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].
TacoGFN: Target-conditioned GFlowNet for Structure-based Drug Design
Authors: Tony Shen, Seonghwan Seo, Grayson Lee, Mohit Pandey, Jason R Smith, Artem Cherkasov, Woo Youn Kim, Martin Ester
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the generative setting for Cross Docked2020 benchmark, Taco GFN attains a state-of-the-art success rate of 56.0% and 8.44 kcal/mol in median Vina Dock score while improving the generation time by multiple orders of magnitude. Fine-tuning Taco GFN further improves the median Vina Dock score to 10.93 kcal/mol and the success rate to 88.8%, outperforming all optimization-based methods. |
| Researcher Affiliation | Collaboration | a School of Computing Science, Simon Fraser University, Burnaby, Canada. b Department of Chemistry, KAIST, Daejeon, Republic of Korea. c Graduate School of Data Science, KAIST, Daejeon, Republic of Korea. d HITS Inc., Seoul, Republic of Korea. e Vancouver Prostate Centre, University of British Columbia, Vancouver, Canada. |
| Pseudocode | No | The paper describes methods using prose and mathematical equations but does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | In this study, we used the open-sourced code for GFlow Net (Bengio et al., 2021), Pharmaco Net (Seo & Kim, 2023) and GVP-GNN (Jing et al., 2021). Our models were implemented using the Pytorch (Paszke et al., 2019) and Py Torch Geometric (Fey & Lenssen, 2019) libraries, which enabled efficient training and evaluation. We utilized RDKit (Landrum et al., 2006), a widely-used chem-informatics library, to handle the molecular structures and compute chemical properties. We employed the Quick Vina 2.1 (QVina) (Alhossary et al., 2015) and Uni Dock (Yu et al., 2023) for docking, and used Openbabel (O Boyle et al., 2011) and Auto Dock Tools (Huey et al., 2012) to generate ready-to-dock files. |
| Open Datasets | Yes | We train and evaluate Taco GFN on the commonly used Cross Docked benchmark (Francoeur et al., 2020). We first apply the splitting and processing protocol on the Cross Docked dataset to obtain the same train and test split of the 100k protein-ligand pairs as previous methods (Luo et al., 2021; Peng et al., 2022; Guan et al., 2023b). (See details in Appendix D.1). We then train our docking score predictor on the training split of the protein-ligand pairs to predict their corresponding Vina Dock scores. |
| Dataset Splits | Yes | We first apply the splitting and processing protocol on the Cross Docked dataset to obtain the same train and test split of the 100k protein-ligand pairs as previous methods (Luo et al., 2021; Peng et al., 2022; Guan et al., 2023b). ... After splitting, 100,000 protein-ligand pairs are randomly drawn for the training set. 100 test pockets are drawn from the remaining pocket clusters. 15,307 unique protein pockets remain in the training set. |
| Hardware Specification | Yes | We report the time taken for fine-tuning and docking molecules for our method, measured on a single A4000 GPU a GPU with less performance than the one used by Evo SBDD (A6000 GPU), or Target Diff+Opt/Decomp Opt (A100 GPU). The training process takes about a day on 4 NVIDIA RTX A4000 GPUs. |
| Software Dependencies | No | In this study, we used the open-sourced code for GFlow Net (Bengio et al., 2021), Pharmaco Net (Seo & Kim, 2023) and GVP-GNN (Jing et al., 2021). Our models were implemented using the Pytorch (Paszke et al., 2019) and Py Torch Geometric (Fey & Lenssen, 2019) libraries, which enabled efficient training and evaluation. We utilized RDKit (Landrum et al., 2006), a widely-used chem-informatics library, to handle the molecular structures and compute chemical properties. We employed the Quick Vina 2.1 (QVina) (Alhossary et al., 2015) and Uni Dock (Yu et al., 2023) for docking, and used Openbabel (O Boyle et al., 2011) and Auto Dock Tools (Huey et al., 2012) to generate ready-to-dock files. |
| Experiment Setup | Yes | Table 7: Hyperparameters used for target conditional GFlow Net Num of training steps 30,000 Learning rate 10-4 Weight decay 10-8 Momentum 0.9 Adam eps 10-8 Sampling τ 0.99 Learning rate Z estimator 10-3 Max nodes 9 Random action prob 0.01 Batch size 8 Training reward temp β Uniform(0, 64) Inference reward temp 64 Pocket cond dim 128 Transformer hidden dim 256 Num of transformer layers 2 QED threshold t QED 0.7 SA threshold t SA 0.8 |