|
a |
|
b/Code/Drug Discovery/pLDM/README.md |
|
|
1 |
## [Original Notebook](https://github.com/lamm-mit/ProteinMechanicsDiffusionDesign/blob/main/notebook_for_colab/pLDM_inference_standalone_colab.ipynb), [Original Paper](https://www.science.org/doi/10.1126/sciadv.adl4000), [Presentation PDF](https://drive.google.com/file/d/1pxL73oXK3Hn434YfERKi2aZ2GQ-HMvyF/view?usp=sharing) |
|
|
2 |
|
|
|
3 |
The following are modifications or improvements to original notebooks. Please refer to the authors' models for the published primary work. |
|
|
4 |
|
|
|
5 |
Conclusions: Total Run Time of Protein language diffusion model pLDM.pt 4.14GB file takes minutes to load; with other installations |
|
|
6 |
NVIDIA V100 High Ram GPU average of 8 protein generations = 10.71 seconds each |
|
|
7 |
NVIDIA A100 40GB GPU average of 8 protein generations = 8.87 seconds each |
|
|
8 |
|
|
|
9 |
Generative AI Cost Efficiency: NVIDIA V100 High Ram GPU average of 8 protein generations = 0.9528 |
|
|
10 |
Generative AI Cost Efficiency: NVIDIA A100 40GB GPU average of 8 protein generations = 0.4777 |
|
|
11 |
Equation: 1/(0.2 * Time average * Hourly gpu cost); V100 = 1/(0.2 * 10.71 * 0.49), A100 = 1/(0.2 * 8.87 * 1.18) |
|
|
12 |
|
|
|
13 |
Summary: Several minutes were needed for loading and installing packages prior to protein generations |
|
|
14 |
V100 High Ram GPU for protein generations was $0.49/hr, only 20% slower, and twice as cost efficient |
|
|
15 |
V100 is the better GPU choice for pLDM in many applications, Protein generation is practical for drug discovery |
|
|
16 |
|
|
|
17 |
ChemicalQDevice research & development of Ni, B., et al. 2024 Science Advances notebook |
|
|
18 |
The number of generations based on highest pull forces was increased to eight proteins |
|
|
19 |
Perplexity decreased with pulling force, likely associated with less alpha helix structural motifs |
|
|
20 |
GPU Cost efficiency study highlighted appropriateness of NVIDIA V100 High Ram GPU, shown above |