Optimizing intervertebral disc cell metabolic phenotyping with machine learning and artificial neural networks Authors: Md Entaz Bahar, Rizi Firman Maulidi, Quang Nhat Ngo, Deok Ryong Kim Published: 2025 ABSTRACT Biological phenotyping of cellular metabolism is essential for deciphering health and disease states. The Seahorse XF analyzer enables direct measurement of oxygen consumption rate (OCR) and extracellular acidification rate (ECAR), providing insight into mitochondrial function and glycolytic activity in living cells. Accurate phenotyping requires meticulous preparation, including forming a uniform cell monolayer and carefully titrating critical reagents such as the uncoupler carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone (FCCP), to ensure precise and reproducible bioenergetic measurements. Artificial neural networks (ANN) and machine learning (ML) algorithms serve as powerful tools for analyzing experimental findings, enhancing the accuracy of result predictions, and reducing the number of experimental treatments and combinations required. This study was designed to optimize seahorse bioenergetics flux analysis using ML and ANN for in vitro metabolic phenotyping attributes. The performance of the ANN and ML models was assessed using root mean square error (RMSE), mean square error (MSE), mean absolute error, and accuracy percentage. We applied supervised ML techniques in the Regression and Classification Learner App to evaluate the predictive capability of various algorithms. Our findings indicated that fine and boosted trees in the regression model, as well as fine and medium trees and linear and quadratic support vector machines (SVM) in the classification model, enhanced data prediction accuracy. Notably, we developed a rigorous biological workflow for bioenergetic phenotyping by integrating unsupervised machine learning with a supervised ANN, which closely aligned with experimental data and minimized discrepancies between predicted and observed values. This approach establishes a robust, unbiased, and data-driven framework for high-resolution metabolic phenotyping. By uniting precise experimental biology with advanced computational analysis, our study highlights a powerful synergy with significant promise for both fundamental research and therapeutic discovery.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-28236-7. Introduction Metabolism is central to cell biology, and disruptions in metabolic processes have been associated with the development and progression of diseases1. Mitochondria play a crucial role in regulating cell metabolism by generating adenosine triphosphate (ATP) through the primary energy-production pathways of mitochondrial respiration and glycolysis2. The balance between these two pathways enables cells to perform essential functions and helps prevent the onset or progression of diseases3,4. Traditionally, understanding cell metabolism, particularly mitochondrial respiration and glycolytic activity, has required extensive experimental testing and comparative analysis to gain insights into cellular energy dynamics5,6. This approach, while informative, is both labor-intensive and resource-intensive due to the complex nature of cellular mechanisms and the need for precise control of experimental conditions7. Additionally, functional validation of these findings has been limited to scientists with expertise in the various assays and data analysis techniques necessary to characterize the metabolic phenotype, which is determined by the relative use of mitochondrial respiration and glycolysis. Consequently, scientists have faced challenges in validating their functional metabolic data due to a lack of specialized skills. All cells utilize various combinations of glycolysis and mitochondrial respiration to produce ATP, meeting their biosynthetic needs. It is now recognized that the most effective way to study cell metabolism in both health and disease is by simultaneously measuring both pathways and understanding their interactions. By examining how these pathways function together, scientists can gain a deeper understanding of metabolic processes, leading to improved disease therapies and a clear understanding of cellular energy production and utilization. For instance, energy production and consumption are crucial in metabolic phenotyping, especially in complex tissues like the intervertebral disc8. Human nucleus pulposus cells (hNPCs), which form the central core of the intervertebral disc, play a vital role in maintaining disc structure and function9. Changes in the metabolism of NP cells are associated with conditions such as intervertebral disc degeneration (IVDD), highlighting the necessity for detailed metabolic studies10. The Seahorse metabolic flux analysis is a widely used technique for assessing cellular metabolism, providing important metrics like the oxygen consumption rate (OCR) and extracellular acidification rate (ECAR)11. These measurements allow researchers to quantify mitochondrial respiration and glycolytic activity, respectively. While the Seahorse XF Analyzer remains a widely used high-throughput tool for cellular metabolic profiling, emerging micro- and nano-sensing technologies are offering alternative approaches. These newer methods, particularly suited for complex biological models such as 3D cell cultures, often provide greater reliability in parameter estimation, albeit with lower throughput12,13. Moreover, optimizing Seahorse assays can be labor-intensive, as manual data adjustment and interpretation frequently require repeated experimental iterations to fine-tune critical variables such as carbonyl cyanide-p-trifluoromethoxyphenylhydrazone (FCCP) concentration, cell seeding density, and assay timing. Multivariate statistical analyses of Seahorse metabolic data, such as those from