Publicly available antioxidant activity data for a diverse range of compounds was retrieved from the PubChem database. To control methodological consistency within the dataset, data was exclusively sourced from well-established assays: ABTS and DPPH. ABTS and DPPH assays quantify a compound's free radical scavenging (ABTS) or reduction (DPPH) capacity, reflected by a measurable color change [4,5]. Selection of these assays was based on their widespread adoption in rapid antioxidant potential screening due to their simplicity and effectiveness. PubChem searches employing the keywords "DPPH" and "ABTS" identified a total of 1,651 DPPH and 366 ABTS assays encompassing 19,454 compounds.

To eliminate redundancy arising from compounds with identical structures but varying identifiers, the International Union of Pure and Applied Chemistry (IUPAC) InChIKeys was adopted as the unique identifier. The RDKit Python package was utilized to convert the Simplified Molecular Input Line Entry System (SMILES) strings for each compound into IUPAC InChIKeys. This process refined the initial 19,454 PubChem CIDs into 10,053 unique compounds. Subsequently, Extended-Connectivity Fingerprints (ECFP) with a radius of 4 (ECFP-4) were generated using the RDKit Python package to represent chemical structures based on each compound’s SMILES string. Compounds were categorized into four activity groups for each assay: ‘Active’, ‘Inactive’, ‘Unspecified’, or ‘Inconclusive’, and those with solely ‘Inconclusive’ and ‘Unspecified’ designations across all assays were excluded. In addition, compounds exhibiting inconsistent activity results between the two assays were eliminated. To enhance the dataset's comprehensiveness, 24 well-known antioxidant compounds documented in prior studies were incorporated and designated as the ‘Active’ set [11].

The Tanimoto coefficient [12] between each pair of all compounds was computed using their ECFP-4 fingerprint via the DataStructs Python package. ECFP-4 fingerprint is a binary vector of 2,048 bits in length, where each bit represents the presence (set to 1) or absence (set to 0) of specific circular substructures within a molecule. The Tanimoto coefficient itself is a value between 0 and 1, representing the degree of structural similarity between two molecules. A higher coefficient indicates a greater degree of structural similarity. The calculation is as follows:

where Nab is the number of common features between the two compounds, Na is the total number of features in compound A, and Nb is the total number of features in compound B. To visualize the relationships between the compounds based on the calculated Tanimoto coefficients, a compound-compound network was constructed. In this network, each compound is represented by a node, and edges connect nodes that exhibit a Tanimoto coefficient exceeding a predefined threshold (set to 0.7 in this case). Cytoscape 3 (version 10.3.1), an open-source software platform for network visualization, was used to generate this network representation.

Five ML algorithms were chosen for their capability to model complex relationships between molecular structure and antioxidant activity: Support Vector Machine (SVM), Logistic Regression (LR), XGBoost (XGB), Random Forest (RF), and Deep Neural Network (DNN). Each model was trained via functions from the scikit-learn Python package with default parameter settings to prevent overfitting and ensure generalizability (Table 1, Supplementary Fig. 1). For robust and unbiased evaluation, five-fold cross-validation was conducted using two data splitting strategies: (i) random splitting and (ii) scaffold splitting. To achieve scaffold splitting, all compounds were classified into scaffold groups using a Scaffold Network Generator [13] implemented in the RDKit Python package. This tool organized compounds into a hierarchical tree structure, with groups ranging from single-ring to a maximum of 15-ring structures. Consequently, 1,931 compounds were categorized into 778 scaffold groups. The training data was then split based on scaffold membership, ensuring the testing set contained compounds with unseen scaffold structures. Using both splitting strategies, the five-fold cross-validation was repeated 100 times. Performance metrics including accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUROC) were computed for each iteration across all five models using the scikit-learn Python package.

This study evaluated well-established ML algorithms for predicting the antioxidant activity of compounds based solely on their molecular structure represented by SMILES strings (Fig. 1). A dataset of 19,454 compounds with antioxidant activity data and corresponding SMILES information was collected from PubChem and literature searches. Following a rigorous data-cleaning process, a final set of 1,931 compounds (1,092 active, 839 inactive) was used to develop antioxidant activity prediction models. The ECFP-4 fingerprints, encoding the chemical environment of each molecule, were used as input features for the five ML algorithms (SVM, LR, XGB, RF, and DNN). A cross-validation scheme was implemented to identify the most suitable model for predicting antioxidant activity. The generalizability of the chosen model was further evaluated using external datasets of natural product compounds.

To quantify the chemical diversity within the dataset and explore potential structural relationships associated with antioxidant activities, pairwise Tanimoto similarity coefficients [12] were calculated for all compounds. A compound-compound network, where nodes represented individual compounds, with edges connecting nodes that shared a Tanimoto coefficient exceeding a threshold of 0.7, was constructed to explore structurally similar compounds in the collected compounds (Fig. 2A). This network revealed that active and inactive compounds formed separate clusters, suggesting a positive correlation between structural similarity and antioxidant activity. However, an intriguing exception was observed within a cluster enriched with flavonoids structurally related to chrysin, a potent antioxidant flavone (Fig. 2B). While all compounds within this cluster shared a flavone backbone similar to the potent antioxidant chrysin, their antioxidant capacities diverged. This highlights the importance of the specific arrangement of functional groups within the flavone structure. Chrysin, with hydroxyl groups at the 5 and 7 positions of the A ring, exhibits strong antioxidant activity due to the well-documented radical scavenging properties of these groups [14]. The addition of hydroxyl group at the 3' position of the B ring (as in apigenin) or at both the 3' and 4' positions (as in luteolin) enhances antioxidative activity. However, the addition of a hydroxyl group at the 2' position of the B ring (as in 5,7,2'-trihydroxyflavone) does not result in antioxidative effects. Substitution of hydroxyl groups with methyl groups diminishes activity; for example, tectochrysin and acacetin (methylated at the 7 position of the A ring and 3' position of the B ring, respectively) and apigenin 7,4'-dimethyl ether (methylated at the 7 position of the A ring and 4' position of the B ring) both lose their antioxidative properties. These observations highlight that the hydroxyl groups at the 5 and 7 positions on the A ring, with an unblocked 3' hydroxyl group on the B ring, are crucial for the antioxidant activity of flavones. This data suggests that while overall structural features are informative in determining antioxidant activity, it is the precise arrangement of functional groups within the molecule that ultimately dictates its efficacy.

Five ML models, including SVM, LR, XGB, RF, and DNN, were employed to learn these subtle yet important structural features for predicting antioxidant activity based on ECFP-4 fingerprints. Model performance was evaluated using two five-fold cross-validation (5-fold CV) schemes: random splitting and scaffold splitting (Fig. 3). Random splitting divided the training data into five equal folds, where each fold is used for testing once while the remaining four are used for training. On the other hand, scaffold splitting grouped structurally similar compounds together in each fold. This approach evaluates the model's ability to predict the activity of compounds with novel scaffold structures not present in the training data, a crucial capability for discovering novel antioxidants with distinct chemical backbones (known as scaffold hopping).

All models achieved commendable performance on the random splitting CV (Fig. 3A and Table 2). RF outperformed other models in all metrics, including accuracy (0.908 ± 0.004), precision (0.912 ± 0.004), recall (0.927 ± 0.005), F1 score (0.919 ± 0.003), and AUROC (0.968 ± 0.002). SVM and XGB showed competitive performance with accuracies exceeding 0.900 and AUROC above 0.955. LR performed similarly but with slightly lower recall and F1 scores. DNN, while achieving a respectable accuracy (0.877 ± 0.015), exhibited higher variability in its metrics, suggesting potential overfitting or sensitivity to the composition of training data.

Notably, all models maintained good performance on scaffold-splitting CVs, albeit their scores were slightly lower compared to random-splitting (Fig. 3B and Table 3). All models except DNN maintained high accuracies above 0.900, with SVM and XGB recording the highest (0.906 ± 0.007). Precision was notably higher for LR (over 0.918), while RF and SVM demonstrated superior recall rates (both exceeding 0.950). F1 scores remained consistent and high for SVM, LR, XGB, and RF. However, DNN's performance significantly declined, suggesting its lower robustness to scaffold-based splits. Similarly, AUROC scores for SVM, LR, XGB, and RF remained high, demonstrating their ability to distinguish classes across diverse data segmentation.

Overall, the results suggest that SVM and RF are well-suited for predicting antioxidant activity based on ECFP-4 fingerprints, exhibiting both high accuracy and generalizability across splitting methodologies.

To assess the ability of RF and SVM models to capture subtle structural features that are important for predicting antioxidant activity, we investigated their performance in differentiating between active and inactive compounds with high structural similarity. We focused on three network modules, each containing reference active compounds (chrysin, eriophorin A, and 4-(1H-indol-2-yl)aniline) (Fig. 4). Within the chrysin module (Fig. 2B), both models accurately classified chrysin and luteolin as active and apigenin as inactive, despite their high Tanimoto coefficient (> 0.7) (Fig. 4A). However, their performance diverged in other modules. Specifically, the RF model exhibited superior discriminatory power in the eriophorin A module (Fig. 4B), correctly classifying eriophorin A as active and eriophorin B and cajanin as inactive, even with their high structural similarity. Conversely, the SVM model, while differentiating active from inactive compounds, underestimated eriophorin A's activity, classifying it as inactive. A similar trend emerged in the 4-(1H-indol-2-yl)aniline module (Fig. 4C). The RF model accurately predicted active compounds, while the SVM model struggled. These findings demonstrate that, while both models can discriminate between some highly similar compounds, the RF model exhibits a stronger ability to distinguish active and inactive compounds with a high degree of structural similarity across diverse scaffold compounds. This suggests that RF models may be better suited to capture subtle structural variations that significantly influence antioxidant activity.

To evaluate the generalizability of SVM and RF models, we performed external validation using a dataset of natural product compounds from the BATMAN (Bioinformatics Analysis Tool for Molecular mechANism of Traditional Chinese Medicine) database [15]. BATMAN offers a comprehensive resource for bioactive compounds found in traditional medicine and other natural products. We retrieved chemical structure data for 1,708 well-defined ingredient compounds extracted from 8,404 medicinal plants. A subset of 1,594 compounds not included in the training set was selected for unbiased evaluation. These natural compounds were then subjected to the SVM and RF models to predict their potential antioxidant activity (Supplementary Table 1). Subsequently, the top candidates with the highest average scores obtained from both models were shortlisted for further investigation. Notably, this shortlist was significantly enriched for highly hydroxylated flavonoids (Table 4). These specific classes of compounds are recognized for exhibiting antioxidant activity through free radical scavenging, metal chelation, and involvement in redox reactions [16-23].

To explore natural compounds with novel antioxidant bioactivities, a literature search was conducted using SciFinder [24]. This search yielded publication counts for each compound, providing an indicator of their prior investigation in the context of biological activity. Most of the top 10 compounds are well-studied (median publication count = 841), suggesting their high potential for bioactivity. Among them, we focused on two particularly intriguing compounds, ellagic acid dihydrate and strictinin, which had less than 30 references each, potentially indicating a lack of previous exploration regarding their antioxidant properties (Fig. 5). Ellagic acid dihydrate is a crystalline form of ellagic acid, a well-known polyphenol found in various fruits and nuts, containing two water molecules within its structure. While ellagic acid itself exhibits potent antioxidant activity, exceeding established antioxidants like butylated hydroxytoluene and vitamin E in inhibiting lipid peroxidation [25], the specific activity of its dihydrate form remains unexplored. Given the established bioactivity of ellagic acid, its dihydrate form is a promising candidate for further investigation with a high probability of exhibiting antioxidant properties. Strictinin, a hydrolyzable ellagitannin, has also been demonstrated to possess significant antioxidant properties [26]. Multiple studies have demonstrated that strictinin possesses potent antioxidant properties that can inhibit lipid peroxidation and scavenge free radicals [27,28]. Collectively, our ML approach effectively prioritizes promising candidates for further investigation of their potential antioxidant bioactivities.