LinkerNet: Fragment Poses and Linker Co-Design with 3D Equivariant Diffusion
Authors: Jiaqi Guan, Xingang Peng, PeiQi Jiang, Yunan Luo, Jian Peng, Jianzhu Ma
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical studies on ZINC and PROTAC-DB datasets demonstrate that our model can generate chemically valid, synthetically-accessible, and low-energy molecules under both unconstrained and constrained generation settings. |
| Researcher Affiliation | Academia | Jiaqi Guan University of Illinois Urbana-Champaign jiaqi@illinois.edu Xingang Peng Peking University xingang.peng@gmail.com Peiqi Jiang Tsinghua University jpq20@mails.tsinghua.edu.cn Yunan Luo Georgia Institute of Technology yunan@gatech.edu Jian Peng University of Illinois Urbana-Champaign jianpeng@illinois.edu Jianzhu Ma Tsinghua University majianzhu@tsinghua.edu.cn |
| Pseudocode | Yes | Algorithm 1 Training Procedure of Linker Net ... Algorithm 2 Sampling Procedure of Linker Net |
| Open Source Code | Yes | Reproducibility Statements The model implementation, experimental data and model checkpoints can be found here: https://github.com/guanjq/Linker Net |
| Open Datasets | Yes | We use a subset of ZINC [43] for the unconstrained generation. ... For the constrained generation, we use PROTAC-DB [45], a database collecting PROTACs from the literature or calculated by programs. |
| Dataset Splits | Yes | We use the same procedure as [21] to create fragments-linker pairs and randomly split the dataset, which results in a training/validation/test set with 438,610 / 400 / 400 examples. For the constrained generation, we use PROTAC-DB [45], a database collecting PROTACs from the literature or calculated by programs. The same procedure is applied to obtain reference conformations and create data pairs. We select 10 different warheads as the test set (43 examples) and the remaining as the training set (992 examples). |
| Hardware Specification | Yes | We trained our model on one NVIDIA RTX A6000 GPU |
| Software Dependencies | No | The paper mentions "Adam W [30]" as an optimizer but does not provide specific software versions for libraries, frameworks, or programming languages (e.g., PyTorch version, Python version). |
| Experiment Setup | Yes | The model is trained via Adam W [30] with init_learning_rate=5e-4, betas=(0.99, 0.999), batch_size=64 and clip_gradient_norm=50.0. To balance the scales of different losses, we multiply a factor λ = 100 on the atom type loss and bond type loss. During the training phase, we add a small Gaussian noise with a standard deviation of 0.05 to linker atom coordinates as data augmentation. We also schedule to decay the learning rate exponentially with a factor of 0.6 and a minimum learning rate of 1e-6. The learning rate is decayed if there is no improvement for the validation loss in 10 consecutive evaluations. The evaluation is performed for every 2000 training steps. |