A Forecasting System under the Multiverse ansatz via Machine Learning and Cheminformatics

Specific scope : Computational Time of DFT/TDDFT Calculations for this version

• Thank PARATERA company (https://www.paratera.com/) for the cooperations

1. Catalogue

• database :
  • DrugBank : DrugBank all-in-one sdf file, and the scripts for generating the training and the testing suits
  • polyfitted : Fitted ploynormal equations for selected 89 reference DFT functionals
  • rawdata : The assembled Gaussuian09-D.01 timing data, and the separated sdf files with added H atoms
  • trained-models : Trained models for few DFT functional/basis set combinations
• example : The sample molecule to be predicted
• src : source code folder
• tools : Independent scripts
  • experimental : Some scripts in developing or experimental stage
• TRmod_kernel_A1.py : Training script sample
• Fcst_kernel_A1.py : Predicting script sample

2. Installation

• Prerequisities
  • python3 with numpy, scipy, scikit-learn
  • pytorch, with CUDA, cudatoolkit, torchvision, dgl, gensim
  • basis_set_exchange, libxc
  • rdkit, openbabel
  • optional: xlsxwriter, pillow
• Installation example (recommended with conda):
  • git clone git@github.com:yingjin-ma/Fcst_sys_public.git Fcst_sys_public
  • cd Fcst_sys_public
  • conda create -n Fcst_sys_public python=3.8 (3.7 is also tested)
  • conda activate Fcst_sys_public
  • conda install rdkit pytorch gensim torchvision numpy scipy xlsxwriter scikit-learn basis_set_exchange libxc matplotlib tqdm
    • Suggested installing order:
    • conda install rdkit
    • conda install pytorch=1.11.0=cuda112py38habe9d5a_1
    • conda install gensim torchvision numpy scipy xlsxwriter scikit-learn basis_set_exchange libxc matplotlib tqdm
  • install dgl cuda version (notice at least the major version should match that of installed cudatoolkit)
    • conda install -c dglteam dgl-cuda11.3
    • now: python TRmod_kernel_A1.py should work
  • Install the pylibxc (Please see https://www.tddft.org/programs/libxc/installation/)
    • now: python Fcst_kernel_A1.py should work

3. Usage

• python TRmod_kernel_A1.py for training
• python Fcst_kernel_A1.py for predicting
• python Fcst_kernel_A1_LB_wrapper.py for load-balancing
  • The "Predicted_Loads.txt" will be generated for later usage
• More will be added

4. Citation

• https://arxiv.org/abs/1911.05569v3
• https://pubs.acs.org/doi/10.1021/acsomega.0c04981

5. Acknowledgement

• National Key Research and Development Program of China (Grant No.2018YFB0203805)
• National Natural Science Foundation of China (Grant No.21703260)
• Informationization Program of the Chinese Academy of Science (Grant No.XXH13506-403)
• Guangdong Provincial Key Laboratory of Biocomputing (Grant No.2016B030301007)

6. Corresponding authors

• yingjin_ma@163.com and zjin@sccas.cn