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..

Manipulating 3D Molecules in a Fixed-Dimensional E(3)-Equivariant Latent Space

Authors: Zitao Chen, Yinjun Jia, Zitong Tian, Wei-Ying Ma, Yanyan Lan

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments We train and evaluate Mol FLAE on three datasets: QM9 [38], GEOM-Drugs [39] and ZINC-9M (the in-stock subset of ZINC [40] with 9.3M molecules). Unconditional Molecule Generation To assess the capability of Mol FLAE generate stable, diverse molecules, we first focus on 3D molecule generation task following the setting of prior works [15, 16, 17]. We conduct 10,000 random samplings in the latent space, then decode them into molecules using Mol FLAE decoder, subsequently evaluating qualities of these molecules. We sample the atom number from the prior of the training set as previous works like [15]. Table 1 illustrates the benchmark results of unconditional generation with Mol FLAE.
Researcher Affiliation Academia 1 Institute for AI Industry Research (AIR), Tsinghua University 2 Department of Computer Science and Technology, Tsinghua University 3 Qiuzhen College, Tsinghua University 4 Beijing Frontier Research Center for Biological Structure, Tsinghua University 5 Beijing Academy of Artificial Intelligence
Pseudocode Yes Algorithm 1: Inference of General Bayes Flow Networks Algorithm 2: Forward Pass of Mol FLAE Algorithm 3: Decoder: sampling Molecules conditioned on latent code
Open Source Code No Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We will release our codes to reproduce the main experiment soon.
Open Datasets Yes We train and evaluate Mol FLAE on three datasets: QM9 [38], GEOM-Drugs [39] and ZINC-9M (the in-stock subset of ZINC [40] with 9.3M molecules).
Dataset Splits Yes Table 10: Hyperparameters for training. Parameter Value or description Train/Val/Test Splitting 6921421/996/remaining data for GEOM-DRUG 9322660/932/remaining data for ZINC-9M 100000/17748/remaining data for QM9
Hardware Specification Yes Table 9: Training costs. Dataset GPUs Time Max Epoch GEOM-DURG 4 Nvidia A100s(80G) 6 days 15 QM9 4 Nvidia A100s(80G) 16h 250 ZINC-9M 8 Nvidia A800s(80G) 3 days 25
Software Dependencies No In details, we sample 10,000 molecules from each model with their default settings, obtaining atom positions and types, and then inferred bond types using Open Babel. We then fix the bond order using Schordinger due to some bugs in Open Babel. The final molecules are then be evaluated using RDKit for the following metrics: QED, SA, Lipinski, and Strain Energy.
Experiment Setup Yes Table 10: Hyperparameters for training. Parameter Value or description Train/Val/Test Splitting ... Batch size 100 for GEOM-DRUG,200 for ZINC-9M,400 for QM9 Optimizer Adam β1 0.95 β2 0.99 Lr 0.005 Weight decay 0 Learning rate decay policy Reduce LROn Plateau Learning rate factor 0.4 for GEOM-DRUG, 0.6 for QM9 and ZINC-9M Patience 3 for GEOM-DRUG, 10 for QM9 and ZINC-9M Min learning rate 1.00E-06 Embedding dimension Df 128 Head number 16 Layer number 9 k (knn) 32 Activation function Re LU NZ 10 DZ 32 varx 100 varh 1 Reconstruction loss weight 1 Regularization loss weight 0.1