BONAFIDE: a Python framework for the calculation of local features for atoms and bonds in molecules Authors: Lukas M Sigmund, Michele Assante, Matthew Ball, Mikhail Kabeshov Published: 2026 ABSTRACT AbstractData-driven chemistry relies on robust, efficient, and reproducible pipelines for generating digital molecular representations. While comprehensive software for global molecular descriptors is widely available, the calculation of local atom and bond features remains fragmented across disparate tools with inconsistent interfaces. Here, we introduce the BOnd aNd Atom FeaturIzer and Descriptor Extractor (BONAFIDE), a Python package designed to standardize the generation of local descriptors. BONAFIDE provides a unified application programming interface that integrates 11 external computational chemistry and cheminformatics packages, including RDKit, MORFEUS, and Multiwfn as well as quantum chemical engines like xtb and Psi4 for single-point energy computations. The software currently supports the calculation of 590 distinct atom and bond features derived from either 2D or 3D molecular representations and electronic structure data. Key functionalities include flexible input format processing as well as output customization, seamless atom and bond feature calculation across all implemented descriptors, transparent hyperparameter tracking, comprehensive logging, and the possibility to implement custom featurizers within the BONAFIDE ecosystem. We also include runtime statistics for featurization and a case study that demonstrates how the tool is used for the featurization of amines through a multistep featurization pipeline. Current limitations and future development directions are discussed.Graphical Abstract Scientific contribution This work introduces a Python software package denoted BONAFIDE, which facilitates the calculation of descriptors for atoms and bonds in molecules. By providing a unified interface that bridges diverse cheminformatics and quantum chemistry tools, this package streamlines data generation pipelines and accelerates the development of robust machine learning models. BONAFIDE is available free of charge as open-source software under the Apache 2.0 license and can be obtained from GitHub (https://github.com/MolecularAI/atom-bond-featurizer).Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13321-026-01283-6.Keywords: Atom and bond descriptors, Local features, Reactivity prediction, Open-source software Introduction The paradigm of chemical research is increasingly shifting toward data-driven approaches [1, 2]. The underlying methods range from quantum mechanical (QM) calculations and molecular dynamics simulations to cheminformatics and statistical models obtained by the means of machine learning (ML) [3]. Going even further, computational systems based on large language models have been developed to orchestrate individual digital chemistry tools and eventually combine them with automated experimental laboratories to accelerate and drive chemical innovation within the realm of agentic artificial intelligence (agentic AI) [4, 5]. Besides the exact model architecture, the key enabler for every statistical tool is data of sufficient quality and quantity [6]. Consequently, robust, efficient, standardized, and reproducible data generation pipelines are desired. This is especially true for agentic systems and highly automated workflows as they depend on reliable and well-structured application programming interfaces (APIs). However, this also holds for small-scale exploration studies relying on manual data curation and analysis, which is otherwise very labor-intensive, prone to errors, and slow. In the context of chemistry, digital data generation refers to creating expressive machine-readable molecular representations for building (statistical) models. Over the years, a large number of descriptors and featurization approaches have been developed for this [7–13]. They differ from each other, for example, with respect to the underlying molecular representation (e.g., a 2-dimensional, 2D, molecular graph vs. a 3-dimensional, 3D, molecular structure), or whether or not they are derived from an electronic structure simulation such as a density functional theory (DFT) calculation. As deep learning methods have developed, increasingly informative learned representations can be obtained from neural networks [14, 15]. Which exact descriptor or a respective collection thereof is best suited for a given problem depends on various factors such as the size of the dataset at hand, the employed model architecture along with the exact definition of its endpoint, the available computational resources, and the desired degree of model accuracy and interpretability. Per se, it is not straightforward to easily identify the best feature space for a given problem, and often a combination of features from different sources leads to the best performing models [16–19]. Another important point for the differentiation of features for molecules is their classification into global and local atom and bond descriptors, respectively (Fig. 1). Global descriptors relate the entirety of a given molecule to a single numeric value or to a vector of predefined length. On the other hand, local descriptors describe the properties of an atom or bond, either through a single value or fixed-size vector [20]. While global descriptors are certainly very important for many modeling tasks, several scenarios can be envisioned for which local information, that is, descriptors for atoms and bonds, is indispensable. For example, the modeling