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..
DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules
Authors: Junjie Xu, Jiahao Zhang, Mangal Prakash, Xiang Zhang, Suhang Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Dual Equi Net across established RNA and protein property prediction datasets, demonstrating consistent improvements over existing geometric baselines (with average performance gains of 8.4% 33.5% across tasks). To further probe the model s ability to capture nuanced 3D structural features, we developed two new benchmark datasets for Solvent-Accessible Surface Area (SASA) Prediction and Torsion Angle Prediction...Dual Equi Net achieves improvements of 3.1%-28.8% and 2.5%, respectively, over the strongest prior methods on these benchmarks as well. Code and datasets will be released publicly upon paper acceptance. |
| Researcher Affiliation | Academia | Junjie Xu1, Jiahao Zhang1, Mangal Prakash2, Xiang Zhang1, Suhang Wang1 1The Pennsylvania State University 2Independent Researcher EMAIL, EMAIL |
| Pseudocode | No | The paper includes architectural diagrams (Figure 1) and mathematical equations describing the methodology, but no explicitly labeled 'Pseudocode' or 'Algorithm' blocks are present. |
| Open Source Code | Yes | The source code is available at https://github.com/junjie-xu/dualequinet. |
| Open Datasets | Yes | We use three RNA datasets consisting of thousands of atoms: (i) Covid Vaccine [74], which provides nucleotide-level reactivity and degradation labels relevant for vaccine stability; (ii) Ribonanza [37], which includes nucleotide-level reactivity under different compounds to assess RNA folding and flexibility; and (iii) Tc-ribo [33], which contains RNA-level regulatory behavior in response to tetracycline. ... For protein SASA prediction, we use the existing m RFP protein dataset from [68]... Additionally, we introduce a new SASA benchmark dataset for RNA, sourced from RNASolo [2]... The Torsion Angle dataset uses ground-truth RNA sequences and structures from RNASolo... |
| Dataset Splits | Yes | All models are evaluated using consistent train-validation-test splits (8:1:1), with hyperparameters selected based on the best validation performance, optimized via Optuna [3] to ensure fair comparisons. ... For each dataset, we adopt an 8:1:1 split for training, validation, and testing. |
| Hardware Specification | Yes | All experiments were conducted on a single NVIDIA A6000 GPU with a number of parameters limit up to 4 × 10^6. |
| Software Dependencies | No | Our implementation is built on PyTorch and PyTorch Geometric [25]. While the software packages are mentioned, specific version numbers are not provided, which is necessary for full reproducibility of software dependencies. |
| Experiment Setup | Yes | Table 12: Optimal hyperparameters of Dual Equi on each dataset. Dataset Covid Ribonanza Tcribo RNASolo SASA m RFP Protein SASA Torsion Angle Latom 4 1 3 4 1 3 Lnt 3 4 3 1 1 4 # heads 2 8 4 6 3 5 hidden 72 128 84 150 174 140 atom d EU 4.82 3.93 3.03 10.44 3.68 8.93 residue d EU 26.4 30.61 122.64 47.02 116.39 42.45 lr 0.0030 0.0015 0.0002 0.0002 0.0005 0.0013 weight decay 1.37e-6 5.65e-8 1.17e-7 1.37e-5 3.00e-7 2.10e-7 lmax 2 2 2 2 3 3 atom d SH 0.97 1.0 0.95 1.00 0.96 0.99 residue d SH 0.97 0.95 0.92 0.95 0.95 0.91 batch size 32 25 32 16 16 16 epochs 1000 1000 1000 1000 1000 1000 |