Diffusion-based Molecule Generation with Informative Prior Bridges

Authors: Lemeng Wu, Chengyue Gong, Xingchao Liu, Mao Ye, Qiang Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental With comprehensive experiments, we show that our method provides a powerful approach to the 3D generation task, yielding molecule structures with better quality and stability scores and more uniformly distributed point clouds of high qualities.
Researcher Affiliation Academia Lemeng Wu University of Texas at Austin lmwu@cs.utexas.edu Chengyue Gong University of Texas at Austin cygong@cs.utexas.edu Xingchao Liu University of Texas at Austin xcliu@cs.utexas.edu Mao Ye University of Texas at Austin my21@cs.utexas.edu Qiang Liu University of Texas at Austin lqiang@cs.utexas.edu
Pseudocode Yes Algorithm 1 Learning diffusion generative models.
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes Dataset Settings QM9 [36] molecular properties and atom coordinates for 130k small molecules with up to 9 heavy atoms with 5 different types of atoms. ... GEOM-DRUG [4] is a dataset that contains drug-like molecules. ... We use the Shape Net [6] dataset for point cloud generation.
Dataset Splits Yes We follow the common practice in [19] to split the train, validation, and test partitions, with 100K, 18K, and 13K samples.
Hardware Specification Yes It takes approximately 10 days to train the model on these two datasets on one Tesla V100-SXM2-32GB GPU.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers).
Experiment Setup Yes On QM9, we train the EGNNs with 256 hidden features and 9 layers for 1100 epochs, a batch size 64, and a constant learning rate 10^-4, which is the default training configuration. We use the polynomial noise schedule used in [19] which linearly decay from 10^-2/T to 0. We linearly decay from 10^-3/T to 0 w.r.t. time step. We set k = 5 (7) by default. On GEOM-DRUG, we train the EGNNs with 256 hidden features and 8 layers with batch size 64, a constant learning rate 10^-4, and 10 epochs.