of regio- and site-selectivity of chemical reactions requires local features to distinguish the potential sites of a molecule a given reaction can take place at [16, 19, 21–25]. More specifically and in the context of medicinal chemistry, site-of-metabolization predictions [26, 27] or the description of protein–ligand interactions can be based on local atom and bond descriptors [28, 29]. Furthermore, the influence of atom and bond descriptors was investigated in the context of chemical reaction condition [30] and molecular property predictions [31]. In total, this shows that the calculation of local features respecting the above-mentioned data generation guidelines is of high relevance for the field of data-driven chemistry. Comparison of global molecule-level descriptors and local atom- and bond-level descriptors, with the latter being the subject of this paper Several comprehensive (both open- and closed-source) software packages have been developed for the calculation of global features. Prominent tools are Mordred [32], RDKit [33], PaDEL-descriptor [34] along with its Python wrapper PaDELPy [35], alvaDesc [36], Cinfony [37], and more recently DOPtools [38] among other examples [39–42]. These tools offer rather straightforward access to thousands of molecule-level descriptors. Additionally, all major quantum chemistry programs directly output key quantum mechanical descriptors such as molecular orbital energies or the dipole moment along with the electronic structure data for further processing. Significant work has also been invested into increasingly streamlining and automating molecule featurization and ML model training, deployment, and evaluation [38, 43–47]. When it comes to the descriptor calculation of atoms and bonds, the situation is less well organized. Several separate software tools have been developed for the calculation of local features. However, they follow variable software design paradigms, require different input data, and produce non-uniform output. Moreover, it is the lack of a consistent access point that hampers the combination of features from different sources. Such a holistic API should also have a modular architecture to allow straightforward extension with new featurization tools. Also, there is no comprehensive overview of available atom and bond features for practitioners to choose from. There is also a lack of consistent logging during featurization pipeline execution including error recording and handling or the tracking of applied hyperparameters. To address the points raised, we have developed the BOnd aNd Atom FeaturIzer and Descriptor Extractor (BONAFIDE, Fig. 1). It is a Python package that provides a uniform API to 11 featurization dependencies, combined with a collection of from-scratch feature implementations. In the following sections, we give a general overview of the software, outline the included descriptor calculation dependencies, and describe the package’s individual functionalities. After that, runtime statistics for atom and bond featurization are given as well as a case study that showcases how BONAFIDE can be used to implement a multistep featurization pipeline. The BONAFIDE package includes detailed online documentation and is available from GitHub under the Apache 2.0 license. All data reported herein refers to the 0.2.4 version of the package. Results and discussion General overview The bonafide package was implemented in Python 3.12 and was developed for Linux-based operating systems. A general overview of the codebase is shown in Fig. 2. The AtomBondFeaturizer class delivers all functionalities to the user by exposing 20 public methods. They are described in more detail further below. All data, including the input and meta data as well as calculated features are stored in a MolVault dataclass object from which also the output data is extracted after feature calculation. A given feature is obtained by calling a feature factory class (one for each feature) that inherits functionalities common for all feature factories from the BaseFeaturizer class. Every feature is associated with a unique index through which it can be requested from BONAFIDE. After feature calculation, the data is stored as atom and bond properties of RDKit molecule objects, which themselves are stored within the MolVault dataclass object. All hyperparameters of the individual feature factories were compiled in a single hierarchical TOML file which is processed by the AtomBondFeaturizer class. The default values of the respective dependencies have been adopted with a few exceptions. The user can programmatically modify and export the hyperparameters, which allows transparent and reproducible atom and bond featurization. Every user input is strictly controlled for its correct data type and format, which also includes the featurization hyperparameters. All activities of the software package are collected in a comprehensive log file to ensure full transparency and to facilitate error handling. The functionalities of BONAFIDE were extensively tested with an overall test coverage of 93%, including the explicit end-to-end testing of all implemented atom and bond features. High-level schematic summary of the bonafide package. The AtomBondFeaturizer class stores all input data in a MolVault dataclass object. It also requests and receives feature data from the individual feature factories and stores them in the MolVault object, from which the final output data provided to the user is ultimately extracted BONAFIDE can process input in form of Simplified Molecular Input Line Entry System (SMILES) strings, XYZ and structure-data files (SDF), containing either one or multiple conformers of a given molecule. RDKit molecule objects can also be directly read in. The featurization of hydrogen atoms and respective X–H bonds depends on the user input: explicit hydrogens are preserved and featurized normally, while implicit hydrogens are not added to the system and remain unconsidered and unfeaturized. Generally, the molecule can either have a closed- or open-shell electronic structure. Consistent atom and bond indexing is ensured when combining different input formats. The final output can be obtained as a pandas DataFrame [48], a hierarchical dictionary, or as RDKit molecule object(s) with associated atom and bond properties. The data type of a given feature is either a string, Boolean, integer, or floating-point number. Feature vectors (e.g., atom-rooted structural fingerprints) are returned as strings with comma-separated values. In case a feature is not available or defined for an atom or bond it was requested for, it is returned as “_inaccessible”, with the leading underscore distinguishing this label from the rest of the descriptors. This can, for example, happen for atom features that are calculated on an iso-value molecular surface if a respective atom does not contribute to the surface as it was defined. Contrarily, if a feature is requested to be returned for a specific atom or bond but has not been calculated so far, a Not a Number (NaN) value is obtained. Besides providing access to local features, BONAFIDE also seamlessly interfaces two quantum chemistry software packages, which are xtb [49] and Psi4 [50–52], for carrying out single-point energy calculations. The interface follows a similar modular design principle as the feature factories (vide supra), which facilitates the potential addition of additional QM programs in future releases. For now, xtb and Psi4 were selected as open-source programs to cover a wide range of quantum chemical methods; from semi-empirical DFT to (hybrid) DFT and various wavefunction-based approaches. The QM interface allows the calculation of features from electronic structure data as well as Boltzmann-weighting of features across a conformational ensemble. Moreover, xtb can independently calculate a collection of atom features which can be directly requested from BONAFIDE. No atom or bond features are directly obtained from Psi4, but instead from Multiwfn (vide infra) [53, 54]. Both programs also support the inclusion of implicit solvation models for various solvents; by default, no solvation model is applied. Of note, as a featurization tool for atoms and bonds, BONAFIDE cannot be used to generate conformer ensembles of molecules or to run structure optimizations or frequency calculations. This must be achieved with separate software such as RDKit, OpenBabel [55], or CREST [56] combined with respective quantum chemical programs if required [57, 58]. Importantly, previously generated electronic structure data as well as molecular (Gibbs free) energies can be injected into BONAFIDE and can be used for descriptor calculation to prevent redundant re-computations. The entire feature generation routine within BONAFIDE has a highly modular structure. This facilitates the implementation of new feature factories in future releases of the package. At the same time, the AtomBondFeaturizer class also allows the incorporation of custom feature factories at run time to combine dedicated user features with the pre-implemented descriptors and BONAFIDE’s data processing routines. Notably, the package also provides functionality to enable access to Python packages that are incompatible with the BONAFIDE environment and therefore must be installed in separate environments. In its current form, BONAFIDE includes 590 features as defined by a unique feature identification string (Fig. 3). A full list of all features together with their metadata is part of the BONAFIDE web documentation. A total of 466 atom and 124 bond features can be calculated. For two thirds of the features, a 3D molecular representation is required. The Multiwfn dependency contributes with almost 50%, the majority of the descriptors. This means also that approximately half of the features demand the presence of electronic structure data, as all features from Multiwfn (excluding a few exceptions) are extracted from molden, formatted checkpoint (fchk), or similar electronic structure data files. A Distribution of atom and bond features that can be calculated with BONAFIDE for molecules for which 2D or 3D structural data is available. B Contributions of the individual dependencies for feature calculation. Psi4 is also a dependency of BONAFIDE but was excluded from the chart as no features are directly calculated with it. In both panels, the numbers in parenthesis refer to the absolute number of features Dependencies and from-scratch feature implementations BONAFIDE relies on 11 dependencies for descriptor calculation in its current form (cf. Figure 3B), complemented with a collection of features implemented from scratch (described at the end of this section). RDKit contains a pool of mostly 2D features for both atoms and bonds that have been implemented into BONAFIDE [33]. It can also be used to obtain various atom-rooted structural fingerprints [59, 60]. Besides for actual feature calculation, RDKit was used for all cheminformatics tasks within BONAFIDE as well as for storing calculated features as properties of atom or bond objects. The mendeleev Python package encompasses an extensive database of the properties of