A Content-Addressable DNA Database with Learned Sequence Encodings

We present strand and codeword design schemes for a DNA database capable of approximate similarity search over a multidimensional dataset of content-rich media. Our strand designs address crosstalk in associative DNA databases, and we demonstrate a novel method for learning DNA sequence encodings from data, applying it to a dataset of tens of thousands of images. We test our design in the wetlab using one hundred target images and ten query images, and show that our database is capable of performing similarity-based enrichment: on average, visually similar images account for 30% of the sequencing reads for each query, despite making up only 10% of the database.