3D molecule generation by denoising voxel grids

Authors: Pedro O. O. Pinheiro, Joshua Rackers, Joseph Kleinhenz, Michael Maser, Omar Mahmood, Andrew Watkins, Stephen Ra, Vishnu Sresht, Saeed Saremi

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments In this section, we evaluate the performance of our model on the task of unconditional 3D molecule generation. Our approach is the first of its kind and therefore the objective of our experiments is to show that (i) Vox Mol is a feasible approach for unconditional generation (this is non-trivial) and (ii) it scales well with data, beating a established model on a large, drug-like dataset.
Researcher Affiliation Industry Prescient Design, Genentech
Pseudocode Yes Algorithm 1: Walk-jump sampling [1] using the discretization of Langevin diffusion by [85].
Open Source Code No The paper mentions using 'the official implementation from https://github.com/cvignac/Mi Di' for baselines, but does not provide a statement or link for the open-source code of their own method, Vox Mol.
Open Datasets Yes We consider two popular datasets for this task: QM9 [40] and GEOM-drugs [41].
Dataset Splits Yes We perform the same pre-processing and dataset split as [20] and end up with 100K/20K/13K molecules for QM9 and 1.1M/146K/146K for GEOM-drugs (train, validation, test splits respectively).
Hardware Specification Yes All experiments and analysis on this paper were done on A100 GPUs and with Py Torch [97]. The models on QM9 were trained with 2 GPUs and the models on GEOM-drugs on 4 GPUs.
Software Dependencies No The paper mentions software like PyTorch [97], RDKit [90], Open Babel [89], and Py UUL [94], but does not provide specific version numbers for any of these dependencies.
Experiment Setup Yes Our models are trained with noise level σ=.9, unless stated otherwise. We train our models with batch size of 128 and 64 (for QM9 and GEOM-drugs, respectively) and we use Adam W [88] (learning rate 2 10 5, weight decay 10 2) to optimize the weights. The weights are updated with exponential moving average with a decay of .999. We use γ =1.0, u=1.0 and δ=.5 for all our MCMC samplings.