ADAPTIVE MOLECULAR GRAPH ABSTRACTION VIA REINFORCEMENT LEARNING ACROSS GEOMETRIC SCALES

Molecular graph representation learning has emerged as a fundamental challenge in computational chemistry and drug discovery, where the ability to capture multi-scale structural information from atomic to molecular levels is crucial for accurate property prediction and molecular design. Traditional graph neural network approaches typically operate at a single resolution, failing to capture the hierarchical nature of molecular structure that spans multiple geometric scales from local atomic environments to global molecular topology. This paper presents a novel framework that employs Adaptive Molecular Graph Abstraction (AMGA) through reinforcement learning to dynamically learn optimal abstraction strategies across different geometric scales. The proposed framework combines hierarchical graph pooling mechanisms with reinforcement learning agents that adaptively select abstraction levels and pooling strategies based on molecular characteristics and task requirements. Our approach employs differentiable clustering algorithms that learn to group atoms into chemically meaningful motifs, progressing through multiple pooling levels to create increasingly abstract molecular representations. The reinforcement learning component formulates abstraction strategy selection as a sequential decision-making problem, enabling dynamic adaptation to different molecular families and property types. Experimental evaluation demonstrates superior reconstruction accuracy compared to existing approaches, with our method maintaining over 85% accuracy across all molecule sizes while baseline methods like CG-VAE deteriorate significantly for larger molecules. The adaptive abstraction mechanism enables automatic discovery of optimal representation granularities for different chemical contexts, providing interpretable insights into structure-property relationships while maintaining computational efficiency suitable for large-scale molecular screening applications.