Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation

Authors: Jiwoong Park, Yang Shen

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate our hierarchical molecular conformer generation framework via Equivariant Blurring Diffusion (EBD) on molecular conformer generation task. We conducted experiments to answer the following questions: i) Ablation studies (Sec. 5.2): What are the effects of granularity of the fragment vocabulary, loss reparameterization, and data corruption processes of diffusion models? ii) Geometric evaluation (Sec. 5.3): Can EBD generate more diverse and accurate molecular conformers in Euclidean space than previous deep generative approaches? iii) Property prediction (Sec. 5.4): Can EBD generate low-energy, stable conformers?
Researcher Affiliation Academia Jiwoong Park, Yang Shen Department of Electrical and Computer Engineering Texas A&M University ptywoong@gmail.com, yshen@tamu.edu
Pseudocode Yes In this subsection, we provide the Pytorch-style [37] pseudo-codes. The RDKit conformer generator to obtain the approximate fragment structure, linear interpolation blurring schedule, training process, and sampling process were given in Pseudo-codes 1, 2, 3, and 4, respectively.
Open Source Code Yes Codes are released at https://github.com/Shen-Lab/EBD.
Open Datasets Yes We use GEOM-QM9 (QM9) [39] and GEOM-Drugs (Drugs) [1] which are small molecules and drug-like molecules, respectively. We obtained the raw data, the pre-processed data and the data split at https://github.com/Deep Graph Learning/Conf GF.
Dataset Splits Yes Each dataset comprises 40,000 molecules for the training set and 5,000 molecules for the validation set, with each molecule containing 5 conformers following data split of [46].
Hardware Specification Yes We used a single NVIDIA A100 GPU for every training and generation tasks.
Software Dependencies No The paper mentions using PyTorch and RDKit but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes For EBD, we use the T = 50, a noise scale of 0.01 for the forward process (σ in Eq. 6) and 0.0125 for the reverse process (δ in Eq. 8) in every experiments. For training, we used a learning rate 10 4 with the Adam W optimizer [33]. Table 7: Hyperparameters of EBD. Dataset T # l # d # of hops cutoff batch size training iter. Drugs 50 6 128 3 10 Å 32 650k QM9 50 6 128 3 10 Å 64 650k