Reinforcement Learning Based Design of MDM2-Binding Anticancer peptides

Abstract

Conventional cancer therapies, particularly drug based treatment, pharmacotherapy, are often limited by rapid emergence of drug resistance. To overcome these hurdles, Anticancer peptides (ACPs) have gained attention for their ability to selectively disrupt cancer cell membranes via electrostatic interactions, a mechanism that makes acquiring resistance difficult. Extending this therapeutic potential, we propose a synergistic strategy: designing dual-functional peptides that combine this membrane-lytic activity with the intracellular inhibition of the MDM2-p53 interaction. This approach aims to attack cancer cells through physical disruption from the outside and apoptotic activation from the inside. However, the de novo design of such peptides presents a formidable challenge due to the vast chemical space and the scarcity of known sequences that satisfy these distinct structural requirements. To overcome this data scarcity and navigate the vast chemical space, this thesis presents a novel Reinforcement Learning (RL) framework. We employed a pre-trained GPT-2 agent optimized via protein language model-based reward models to generate dual-functional candidates. Furthermore, a diversity driven experience replay mechanism was integrated to ensure robust exploration and prevent mode collapse. The proposed framework successfully generated a diverse library of peptides that satisfy both objectives. Computational validation confirmed that the generated sequences possess optimal physicochemical profiles for ACP activity while maintaining high structural compatibility with MDM2. Notably, competitive binding assays using AlphaFold-Multimer demonstrated that 75% of the top candidates exhibited superior binding affinity to MDM2 compared to the native p53 protein. These results validate the effectiveness of our RL-based approach in accelerating the discovery of complex therapeutic peptides.