the chemical elements (e.g., various atomic radii or electronegativity scales), which can be accessed through BONAFIDE [61]. MORFEUS [62], DBSTEP [63], and kallisto [64] are packages for the calculation of local descriptors from 3D molecular structures, in particular targeting features for describing sterics such as the buried volume or solvent accessible surface area. Kallisto can also be used to rapidly calculate atomic partial charges based on the electronegativity equilibration method. Similarly, DScribe [65, 66] calculates features from 3D molecular representations, however, generates fixed-length numerical fingerprints, which can serve as input to machine learning models. Examples are Atom-Centered Symmetry Functions (ACSF) or Smooth Overlap of Atomic Positions (SOAP). The pretrained ML models ALFABET [67, 68] and qmdesc [22, 69] calculate bond dissociation energies and other bond properties as well as various atom descriptors important for reactivity modeling such as condensed Fukui coefficients or atomic partial charges through graph neural network inference. The obtained ML descriptors can then be combined with other features for downstream prediction tasks. Alternatively, QM analogs of the ML descriptors can be calculated with BONAFIDE to benchmark the statistical ML model for a given dataset (vide infra). Multiwfn is an extensive analysis tool for electronic wavefunctions or related data, obtained from a quantum chemical calculation, which can be operated through a modular text-based user interface (TUI; a graphical user interface is also available) [53, 54]. The TUI is highly ramified due to the plethora of available descriptors and their associated hyperparameters, which can complicate its usage. It is also challenging to survey all available descriptors that might be useful for a given task. Furthermore, the exact output format depends on the input to the program and on the respective feature. BONAFIDE resolves these challenges and integrates Multiwfn seamlessly into its featurization routine. The atom and bond features were grouped in subcategories, can easily be accessed through their unique feature index, all hyperparameters were compiled along with meaningful default values, and respective output parsing was implemented. Multiwfn is used within BONAFIDE to calculate features as the results, e.g., of population, bond order, molecular surface, topological, orbital composition, or atoms in molecules analyses, as well as for descriptors from conceptual DFT (C-DFT) [53, 54]. As quantum chemistry engines xtb [49] and Psi4 [50–52] are available in BONAFIDE for single-point energy calculations. Both programs can be used to generate input files for Multiwfn (vide supra) or to annotate the members of a conformational ensemble with their electronic energy. Xtb includes different extended semi-empirical tight-binding DFT models, which can calculate the electronic structure of typical small (organic) molecules in less than a second, which makes it attractive for the processing of large datasets. More accurate but also several orders of magnitude slower are the DFT or wavefunction calculations offered by Psi4. In addition to pre-existing software to generate features, we also implemented several independent features into BONAFIDE. The majority of them are not necessarily meant to provide direct input, e.g., for ML models but instead to assist in building efficient featurization pipelines (vide infra). A large number of the implemented features focus on the structural analysis of atoms and bonds relative to the rest of the atoms and bonds, respectively, for example, distance relationships in 2D (number of chemical bonds) and 3D (Euclidean distance) space. Of particular importance is the “functional_group_match” feature. It analyzes whether a given atom is part of a set of predefined functional group SMILES arbitrary target specification (SMARTS) patterns. For that, we used SMARTS-RX, a hierarchical set 406 SMARTS strings of typical organic functional groups tailored for reactivity analysis [70]. This allows to identify a certain target atom within a molecule and the subsequent calculation of features for this atom or its chemical vicinity accessible through the structural analysis features as described above. BONAFIDE makes it also possible to extend SMARTS-RX with custom SMARTS patterns. Furthermore, we implemented a “is_symmetric_to” feature for both atoms and bonds which allows to determine if a given atom or bond is chemically equivalent to one or multiple other atoms or bonds due to symmetry (including the optional consideration of resonance structures; cf. the oxygen atoms of the NO2 group.). This is important for building efficient featurization workflows by avoiding the calculation of symmetry-redundant feature vectors. Lastly, we also implemented atom-centered autocorrelation vector descriptors calculated from 2D molecular graphs which can be computed based on every numeric feature available within BONAFIDE, e.g., data from the mendeleev package (vide infra) [71, 72]. Individual functionalities In total, all of BONAFIDE’s functionality is contained in the AtomBondFeaturizer class, which exposes 20 public methods (cf. Figure 2). The list_atom_features and list_bond_features methods allow to list all available atom and bond features, respectively. The returned feature data frames can be filtered with respect to various criteria, e.g., whether a feature comes from a certain dependency, has a specific data type, or requires 3D molecular information or electronic structure data to be calculated. These