Generative Flows on Synthetic Pathway for Drug Design
Authors: Seonghwan Seo, Minsu Kim, Tony Shen, Martin Ester, Jinkyoo Park, Sungsoo Ahn, Woo Youn Kim
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experimentally demonstrate that RXNFLOW outperforms existing reaction-based and fragment-based models in pocket-specific optimization across various target pockets. Furthermore, RXNFLOW achieves state-of-the-art performance on Cross Docked2020 for pocket-conditional generation, with an average Vina score of 8.85 kcal/mol and 34.8% synthesizability. The results for the first five targets are shown in Tables 1 and 2, and additional results for the 10 remaining targets are reported in Sec. D.1. The property distribution for each target are reported in Sec. D.2. |
| Researcher Affiliation | Collaboration | Seonghwan Seo1 , Minsu Kim1, Tony Shen2, Martin Ester2, Jinkyoo Park1,3, Sungsoo Ahn4, Woo Youn Kim1,5 1KAIST 2Simon Fraser University 3OMELET 4POSTECH 5HITS Correspondence to EMAIL |
| Pseudocode | Yes | Algorithm 1 Training GFlow Nets with action space subsampling 1: input Entire action space A, Maximum trajectory length N 2: repeat 3: Sample partial action spaces A 0, A 1, ..., A N 1 from subsampling policy P(A) 4: Sample trajectory τ from sampling policy πθ( | ; A ( )) 5: Update model θ θ η ˆLTB(τ) 6: until model converges |
| Open Source Code | Yes | Source code available at https://github.com/Seonghwan Seo/Rxn Flow. |
| Open Datasets | Yes | We generate 100 molecules for each of the 100 test pockets in the Cross Docked2020 benchmark (Francoeur et al., 2020) and evaluate them with the following metrics: Vina (kcal/mol)... |
| Dataset Splits | Yes | We generate 100 molecules for each of the 100 test pockets in the Cross Docked2020 benchmark (Francoeur et al., 2020) and evaluate them with the following metrics: Vina (kcal/mol)... |
| Hardware Specification | Yes | All experiments were performed on a single NVIDIA RTX A4000 GPU. |
| Software Dependencies | Yes | For a fair comparison with the baseline model, we used Uni Dock (Yu et al., 2023) for target-specific generation and Quick Vina 2.1 (Alhossary et al., 2015) for SBDD. The initial ligand conformer is generated with sr ETKDG3 (Wang et al., 2020) in RDKit (Landrum et al., 2013). To evaluate the synthetic accessibility of molecules, we used the retrosynthesis planning tool Ai Zynth Finder (Genheden et al., 2020). |
| Experiment Setup | Yes | Table 7: Default hyperparameters used in RXNFLOW training. Hyperparameters Values Minimum trajectory length 2 (minimum reaction steps: 1) Maximum trajectory length 4 (maximum reaction steps: 3) GFN temperature β Uniform(0, 64) Train random action probability 0.05 (5%) Action space subsampling ratio 1% Building block embedding size 64. For hyperparameters, we set the pocket embedding dimension to 128 and the training GFN temperature to Uniform(0, 64) which are used in Taco GFN. We trained the model with 40,000 oracles whereas Taco GFN is trained for 50,000 oracles. |