Aligning Target-Aware Molecule Diffusion Models with Exact Energy Optimization
Authors: Siyi Gu, Minkai Xu, Alexander Powers, Weili Nie, Tomas Geffner, Karsten Kreis, Jure Leskovec, Arash Vahdat, Stefano Ermon
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical studies on the Cross Docked2020 benchmark show that ALIDIFF can generate molecules with state-of-the-art binding energies with up to -7.07 Avg. Vina Score, while maintaining strong molecular properties. ... We conduct comprehensive comparisons and ablation studies on the Cross Docked2020 benchmark to justify the effectiveness of ALIDIFF. |
| Researcher Affiliation | Collaboration | Siyi Gu 1, Minkai Xu 1 , Alexander Powers1, Weili Nie2, Tomas Geffner2 Karsten Kreis2, Jure Leskovec1, Arash Vahdat2, Stefano Ermon1 1 Stanford Univeristy 2 NVIDIA |
| Pseudocode | Yes | The pseudo-code for ALIDIFF and ALIDIFF-T are provided below. Algorithm 1 Training Procedure ALIDIFF; Algorithm 2 Training Procedure for ALIDIFF-T. |
| Open Source Code | Yes | Code is available at https://github.com/Minkai Xu/Ali Diff. |
| Open Datasets | Yes | We train and evaluate ALIDIFF using the Cross Docked2020 dataset [Francoeur et al., 2020]. |
| Dataset Splits | No | The paper states 'The final dataset uses a train and test split of 65K and 100,' but does not explicitly provide information about a separate validation split or its size/methodology. |
| Hardware Specification | Yes | Our parameterized diffusion denoising model, IPDiff was trained on a single NVIDIA A6000 GPU and achieved convergence within 200k steps. ... We trained our model with one NVIDIA Ge Force GTX A100 GPU, and it could converge within 30k steps. |
| Software Dependencies | No | The paper mentions using the 'Adam optimizer' but does not specify version numbers for any key software components or libraries required to replicate the experiments (e.g., Python, PyTorch, specific ML frameworks and their versions). |
| Experiment Setup | Yes | We adopted the Adam optimizer with a learning rate of 0.001 and parameters β values of (0.95, 0.999). The training was conducted with a batch size of 4 and a gradient norm clipping value of 8. To balance the losses for atom type and atom position, we applied a scaling factor λ of 100 to the atom type loss. Additionally, we introduced Gaussian noise with a standard deviation of 0.1 to the protein atom coordinates as a form of data augmentation. ... For finetuning, the pre-trained diffusion model is further fine-tuned via the gradient descent method Adam with init learning rate=5e-6, betas=(0.95,0.999). We keep other setting the same as pretraining. We use β = 5 in Equation (5). |