methods enable to identify the indices of a desired set of features, either in a manual or programmatic way. The print_options and set_options methods allow to inspect and adjust the configuration settings of the individual featurization tools. This provides full transparency and flexibility over the exact feature calculation process. Meaningful default settings for all features are implemented into BONAFIDE, and the validity (i.e., datatype or numeric value) of a modified configuration setting is ensured before feature calculation. Every feature calculation with BONAFIDE must start by providing input to the read_input method. A valid input can be a SMILES string (2D structural data) or either an XYZ or SD file (3D structural data either with or without information on the bonds of the considered molecule) with one or multiple conformers per molecule. For the seamless integration into larger cheminformatics workflows, it is also possible to directly read in an RDKit molecule object through the read_input method. Additionally, the user must provide a name for the molecule for consistent logging and output file generation. For file and RDKit molecule inputs, it is also possible to directly read in the energy of the respective conformers to use them to calculate Boltzmann-weighted feature values. Alternatively, this is also possible through the attach_energy method. This method not only allows to define the energies of the actual molecule under consideration but also of its one-electron oxidized or one-electron reduced form. This permits the calculation of certain features from C-DFT based on the vertical ionization potential and electron affinity [73, 74]. Moreover, the charge and spin multiplicity of the molecule of interest can be specified though the set_charge and set_multiplicity method, respectively. A common scenario in computational chemistry is the availability of an XYZ file (no bond information contained) and a corresponding SMILES string of a given molecule. BONAFIDE facilitates the combination of both by first reading in the XYZ file (vide supra) followed by attaching the SMILES string through the attach_smiles method. This defines the chemical bonds within the molecule for feature calculation. The method also resolves the atom indexing problem, that is, a potentially different atom order in the XYZ and SMILES string, and the user can decide whether the atom order from the XYZ file or the SMILES string should be applied. Alternatively, if no SMILES string is available, the determine_bonds method can be used to define the chemical bonds of a molecule read from an XYZ file [75]. Almost all features from the Multiwfn program require access to electronic structure data files for descriptor calculation. BONAFIDE offers two options to meet this requirement: calculation from scratch or the use of precomputed data. The calculate_electronic_structure method initiates single-point energy calculations. For that, the user can decide between fast semi-empirical (DFT) calculations with xtb or much slower but more accurate DFT or wavefunction calculations with Psi4. The interaction between BONAFIDE and xtb and Psi4, respectively, was fully automated and no further interventions other than calling the calculate_electronic_structure method with the desired input and hyperparameters are required. Alternatively, already available data can be used through the attach_electronic_structure method. There is no restriction on the exact file format (e.g., molden or fchk) as long as it is processable by Multiwfn. As for the attach_energy method, both the calculate_electronic_structure and attach_electronic_structure method can be used to obtain or introduce data for the actual molecule as well as for its one-electron oxidized or reduced form. The actual calculation of atom and bond features is achieved through the featurize_atoms and featurize_bonds methods, respectively. The only required input for both methods are the indices of the target atoms or bonds as well as the indices of the desired features. After feature computation, all available data is cached to accelerate potential following featurizer calls. The requested data is stored as properties of atoms and bonds, respectively, of RDKit molecule objects. It is possible to clear the feature caches with the clear_atom_feature_cache and clear_bond_feature_cache methods. This is required if a feature should be recalculated with a changed configuration setting, e.g., within a more elaborate data generation pipeline or for benchmarking purposes. The return_atom_features and return_bond_features methods are used to obtain the calculated feature values. For the output format, the user can choose from a pandas DataFrame, a hierarchical dictionary, and an RDKit molecule object with the calculated data as properties of atoms or bonds. In the case of the featurization of an ensemble of conformers, it can be decided if all data for each conformer should be returned or if the minimum, maximum, and mean of a given feature across all conformers, and in case the energies of the conformers are available, additionally the data of the lowest- and highest-energy conformer as well as the Boltzmann-weighted mean should be calculated. By combining several of the methods presented so far, a working example of BONAFIDE is depicted in Fig. 4. Basic notebook example of using BONAFIDE for the calculation of all 2D RDKit features for all atoms in an example molecule starting from its SMILES string. The full output pandas DataFrame in the bottom has 19 rows and 45 columns but was truncated due to