TurboHopp: Accelerated Molecule Scaffold Hopping with Consistency Models
Authors: Kiwoong Yoo, Owen Oertell, Junhyun Lee, Sanghoon Lee, Jaewoo Kang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we compare 3D scaffold generation qualities between Turbo Hopp and Diff Hopp (Torge et al. [2023]), a DDPM-based scaffold-hopping model. For Turbo Hopp, we also vary the number of steps in multi-step generation. |
| Researcher Affiliation | Collaboration | Kiwoong Yoo1, Owen Oertell2, Junhyun Lee3, Sanghoon Lee1,3, and Jaewoo Kang 1,3 1AIGEN Sciences 2Cornell University 3Korea University |
| Pseudocode | Yes | A Algorithm Pseudocode In this section we present the pseudocode for consistency model training (Algorithm 1), consistency model sampling (Algorithm 2), RL training (Algorithm 3), and inpainting. (Algorithm 4) |
| Open Source Code | Yes | The code is provided at https://github.com/orgw/Turbo Hopp |
| Open Datasets | Yes | Dataset We follow the dataset preprocessing scheme regarding filtering of compounds and determining of scaffolds as done in Torge et al. [2023], filtering those above QED of 0.3, training on 19,378 protein-ligand complexes in PDBBind. We also adopt the same scaffold extraction method, using Murko-Bemis method (Bemis and Murcko [1996]). |
| Dataset Splits | No | The paper mentions training on PDBBind and following a train-test split suggested in another paper, but does not explicitly state its own validation split percentages, sample counts, or methodology. |
| Hardware Specification | Yes | Device: 4x NVIDIA A100 GPUs (for Turbo Hopp training) and Device: 8x NVIDIA A100 GPUs (for RLCM training). |
| Software Dependencies | No | The paper mentions specific tools like 'Autodock GPU' and 'QVina2' and an optimizer 'Adam', but it does not list specific versions for general software dependencies like programming languages or libraries (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | B Information on Hyperparameters and Experiment Details Parameter setting for Turbo Hopp Setting Parameters... timesteps: 150, 100, 50, 25 batch size: 256 lr: 1e-4 schedule: Reduce LROn Plateau (min: 1e-6, factor: 0.9) num epochs: 5500 σmin: 0.002 σmax: 80.0 σdata: 0.5 ρ: 7.0 |