Fei Zhu

Most existing adversarial training methods in reinforcement learning (RL) offer limited robustness and remain vulnerable to novel attacks. To address this limitation, an approach that enhances policy robustness by leveraging the historically optimal policy to guide policy optimization and generating diverse adversarial perturbations, termed robust RL via leveraging historically optimal policy with regulation of performance (HORP), is proposed. Unlike other approaches that rely solely on trial-and-error interactions, HORP constructs a guidance value function by simultaneously considering value gaps and policy distribution divergence, thereby focusing on prioritized learning in promising action spaces. It also incorporates an adaptive performance-aware optimization mechanism to trigger timely corrections, preventing the agent from deviating from optimal performance. Furthermore, HORP dynamically modulates perturbation entropy through controlled uncertainty injection, thereby improving the agent’s generalized defensive capabilities. Experiments demonstrate that HORP achieves superior performance in most cases regarding both natural performance and robustness against various state attacks.

Accurately predicting molecular properties can identify more promising drug candidates and facilitate the process of drug discovery. There are advances in methods for molecular property prediction using self-supervised learning. However, most methods tend to focus only on a certain representation, resulting in reduced information diversity. Moreover, less attention is paid to the functional groups in the molecule, which are highly related to the molecular properties. To solve the problem, a multimodal contrastive learning framework called MMCL for molecular property prediction is proposed, which learns common features among similar molecules by utilizing different views of molecules for predicting molecular properties. In addition, by explicitly representing functional groups as nodes in the molecular graph, the model is prompted to learn features related to molecular properties. Prediction results on benchmark datasets and a series of case studies show that MMCL has optimal performance in most property prediction tasks and can capture information related to chemical semantics. Overall, MMCL can be used as an efficient deep learning tool for molecular-related tasks to facilitate the drug discovery process.

Many complex diseases involve intricate and interdependent biological mechanisms, making multi-target drug design therapeutically important. However, existing molecular generation models often struggle to balance conflicting property requirements, maintain structural exploration, and avoid premature convergence during reinforcement learning. To address these challenges, we propose CMIR-RL, a reinforcement learning-based molecular generation model that integrates cooperative crossover-mutation exploration, structure-based intrinsic rewards, and Pareto-based multi-objective optimization. In CMIR-RL, mutation and crossover networks synergistically guide the agent to balance exploration and exploitation, while Pareto-front-guided reward assignment updates the generation policy under multiple property constraints. Experiments on a three-objective polypharmacology task show that CMIR-RL significantly improves desirability from 0.8868 to 0.9542 over the baseline method, while maintaining competitive validity, uniqueness, and diversity. Additional GuacaMol results further indicate competitive generalization across goal-directed molecular generation tasks. These results show that CMIR-RL is an effective reinforcement learning strategy for early-stage multi-objective molecular design.