AIMNET-X2D: Universal and Scalable Framework for Chemical Property Transfer Learning

Accurate molecular property prediction is fundamental to drug discovery and materials science, where rapidly expanding datasets now challenge existing computational approaches. We present AIMNet-X2D, a highly scalable Graph Neural Network framework that processes molecular datasets of unprecedented size while maintaining state-of-the-art accuracy. By overcoming critical bottlenecks that have limited previous architectures, AIMNet-X2D achieves exceptional scalability through three key innovations: (1) optimized multi-hop edge precomputation, (2) an explicit shell convolution mechanism that processes multiple structural contexts simultaneously, and (3) dynamic attention based weighting that captures both local and long-range molecular interactions. Benchmarked on datasets of unprecedented size—up to 20 million molecules—AIMNet-X2D demonstrates near-linear scalability (R2 ¿ 0.99) that enables efficient processing of chemical datasets orders of magnitude larger than previously feasible. The framework achieves excellent predictive performance across diverse tasks, including quantum mechanical properties (QM9, R2 = 0.97), physical properties (Tetko melting points, R2 = 0.77), and pharmacological properties (LogP, R2 = 0.93). Our framework transforms high-throughput virtual screening capabilities and accelerates AI-guided discovery across pharmaceuticals, materials science, and beyond by enabling chemical AI to scale in ways previously thought unattainable.