colliner/fraggraph

FragGraph: A Fragmentation-Based Graph Embedding Framework for QM/ML

FragGraph is a framework of molecular descriptors for deep learning combining ideas from graph theory, fragmentation, error-cancellation, and machine learning. These representations can be derived from chemical graph theory with node and edge attributes from standard fingerprinting techniques on atom-centric fragmention from the fragmentation scheme of the Connectivity Based Hierarchy (CBH) of generalized isodemic reactions.

This repository provides scripts to generate the FragGraph attributed graph representation.

Authors:

• Eric M. Collins (Indiana University) [email protected]
• Krishnan Raghavachari (Indiana University) [email protected]

Performance of QM/ML models on 13k Hold-out set from GDB9-G4(MP2)

Baseline (QM)	Model (GN)	MAE (kcal/mol)
PM7	FG(CBH-2)	0.50
PM7	FC-FG(CBH-2)	0.38
B3LYP	FG(CBH-2)	0.16
B3LYP	FC-FG(CBH-2)	0.12

QM/ML Learning Curve

Installation & Usage

Requirements

• python (version>=3.6)
• rdkit (version>=2020.03.4)
• numpy
• networkx
• mol2vec
• xyz2mol
• pytorch
• torch-geometric

Installation via conda

Clone the repository: git clone https://github.com/colliner/FragGraph.git

Run make conda_env to create the conda environment.

Testing the installation

Activate the conda environment: conda activate FragGraph

Evaluate pretrained model

cd model; python eval_model.py