De novo Protein Design Using Geometric Vector Field Networks
Authors: Weian Mao, Muzhi Zhu, Zheng Sun, Shuaike Shen, Lin Yuanbo Wu, Hao Chen, Chunhua Shen
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results consistently demonstrate the remarkable performance of VFN. For protein diffusion, VFN-Diff significantly outperforms the prior solid baseline, Frame Diff (Yim et al., 2023), in terms of designability (67.04% vs. 53.58%) and diversity (66.54% vs. 51.98%). |
| Researcher Affiliation | Collaboration | Weian Mao1,2 , Muzhi Zhu1 , Zheng Sun3 , Shuaike Shen1, Lin Yuanbo Wu3, Hao Chen1, Chunhua Shen1,4 1 Zhejiang University, China 2 The University of Adelaide, Australia 3 Swansea University, UK 4 Ant Group |
| Pseudocode | Yes | In this section, we first present the pseudocode for IPA and VFN to provide an intuitive comparison, as shown in Subsection A.1.1. ... Algorithm 1 The pseudo-code for the IPA module. ... Algorithm 2 The pseudo-code for the VFN module. |
| Open Source Code | Yes | Code is available at https://github.com/aim-uofa/VFN |
| Open Datasets | Yes | In inverse folding, we followed the settings of (Gao et al., 2022) and tested the sequence recovery performance of VFN on the CATH 4.2 (Orengo et al., 1997), TS50, and TS500 datasets (Li et al., 2014). ... The training dataset consisted of proteins from the PDB database (Berman et al., 2000) in August 2023, encompassing 21,399 proteins with lengths ranging from 60 to 512 and a resolution of < 5 A. |
| Dataset Splits | Yes | The dataset consists of 18,024 proteins for training, 608 for validation, and 1120 for testing. |
| Hardware Specification | Yes | Frame Diff was re-trained using the same standards as in (Yim et al., 2023), while VFN-Diff was trained on four NVIDIA 4090 GPUs for a total duration of 10 days and 14 hours. |
| Software Dependencies | No | Our training regimen employs the Adam optimizer with the following hyperparameters: a learning rate of 0.0001, β1 set to 0.9, and β2 set to 0.999. ... VFN-IF are trained with batch size 8 and are optimized by Adam W with a weight decay of 0.1. No specific software versions for libraries like PyTorch, TensorFlow, or CUDA are mentioned. |
| Experiment Setup | Yes | Our training regimen employs the Adam optimizer with the following hyperparameters: a learning rate of 0.0001, β1 set to 0.9, and β2 set to 0.999. ... VFN-IF are trained with batch size 8 and are optimized by Adam W with a weight decay of 0.1. We apply a One Cycle scheduler with a learning rate of 0.001 and train our model for a total of 100,000 iterations. |