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 [1].
SE3Set: Harnessing Equivariant Hypergraph Neural Networks for Molecular Representation Learning
Authors: Hongfei Wu, Lijun Wu, Guoqing Liu, Zhirong Liu, Bin Shao, Zun Wang
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | SE3Set has shown performance on par with state-of-the-art (SOTA) models for small molecule datasets like QM9 and MD17. It demonstrates outstanding performance on the MD22 dataset, achieving a remarkable 20% improvement in accuracy across all molecules. Furthermore, on the OE62 dataset, SE3Set outperforms all short-range models. We also conducted a detailed analysis of OE62, highlighting the prevalence of complex many-body interactions in large molecules. [...] We tested our equivariant hypergraph neural network on QM9 (Ruddigkeit et al., 2012; Ramakrishnan et al., 2014), MD17 (Chmiela et al., 2017) (see Appendix G), MD22 (Chmiela et al., 2023), and OE62 (Stuke et al., 2020) to assess its molecular representation learning. QM9 and MD17 gauge small molecule property prediction, while MD22 and OE62 evaluate larger systems with complex many-body interactions (Wang et al., 2023). An ablation study was also conducted to pinpoint the contributions of fragmentation and architecture to our method s performance, offering insights into the network s efficacy and areas for enhancement. |
| Researcher Affiliation | Collaboration | Lijun Wu Microsoft Research AI4Science, Beijing, 100084, China; Zhirong Liu Liu Zhi EMAIL College of Chemistry and Molecular Engineering, Peking University, Beijing, 100871, China; Bin Shao EMAIL Microsoft Research AI4Science, Beijing, 100084, China |
| Pseudocode | Yes | Algorithm 1 Pseudo code of fragmentation merge step. Input: groups {G}, minimum atoms number nmin, maximum atoms number nmax, Topological bond order matrix B, isolated threshold cis m = |{G}| Isolate groups {GI} = {} Calculate fragmentation bond order matrix WGi Gj. Sort {G} in descending order based on the following attributes: number of atoms, P G ,G =G WGG . repeat Pop last fragment as Gk from {G} for i = m 1 to 1 do a = MAX_INT, merge_idx = 1 if any Bij 1, i Gij Gk and |Gi| < a and a + |Gi| nmax then a = |Gi|, merge_idx = i end if end for if merge_idx == 1 then for i = m 1 to 1 do if any WGi Gk cis and |Gi| < a then a = |Gi|, merge_idx = 1 end if end for end if if merge_idx = 1 then Merge Gk to Gmerge_idx Resort {G} by the same priority and update W. else Add {Gk} to {GI} end if until |{Gk}| nmin {F} = {G} {GI} |
| Open Source Code | Yes | The code of this work is available at https://github.com/Navantock/SE3Set. |
| Open Datasets | Yes | We tested our equivariant hypergraph neural network on QM9 (Ruddigkeit et al., 2012; Ramakrishnan et al., 2014), MD17 (Chmiela et al., 2017) (see Appendix G), MD22 (Chmiela et al., 2023), and OE62 (Stuke et al., 2020) to assess its molecular representation learning. |
| Dataset Splits | Yes | SE3Set, after training on 110k QM9 molecules and validation on 10k, achieves low mean absolute errors (MAEs) in 12 tasks, performing on par with leading models, as detailed in Table 1. [...] We partition the training/test set following Quin Net (Wang et al., 2023). [...] The OE62 dataset (Stuke et al., 2020) contains about 62k large organic molecules with annotated DFTcomputed energies in the unit of e V. We randomly selected 50,000 data points as the training set and 6,000 data points as the validation set following the work of Kosmala et al. (2023). [...] Table 2: Molecule # Train/Val s Ac-Ala3-NHMe 5500/500 Energy [...] DHA 7500/500 Energy [...] Stachyose 7500/500 Energy [...] AT-AT 2500/500 Energy [...] AT-AT-CG-CG 1500/500 Energy |
| Hardware Specification | Yes | On the QM9 and MD17 datasets, our model was trained using a single Tesla V100 GPU with 32GB of memory, except for the 6-layer model employing the exponential bond order on the QM9 dataset, which was trained on two Tesla V100 GPUs with 32GB each. For MD22 dataset and OE62 dataset, our model was trained on a single Tesla A100 GPU with 80GB of memory. [...] All tests were performed on a single Tesla A100 80G. |
| Software Dependencies | No | Our dataset construction is founded on Py Torch Geometric (Fey & Lenssen, 2019) augmented with our fragmentation process (Sec. 4.1). Due to inconsistencies in molecular topology identified through RDKit s sanitization routine (Landrum et al., 2020), 1,403 data points were excised from the original dataset. The paper mentions software (PyTorch Geometric, RDKit) but does not provide specific version numbers for them. |
| Experiment Setup | Yes | This section outlines the training specifics, encompassing the fragmentation parameters, SE3Set hyperparameters, and certain implementation nuances utilized in our experimental setup. [...] Table 9: Hyper-parameters for training SE3Set model. In the context of hyperparameter settings for dimensions, the symbols e and o are used to denote even and odd parity, respectively. [...] Table 10: Hyper-parameters for fragmentation. The expanded threshold does not work for models training on the MD22 dataset because they adapt implicit overlap scheme. [...] Table 11: Hyper-parameters for step 4* of implicit overlap scheme in MD22 and OE62 experiments. |