About

This repository provides a bucket of models from "Neural Mulliken Analysis: Molecular Graphs from Density Matrices for QSPR on Raw Quantum-Chemical Data" paper (also available as a preprint).

A notebook with minimal inference example is also available (see rhnet2.ipynb). For general RhNet2 model implementation see the corresponding repository.

Note: provided here, ensemble of models is probably to heavy to play with. For a lighter model trained on a larger dataset see another repository

Contents

• Repo structure
• Model

Repo structure

+-- rhnet2.ipynb                # Data preprocessing, model loading and inference example (No TF-GNN needed)
+-- rhnet2_with_tfgnn.ipynb     # Data preprocessing, model loading and inference example (with TF-GNN)
+-- data_utils.py               # Data preprocessing utilities
+-- graph_utils.py              # TF-GNN graph data preprocessing utilities
+-- models.config               # Tensorflow Model Server config
+-- rhnet2.pbtxt                # TF-GNN graph schema
+-- models                      # RhNet2SC1 models
|   +-- LOO_G8D1W1Hd0Hw0_GKANTrueWKANFalseHKANFalseL2H0.08L2W0.08_batch8_lr0.0001
|   |   +-- 50-06-6             # Single model trained on the SC-1 set. Phenobarbital (CASRN 50-06-6) excluded.
|   |   +-- 50-33-9             # Single model trained on the SC-1 set. Phenylbutazone (CASRN 50-33-9) excluded. 
|   |   +-- ...
+-- data_example                # Examples of DFT calculations of the SC-1 test set compounds
|   +-- dft                     # ORCA outputs
|   |   +-- Acebutolol.zip      # ORCA calculation for Acebutolol
|   |   +-- ...
|   +-- tfrecords               # Same data converted to TF-GNN graphs
|   |   +-- Acebutolol.tfrecord # Acebutolol density graph

Model

The model provided here was trained and tested using the training and test sets from the First Solubility Challenge (SC-1, link, link). For model and fitting details see the corresponding paper (also available as a preprint).