Plain language

What this result means

A DNA storage code is useful only if many words can be packed into the same constraint set. More words means more possible stored symbols at the same word length and error tolerance. The hard part is comparison: best-known is a literature claim, while optimal is a proof claim. This page separates those ideas and also separates strict GC from wider GC bands, because those are different problems.

  • The largest margin is GCNL (10,4): ROEAO's 581 to 657 under the strict-GC no-runlength/Hamming convention. CLGBO reports 580 for the same cell; Limbachiya's original value is 525.
  • The largest all-convention row is King's reverse-complement (8,3): 390 to 426.
  • IAOA's traditional GCNL table was checked explicitly: it raises the strict (6,3) baseline to 61, but does not erase any claimed row.
  • Eight rows beat every published value under any GC convention. The remaining eight are strict-GC improvements only, and the table says so instead of blurring the comparison.

Visual notes

How to read the result

Horizontal bar chart of sixteen DNA storage code improvements, with white bars for all-convention rows and gray bars for strict-GC rows.
Record marginsWhite bars are rows that beat every published GC convention. Gray bars are strict-GC improvements, where wider GC-band tables are a different comparison.
Bar chart showing how many length ten no-runlength words exist at each GC count, with GC equals five highlighted for strict GC and GC equals four through six highlighted for the band.
The GC convention trapFor length 10, strict GC uses only GC=5 words. A GC band also includes GC=4 and GC=6 words, so it is a larger search space.
Dense barcode-style visualization of the 657 codewords in the GCNL length ten distance four DNA code.
Actual codewordsThe largest saved code in the set has 657 length-10 DNA words. Each row is one word, drawn base by base.

Result table

Sixteen DNA code-size improvements, with the baseline source named row by row.

CellBaselineNumaroDeltaNote
GCNL (10,4)581657+76strict GC; ROEAO source; CLGBO has 580
GCRC (8,3)390426+36all conventions; King source
GCNL (10,5)168190+22strict GC; CLGBO source
GCNL (12,9)920+11all conventions; MFOS band source
GCNL (10,6)5767+10strict GC; ROEAO source
GCNL (8,3)373383+10strict GC; LEO/IAOA source
GCNL (8,4)119128+9strict GC; LEO source
GCNL (10,8)1016+6all conventions; ROEAO/IAOA/MFOS source
GCNL (10,7)2328+5all conventions; CLGBO source
GCNL (12,10)69+3all conventions; MFOS band source
GCNL (6,3)6162+1strict GC; IAOA Table 4 source; proven optimal
GCNL (8,5)4042+2strict GC; LEO source; ties widest band
GCRC (6,3)4143+2all conventions; King source
GCRC (8,5)2628+2all conventions; King source
GCNL (6,4)2728+1all conventions; ROEAO/MFOS source; proven optimal strict
GCNL (8,6)1617+1strict GC; LEO/IAOA source

Method

How it was found

Each cell becomes a graph problem. A vertex is an admissible DNA word. Two vertices are connected only if the two words can live in the same code. A large code is a large clique in that graph.

  • Compiled the prior values from King, Limbachiya-Gupta-Aggarwal, NOL-HHO, CLGBO, EORS, QRSS-MPA, LEO, ROEAO, IAOA, and MFOS while recording the exact GC rule used by each source.
  • Used ROEAO Table 5 only for the traditional S^GC,NL Hamming rows; its enhanced RTSC/DTW table is a different metric and is not used as a Hamming baseline.
  • Used IAOA Table 4 only for the traditional A^GC,NL Hamming rows; its double-matching and error-pairing tables are extra-constraint families and are not used as Hamming baselines.
  • Generated every admissible word for the requested length, distance, GC rule, and no-runlength or reverse-complement rule.
  • Searched the compatibility graph with a bitmask clique solver; small cells were also closed with exact CP-SAT.
  • Dropped apparent wins when the baseline used a wider GC band and the numbers were not comparable.

Verification

How it was checked

The checker rereads each saved code and tests the rules directly: every word has the right length and alphabet, the GC count is correct, no adjacent bases repeat where that rule applies, all words are distinct, every pair is far enough apart, and reverse-complement distances are checked for the reverse-complement family.

Scope

What is not being claimed

Only rows compared under the same convention are claimed. Best-known means larger than the published lower bounds we found for that matched constraint; optimal means proven no larger code exists. Most rows are best-known improvements, not optimality proofs. The proven-optimal statements are only for the strict-GC n=6 cells. Wider GC bands are a different, easier problem because they allow more candidate words.

References

Baseline sources

Citation

How to cite

Numaro AI Autoresearch Team. "DNA storage codes with larger codebooks." Numaro Research Report NUMARO-2026-011, 2026.

@techreport{numaro2026DnaStorageCodes,
  title = {DNA storage codes with larger codebooks},
  author = {Numaro AI Autoresearch Team},
  institution = {Numaro},
  number = {NUMARO-2026-011},
  year = {2026},
  url = {https://numaro.tech/research/dna-storage-codes-2026/}
}