GTMGC: Using Graph Transformer to Predict Molecule’s Ground-State Conformation
Authors: Guikun Xu, Yongquan Jiang, PengChuan Lei, Yan Yang, Jim Chen
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method has been evaluated on the Molecule3D benchmark dataset and the QM9 dataset. Experimental results demonstrate that our approach achieves remarkable performance and outperforms current state-of-the-art methods as well as the widely used open-source software RDkit. |
| Researcher Affiliation | Academia | 1School of Computing and Aritifical Intelligence, Southwest Jiaotong University, Chengdu, China 2Institute of Aritifical Intelligence, Southwest Jiaotong University, Chengdu, China 3Engineering Research Center of Sustainable Urban Intelligent Transportation, Ministry of Education, China 4Department of Computer Science, George Mason University, Fairfax, VA, USA |
| Pseudocode | No | The paper does not contain any sections explicitly labeled "Pseudocode" or "Algorithm", nor does it present any code-like structured steps for a procedure. |
| Open Source Code | Yes | (v). The source code of our method is available at https://github.com/Rich-XGK/GTMGC. |
| Open Datasets | Yes | Our method has been evaluated on the Molecule3D benchmark dataset and the QM9 dataset. ... Molecule3D. The first benchmark introduced by (Xu et al., 2021d) ... QM9. A small-scale quantum chemistry dataset (Ramakrishnan et al., 2014; Wu et al., 2018) |
| Dataset Splits | Yes | Molecule3D. ...Both partitions use a 6:2:2 ratio for training, validation, and testing. ... QM9. ...We adopt the identical data split as described in (Liao & Smidt, 2022), where 110k, 10k, and 11k molecules are allocated for training, validation, and testing, respectively. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions software components like "Adam W optimizer" and "Mole BERT Tokenizer", but it does not specify version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | For our primary task of predicting the molecular ground-state conformation, both the encoder and decoder of GTMGC are configured with 6 transformer blocks and 8 attention heads. Notably, we set our dmodel to 256, dffn to 1024, and dhidden in the prediction head to 768, resulting in a lean model with only 9M parameters. During the training phase, we employ the Adam W optimizer with a learning rate of 5e-5 and a batch size of 100. We initially warm up the learning rate from 0 to 5e-5, followed by a linear decay to 0, over a total of 20 epochs. For our auxiliary task of molecular properties prediction, we have constructed three versions of GTMGC: small, base, and large. The specific hyperparameters for these three versions are detailed in Table 7. For this task, we apply the same training strategy across all three versions of the model, but increase the number of training epochs to 60, with a larger learning rate warm up of 0.3 and a cosine decay strategy. |