Independent SE(3)-Equivariant Models for End-to-End Rigid Protein Docking
Authors: Octavian-Eugen Ganea, Xinyuan Huang, Charlotte Bunne, Yatao Bian, Regina Barzilay, Tommi S. Jaakkola, Andreas Krause
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we achieve significant running time improvements and often outperform existing docking software despite not relying on heavy candidate sampling, structure refinement, or templates. |
| Researcher Affiliation | Collaboration | Octavian-Eugen Ganea MIT Xinyuan Huang ETH Zurich Charlotte Bunne ETH Zurich Yatao Bian Tencent AI Lab Regina Barzilay MIT Tommi Jaakkola MIT Andreas Krause ETH Zurich |
| Pseudocode | No | The paper provides mathematical equations and descriptions of its model components, such as the IEGMN in equations (5)-(11), but these are not presented in a clearly labeled "Pseudocode" block or "Algorithm" figure. |
| Open Source Code | Yes | Our code is publicly available: https://github.com/octavian-ganea/equidock_public. |
| Open Datasets | Yes | Datasets. We leverage the following datasets: Docking Benchmark 5.5 (DB5.5) (Vreven et al., 2015) and Database of Interacting Protein Structures (DIPS) (Townshend et al., 2019). DB5.5 is obtained from https://zlab.umassmed.edu/ benchmark/, while DIPS is downloaded from https://github.com/drorlab/DIPS. |
| Dataset Splits | Yes | Datasets are then randomly partitioned in train/val/test splits of sizes 203/25/25 (DB5.5) and 39,937/974/965 (DIPS). |
| Hardware Specification | Yes | Hardware specifications are as follows: ATTRACT was run on a 6-Core Intel Core i7 2.2 GHz CPU; HDOCK was run on a single Intel Xeon Gold 6230 2.1 GHz CPU; EQUIDOCK was run on a single Intel Core i9-9880H 2.3 GHz CPU. |
| Software Dependencies | No | The paper mentions using "Adam (Kingma and Ba, 2014)" and "Py Torch" for implementation, and the "POT library (Flamary et al., 2021)" for optimal transport, but it does not specify version numbers for these software components. |
| Experiment Setup | Yes | Training Details. We train our models on the train part of DIPS first, using Adam (Kingma and Ba, 2014) with learning rate 2e-4 and early stopping with patience of 30 epochs. We update the best validation model only when it achieves a score of less than 98% of the previous best validation score, where the score is the median of Ligand RMSD on the full DIPS validation set. The best DIPS validated model is then tested on the DIPS test set. For DB5.5, we fine tune the DIPS pre-trained model on the DB5.5 training set using learning rate 1e-4 and early stopping with 150 epochs patience. Table 3: Hyperparameter choices. |