Fig. 1: Overview of the InstructNA framework.
From: De novo design of functional nucleic acids of aptamers
a, Architecture of the InstructNA framework: by employing a two-stage training approach, InstructNA transfers the domain knowledge of HT-SELEX FNA sequences to an existing pretrained NA-LLM to produce a domain-adapted FNA-LLM, and then generates new FNA sequences through the domain-adapted FNA decoder. Use of the HC-HEBO algorithm in the latent space enables function-guided evolution of FNAs to achieve better functionality. b, The fraction of 30 DNA sequences generated by InstructNA and RaptGen that have higher binding specificity scores than the ten top-frequency sequences in the original HT-SELEX datasets of Ar, Dbp and Srebf1. c, Scatter plot of the relative binding specificity versus maximum sequence similarity of the DNA sequences generated by InstructNA and RaptGen on the Srebf1 dataset. n = 30. The relative binding specificity is normalized to a range from 0 to 100 using min–max normalization of binding specificity scores. The maximum sequence similarity refers to the highest sequence similarity between the generated DNA sequence and ten top-frequency sequences in the original HT-SELEX dataset. The dashed line indicates the highest relative binding specificity among the ten top-frequency sequences in the original HT-SELEX dataset. Icons in a,b created in BioRender; Guo, P. https://biorender.com/xqmjipu (2026).
