DEEP LEARNING-BASED SEQUENCE GENERATION OF SINGLE-STRANDED DNA APTAMERS

Aptamers are short single-stranded DNA or RNA sequences that can specifically bind to target molecules, making them useful for various applications such as diagnostics, therapeutics, and biosensors. However, designing aptamers with high binding affinity and specificity is a challenging task due to the complexity and diversity of the target molecules. This project proposes a novel approach for generating novel ssDNA aptamer sequences using a variational autoencoder (VAE) model. The VAE is a deep generative model that can learn the underlying distribution of a given dataset and generate new samples from the learned distribution. The aim is for the VAE model to learn the sequence and structural features of the ssDNA aptamers and generate novel ssDNA sequences. The proposed method consists of three main steps: (1) preprocessing and feature engineering of the ssDNA aptamer sequences, (2) training of the VAE model on the preprocessed sequences, and (3) generation of novel ssDNA aptamer sequences by sampling from the learned latent space of the VAE model. The model is trained on a dataset of ssDNA aptamers targeting various molecules like steroids, toxins, proteins, and drugs. Additionally, several software tools for aptamer sequence design and evaluation are reviewed (NUPACK, Vienna, MFOLD, and MEME). These tools provide informative characteristics on ssDNA sequences that facilitate the design of feature vectors. The model’s performance metrics demonstrate that the VAE model could capture the sequence and structural features of the ssDNA aptamers and generate novel sequences. The unsupervised, generative nature, and potential for transfer learning to other aptamer datasets, makes this method promising. The model could serve as a tool for generating novel ssDNA aptamer sequences, which can facilitate the discovery and development of new aptamers for various applications.