the Mito Stress Test, have revealed that biological factors (e.g., cell line) often account for variations in metabolic activity more than environmental variables (e.g., spheroid size)14. While these approaches provide valuable insights, they remain limited in scalability and in their ability to capture complex, non-linear interactions due to their reliance on classical statistics. Machine learning and artificial neural networks present promising avenues to enhance Seahorse metabolic flux analysis by simultaneously integrating multiple features. Although some studies have explored ML for clustering or interpreting Seahorse data, a comprehensive and systematic application for recurrent bioenergetic phenotyping is still missing. ANNs, in particular, hold significant potential to assist experimentalists by automating phenotype classification, improving interpretability, and enhancing reproducibility across laboratories. Our study aims to refine Seahorse bioenergetics flux analysis using ML and ANN models for the metabolic phenotyping of human nucleus pulposus cells. Our initial modeling efforts concentrated on optimizing key experimental variables, including seeding density and FCCP concentration, as a baseline application of machine learning to Seahorse data. While these analyses provide useful technical validation, their direct biological relevance is limited. To address this limitation, we developed an approach designed to classify predefined bioenergetic phenotypes. Specifically, unsupervised clustering was first employed to identify reproducible phenotypes from Seahorse parameters, followed by the use of ANN for supervised classification based on these biologically meaningful categories. This strategy enables automatic phenotype detection from metabolic readouts, improving both efficiency and accuracy. By saving time and resources, it enhances metabolic phenotyping of NP cells and supports deeper insights into cellular bioenergetics under healthy and diseased conditions. Materials and methods Human nucleus pulposus cells culture Human nucleus pulposus cells were purchased from ScienCell Research Laboratories (4800, ScienCell, Texas, USA) and cultured in Nucleus Pulposus Cell Medium (NPCM, ScienCell, Texas, USA) supplemented with 2% fetal bovine serum (FBS; ScienCell, Carlsbad, CA, USA), 1% penicillin/streptomycin (P/S; ScienCell, Carlsbad, CA, USA), and nucleus pulposus cell growth supplement (NPCGS; ScienCell, Carlsbad, CA, USA) in a poly-D-lysine-coated culture plate. The cells were kept in an incubator with 5% CO₂ at 37 °C. The medium was changed every three days until the cells reached about 70% confluency, and then every other day until the cells reached about 90% confluency. Using a mild DPBS rinse, 0.05% trypsin/EDTA dissociation, neutralization, and replating at 5,000 cells/cm², the subculture was carried out at 90–95% confluency. To reduce stress and prevent centrifugation and refreezing, cryopreserved cells were quickly thawed and handled with care. Every process was carried out in an aseptic environment while adhering to accepted biosafety guidelines. Data collection of seahorse XF96 real-time cell metabolic analyses Cells were seeded on Seahorse XF96 Cell Culture microplates (10000, 20000, 30000, and 40000 cells/well) and were subjected to metabolic profiling in sextuplicate two days after seeding. For Mito stress tests, the growth medium was replaced with Seahorse XF DMEM Base Medium (#103575-100, Agilent Technologies, pH adjusted to 7.4), supplemented with 10 mM D-glucose, 2 mM L-glutamine, and 1 mM sodium pyruvate. Cells were then cultured for 45 min in a CO2-free incubator at 37 °C. OCR and ECAR were monitored at basal conditions and, after sequential injections of 2.5 µM oligomycin (Oligo) to block the mitochondrial ATP synthase, FCCP (1.25, 2.5, and 5 µM) to uncouple oxidative phosphorylation and then 0.5 µM rotenone and antimycin A (Rot/AA) was used to inhibit mitochondrial respiration fully. The groups were designed with respective cell density and FCCP concentration (Table 1). Measurements of OCR and ECAR were performed in a 3-minute mix and 3-minute measured cycles at 37 °C on a Seahorse XFe96 Analyzer (Agilent Technologies). Wave software (Agilent Technologies) was used to analyze the datasets. OCR and ECAR were represented as pmol/min and mpH/min, respectively, and normalized to the protein concentration of each well measured by BCA assay. The Pierce BCA Protein Assay Kit (89900, Thermo Fisher Scientific) was used to measure the total protein concentration after completing the Seahorse run. The cells were lysed using RIPA lysis and extraction buffer (#89900, Thermo Fisher Scientific) supplemented with 1× Halt Protease and Phosphatase Inhibitor Cocktail (#78441, Thermo Fisher Scientific) in accordance with the manufacturer’s instructions. The Seahorse Wave software was used to enter protein data. It automatically adjusted the OCR and ECAR values to the protein content of each well and gave the results as OCR or ECAR per µg of protein, taking into consideration variations in the number of cells in each well. Table 1Group design with cell seeding density and dose of oligomycin, FCCP, and Rot/AA.GroupCells/wellOligo (µM)FCCP (µM)Rot/AA (µM)G11 × 1042.51.250.5G21 × 1042.52.50.5G31 × 1042.550.5G42 × 1042.51.250.5G52 × 1042.52.50.5G62 × 1042.550.5G73 × 1042.51.250.5G83 × 1042.52.50.5G93 × 1042.550.5G104 × 1042.51.250.5G114 × 1042.52.50.5G124 × 1042.550.5Oligo: Oligomycin; Rot/AA: Rotenone/antimycin A;.FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone. Table 1Group design with cell seeding density and dose of oligomycin, FCCP, and Rot/AA.GroupCells/wellOligo (µM)FCCP (µM)Rot/AA (µM)G11 × 1042.51.250.5G21 × 1042.52.50.5G31 × 1042.550.5G42 × 1042.51.250.5G52 × 1042.52.50.5G62 × 1042.550.5G73 × 1042.51.250.5G83 × 1042.52.50.5G93 × 1042.550.5G104 × 1042.51.250.5G114 × 1042.52.50.5G124 × 1042.550.5Oligo: Oligomycin; Rot/AA: Rotenone/antimycin A;.FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone. Group design with cell seeding density and dose of oligomycin, FCCP, and Rot/AA. Oligo: Oligomycin; Rot/AA: Rotenone/antimycin A;. FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone. Normalization of all input data The input data is normalized to adjust the data features to a specified range, typically between 0 and 1 or −1 and 1. This normalization ensures that all features contribute equally to the machine learning model, preventing features with larger ranges from dominating the learning process. We utilized the widely used normalization method, Min-Max scaling, which rescales each feature individually based on its minimum and maximum values, as described in Eq. (1)1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X = \frac{{x - {x_{min}}}}{{{x_{max}} - {x_{min}}}} \times \:\left( {u - l} \right) + l\:$$\end{document} In this context, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}$$\end{document} represents the normalized data, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}$$\end{document} denotes the original data. The symbols (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{x}}_{{m}{a}{x}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{x}}_{{m}{i}{n}}$$\end{document}) indicate the maximum and minimum values of the original vector, respectively. The variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}$$\end{document} represent the upper and lower bounds of the new range for the normalized data, with typical values being [\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}$$\end{document} = 1, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}$$\end{document} = 0]. Automated analysis by supervised machine learning toolbox Classification algorithms We evaluated the initial accuracy of various supervised learning algorithms using a standard workflow for training classification models, or classifiers, in the MATLAB “Classification Learner App” (Fig. 3a). The algorithms examined include decision trees, discriminant analysis, ensemble methods, support vector machines (SVM), K-nearest neighbors (KNN), and ensemble classifiers. To ensure an accurate estimation of performance and to prevent overfitting, we employed a 5-fold cross-validation technique. Cross-validation is a machine learning method used to evaluate a model’s overall performance and its ability to generalize. In 5-fold cross-validation, the dataset is divided into five equal parts, known as folds. The model is trained on four of these folds and tested on the fifth fold. This process is repeated five times, ensuring each fold is used as the test set exactly once. The accuracy is then calculated as the average over the five iterations, providing a reliable assessment of the model’s effectiveness. We assessed the accuracy of several techniques by analyzing an experimental dataset comprising 144 samples. The Classification Learner was used to train models for the following classifiers: decision trees, discriminant analysis, naive Bayes, support vector machine (SVM), K-nearest neighbors (KNN), and ensembles. After training the classification models, the application displays the model accuracy, reflecting the validation results. The formula for calculating the model accuracy percentage is applied as shown in Eq. (2). 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{Accuracy}}\frac{{{\text{True negative + True negative}}}}{{{\text{Total instance}}}}$$\end{document} Where total instance is a sum of True positive True + True negative + False positive + False negative. Regression algorithms As part of a thorough assessment, we evaluated the predictive accuracy of various regression algorithms for OCR and ECAR using large numerical datasets. The MATLAB “Regression Learner App” was employed to develop regression models and assess their predictive capabilities through a standard workflow for training regression models (Fig. 1Sa). The regression techniques explored included linear regression, decision tree regression, SVM regression, ensemble regression, and Gaussian process regression (GPR). To evaluate the effectiveness of these models and compare them, we utilized the model statistics available in the Current Model Summary pane. These statistics were gathered after training the regression models in the Regression Learner App. To prevent overfitting and ensure accurate performance evaluation, we implemented cross-validation. This technique involves splitting the dataset into distinct folds and calculating the accuracy for each fold, helping to assess the model’s ability to generalize. The dataset used for this evaluation consisted of 144 samples. After training the models in the Regression Learner App, we reviewed the model’s window to select the one with the highest overall score. This score was determined using the validation dataset’s optimal R-squared (R²), root mean square error (RMSE), mean squared error (MSE), and mean absolute error (MAE) metrics, as detailed in Eq. (3-6). These metrics serve as indicators of how well the trained model is likely to perform on new testing data. The main goal of this assessment was to identify the best regression strategy for accurately predicting OCR and ECAR based on the provided dataset. We utilized a combination of model statistics, cross-validation, and performance metrics to make well-informed evaluations and decisions among the tested regression models.