display space constraints A show_molecule utility method is included in the package to inspect the atom or bond indices either with a Lewis structure or a 3D representation of the molecule (in case available) within a Jupyter Notebook. This can help to identify certain atom or bond indices before feature calculation or to correlate descriptors with specific atoms or bonds after feature calculation. Ultimately, BONAFIDE provides an interface for the introduction of a custom feature factory through its add_custom_featurizer method. Therefore, features currently not implemented in BONAFIDE can be easily integrated into the featurization routine and processed together with already supported descriptors. A step-by-step tutorial on how to use the add_custom_featurizer method is provided on BONAFIDE’s documentation web page. Applications Runtime analysis First, the runtime of BONAFIDE for featurizing a given dataset of molecules was investigated. While most scenarios do not require the exhaustive calculation of all local descriptors, this investigation still provides an upper bound time estimate to define the limits of computational cost. To this end, we calculated all 2D or 3D features for all atoms and bonds, respectively, with the exclusion of a few descriptors that could not be routinely calculated for all atoms or bonds (see Chapter S2 in the SI for details). As the dataset, 100 randomly selected molecules from PubChem [76] were used totaling 4,830 atoms and 4,999 bonds (including explicit hydrogen atoms). The default hyperparameter settings implemented in BONAFIDE were maintained, and four featurization processes were run in parallel to speed-up the overall runtime of the feature calculations (for further details, see Chapter S2 in the SI). For the 2D atom and bond descriptors (149 and 45 features, respectively), the featurization was completed after 6 min and 3 s of wall-clock time. The mean effective central processing unit (CPU) time for featurization (that is, excluding parallelization) was 14.2 (2.2) seconds (mean with standard deviation in parenthesis) in total. This splits up into 3.9 (1.5) seconds for the atoms and 10.3 (0.8) seconds for the bonds. The higher time demand for bond featurization can be attributed to the ALFABET ML model, which is independently executed for every molecule. Next, featurization was performed for the respective 3D molecular structures. Both the 2D and 3D descriptors were calculated to cover the entire feature space implemented within BONAFIDE. One conformer per molecule was considered, and electronic structure calculations (B3LYP/6-31G* with Gaussian [77]) were conducted independently to decouple the computational cost of structure optimization and single-point energy calculations from the actual featurization process; all computational details are given in the SI. The calculation of all atom and bond descriptors (457 and 122 features, respectively) is significantly more computationally expensive. For the chosen 100 molecule dataset, the total wall-clock time increased to 35 h and 34 min, which corresponds to a mean CPU time of 68.4 (123.6) minutes per molecule. The respective processing times were 63.9 (117.6) minutes for atoms and 4.5 (6.5) minutes for bonds. This shows that the calculation of descriptors derived from 3D structures as well as electronic structure data is significantly more time consuming, but also heavily depends on the size of the molecule under consideration. The molecule with the lowest CPU time demand (15 atoms, 16 bonds) was completed after less than 5 min whereas the compound with the highest computational requirement (175 atoms, 174 bonds) was only done after 17 h. The features that took the longest to be calculated were 3D structural analysis features from DBSTEP and MORFEUS because they must be calculated separately for every atom as well as different surface descriptors and partial charges from Multiwfn; more details are given in Table S1 of the SI. Amine featurization workflow Ultimately, we used BONAFIDE to construct an entire data processing and featurization workflow. For this, we chose to target the chemical space of amines (primary and secondary, aromatic and aliphatic), which constitutes an important class of compounds in synthetic chemistry (Fig. 5A). Molecules containing aliphatic halogen groups were excluded. The workflow started by selecting the desired compounds from a set of 1,000 molecules randomly sampled from PubChem by using BONAFIDE’s “functional_group_match” feature (vide supra). For the resulting 247 molecules, the desired atom and bond indices were identified while considering the equivalency of atoms and bonds due to symmetry (“is_symmetric_to” feature, vide supra). Then, a selected list of atom and bond descriptors was calculated, which was successful for 225 molecules. For the full list, see Chapter S3 in the SI. A reason for failure was, for example, the non-convergence of a single-point energy calculation. The failed data points were excluded from further analyses. A Workflow for the featurization of amines implemented with BONAFIDE. B, C Comparison of ML and DFT data of the amines selected by the workflow. The values in parenthesis refer to the number of datapoints per subclass and the subclass-specific squared Pearson correlation coefficient. Dimensionality-reduced chemical space of the selected amines based on D a 20D mixture of different 2D and 3D atom and bond features and E atom-rooted Morgan fingerprints of size 2,048 and radius 3. In all cases of 3D features, Boltzmann-weighted properties were used. For further