Skip to main content
In-DNA computing turns the archive itself into an associative memory. Computation is performed directly on stored molecules, so large collections can be searched and filtered in place before any bulk sequencing. The result is massive parallelism with very low energy cost and dramatic reductions in read volume.

​
Data Model

The archive is written as identifiers DNA molecules assembled by choosing one component from each of several layers and ligating them in order.

Combinatorial address space

The Cartesian product of layers creates an address space with a well-defined rank for every identifier. Data are mapped to codewords with uniform weight.

Trie organization

A subset of layers is reserved as a key so identifiers can be addressed chemically by their components. This induces a trie over the identifier space, enabling dictionary-like operations in the wet lab.

​
Instruction Set

Two primitive operations form the core of in-DNA compute:

Select

Takes a key defined by a sequence of components across a contiguous set of layers. Returns the subset of identifiers that contain that key. Implemented via recursive selective PCR after m steps only identifiers containing the full key remain enriched.

Quotient

Takes a layer index q. Truncates all enriched molecules at layer q so that identifiers differing only in deeper layers collapse to the same sub-identifier. Signals from many descendants in the trie are summed into their common ancestor.
Select isolates by content. Quotient aggregates by structure. Together they support exact match queries, approximate similarity queries, and multi-step pipelines without scanning the full archive.

​
Exact Search Workflow

1

Derive bit-position keys

Obtain the codeword for the query object. With weight w, it defines w distinct bit-position keys.
2

Apply select for each key

Run select on each key. Merge the enriched products.
3

Apply quotient

Quotient at the layer adjacent to the bit-position key. Each true occurrence contributes one strong sub-identifier; non-targets contribute weak residual fragments.
4

Sequence the enriched subset

A short sequencing run over this enriched library calls occurrences with high signal-to-noise ratio reading orders of magnitude fewer molecules than a full scan.

​
Similarity Search Workflow

Real-valued embeddings are mapped to uniform-weight codewords so that similar vectors share more one-bit positions.
1

Encode embeddings

Choose a small set of reference points in embedding space. Assign a one to the k nearest references for each item. Two unrelated items share w²/L bits on average; related items share significantly more.
2

Run select on query keys

Apply select on the union of the query’s bit-position keys.
3

Apply quotient and measure signal

Signal after quotient is proportional to shared one-bit positions directly correlating with semantic similarity.
4

Sequence top candidates

A small sequencing budget recovers top candidates for precise digital ranking.

​
Parallelism & Complexity

Select and quotient run in single-instruction, multiple-data fashion across the entire library. The number of wet-lab steps for a select equals the number of layers in the key it does not depend on total stored items. Total reads can drop by two to three orders of magnitude relative to a full scan, while maintaining strong signal separation between targets and non-targets.
Multiple selects can execute in parallel in separate reactions. Large keys can be split across parallel paths and recombined. With more sophisticated chemistry, selection over a multi-layer key can be compressed into a constant number of operations. The same library supports additional molecular instructions: These micro-instructions compose into higher-level operations filtering, set logic, and associative arithmetic directly in the archive.

​
Design Parameters

​
Scope and Limits

In-DNA computing excels at selection, association, and reduction over very large collections with extremely low energy. It is not a replacement for floating-point numerical computation or model training. The intended pattern is molecular filtering and recall in place, followed by compact digital sequencing and precise ranking. This hybrid approach combines molecular-scale parallelism with digital precision the most efficient way to work with archives at petabyte and exabyte scale.