De novo design of protein-binding aptamers through deep reinforcement learning assembly of nucleic acid fragments

Abstract Nucleic acid aptamers targeting proteins are becoming increasingly important in biopharmaceuticals and molecular diagnostics. Traditionally, aptamers are discovered through labor-intensive screening of nucleic acid libraries using the SELEX method. However, de novo design approaches without experimental screening remain a significant challenge. Here, we employed deep reinforcement learning to develop an artificial intelligence (AI) agent capable of de novo aptamer design, termed AiDTA (AI-driven Docking-Then-Assembling). First, nucleic acid fragments were docked to the target protein to identify target-binding fragments. Then, AiDTA automatically assembled these fragments into aptamers using the Monte Carlo tree search algorithm and a policy-value neural network to guide the agent in generating aptamers with secondary structures similar to the original constituent fragments. Experimental validation demonstrated that the AiDTA-designed DNA aptamers targeting disease-related proteins exhibited high binding affinities in the nanomolar range, achieving the de novo design of protein-binding aptamers for the first time. Our study establishes a new approach to obtaining protein-binding aptamers for potential applications in biopharmaceuticals and diagnostics.