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R^2} = 1 - \frac{{\sum {\:_{i = 1}^n} {{\left( {Xi - Yi} \right)}^2}}}{{\sum {\:_{i = 1}^n} {{\left( {Xi - \mathop X\limits^ - } \right)}^2}}}$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$RMSE = \sqrt {\frac{{\sum {\:_{i = 1}^n} {{\left( {Xi - Yi} \right)}^2}}}{n}\:\:\:}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:MAE = \frac{1}{n}\sum \: \left( {Xi - Yi} \right)^2$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:MAE = \frac{1}{n}\sum \: \left| {Xi - Yi} \right|\:$$\end{document} In the equation, , , and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}$$\end{document} denote the observed (experimental) value of the evaluation model, the associated anticipated (predicted) data, and the number of experimental data, respectively, and X̄ represents mean of the observed values. Multilayer perceptron neural Network-Based predictive modeling The ANN model was designed with a structured architecture, including an input layer, hidden layers, and an output layer. In this study, the Multilayer Perceptron (MLP) neural network employs the backpropagation algorithm to improve accuracy by minimizing the error between predicted and actual outputs (Fig. 1a). Initially, the model calculates the output using the current network weights and compares it with the actual target, generating an error term. This error, representing the difference between the network’s prediction and the true output, is propagated backward from the output layer to the hidden layers. During this backward pass, the algorithm adjusts the network weights to reduce errors. This iterative process continues as the model recalculates the output with updated weights, checks the error, and further refines the weights until the error is minimized or reaches an acceptable level. The backpropagation algorithm optimizes the neural network’s weights through repeated iterations, enhancing its predictive accuracy. The ANN structure was developed using MATLAB’s Neural Network Toolbox (nntool) to explore the relationship between input factors—cell density, FCCP dose, and time interval—and output variables—oxygen consumption rate and extracellular acidification rate (Fig. 1b). A total of 144 data samples were used to train and test the neural networks; 108 samples (75%) were selected for the training subset and 36 samples (25%) for the test subset. Various network topologies (Supplementary Table 1) were evaluated, featuring one, two, or three hidden layers with different neuron counts per layer. The hyperbolic tangent sigmoid (tansig) activation function Eq. (7) was applied in the hidden layers to introduce non-linearity into the model.Fig. 1Illustration of the Multilayer Perceptron (MLP) neural network architecture utilized for predictive modeling. (a) A typical MLP structure. (b) An overview of the proposed MLP structure, which includes the input layer (cell density, FCCP dose, and time interval), the hidden layer, and the output layer (OCR and ECAR). (c) A comparison of various structures (Network 1–24) based on mean absolute error (MAE). OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone; ANN: Artificial neural network. Illustration of the Multilayer Perceptron (MLP) neural network architecture utilized for predictive modeling. (a) A typical MLP structure. (b) An overview of the proposed MLP structure, which includes the input layer (cell density, FCCP dose, and time interval), the hidden layer, and the output layer (OCR and ECAR). (c) A comparison of various structures (Network 1–24) based on mean absolute error (MAE). OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone; ANN: Artificial neural network. 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f\left( z \right) = \frac{2}{{1 + {e^{ - 2z}}}} - 1$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}$$\end{document} represents the weighted sum of inputs to a neuron and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}\left({z}\right)$$\end{document} donates the function’s output, which is the transformed value based on the input \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}$$\end{document}. Two independent training techniques, Levenberg-Marquardt (TRAINLM) and Bayesian regularization (TRAINBR) backpropagation, were employed to optimize the network. TRAINLM and TRAINBR utilize distinct formulations for adjusting weights, often incorporating second-order derivative information or regularization terms to prevent overfitting. Both training methods were evaluated using the mean absolute error (MAE) Eq. (4) as the performance metric (Fig. 1c), enabling an effective assessment of the model’s predictive accuracy in estimating OCR and ECAR across various experimental scenarios. As illustrated in Fig. 1c, the MLP model with a Network-12 configuration (i.e., 3–12-10-2 structure, with 3 neurons in the input layer, 12 neurons in the first hidden layer, 10 neurons in the second hidden layer, and 2 neurons in the output layer) achieved the lowest MAE of 0.0503. The error for the training and testing data of the proposed ANN model was assessed using mean squared error (MSE) Eq. (5) and root mean square error (RMSE) Eq. (6) (Table 2). Table 2Average of MSE and RSME for all outputs at the proposed ANN structure.NetworkANN StructureMSERSMETrainTestTrainTestOCRECAROCRECAROCRECAROCRECAR123–12-10-20.05930.00860.4860.02340.24340.09270.69720.1529MSE: Mean square error; RSME: Root mean square error; OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate. Table 2Average of MSE and RSME for all outputs at the proposed ANN structure.NetworkANN StructureMSERSMETrainTestTrainTestOCRECAROCRECAROCRECAROCRECAR123–12-10-20.05930.00860.4860.02340.24340.09270.69720.1529MSE: Mean square error; RSME: Root mean square error; OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate. Average of MSE and RSME for all outputs at the proposed ANN structure. MSE: Mean square error; RSME: Root mean square error; OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate. Bioenergetic phenotype identification In our integrated framework, supervised ANN classification for accurate phenotype prediction was combined with unsupervised clustering for phenotype discovery. Seahorse metabolic parameters—including basal respiration (BR), maximal respiration (MT), ATP production (ATP), baseline OCR (bOCR), baseline ECAR (bECAR), stressed OCR (sOCR), and stressed ECAR (sECAR)—were normalized using Z-score transformation Eq. (8) to ensure equal contribution of each parameter to clustering analysis.