SE(3) diffusion model with application to protein backbone generation
Authors: Jason Yim, Brian L. Trippe, Valentin De Bortoli, Emile Mathieu, Arnaud Doucet, Regina Barzilay, Tommi Jaakkola
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we find through experiments in Sec. 5 that Frame Diff can generate designable, diverse, and novel protein monomers up to length 500. |
| Researcher Affiliation | Academia | 1Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Massachusetts, USA 2Department of Statistics, Columbia University, New York, USA 3Center for Sciences of Data, French National Centre for Scientific Research, Paris, France 4Department of Engineering, University of Cambridge, Cambridge, United Kingdom 5Department of Statistics, University of Oxford, Oxford, United Kingdom. |
| Pseudocode | Yes | Algorithm 1 Frame Diff sampling of protein backbones. Algorithm 2 Timestep Batch. Algorithm 3 Training. |
| Open Source Code | Yes | Code: https://github. com/jasonkyuyim/se3_diffusion |
| Open Datasets | Yes | We train Frame Diff over monomers... downloaded from PDB (Berman et al., 2000) on August 8, 2021. |
| Dataset Splits | No | The paper mentions training data but does not explicitly describe train/validation/test splits or a separate validation set for hyperparameter tuning or model selection. |
| Hardware Specification | Yes | Our model comprises 17.4 million parameters and was trained for one week on two A100 Nvidia GPUs. |
| Software Dependencies | No | The paper mentions 'Py Torch Geometric' and 'Adam optimizer' but does not specify their version numbers. |
| Experiment Setup | Yes | Neural network hyperparameters: Global parameters: Dh = 256 Dz = 128 Dskip = 64 L = 4 IPA parameters: heads=8 query points=8 value points= 12 Transformer parameters: heads=4 layers=2. SDE parameters. Translations: schedule=linear βmin = 0.1 βmax = 20 Rotations: schedule=logarithmic σmin = 0.1 σmax = 1.5. We use Adam optimizer (Kingma & Ba, 2014) during training with learning rate 0.0001, β1 = 0.9, β2 = 0.999. |