Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform
Authors: Peizhi Niu, Yu-Hsiang Wang, Vishal Rana, Chetan Rupakheti, Abhishek Pandey, Olgica Milenkovic
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present next the results of running DMol on several benchmarking datasets. For experiments pertaining to other graph models (e.g., SBMs and planar graphs), the reader is referred to Appendix G.1. Benchmarks. We compared DMol to several graph generation methods. The benchmarking models used in the experiments include Graph VAE Simonovsky and Komodakis [2018], GT-VAE Mitton et al. [2021], Set2Graph VAE Vignac and Frossard [2021], SPECTRE Martinkus et al. [2022b], Graph NVP Madhawa et al. [2019], GDSS Jo et al. [2022b], Gru M Jo et al. [2024], Di Gress Vignac et al. [2023], Dis Co Xu et al. [2024b] and De Fo G Qin et al. [2025]. Data. We tested the performance of different generative models on the QM9 Wu et al. [2018], MOSES Polykovskiy et al. [2020], and GUACAMOL Brown et al. [2019] datasets. ... Results. The experimental results are presented in Tables 1, 2, 3 and 4. On QM9, DMol achieves the best performance compared to other baseline models. |
| Researcher Affiliation | Collaboration | Peizhi Niu1, Yu-Hsiang Wang1, Vishal Rana1, Chetan Rupakheti2, Abhishek Pandey2, Olgica Milenkovic1 1University of Illinois Urbana-Champaign 2Abb Vie EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Forward Process Input: G0 = (X0, E0), state transition matrices QX&QE, node&edge selection functions N( )&M( ). Process: Sample t {1 . . . T}. Sample Gt = ( X t, E t) from X0QX E0QE. Calculate u = N(t) and v = M(t). Randomly select u nodes and v edges from the subgraph induced by the u nodes and generate filter matrices Kx and Ke which serve as indicators of the selections. Set Gt = G0 and replace Kx Xt with Kx X t, and Ke Et with Ke E t. Algorithm 2 Sampling Process Input: Denoising Network ϕθ, node&edge class distribution m T x = [px,T i ]&m T e = [pe,T k ], state transition matrices QX&QE. Process: Sample ˆGT from m T x m T e . for t = T to 1 do z f( ˆGt, t) {Structural and spectral features}. ˆp X, ˆp E ϕθ( ˆGt, z, t, Rx, Re) {Prediction}. Sample ˆG0 = (ˆX 0, ˆE 0) from ˆp X ˆp E. Sample Gt 1 = ( X t 1, E t 1) from ˆX 0QX ˆE 0QE. Calculate u = N(t 1), v = M(t 1). Randomly select u nodes and v edges from the subgraph induced by the u nodes and generate corresponding filter matrices Kx and Ke. Set ˆGt 1 = ˆG0 and replace Kx ˆX t 1 with Kx X t 1, KeˆE t 1 with Ke E t 1. end for |
| Open Source Code | Yes | 1The link to the code can be found at: https://github.com/liekon/Discrete-Graph-Generation |
| Open Datasets | Yes | Data. We tested the performance of different generative models on the QM9 Wu et al. [2018], MOSES Polykovskiy et al. [2020], and GUACAMOL Brown et al. [2019] datasets. |
| Dataset Splits | Yes | QM9 is a small dataset: we used 100K molecules for training, 20K for validation, and 13K for testing. Both MOSES and GUACAMOL contain millions of molecules, with the number of heavy atoms inside a single molecule bounded by 26 and 88, respectively. We used 85% of the molecules for training, 5% for validation, and 10% for testing. |
| Hardware Specification | Yes | The experiments were conducted using NVIDIA H100 GPUs. The QM9 runs were executed on a single GPU, while the MOSES and GUACAMOL experiments were processed in parallel using 4 GPUs. |
| Software Dependencies | No | The paper mentions software such as RDKit, Mol2Vec, and Grover, but does not provide specific version numbers for these or any other software dependencies, which would be necessary for full reproducibility. |
| Experiment Setup | Yes | Setup. For each molecule dataset, we choose a different (relatively small) number of motifs to be converted into supernodes. The selected motif structures can be found in the Appendix A. We set the hyperparameters of DMol in both setting to k = 2, r = 0.2, λ1 = 5, λ2 = 1, λ3 = 1. The number of training epochs is set to 500. |