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Xnorm = \frac{{x - \mu \:}}{{\sigma \:}}$$\end{document} Where X represents the raw parameter value, µ is the mean, and σ is the standard deviation across all samples. To determine the optimal number of clusters (K), we employed three complementary methods: the elbow method to minimize within-cluster sum of squares (WSS), silhouette analysis to evaluate average silhouette values, and the adjusted Rand index (ARI) to assess clustering consistency across four algorithms (K-means, Spectral Clustering, GMM, and DBSCAN). Statistical validation was performed using one-way ANOVA with cluster assignment (phenotype) as the independent variable, followed by Tukey’s HSD post-hoc test for pairwise comparisons. Significance was defined as p < 0.05. The three resulting phenotypes were further confirmed through multidimensional scatter plots with convex hulls and correlation pattern analysis. Based on parameter amplitude and their coordinated patterns, clusters were named Energetic (high values, strong associations), Quiescent (low values, atypical patterns), and Intermediate (moderate values). The ANN model was implemented in MATLAB R2019b using the Neural Fitting Toolbox (nftool). The fully connected feedforward architecture (patternnet) consisted of an input layer of seven neurons (normalized metabolic parameters), one hidden layer with ten neurons using hyperbolic tangent activation functions, and an output layer of three neurons with a softmax activation function for probabilistic multi-class classification. Stratified random sampling was applied to split the dataset into training (70%), validation (15%), and testing (15%) subsets, preserving the original phenotype distribution. Training was carried out with the scaled conjugate gradient backpropagation algorithm (trainscg) to minimize cross-entropy loss. To reduce overfitting, L2 weight regularization was applied with a coefficient (λ) of 0.001. Result and discussion Seahorse XF96 cell density and FCCP dose Titration We investigated the metabolic characteristics of cultured human nucleus pulposus cells (hNPCs) using real-time Seahorse technology to simultaneously measure the oxygen consumption rate, indicative of mitochondrial respiration or oxidative phosphorylation, and the extracellular acidification rate, indicative of glycolysis15. The Seahorse XF96 Cell Mito Stress Test (MST) assessed OCR before and after the introduction of a series of electron transport chain (ETC) inhibitors, including oligomycin (which inhibits the mitochondrial ATP synthase complex, Complex V), FCCP (a potent uncoupler of oxidative phosphorylation), and rotenone/antimycin A (which inhibit Complex I and III, respectively) (Fig. 2a, b). To determine the optimal FCCP dose and hNPCs seeding density that produced ideal OCR and ECAR readings, Seahorse XF96 cell density and FCCP dose titration were conducted. Increasing cell density led to linear increases in both OCR and ECAR (Fig. 2c, d). When hNPCs were seeded at 10,000 or 20,000 cells per well with varying FCCP concentrations (1.25, 2.5, and 5 µM), OCR and ECAR exhibited the lowest values with fluctuating spectra, possibly due to cell detachment before or during the analysis. However, OCR and ECAR spectra increased with cell numbers between 30,000 and 40,000 cells per well. This suggested that a seeding density of 30,000 to 40,000 cells per well was optimal for obtaining OCR and ECAR within the linear response range. Following the initial Seahorse XF96 mitochondrial stress analysis, cells seeded at 30,000 or 40,000 cells per well with FCCP 2.5 or 5 µM showed minor to moderate detachment at well edges, whereas those with FCCP 1.25 µM did not. Therefore, a seeding density of 30,000 or 40,000 cells per well and 1.25 µM FCCP were confirmed as the optimal conditions for Seahorse XF96 analysis.Fig. 2Optimization of Seahorse XF96 for hNPCs seeding density and FCCP concentration. (a) Illustration of oxygen consumption rate (OCR) measurement before and after the injection of three compounds—oligomycin (Oligo), carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone (FCCP), and rotenone/antimycin A (Rot/AA) inhibitors—to derive various parameters of mitochondrial respiration. (b) Targets of each compound within the electron transport chain. (c) Real-time OCR reflecting the mitochondrial respiration rate of cultured hNPCs at different cell densities and FCCP concentrations. (d) Real-time extracellular acidification rate (ECAR) showing the glycolysis rate of cultured hNPCs at different cell densities and FCCP concentrations. (e) Illustration of a typical XF cell energy phenotype profile used to measure the metabolic potential of live cells. (f) Metabolic profiling of hNPCs across different cell densities and FCCP concentrations. (g) Metabolic potential of cultured hNPCs, expressed as a percentage of the mean basal OCR/ECAR, at varying cell densities and FCCP concentrations. hNPCs: Human nucleus pulposus cells; ADP: A adenosine diphosphate; ATP: Adenosine triphosphate Optimization of Seahorse XF96 for hNPCs seeding density and FCCP concentration. (a) Illustration of oxygen consumption rate (OCR) measurement before and after the injection of three compounds—oligomycin (Oligo), carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone (FCCP), and rotenone/antimycin A (Rot/AA) inhibitors—to derive various parameters of mitochondrial