Molecular representation learning is a critical task in AI-driven drug development. While graph neural networks (GNNs) have demonstrated strong performance and gained widespread adoption in this field, efficiently extracting and explicitly analyzing functional groups remains a challenge. To address this issue, we propose MolUNet++, a novel model that employs Molecular Edge Shrinkage Pooling (MESPool) for hierarchical substructure extraction, utilizes a Nested UNet framework for multi-granularity feature integration, and incorporates a substructure masking explainer for quantitative fragment analysis. We evaluated MolUNet++ on tasks including molecular property prediction, drug-drug interaction (DDI) prediction, and drug-target interaction (DTI) prediction. Experimental results demonstrate that MolUNet++ not only outperforms traditional GNN models in predictive performance but also exhibits explicit, intuitive, and chemically logical interpretability. This capability provides valuable insights and tools for researchers in drug design and optimization.

Artificial Intelligence-Generated Content (AIGC)—content autonomously produced by AI systems without human intervention—has significantly boosted efficiency across various fields. However, AIGC in material science faces challenges in efficiently discovering novel materials that surpass existing databases, while ensuring the invariance and stability of crystal structures. To address these challenges, we develop T2MAT (text-to-material), an end-to-end agent that transforms user-input text into the inverse design of novel material structures with target properties beyond existing database, enabled by comprehensive exploration of chemical space and fully automated first-principles validation. Furthermore, we propose CGTNet (Crystal Graph Transformer NETwork), a graph neural network specifically designed to capture long-range interactions, which dramatically improves the accuracy and data efficiency of property predictions and thereby strengthens the reliability of inverse design. Through these contributions, T2MAT reduces the reliance on human expertise and accelerates the discovery of high-performance functional materials, paving the way for truly autonomous material design.

Due to the hexagonal structure, thermal stability, and wide bandgap, two-dimensional group-Ⅲ nitrides (h-BN, h-AlN, h-GaN and h-InN) show great promise for electronic and optoelectronic applications. Density functional theory (DFT) and classical molecular dynamics (MD) methods have advantages in calculation accuracy and scale respectively, but they are limited in the application of high-precision large-scale structure and performance research. Herein, we employ deep potential (DP) method to construct a high-precision machine learning potential (MLP) and systematically investigate the lattice dynamics, thermodynamic, mechanical, and thermal transport properties of two-dimensional Group Ⅲ nitrides. The DP method can achieve DFT accuracy in energy and atomic force predictions and accurately reproduce phonon dispersion and thermodynamic functions (free energy, heat capacity, entropy) across the 0–1200 K temperature range. MD simulations of uniaxial tensiles reveal distinct mechanical behavior differences among materials. h-BN exhibits high strength but brittle fracture characteristics, while h-AlN and h-GaN demonstrate good strength and ductility. h-InN shows relatively weak overall mechanical performance. Non-equilibrium MD simulations on thermal conductivity reveal significant length-dependent effects in h-BN and h-AlN, attributed to longer phonon mean free paths. Enhanced phonon scattering in h-GaN and h-InN results in lower thermal conductivities. These findings demonstrate that the DP method combines DFT accuracy with large-scale simulation capabilities can deepen understanding of structures and properties of two-dimensional Group Ⅲ nitrides and provide a computational framework and theoretical foundations for material design and device application.

Atomic simulation is becoming a vital tool in modern science, bridging the gap between theory and experiments. Since its birth in 1950s, the balance between accuracy and speed has been the main theme in simulating atomic world and in recent years machine learning potential based methods emerged as a promising alternative to density functional theory calculations for exploring complex potential energy surface (PES). Here we report our implementation of LASPAI (www.laspai.com), a web-based platform for future atomic simulations, which is built using the generalized global neural network potential for fast PES evaluation as implemented in LASP software, together with a series of general diffusion generative models, stochastic surface walking (SSW) global optimization, and other common simulation tools for the PES exploration of molecules and materials. We show that LASPAI platform offers a task-orientated, user-friendly, web-based graphical user interface (GUI) to greatly simplify and speed-up atomic simulations for a wide range of scientific areas, ranging from molecule and material structure prediction to solid-gas, solid-liquid, solid-solid interface identification, and reaction pathway simulations. It aims to provide a fast chemical knowledge delivery for scientists to design new materials and reactions.