Computational chemists have developed a standardized Python framework named BONAFIDE to streamline the calculation of local atom and bond features from molecular structures. While global molecular descriptors that assign a single value or vector to an entire chemical entity have long been supported by established software suites, local descriptors have historically suffered from fragmentation across specialized packages with inconsistent interfaces. The new software, introduced in the Journal of Cheminformatics, provides a unified application programming interface that consolidates access to 11 external computational chemistry and cheminformatics tools alongside native from-scratch implementations. Designed to support data-driven chemistry and machine learning pipelines, the package standardizes data formats and tracks execution parameters to generate reproducible chemical representations.
The distinction between global and local molecular representations is critical across diverse chemical modeling tasks. Global descriptors capture macroscopic properties of a whole molecule, but they are insufficient for problems that depend on site-specific behavior, such as predicting the regio- and site-selectivity of chemical reactions, identifying sites of metabolization in medicinal chemistry, or detailing protein-ligand binding interactions. In these contexts, distinct atom and bond properties dictate reactivity and interaction geometry. By formalizing the extraction of local numerical vectors and scalar properties, the researchers sought to remove the technical friction associated with preparing input data for machine learning models and automated experimental systems.
At its core, BONAFIDE is built in Python 3.12 for Linux-based operating systems and is architected around an AtomBondFeaturizer class exposing 20 public methods. All user inputs, metadata, and computed values are stored inside an internal MolVault dataclass object, which manages data exchange between individual feature factory classes that inherit from a common base class. The architecture allows inputs in the form of Simplified Molecular Input Line Entry System strings, XYZ coordinates, Structure-Data Files, or native RDKit molecule objects. Explicit hydrogen atoms are featurized normally when supplied, while implicit hydrogens are omitted from calculations. Output can be retrieved as pandas DataFrames, hierarchical Python dictionaries, or RDKit molecule objects annotated with property fields, maintaining consistent indexing across mixed input sources.
In its release version 0.2.4, BONAFIDE supports 590 distinct local features, comprising 466 atom features and 124 bond features. Approximately two-thirds of these descriptors require three-dimensional molecular geometries, and nearly half depend on electronic structure data. The software integrates diverse external dependencies, including RDKit for two-dimensional descriptors and structural fingerprints, mendeleev for elemental atomic properties, and packages such as MORFEUS, DBSTEP, and kallisto for three-dimensional steric measurements like solvent accessible surface area and buried volume. DScribe provides fixed-length numerical representations such as Smooth Overlap of Atomic Positions and Atom-Centered Symmetry Functions, while Multiwfn contributes nearly half of the total descriptor library through detailed wavefunction and population analyses.
To support quantum mechanical property calculations without manual intervention, the platform integrates interfaces to xtb and Psi4. Users can run single-point energy calculations directly through semi-empirical tight-binding density functional theory in xtb or through higher-level density functional theory and wavefunction methods in Psi4. These engines generate wavefunctions that feed Multiwfn or calculate conformer energies used to determine Boltzmann-weighted feature averages across conformational ensembles. Furthermore, precomputed electronic structure files, such as formatted checkpoint or molden files, can be attached directly to avoid redundant calculations, ensuring computational efficiency when extensive quantum chemical datasets already exist.
Beyond interfacing third-party software, the developers implemented several native features directly within BONAFIDE to assist in pipeline construction. A key structural tool is the functional group matching feature, which leverages SMARTS-RX, a curated collection of 406 SMARTS patterns tailored for reactivity analysis, to locate specific substructures within target compounds. Another native tool evaluates symmetry equivalence, enabling algorithms to determine whether specific atoms or bonds are chemically equivalent, including resonance structures such as the oxygen atoms in a nitro group. By filtering out symmetry-redundant positions before initiating expensive calculations, workflows can significantly reduce unnecessary computational overhead.
To establish baseline computational costs, the authors conducted a benchmark featurization across 100 randomly selected molecules from PubChem, comprising 4,830 atoms and 4,999 bonds including explicit hydrogens. When evaluating 2D descriptors using four parallel processes, the calculation of 149 atom and 45 bond features completed in 6 minutes and 3 seconds of wall-clock time, with a mean effective central processing unit time of 14.2 seconds per molecule. The atom calculations averaged 3.9 seconds, while bond calculations required an average of 10.3 seconds due to the independent execution of the ALFABET machine learning model. These measurements demonstrate that two-dimensional featurization remains computationally accessible for moderately sized datasets.
Conversely, evaluating 3D descriptors across the same 100 molecules required substantially greater computational investment. Calculating 457 atom and 122 bond features across pre-optimized geometries and electronic structures obtained from Gaussian at the B3LYP/6-31G* level took 35 hours and 34 minutes of wall-clock time, reflecting an average CPU time of 68.4 minutes per molecule. Atom-level evaluations accounted for the vast majority of this duration, averaging 63.9 minutes, whereas bond evaluations averaged 4.5 minutes. The computational expense scaled steeply with molecular size: a compound containing 15 atoms and 16 bonds finished in under 5 minutes, while a complex structure with 175 atoms and 174 bonds required 17 hours, driven largely by per-atom steric calculations in MORFEUS and DBSTEP and surface analyses in Multiwfn.
