Structure-based drug design by denoising voxel grids

Authors: Pedro O. Pinheiro, Arian Rokkum Jamasb, Omar Mahmood, Vishnu Sresht, Saeed Saremi

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

Reproducibility Variable Result LLM Response
Research Type Experimental Compared to the current state of the art, our model is simpler to train, significantly faster to sample from, and achieves better results on extensive in silico benchmarks the generated molecules are more diverse, exhibit fewer steric clashes, and bind with higher affinity to protein pockets.
Researcher Affiliation Industry 1Prescient Design, Genentech. Correspondence to: Pedro O. Pinheiro <pedro@opinheiro.com>, Saeed Saremi <saremis@gene.com>.
Pseudocode Yes Algorithm 1 c WJS, using the BAOAB scheme (Sachs et al., 2017, Algorithm 1) adapted for the conditional ULD (9).
Open Source Code No The paper mentions that the code for the Pose Check tool is available, but does not provide any concrete access to the source code for the Vox Bind method described in the paper.
Open Datasets Yes We benchmark our model on Cross Docked2020 (Francoeur et al., 2020), a popular dataset for pocket-conditional molecule generation.
Dataset Splits Yes 100,000 and 100 ligand-pocket pairs are partitioned into the training and test sets, respectively. We take 100 samples from the training set as our hold-out validation set.
Hardware Specification Yes Both models are trained for 340 epochs on four NVIDIA A100 GPUs (a total of ten days).
Software Dependencies No The paper mentions several software tools like Open Babel, RDKit, Auto Dock Vina, MMseqs2, and PyUUL along with their citations, but it does not explicitly provide specific version numbers for these software dependencies in the text.
Experiment Setup Yes The models are trained with batch size of 64, learning rate 10 5 and weight decay 10 2. The weights are updated with Adam W optimizer (Loshchilov & Hutter, 2019) (β1 = .9, β2 = .999) and exponential moving average with decay .999. Both models are trained for 340 epochs... We use δ = σ/2 and fix γ = 1/δ, i.e., the effective friction (Saremi et al., 2023) is set to 1.