Navigating Chemical Space with Latent Flows

Authors: Guanghao Wei, Yining Huang, Chenru Duan, Yue Song, Yuanqi Du

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the efficacy of Chem Flow on molecule manipulation and singleand multi-objective molecule optimization tasks under both supervised and unsupervised molecular discovery settings. Codes and demos are publicly available on Git Hub at https://github.com/garywei944/Chem Flow.
Researcher Affiliation Collaboration Guanghao Wei* Cornell University gw338@cornell.edu Yining Huang* Harvard University yininghuang@hms.harvard.edu Chenru Duan Deep Principle, Inc. duanchenru@gmail.com Yue Song California Institute of Technology yuesong@caltech.edu Yuanqi Du Cornell University yd392@cornell.edu
Pseudocode Yes Algorithm 1 Chem Flow Training
Open Source Code Yes Codes and demos are publicly available on Git Hub at https://github.com/garywei944/Chem Flow.
Open Datasets Yes We extract 4,253,577 molecules from the three commonly used datasets for drug discovery including MOSES [46], ZINC250K [27], and Ch EMBL [68].
Dataset Splits Yes The VAE is trained for 150 epochs with 4 restarts on 90% of the training data and validated with the rest 10% data.
Hardware Specification Yes All the experiments including baselines are conducted on one RTX 3090 GPU and one Nvidia A100 GPU. All docking scores are computed on one RTX 2080Ti GPU.
Software Dependencies No The paper mentions software components like Adam W optimizer and Mish activation function but does not provide specific version numbers for any key software dependencies or libraries needed for reproducibility.
Experiment Setup Yes Pre-trained VAE ... a 128 dimension embedding layer, 1024 latent space size, 3-hidden-layer encoder, and 3-hidden-layer decoder both with 1D batch normalization and non-linear activation functions. The hidden layer sizes are {4096, 2048, 1024} for the encoder and reversely for the decoder. The VAE is trained using an Adam W [42] optimizer, 0.001 initial learning rate, and 1,024 training batch size.