Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
3D-GSRD: 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding
Authors: Chang Wu, ZHIYUAN LIU, Wen Shu, Liang Wang, Yanchen Luo, Wenqiang Lei, Yatao Bian, Junfeng Fang, Xiang Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that 3D-GSRD achieves strong downstream performance, setting a new state-of-the-art on 7 out of 8 targets in the widely used MD17 molecular property prediction benchmark. |
| Researcher Affiliation | Academia | 1 University of Science and Technology of China 2 National University of Singapore 3 Sichuan University 4 Institute of Automation, Chinese Academy of Sciences |
| Pseudocode | Yes | C Pseudo Code We present the pseudocode for pretraining (see Algorithm 1) and finetuning (see Algorithm 2) algorithms in this section. |
| Open Source Code | Yes | The code is released at https://github.com/Wu Chang0124/3D-GSRD. |
| Open Datasets | Yes | For pretraining, we use a large-scale molecular dataset PCQM4Mv2 [55], which contains approximately 3.37 million equilibrium 3D molecular graph structures. For downstream tasks, we evaluate our model on two widely used molecular property prediction datasets: QM9 [13] and MD17 [26]. |
| Dataset Splits | Yes | Specifically, QM9 is a quantum chemistry dataset comprising 134k small molecules... Following prior works [30, 31], we split the dataset into 11,000/1,000/10,831 molecules for training, validation, and testing, respectively. For finetuning, we split the dataset into 9500/950 samples for training and validation, and use the remaining samples for testing. |
| Hardware Specification | Yes | All experiments are conducted on NVIDIA A6000-48G GPUs. |
| Software Dependencies | No | The paper describes the software components used (3D-Re Trans, 2D-PE, Transformer decoder) but does not provide specific version numbers for general software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | The 3D graph encoder is configured with a hidden dimension of 256, 8 attention heads, and 12 layers. The 2D-PE shares most of its configuration with the 3D graph encoder, except for a hidden dimension of 64 and 4 attention heads. The decoder consists of 2 layers, with the hidden dimension and number of attention heads same as the 3D graph encoder. The detailed hyper-parameters configuration for pretraining and finetuning are shown in Table 7, Table 8, and Table 9, respectively. |