Energy-Inspired Molecular Conformation Optimization

Authors: Jiaqi Guan, Wesley Wei Qian, qiang liu, Wei-Ying Ma, Jianzhu Ma, Jian Peng

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, these new variants show superior performance in molecular conformation optimization comparing to existing SE(3)-equivariant neural networks. Moreover, our energy-inspired formulation is also suitable for molecular conformation generation, where we can generate more diverse and accurate conformers comparing to existing baselines.
Researcher Affiliation Collaboration Jiaqi Guan University of Illinois at Urbana-Champaign jiaqi@illinois.edu Wesley Wei Qian University of Illinois at Urbana-Champaign weiqian3@illinois.edu Qiang Liu University of Texas at Austin lqiang@cs.utexas.edu Wei-Ying Ma AIR, Tsinghua University maweiying@air.tsinghua.edu.cn Jianzhu Ma Peking University Beijing Institute for General Artificial Intelligence majianzhu@pku.edu.cn Jian Peng University of Illinois at Urbana-Champaign AIR, Tsinghua University Heli Xon Limited jianpeng@illinois.edu
Pseudocode No The paper describes the model's architecture and updates using equations and diagrams, but it does not include formal pseudocode or algorithm blocks.
Open Source Code Yes The model implementation, experimental data and model checkpoints can be found here: https://github.com/guanjq/confopt_official
Open Datasets Yes We test molecular conformer optimization on the QM9 dataset with small molecules (up to 9 heavy atoms) (Ramakrishnan et al., 2014) as well as the GEOM-Drugs dataset with larger molecules (up to 91 heavy atoms) (Axelrod & Gomez-Bombarelli, 2020).
Dataset Splits Yes We randomly split the QM9 dataset into 123k, 5k, 5k, and the GEOM-Drugs dataset into 270K, 10K, and 10K for respective training, validation, and testing.
Hardware Specification Yes We train all conformer optimization models for the QM9 dataset with one NVIDIA Ge Force GTX 1080 GPU and use one Ge Force GTX 3090 GPU for the GEOM-Drugs dataset.
Software Dependencies No The paper mentions optimizers like Adam and libraries like RDKit and Psi4, but does not provide specific version numbers for software dependencies needed to reproduce the experiments.
Experiment Setup Yes The hyper-parameter settings for the optimizer are init_learning_rate=0.001, betas=(0.95, 0.999), and clip_gradient_norm=8. We also schedule to decay the learning rate exponentially with a factor of 0.5 and a minimum learning rate of 0.00001. The learning rate is decayed if there is no improvement for the validation loss in 8 consecutive evaluations, and the training will be completely stopped if no improvement is found for 20 consecutive evaluations. The evaluation is done for every 2000 training steps/batches.