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..
Symmetry-Preserving Conformer Ensemble Networks for Molecular Representation Learning
Authors: Yanqiao Zhu, Yidan Shi, Yuanzhou Chen, Fang Sun, Yizhou Sun, Wei Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present comprehensive experimental evaluations of SPi CE across multiple molecular datasets and tasks. We first describe our datasets and experimental setup, then present results and analyses. We aim to answer the following research questions: RQ1: Performance. How does SPi CE perform across different molecular property prediction tasks compared to state-of-the-art conformer ensemble methods? RQ2: Scalability. How does the performance of SPi CE scale with dataset size? RQ3: Architecture design. What is the impact of key architectural choices in SPi CE? |
| Researcher Affiliation | Academia | Yanqiao Zhu Yidan Shi Yuanzhou Chen Fang Sun Yizhou Sun Wei Wang Department of Computer Science, University of California, Los Angeles EMAIL EMAIL |
| Pseudocode | Yes | A Py Torch-like pseudocode of SPi CE is provided in Appendix B. Algorithm S1 Pseudocode of SPi CE in a Py Torch-like style. |
| Open Source Code | Yes | The implementation of this work can be found in this repository: https://github.com/DannieSYD/SPiCE. |
| Open Datasets | Yes | We evaluate SPi CE on four datasets spanning both regression and classification tasks: (1) Drugs-7.5K, obtained by downsampling 10% of Drugs-75K [51] with a fixed random seed due to computational constraints, with three quantum mechanical properties: IP, EA, and χ, (2) Kraken [52] with four 3D ligand descriptors: (Sterimol B5, Sterimol L, Bur B5, Bur L), and (3-4) Co V2 and Co V23CL from GEOM-Drugs [53]: Co V2 measures general inhibition in human cells, while Co V2-3CL specifically targets the 3CL protease inhibition. |
| Dataset Splits | Yes | Following prior works, for regression datasets, we randomly partition data into training, validation, and test sets with a 7:1:2 ratio, while classification datasets use fixed public splits. We present detailed statistics and descriptions of datasets in Appendix E. Table S2: Dataset statistics for all four datasets and dataset splits for COVID-related datasets. Split Co V2-3CL Co V2 Train 50 (485) 53 (3,294) Validation 15 (157) 17 (1,096) Test 11 (162) 22 (1,086) Total 76 (804) 92 (5,476) (b) Dataset partitioning showing number of active compounds (total compounds in parentheses) |
| Hardware Specification | No | The paper discusses computational constraints and efficiency but does not specify the exact hardware (e.g., GPU/CPU models) used for running the experiments. |
| Software Dependencies | No | The paper mentions several software components like PyTorch (implied), SiLU activation, GCN layers, Gumbel-Sigmoid sampling, GIN layers, AdamW optimizer, Open Babel, and RDKit. However, specific version numbers for these components are not provided. |
| Experiment Setup | Yes | For regression tasks, each backbone model with a different aggregation strategy is trained with the same set of hyperparameters, optimized through 20 iterations of Bayesian Optimization. Specific settings are summarized in Table S1 for regression datasets. For classification tasks, all experiments use the same settings as Pai NN on Drugs-7.5K, as classification is less sensitive to hyperparameters than regression and Drugs-7.5K is a reliable molecular property prediction benchmark. Table S1: Hyperparameters for each backbone model on dataset Drugs-7.5K and Kraken. Dataset Backbone Epochs Batch LR Patience Experts Act. Experts Upcycle τ β Drugs-7.5K Pai NN 2000 32 2e-4 400 8 2 100 0.1 1e-3 ... (table continues with specific values) |