The utility of the framework was further demonstrated through a multi-step featurization workflow targeting primary and secondary amines extracted from a 1,000-molecule PubChem sample. Using the automated SMARTS filter, the team isolated 247 amine-containing compounds, pruned symmetry-equivalent positions, and successfully executed descriptor calculations for 225 structures after discarding cases where quantum mechanical single-point calculations failed to converge. The workflow enabled direct benchmarking of graph neural network predictions from the qmdesc model against quantum chemical DFT calculations, yielding squared Pearson correlation coefficients of 0.738 for nitrogen electrophilic Fukui coefficients and 0.798 for N–H bond orders, illustrating how the framework facilitates validation between fast machine learning approximations and rigorous quantum mechanics.
The amine dataset was also examined via dimensionality reduction using Uniform Manifold Approximation and Projection to contrast different feature spaces. While standard 2,048-bit Morgan fingerprints clustered molecules predominantly by broad structural classes such as ring size and aliphatic nature, a 20-dimensional mixed descriptor space generated through BONAFIDE captured continuous electronic properties, such as average local ionization energy. Additionally, the mixed space successfully resolved subtle physical phenomena, separating cyclic amines of different ring sizes and identifying aliphatic amines where intramolecular hydrogen bonding between amino and hydroxyl groups produced distinct conformers, which was confirmed by conformational analysis.
Despite its broad feature coverage, the authors noted current operational boundaries within the software. BONAFIDE version 0.2.4 calculates bond features exclusively for established chemical bonds rather than arbitrary pairs of atoms, which presents limitations when characterizing non-equilibrium geometries or transition state structures in reaction modeling. Furthermore, the package does not yet include a standalone command-line interface or dedicated batch-processing utilities, and conformation generation or geometric optimizations must be carried out externally. The package has been released under an open-source Apache 2.0 license, providing a foundation for community contributions and integration into broader cheminformatics platforms.
To ensure robust execution across automated pipelines, the developers implemented rigorous software verification and configuration management throughout the framework. All feature factory hyperparameters are centralized within a hierarchical TOML file processed directly by the central featurization class, adopting dependency default settings while allowing users to inspect, adjust, and export configurations programmatically. The software enforces strict data validation on user inputs, including data types and numerical ranges, and writes operational events to a comprehensive log file to facilitate debugging. Across the full codebase, the authors achieved an overall test coverage of 93 percent, which included end-to-end automated testing for each implemented atom and bond descriptor.
The framework also incorporates specialized utilities designed to handle custom computational environments and novel feature descriptors. Through dedicated methods, researchers can introduce custom feature factories at runtime or bridge external Python libraries that require isolated environments due to dependency incompatibilities with the main package. In addition to scalar properties, the software processes fixed-length representations such as smooth overlap of atomic positions and fixed-size symmetry functions, formatting vector-based outputs as comma-separated string values. If a requested descriptor cannot be evaluated for a specific atom—such as an atom positioned outside a defined isovalue molecular surface—the framework explicitly assigns a dedicated inaccessible status label to prevent silent data corruption.
Looking ahead, the authors outlined potential avenues for extending the platform’s computational breadth and functionality. Planned enhancements include incorporating additional electronic structure engines such as Gaussian, ORCA, and g-xTB, integrating machine-learned interatomic potentials, and interfacing the Natural Bond Orbital program to extract specialized bonding descriptors. By facilitating transparent data provenance and reproducible descriptor extraction, the open-source toolkit provides an adaptable foundation for bridging quantum mechanical computations with emerging automated laboratory workflows and predictive modeling architectures in chemical data science.
Subject of Research: Cheminformatics software for local atom and bond descriptor calculation
Article Title: BONAFIDE: a Python framework for the calculation of local features for atoms and bonds in molecules
Article References: Sigmund, L. M., Assante, M., Ball, M., & Kabeshov, M. (2026). BONAFIDE: a Python framework for the calculation of local features for atoms and bonds in molecules. Journal of Cheminformatics, 18(1), Article 127. https://doi.org/10.1186/s13321-026-01283-6
Image Credits: AI Generated
DOI: 10.1186/s13321-026-01283-6
Keywords: BONAFIDE, Cheminformatics, Python Package, Molecular Descriptors, Atom Features, Bond Features, Machine Learning in Chemistry, Quantum Chemistry, Multiwfn, RDKit, xtb, Psi4, MORFEUS, DBSTEP, Feature Engineering, Density Functional Theory
Cite Scienmag News
Bethany Barker. (November 20, 2026). New Open Source Python Package Standardizes Molecular Atom and Bond Descriptors. Scienmag. https://scienmag.com/new-open-source-python-package-standardizes-molecular-atom-and-bond-descriptors/
Bethany Barker. "New Open Source Python Package Standardizes Molecular Atom and Bond Descriptors." Scienmag, 20 November 2026, https://scienmag.com/new-open-source-python-package-standardizes-molecular-atom-and-bond-descriptors/. Accessed 20 November 2026.
Bethany Barker. "New Open Source Python Package Standardizes Molecular Atom and Bond Descriptors." Scienmag. November 20, 2026. https://scienmag.com/new-open-source-python-package-standardizes-molecular-atom-and-bond-descriptors/








