that this approach allows researchers to identify the specific "levers" for improvement:

For Illumina/array systems, gains are likely to come from more uniform synthesis, longer strands (to dilute index costs), or receivers that utilize every read, rather than stronger codes.
For aged DNA, the primary loss is the "full-read gap" (truncation), suggesting a need for molecular repair or receivers that utilize partial reads.
For error-prone chemistries (Nanopore/Photolithographic), the receiver design is the critical factor, as soft decoding can recover a significant portion of the "genie-to-AIR" gap.

The authors remain modest regarding the absolute capacity of physical pools, noting that their bounds apply to fitted models and that the decomposition of losses is a model-based decomposition rather than a unique causal attribution of physical causes. They emphasize that their framework provides a statistically controlled benchmark to determine how close current systems are to the limits of their specific operating conditions.

Technical Summary: "Where DNA data storage loses its bits"

More like this