respiration. (b) Targets of each compound within the electron transport chain. (c) Real-time OCR reflecting the mitochondrial respiration rate of cultured hNPCs at different cell densities and FCCP concentrations. (d) Real-time extracellular acidification rate (ECAR) showing the glycolysis rate of cultured hNPCs at different cell densities and FCCP concentrations. (e) Illustration of a typical XF cell energy phenotype profile used to measure the metabolic potential of live cells. (f) Metabolic profiling of hNPCs across different cell densities and FCCP concentrations. (g) Metabolic potential of cultured hNPCs, expressed as a percentage of the mean basal OCR/ECAR, at varying cell densities and FCCP concentrations. hNPCs: Human nucleus pulposus cells; ADP: A adenosine diphosphate; ATP: Adenosine triphosphate Generation of a metabolic phenotype to validate OCR and EACR Metabolic phenotypes indicate whether cells predominantly depend on glycolysis or OXPHOS for energy production and how this preference changes across different environments. Measurements of OCR and ECAR shed light on mitochondrial respiration and glycolysis, respectively, and can confirm these metabolic pathways (Fig. 2e). To determine the energy phenotype of hNPCs, we assessed mitochondrial respiration and glycolysis under both basal and stress conditions. A dedicated report generator automatically calculates and displays the energy phenotype and metabolic potential of our different groups, thereby simplifying data analysis and interpretation (Fig. 2f, g). Classification modeling The classification model outcomes are represented through diagnostic measures, such as model accuracy, and visualizations like scatter plots or confusion matrix charts. We employed various machine learning methods for classification and evaluated their accuracy in categorizing the dataset (Table 1). Among the methods tested, Fine Tree, Medium Tree, and Linear SVM achieved the highest accuracy rate of 98.1%, indicating their effectiveness in accurately categorizing data points within the dataset (Table 1). In the Fine Tree algorithm, two samples from Group 3 were mistakenly labeled as Group 1, representing a specific error type (Fig. 3b). The model misclassified 2 out of 9 samples that should have been categorized as Group 3, resulting in a 22% misclassification rate for that group (Fig. 3c). An AUC of 0.99 suggests that our model is highly effective at distinguishing between the positive and negative classes, with nearly perfect classification performance (Fig. 3d). A parallel coordinate plot is a common visualization method for displaying high-dimensional data and identifying patterns or outliers across multiple datasets simultaneously. In the plot, the two misclassified samples are shown as distinct lines, colored differently to highlight their misclassification (Fig. 3e). Furthermore, the Quadratic SVM achieved a slightly lower accuracy rate of 97.2%, yet it still demonstrated a significant level of effectiveness in classification tasks. The Bagged Tree model, one of the top-performing algorithms, reached an accuracy of 92.6%, indicating excellent performance, albeit slightly below the top three methods. Similarly, the Fine K-Nearest Neighbors (KNN) model achieved an accuracy of 83.3%, reflecting a reasonable but somewhat lower level of accuracy for this dataset. Additionally, the analysis revealed that nearly all the linear discriminant and Naïve Bayes algorithms failed to classify the data correctly, suggesting their ineffectiveness for this dataset’s data points. Based on the trial results, we selected the complex decision Fine Tree method due to its superior accuracy in classifying the dataset. The findings underscore the importance of algorithm selection and cross-validation in machine learning experiments to ensure reliable and robust performance evaluation.Fig. 3Model accuracy and visualizations for the Fine Tree classification model. (a) The flowchart illustrates a typical workflow for training a classification model or classifier using the Classification Learner App. (b, c) The confusion matrix displays the misclassified samples and classification rate of the Fine Tree classification model. (d) A receiver operating characteristic (ROC) curve depicts the performance of a multiclass classifier model at various threshold settings within the Fine Tree classification model. (e) A parallel coordinate plot represents data points as lines intersecting parallel axes in the Fine Tree classification model. OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone; TPR: True positive rate; FNR: False positive rate. Model accuracy and visualizations for the Fine Tree classification model. (a) The flowchart illustrates a typical workflow for training a classification model or classifier using the Classification Learner App. (b, c) The confusion matrix displays the misclassified samples and classification rate of the Fine Tree classification model. (d) A receiver operating characteristic (ROC) curve depicts the performance of a multiclass classifier model at various threshold settings within the Fine Tree classification model. (e) A parallel coordinate plot represents data points as lines intersecting parallel axes in the Fine Tree classification model. OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone; TPR: True positive rate; FNR: False positive rate. Regression modeling The results of the regression models are illustrated through diagnostic metrics such as model accuracy, and visualizations like response plots and residual plots, which showcase the performance of the validated models. The Regression Learner App uses several regression algorithms to model a target variable based on input variables. The models were evaluated using key performance indicators in the app, including R², RMSE, MSE, and MAE (Table