Discovery of mitochondrial modulators using target-aggregated machine learning

Graphical abstract

Highlights

• •A target-agnostic and data-integrated machine learning strategy was developed to overcome data scarcity and enable robust predictive modeling.
• •Multiple machine learning and deep learning algorithms were systematically evaluated, with XGBoost achieving a top ROC AUC of 0.943.
• •Virtual screening of a large chemical library identified six novel mitochondrial modulators, all experimentally validated in human cell assays.
• •A scalable and target-independent framework was established for phenotypic drug discovery in complex biological systems.

Mitochondria play a central role in cellular metabolism, and their dysfunction is implicated in a range of neurodegenerative, metabolic, and cardiovascular diseases [1]. Current artificial intelligence (AI) approaches in mitochondrial drug discovery are often restricted to narrow mechanistic contexts, such as mitophagy [2], and are limited by sparse and heterogeneous bioactivity data. To address these challenges, we developed a generalized, target-agnostic machine learning strategy to identify diverse mitochondrial modulators. By integrating bioactivity data across address mitochondrial proteins and applying advanced preprocessing and modeling techniques, we constructed a highly predictive classification model. This model successfully identified novel modulators from a large chemical library, underscoring the potential of computational methods in phenotypic drug discovery for complex biological systems. Our target-aggregated machine learning framework supports broad drug activity profiling across mitochondrial mechanisms, improves screening efficiency, and enriches high-confidence hits from extensive libraries while preserving structural novelty.

The foundation of our approach was the assembly of a high-quality, consolidated dataset. We extracted 15,494 compound records associated with 63 mitochondrial targets from the ChEMBL (Raw data 1 and 2 in Supplementary data). Given the complexity of mitochondrial pathways, a reliably defined true-negative dataset is unavailable. We therefore applied a stringent activity threshold to classify compounds as active (half maximal inhibitory concentration (IC₅₀) ≤ 1000 nM) or inactive (IC₅₀ > 1000 nM), ensuring high functional relevance (Figs. 1A and S1). To reduce ambiguity near this decision boundary, we implemented a proximity-aware refinement through selective removal of borderline entries: compounds with IC₅₀ values between 750 and 1500 nM were excluded, while those in the 500–750 nM and 1500–2000 nM ranges were filtered using a probabilistic Gaussian mask (Figs. S2 and S3A). As assay results could vary across targets, a compound was labeled active if any recorded IC₅₀ was at or below the threshold; it was labeled inactive only if all its IC₅₀ values were above the threshold. Structurally redundant molecules (>95% Tanimoto similarity based on 1024-bit Morgan fingerprints) were removed, though compounds with similar scaffolds but divergent activity profiles were retained to capture nuanced structure-activity relationships. This process, supplemented by known active compounds from external sources (Raw data 3 in Supplementary data), yielded a refined and diverse training set of 13,392 compounds (Fig. S3B). t-distributed stochastic neighbor embedding (t-SNE) visualization revealed that compounds associated with the same targets formed distinct clusters, confirming that the integrated dataset preserved structural diversity and extended chemical space coverage (Fig. 1B). All molecular structures were encoded as 1024-bit Morgan fingerprints using RDKit, providing a robust featurization scheme for machine learning.

Fig. 1.

Dataset characteristics, model performance, predicted compound scores, and experimental evaluation of the mitochondrial modulator screening method. (A) Distribution of the initial training data based on the logarithmic scale of the half maximal inhibitory concentration (IC₅₀) values. (B) t-distributed stochastic neighbor embedding (t-SNE) visualization of the training data categorized by targets (top 7) with different colors. (C) Receiver operating characteristic (ROC) curves and area under curve (AUC) values of five models validated on the independent test set for model training and comparison. (D) Predicted score distribution of more than 140,000 compounds from the ZINC library calculated with the trained model. (E) Oxygen consumption rate (OCR) of human aortic endothelial cells (HAECs) after candidate compounds treatment and analysis of parameters related to mitochondrial respiration (n = 4). Statistical significance was calculated by t-test, ^∗P < 0.05, ^∗∗P < 0.01, ^∗∗∗P < 0.001. SVM: Support Vector Machine; B3ODG: bayogenin 3-O-beta-D-glucopyranoside; M5M: methyl 5-methylnicotinate; Ivo: ivosidenib; 2C5MA: 2-chloro-5-methylnicotinic acid; 3M5P: 3-methyl-5-phenylpyridine; MA: momordicoside A; ATP: adenosine triphosphate.

