Interpretable prediction and generation of ASC-speck aptamers using multiscale deep biological learning models.

Motivation: Aptamers are functional nucleic acids that can bind to corresponding ligands and effectively replace monoclonal antibodies. However, some target proteins may lack binding candidate DNA sequence, necessitating methods to generate new aptamer libraries for out-of-donor detection. Therefore, we propose an adaptive prediction and design method, ASC2BF for DNA aptamer generation. Results: Based on the multiscale residual network, ASC2BF not only predicts the aptamers simultaneously based on the characteristics of DNA-protein aptamers, but also uses the bacterial foraging optimization algorithm (BFOA) to generate new DNA aptamer sequences based on biophysical constraints, with initial seeds screened by the neural network predictor. As a case study, ASC2BF was applied to generate aptamer pools against apoptosis-associated speck-like proteins (ASC-speck). Furthermore, we demonstrate the ability of deep learning to capture sequential and functional semantic information. And through interpretability analysis, our results demonstrate what the model learns, helping us build a map from discovering important information to analyzing its biological function. Availability and implementation: The code and data sets are obtainable at https://github.com/nmt315320/aptamer.git.