Conformation-Guided Molecular Representation with Hamiltonian Neural Networks
Authors: Ziyao Li, Shuwen Yang, Guojie Song, Lingsheng Cai
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that the Hamiltonian Engine can well preserve molecular conformations, and that the fingerprints generated by Ham Net achieve stateof-the-art performances on Molecule Net, a standard molecular machine learning benchmark. We also evaluate Ham Net on several datasets with different targets collected in a standard molecular machine learning benchmark, Molecule Net (Wu et al., 2017), all following the same experimental setups. Ham Net demonstrates state-of-the-art performances, outperforming baselines including both 2D and 3D approaches. |
| Researcher Affiliation | Academia | Ziyao Li1, Shuwen Yang2 , Guojie Song2 , Lingsheng Cai2 1Center for Data Science, Peking University, Beijing, China 2Key Laboratory of Machine Perception and Intelligence (MOE), Peking University, Beijing, China {leeeezy,swyang,gjsong,cailingsheng}@pku.edu.cn |
| Pseudocode | No | The paper describes its methods using text and equations but does not include any pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing the source code for its own methodology, nor does it provide a link to a code repository for Ham Net. |
| Open Datasets | Yes | Five molecular datasets are used to evaluate Ham Net, including a Quantum Mechanics dataset (QM9) and four biomedical datasets, namely Tox21, Lipop, Free Solv, and ESOL. All datasets are referred in Molecule Net, and the same metrics 2, data split ratios 3, and multi-task scheme (for QM9 and Tox21) 4 are used in our paper. |
| Dataset Splits | Yes | Data are randomly split to 8 : 1 : 1 as training, validation and test sets. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., CPU, GPU models, or cloud computing instances) used for running the experiments. |
| Software Dependencies | Yes | We use the 2020.03.1.0 version of the RDKit package. See http://www.rdkit.org/ |
| Experiment Setup | Yes | As a default setup, we use a 20-step (T = 20) Hamiltonian Engine with df = 32, and L = 2, M = 2 with 200-dimensional hidden representations (dim(hi) = dim(hg) = 200) in the Fingerprint Generator. For the training of Ham Net, we first train the Hamiltonian Engine with known conformations,5 and use the output to train the Fingerprint Generator, with mean-squarederror losses for regression tasks with RMSE metric, mean-absolute-error losses for those with MAE metric, and cross-entropy losses for classification tasks. Other implementation details, including the choices of hyperparameters on different datasets and the training setup are available in the Appendix. |