S2). R² indicates the proportion of variance in the target variable explained by the model, with values close to 1 representing superior performance. RMSE measures the average prediction error, with lower values indicating higher accuracy. MSE represents the average squared difference between observed and predicted values, where lower values suggest better model performance. Similarly, MAE reflects the mean absolute difference between actual and predicted values, with lower values indicating superior performance, similar to RMSE and MSE. Two specific regression methods, the fine tree and boosted ensemble regression models, excelled in predicting OCR and ECAR during the evaluation (Table S2). The fine tree regression model outperformed others with the lowest error metrics for predicting OCR (Table S2). It demonstrated exceptional performance for OCR, with an R² value of 0.92, a minimal RMSE of 0.41197, an MSE of 0.16972, and an MAE of 0.31523 (Figure S1b, c). These metrics indicate that the fine tree regression model provided excellent predictions of OCR values, showing high precision and a strong fit to the data. However, for predicting ECAR, the boosted ensemble regression model performed better than other methods, achieving the lowest error metrics (Table S2). This model exhibited outstanding performance for ECAR, with an R² value of 0.90, a minimal RMSE of 0.21679, an MSE of 0.046998, and an MAE of 0.16953 (Figure S1d, e). These measures suggest that the boosted ensemble regression model offered precise ECAR forecasts, demonstrating robust model performance and a good fit with the data. Overall, these results reveal that both models possess significant predictive abilities for their respective tasks, with minimal error margins across the evaluated metrics. ANN modeling approach Table 3; Fig. 4a present the configuration of a Multilayer Perceptron (MLP) neural network. This network processes 3 input features—cell density, FCCP dose, and time interval—and outputs the variables OCR and ECAR, with 3 neurons in the input layer. It includes two hidden layers, with 12 neurons in the first layer and 10 neurons in the second, enabling the capture of complex patterns within the data. The output layer comprises 2 neurons, making it suitable for binary classification or multi-output regression tasks related to OCR and ECAR predictions. The learning rate is set at 0.5, facilitating faster learning but with a potential risk of overshooting. The network is trained over 1000 epochs using the Bayesian Regularization algorithm (trainbr), which helps to minimize errors and prevent overfitting, thereby enhancing generalization to new data.Fig. 4The proposed model and its performance. (a) The proposed ANN model features three inputs (cell density, FCCP dose, and time interval), two hidden layers (with 12 neurons in the first and 10 in the second), and two output layers (OCR and ECAR). (b) The regression curve generated in nntool illustrates the performance of the ANN model.OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone; ANN: Artificial neural network;Table 3Configuration of the proposed MPL neural network model.NetworkMPLNumber of hidden layers2Number of neurons in the input layer3Number of neurons in the first hidden input layer12Number of neurons in the second hidden input layer10Number of neurons in the output layer2Learning rate0.5Number of epochs1000Adaptation learning functiontrainBRActivation functiontansigMPL: Multilayer perceptron; trainbr: Bayesian regularization backpropagation;.tansig: Hyperbolic tangent sigmoid transfer function. The proposed model and its performance. (a) The proposed ANN model features three inputs (cell density, FCCP dose, and time interval), two hidden layers (with 12 neurons in the first and 10 in the second), and two output layers (OCR and ECAR). (b) The regression curve generated in nntool illustrates the performance of the ANN model.OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone; ANN: Artificial neural network; Table 3Configuration of the proposed MPL neural network model.NetworkMPLNumber of hidden layers2Number of neurons in the input layer3Number of neurons in the first hidden input layer12Number of neurons in the second hidden input layer10Number of neurons in the output layer2Learning rate0.5Number of epochs1000Adaptation learning functiontrainBRActivation functiontansigMPL: Multilayer perceptron; trainbr: Bayesian regularization backpropagation;.tansig: Hyperbolic tangent sigmoid transfer function. Configuration of the proposed MPL neural network model. MPL: Multilayer perceptron; trainbr: Bayesian regularization backpropagation;. tansig: Hyperbolic tangent sigmoid transfer function. The model’s performance was assessed using the regression curve generated in nntool (Fig. 4b), yielding R-values of 0.99286 for the training set, 0.9955 for the test set, and 0.99329 overall. These R-values evaluate the neural network’s ability to predict continuous output values, with values approaching 1 indicating strong predictive accuracy. An R-value of 0.99286 for the training set suggests a strong positive correlation between predicted outputs and actual target values, demonstrating that the model has effectively learned the relationship between the input and output data with minimal errors. The proximity of data points to the regression line indicates that the model has accurately captured the patterns in the training data. An R-value of 0.9955 for the test set highlights excellent generalization, as the model makes highly accurate predictions on new data, confirming the absence of overfitting. The close clustering of data points around the regression line for the test set further underscores the model’s reliable predictive capabilities. The overall R-value of 0.99329 reflects consistently strong performance across