We evaluated multiple algorithmic approaches, including Decision Tree, Random Forest, Support Vector Machine (SVM), XGBoost, and a fully connected Neural Network. The dataset was partitioned into 90% for training and 10% as a held-out test set. Model hyperparameters were optimized via 10-fold cross-validation on the training set. Performance was assessed using the area under the receiver operating characteristic curve (ROC AUC) and area under the precision-recall curve (PR AUC), along with accuracy, precision, recall, and F1-score. XGBoost consistently outperformed other models, achieving the highest ROC AUC of 0.943 on the independent test set, closely followed by the Neural Network (Figs. 1C, S4–S7, and S8A; Tables S1 and S2). In confusion matrix analysis, XGBoost demonstrated a better balance between true and false positives than other models (Fig. S8B). To rigorously assess generalization to novel chemotypes, we performed additional validation using both scaffold splits (Fig. S8C and Table S3) and uniform manifold approximation and projection (UMAP)-based clustering splits [3] (Fig. S9). Under these challenging conditions, XGBoost consistently demonstrated superior robustness compared to the Neural Network. We also evaluated a message-passing neural network (MPNN), a state-of-the-art graph-based model [4]. Although it performed competitively (best ROC AUC = 0.940) (Table S4), it did not surpass XGBoost, likely due to the limited and unevenly distributed dataset across targets (Raw data 1 in Supplementary data), which can constrain deep learning models that typically require larger, more balanced datasets. Early-recognition metrics, including the Boltzmann-enhanced discrimination of receiver operating characteristic (BEDROC), Matthews correlation coefficient (MCC), enrichment factors (EF), and Robust Initial Enhancement (RIE), were evaluated and further confirmed the superiority of XGBoost (Figs. S10 and S11). Moreover, we identified nine consensus bits by intersecting activity-enriched pharmacophore features with the model’s top 5% SHapley Additive exPlanations (SHAP)) features, indicating alignment with activity-related substructures (Fig. S12 and Tables S5–S7). We further profiled five-consensus-bit molecules (n = 297) against the full dataset (n = 13,392) across 19 functional patterns to inform future biological investigations (Fig. S13 and Table S8).

The optimized XGBoost model was deployed to screen 147,766 molecules from the ZINC in vitro library (Raw data 4 in Supplementary data). Each compound was represented by its Morgan fingerprint and received a prediction score representing its probability of being a mitochondrial modulator. The screen identified 3541 compounds (2.4%) with high prediction scores (Figs. 1D and S14, Raw data 5 in Supplementary data, and Table S9). Candidates were prioritized via a multi-step funnel: first, compounds with known mitochondrial activity, toxicophores, or high similarity to training set actives were excluded; second, molecules with structural features associated with beneficial properties, such as anti-inflammatory or antioxidant activity, were retained. This process yielded six structurally diverse candidates for experimental validation, including bayogenin 3-O-β-D-glucopyranoside, methyl 5-methylnicotinate, momordicoside A, 2-chloro-5-methylnicotinic acid, 3-methyl-5-phenylpyridine, and ivosidenib (Figs. S15–S17, Tables S10–S13). These compounds collectively captured key pharmacophoric features of the training set while maintaining structural novelty (Figs. S18 and S19, Tables S14 and S15).

In human aortic endothelial cells (HAECs), all six compounds significantly altered mitochondrial function (Figs. 1E, S20, and S21; Table S16). Five compounds consistently decreased mitochondrial membrane potential, increased mitochondrial reactive oxygen species (mtROS), and suppressed the oxygen consumption rate (OCR), with four of them also reducing adenosine triphosphate (ATP) production. In contrast, momordicoside A enhanced these metrics, consistent with pharmacophore mapping to known mitochondrial enhancers [5] (Fig. S22), confirming its role as a mitochondrial activator. Compound-membrane interaction was verified via mitochondrial membrane chromatography (Fig. S23). These results validate the computational predictions and demonstrate the model's ability to identify both novel inhibitors and activators.

This study establishes a powerful, target-agnostic machine learning framework for identifying mitochondrial modulators. By integrating multi-target bioactivity data, implementing meticulous data curation, and employing a robust modeling workflow centered on XGBoost, we overcome key challenges related to data scarcity and mechanistic complexity. Furthermore, this approach extends existing target-agnostic phenotypic screening [4] by leveraging publicly available target-ligand databases and by operating effectively under sparse phenotype annotations. Although future work is needed to identify specific targets and elucidate mechanisms of action, this study presents an efficient and scalable strategy that accelerates the discovery of compounds for complex phenotypes. The framework is readily applicable to virtual screening scenarios with similar constraints, offering a generalizable blueprint for future drug discovery efforts.

CRediT authorship contribution statement

Lizhuo Wang: Writing – review & editing, Writing – original draft, Visualization, Methodology, Investigation, Conceptualization. Peng Yang: Writing – review & editing, Visualization, Methodology, Investigation, Data curation. Sirui Wu: Writing – review & editing, Methodology. Fanding Xu: Writing – review & editing, Methodology. Wu Su: Methodology. Le Shi: Methodology, Funding acquisition. Hui Guo: Writing – review & editing, Visualization. Li Cui: Writing – review & editing, Visualization. Jiankang Liu: Writing – review & editing, Supervision. Zhiwei Yang: Writing – review & editing, Supervision, Project administration, Methodology, Conceptualization. Jiangang Long: Writing – review & editing, Supervision, Methodology, Funding acquisition, Conceptualization.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Sirui Wu is currently employed by China Mobile Communications Group Shaanxi Co., Ltd. Other authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article: