graphpancake: a Python package for representing organic molecules as molecular graphs utilizing electronic structure theory

Computational methods for predictive modeling have been increasingly utilized in the early stages of drug discovery to supplement high-throughput screening. The advent of highly efficient and complex machine learning architectures necessitates new methods of collating the plethora of topological, geometrical, and quantum chemical data from small molecule drug candidates to interface with these advanced cheminformatics methods. Graph-based molecular representations encoding atomic and bonding features allow physicochemical meaning to be leveraged by computational methods to predict quantitative structure–activity relationships (QSAR) with high precision and accuracy. We present graphpancake, an open-source Python package that translates quantum mechanical data from density functional theory (DFT) and post-Hartree Fock (HF) wavefunction calculations into molecular graphs of small organic molecules. graphpancake is a quantum chemical graph generator intended for cheminformatics pipelines featuring command line utility, hierarchical graph types with varying feature complexity, and thorough user documentation. To test graphpancake’s utility, we have curated several datasets (regression and classification) from the literature. Using random forest and message passing neural network architectures, we demonstrate how graphpancake’s quantum-enriched features can predict quantitative properties more accurately (R2 values 0.3–0.5 higher) compared to SMILES-generated features. Our code is available on GitHub (MIT license). We present an innovative solution toward the reconciliation of chemical “intuition” with computational methods to construct more informative translations of chemical data for cheminformatics software in drug discovery. graphpancake is an efficient, open-source Python package that generates quantum chemical molecular graphs and features command line utility and database functionality, allowing for integration into custom cheminformatics pipelines. We demonstrate how graphpancake's utilization of electronic stucture theory and subsequent derivation of molecular graphs from optimized quantum chemical, geometric, and computed properties allows for the facilitation of quantitative property prediction using machine learning.