How AI checks whether DNA keeps its extra letters

Inventing another letter for DNA means making another molecule. In 2019, Shuichi Hoshika and colleagues assembled hachimoji DNA with eight building blocks: the familiar A, T, G and C, plus B, S, P and Z. The added letters formed two additional pairs across the double helix. Three crystal structures showed that the expanded alphabet could retain familiar DNA geometry. An engineered polymerase also transcribed it into RNA that produced a fluorescent signal.[2]

The chemistry had to be predictable as well as possible. The researchers measured the melting behavior of 94 DNA duplexes and developed a model that predicted their melting temperatures within 2.1°C on average. Changing the sequence did not have to destroy the rules that hold the two strands together.[2]

That is a foundation for an alphabet: different messages can share the same physical writing system. The next question is persistence. A tube can contain much more DNA after amplification while a special letter becomes progressively less common. Quantity alone would conceal the change.

Nicholas Kaplan and colleagues’ Xenoquant uses 33 neural classifiers for six unnatural bases in specified sequence contexts. Most exceeded 98% recall and specificity. They interpret nanopore currents to measure synthetic-letter fractions, without providing a general basecaller for arbitrary expanded DNA.[1]

A letter changes an electrical neighborhood

Nanopore sequencing measures ionic current as DNA passes through a pore. The signal depends on neighboring bases as well as the letter of interest. In a separate 2025 study, Ashley Stephenson and colleagues tested current-prediction models on ten synthetic bases, withholding one base type at a time. Sequence information explained much of the variation; incorporating molecular structure helped some predictions. The measured current is therefore richer, and more contextual, than a barcode assigned to one isolated letter.[3]

An analogy is recognizing a changed note within a chord. A difference can be detectable without each note having an independent signature. Context helps recognition, while also making it harder to infer what an unfamiliar combination will sound like.

For B/S copying, Xenoquant measured 89.0 ± 0.4% effective fidelity in a baseline Taq-polymerase assay and 98.3 ± 0.2% under optimized Q5 conditions; both intervals are 95% confidence intervals. Optimization changed the enzyme, sequence context and relative supplies of building blocks.[1]

Why 98.3% is not the multiplier

The paper’s simple retention model assumes constant fidelity, loss of synthetic bases and no reverse creation. The average synthetic-letter fraction after n doublings is [(1 + fidelity) / 2]ⁿ, with fidelity expressed as a fraction. A doubling is an amplification quantity, distinct from a programmed thermal cycle.[1]

Here is an illustrative way to see the arithmetic. Assume the two complementary strands copy equally well. Start with 100 correct old strands. They remain after copying, alongside 100 new strands. At 98.3% fidelity, an expected 98.3 new strands are correct, making 198.3 correct strands out of 200. The retained fraction is 99.15%. Repeating that multiplication ten times gives about 91.8%. This calculation concerns the average letter fraction, not the fraction of intact, correctly paired duplexes.[1]

The difference grows with repetition: the same model gives about 32.3% and 84.3% retention after twenty doublings. Those longer-run values are illustrations, not predictions that reaction conditions will remain ideal indefinitely. They show why a few percentage points in a copying assay can alter the composition of the final population so substantially.[1]

A synthetic alphabet earns its usefulness through both chemistry and continuity. Making a new letter gives researchers another possibility; measuring how that letter survives lets them judge whether the possibility remains present after the material has multiplied.

Report an error or suggest a correction