details, see Chapter S3 in the SI First, the obtained data was used to benchmark the output from the ML model qmdesc against DFT data (B3LYP-D3(BJ)/def2-mSVP). This was done for the nitrogen atom-centered Fukui coefficient for electrophilic attack, f( −) (Fig. 5B), and for the bond order of the N–H bond (Fig. 5C). In both cases, the linear correlation is good, with R2 values of 0.738 and 0.798. Interestingly, for both primary and secondary aromatic amines, the correlation in f( −) is notably higher than for the aliphatic subgroups. Conceivable explanations could be the higher conformational flexibility of aliphatic amines, which was not recovered by the ML model, or the possibly higher prevalence of aromatic amines in the training set of qmdesc. In contrast, for the bond order, this trend could not be detected and only the primary aliphatic amines stood out with a higher R2 value. Analyses like these are important to monitor the performance of much faster ML models against the respective QM data to assert the applicability of the model for a given dataset and to ensure that the gain in computational efficiency does not come at the cost of predictive reliability of the descriptors themselves or of downstream models. BONAFIDE facilitates such analyses. Second, the calculated features were used to visualize the investigated chemical space by means of dimensionality reduction. For that, the Uniform Manifold Approximation and Projection (UMAP)[78] was chosen to compare a 20D atom and bond descriptor space sourced through BONAFIDE from six individual featurization tools (Fig. 5D) with an atom-rooted Morgan fingerprint representation (Fig. 5E, see Chapter S3 in the SI for details). The latter embedding is sparse, mostly reproducing the basic chemical categories of amines (cyclic vs. non-cyclic, primary vs. secondary, and aromatic vs. aliphatic). The former also resembles these categories to a certain extent but is more continuous. At the same time, it maps out, for example, the average local ionization energy (ALIE) as an important property of amines for describing their nucleophilicity/basicity by continuously increasing from the bottom right to the top left of the plot. To illustrate the difference between the two representations with a concrete example, the relative location of the cyclic amines I and II in the two 2D-UMAP spaces were compared. In panel E of Fig. 5, they are placed closely together due to their high structural similarity, whereas in D they are far away from each other. In fact, the ring size of cyclic amines can have a significant influence on the properties and reactivity of the amine nitrogen as for example demonstrated in organocatalysis [79]. This could be reflected by the underlying chosen features of Fig. 5D. Furthermore, the primary aliphatic amines III and IV were found clearly separated from their cluster located in the bottom of panel D. In both cases, the formation of a hydrogen bond with free hydroxyl groups in close proximity to the amine nitrogen atom is conceivable (cf. the structures of III and IV in box at the bottom of Fig. 5). Indeed, visual inspections of the 3D molecular structures after conformer generation and structure optimization confirmed this hypothesis with the conformers exhibiting the intramolecular hydrogen bond being clearly the lowest energy ones for both compounds. This important structural detail is reproduced by the selected atom and bond features, which demonstrates their usefulness in effectively exploring amine chemical space. For comparison, in panel E, compounds III and IV were found clustered with the rest of the primary aliphatic amines. In summary, these examples show how different feature spaces can lead to distinct molecular representations, which must be selected based on the given problem at hand as well as with respect to, e.g., availability of computational resources or model runtime restrictions. With BONAFIDE, the individual representations can easily be accessed, benchmarked, and assessed for their application in downstream tasks. Conclusion Dedicated features for atoms and bonds in molecules are important local descriptors, e.g., for modelling the selectivity of chemical reactions [25]. They are a valuable complement to global molecular features widely used in digital chemistry. This work presents the Python-based BOnd aNd Atom FeaturIzer and Descriptor Extractor (BONAFIDE). BONAFIDE provides a consistent API for local descriptor generation and supports, in its current state (version 0.2.4), 11 external dependencies for feature calculation resulting in a total of 590 features. It also includes direct access to quantum chemistry software for single-point energy calculations at various levels of theory. Features can be obtained from 2D or 3D molecular structures or from the results of electronic structure simulations. Therefore, BONAFIDE brings together structure/steric descriptors and data from quantum chemistry – an often-recognized powerful combination, especially for smaller datasets [17, 22, 31, 80–82]. With the development of the presented software package, we aim to facilitate and standardize the combination of features from different sources to accelerate the construction of robust and increasingly generalizable ML models in chemistry across a variety of use case scenarios [83]. Flexible input and output formats as well as the possibility of implementing custom featurizers increases the amenability of BONAFIDE for the integration into (already existing) overarching data workstreams and emerging agentic AI architectures. Moreover, the broad diversity of descriptors will allow the investigation of less explored features for certain research questions beyond the typically deployed features; for example driven by fine-grained mechanistic understanding in the context of chemical reactivity modeling [3]. We investigated the runtimes of BONAFIDE for different descriptor calculation scenarios and demonstrated how it can be used to target and featurize a predefined area of chemical space with a multi-step featurization pipeline. A current limitation of BONAFIDE is that bond features can only be calculated for actual chemical bonds rather than for an arbitrary pair of atoms, which might be disadvantageous for specific use cases. Due to that, the featurization of non-equilibrium configurations such as transition structures could be complicated as well, depending on the desired set of features [17, 82]. Future work can focus on the inclusion of additional dependencies such as Gaussian [77], ORCA [84], or g-xTB [85] as quantum chemistry engines, the Natural Bond Orbital (NBO) program [86] for feature generation, or machine-learned interatomic potentials (MLIP) [87, 88]. BONAFIDE already includes two ML models for atom and bond property prediction, and we believe that the continuous integration of ML models into BONAFIDE (and overarching frameworks in general) in a sustainable and well-documented manner will decrease the hurdles of employing such tools and will therefore facilitate their widespread adoption across the entire chemical research landscape. Future releases of the package can also work toward the inclusion of a command line interface (CLI), which is currently not implemented or to dedicated functionality for batch-processing. The BONAFIDE source code is publicly available, and we encourage the scientific community to make active use of the package as well as to report potential bugs, request new features, or contribute ideas for further extensions. We are planning to actively maintain the online documentation and GitHub repository of BONAFIDE. The package is already utilized in several internal AstraZeneca projects, and we are confident that it will prove equally valuable to the broader scientific community, fostering further innovation in data-driven chemistry. Supplementary Information Supplementary Material 1. Acknowledgements L.M.S. and M.A. are/were part of the AstraZeneca Postdoc Program and are grateful for its support. The authors thank C. Kannas for code review as well as K. Jorner and P.-O. Norrby for interesting discussions. Author contributions All authors contributed to devising the project. L.M.S. developed the software with support of M.A. and M. B.. L.M.S compiled the online documentation, conducted and analyzed the case studies, and wrote the manuscript. All authors contributed to the finalization of the manuscript and agreed to publish the submitted content. The AI tools GPT-5 and Gemini 3 Pro were used for grammar, wording, and punctuation review as well as for the rephrasing of individual sentences and paragraphs. GitHub Copilot (GPT-4.1 and Claude Sonnet 4.5) was used throughout the development process of BONAFIDE. All AI generated text or code was carefully reviewed by the authors before incorporation into the manuscript, Python scripts, or Jupyter Notebooks. Funding M.B. is supported by the European Union's Horizon 2022 research and innovation program under the Marie Skłodowska-Curie Actions, grant agreement “Explainable AI for Molecules - AiChemist” No. 101120466. Data availability The source code of the bonafide package can be found at https://github.com/MolecularAI/atom-bond-featurizer and the online documentation at https://molecularai.github.io/atom-bond-featurizer/. The scripts to reproduce the case studies presented herein are also part of the GitHub repository and can be found in the examples folder along with a getting-started notebook. The version of the code employed for this study is 0.2.4 (see https://github.com/MolecularAI/atom-bond-featurizer/releases for the archived source code, commit ea987d6), and all data reported herein refers to this version. Additional data and explanations supporting the herein presented findings are included in the Supplementary Information. Declarations Competing interests The authors declare no competing interests. Availability and requirements • Project name: BONAFIDE - A Python package for the calculation of features and descriptors for atoms and bonds in molecules • Project home page: https://molecularai.github.io/atom-bond-featurizer/ • Operating system(s): Linux • Programming language: Python • Other requirements: Python 3.12 and listed dependencies in the environment file on GitHub • License: Apache 2.0 Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Michele Assante and Matthew Ball have contributed equally to this work. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The source code of the bonafide package can be found at https://github.com/MolecularAI/atom-bond-featurizer and the online documentation at https://molecularai.github.io/atom-bond-featurizer/. The scripts to reproduce the case studies presented herein are also part of the GitHub repository and can be found in the examples folder along with a getting-started notebook. The version of the code employed for this study is 0.2.4 (see https://github.com/MolecularAI/atom-bond-featurizer/releases for the archived source code, commit ea987d6), and all data reported herein refers to this version. Additional data and explanations supporting the herein presented findings are included in the Supplementary Information.