Biologically Constrained DNA Encoding with Triplet Networks for Similarity Image Retrieval

As the volume of digital data continues to grow exponentially,DNA has emerged as a promising medium for long-term data storagedue to its high density and durability. For enabling data retrieval viaDNA\u2019s biochemical reactions, the encoding strategy plays a criticalrole. This paper proposes a training framework for a DNA encoderthat improves both accuracy and training efficiency in content-basedimage retrieval by incorporating deep metric learning. In addition, weintroduce loss functions that enforce biological constraints, specificallyhomopolymer length and GC content, thereby improving the biochem-ical stability of the generated DNA sequences. To evaluate the effec-tiveness of the proposed method, we conduct quantitative assessmentsbased on image classification performance. Simulations on the CIFAR-10 and CIFAR-100 datasets demonstrate that our method achievesclassification accuracy comparable to CNN-based baselines and a 21-fold speedup over the training time of the existing method. Moreover,the generated DNA sequences enable strict control of homopolymerlength and maintain GC content within the optimal 40-60% range, sig-nificantly improving biological feasibility compared to baseline methods.The source code is publicly available at: https:\/\/github.com\/tkoike-kuee\/DNA-Encoder-under-Bioconstraints.