training, validation, and testing datasets, with few prediction errors. These results demonstrate that the network has successfully internalized the fundamental patterns in the data, generalizes well to new data, and can be trusted for accurate predictions across various datasets. Figure 5 illustrates the comparison between experimental and predicted outcomes using the ANN for OCR and ECAR across all hNPCs data. The regression equations (y = 0.98 * x + 0.10) for the training set and (y = 0.90 * x + 0.47) for the test set show that the ANN’s predictions are closely aligned with the experimental data, with a slight underestimation. Similarly, for ECAR, the equations (y = 1 * x − 0.00) in training and (y = 1.01 * x − 0.01) in testing suggest that the ANN’s predictions are nearly perfect, with a minimal overestimation. The near-perfect correlation indicates that the ANN has successfully identified the fundamental patterns in the data, producing outputs that closely match the experimental results. The small deviations from a slope of 1 and an intercept of 0 suggest that the model is well-calibrated, as expected in practical applications.Fig. 5Comparison between experimental and ANN-predicted OCR and ECAR in hNPCs. (a-b) The regression curve illustrates the prediction accuracy for the training dataset. (c-d) The regression curve illustrates the prediction accuracy for the testing dataset. OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone; ANN: Artificial neural network. Comparison between experimental and ANN-predicted OCR and ECAR in hNPCs. (a-b) The regression curve illustrates the prediction accuracy for the training dataset. (c-d) The regression curve illustrates the prediction accuracy for the testing dataset. OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone; ANN: Artificial neural network. A surface-stem plot combines a three-dimensional (3D) surface plot with a stem plot, providing an effective way to illustrate the correlation among data points in 3D space. Figure 6a-d shows the comparison between actual OCR values and those predicted by the ANN model at various FCCP concentrations and time intervals across different cell densities (10k, 20k, 30k, and 40k). The 3D graph features a fitted ANN surface (mesh) and highlights the individual data points used for model training (red) and testing (blue). Green markers represent the OCR values predicted by the ANN. The strong correlation between predicted and actual values demonstrates.Fig. 6The three-dimensional (3D) plot compares actual and predicted OCR values as a function of FCCP concentration and time points across different cell densities. (a-d) The ANN model’s effectiveness in predicting OCR is demonstrated using training and testing datasets at cell densities of 10 K, 20 K, 30 K, and 40 K. Green circles (ANN predictions) represent the values forecasted by the trained ANN model. Red circles (Training data) depict the actual OCR values measured during model training. Blue circles (Testing data) show real OCR results used to assess the model on previously unobserved data points. The surface mesh (ANN Network) is a smooth, color-coded surface (ranging from blue to yellow) that illustrates the ANN-predicted OCR trends in relation to the input variables (FCCP and time points), with the mesh interpolated using ANN predictions. OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone; ANN: Artificial neural network; 10K: 10000 cells; 20K: 20000 cells; 30K: 30000 cells; 40K: 40000 cells. The three-dimensional (3D) plot compares actual and predicted OCR values as a function of FCCP concentration and time points across different cell densities. (a-d) The ANN model’s effectiveness in predicting OCR is demonstrated using training and testing datasets at cell densities of 10 K, 20 K, 30 K, and 40 K. Green circles (ANN predictions) represent the values forecasted by the trained ANN model. Red circles (Training data) depict the actual OCR values measured during model training. Blue circles (Testing data) show real OCR results used to assess the model on previously unobserved data points. The surface mesh (ANN Network) is a smooth, color-coded surface (ranging from blue to yellow) that illustrates the ANN-predicted OCR trends in relation to the input variables (FCCP and time points), with the mesh interpolated using ANN predictions. OCR: Oxygen consumption rate; ECAR: Extracellular acidification rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone; ANN: Artificial neural network; 10K: 10000 cells; 20K: 20000 cells; 30K: 30000 cells; 40K: 40000 cells. The ANN’s accuracy in estimating OCR dynamics in relation to varying FCCP levels over time. This reflects the model’s capability to predict metabolic responses under different experimental conditions. The ANN model effectively captures the patterns in OCR at various time points and FCCP concentrations, with minor discrepancies between actual data points and the predicted surface, particularly between cell densities of 30 K to 40 K (Fig. 6c, d). The alignment of training data (red) and testing data (blue) with the projected mesh indicates that the ANN model generalizes well across the range of input parameters. Figure 7 presents a comparative analysis of the effects of cell density and FCCP dosage on the experimental outcomes. The bars within the red-dashed box represent specific cell densities, ranging from 30 K to 40 K, indicating the optimal seeding density for the experiment. The purple-dashed box outlines various FCCP concentrations, highlighting 1.25 µM as the optimal value. The results suggest that a combination of 1.25 µM FCCP and a cell density of 40,000 provides the most favorable outcomes. This combination is suitable for further experimental validation, as it offers an ideal balance between cell seeding and FCCP dosage for consistent and reproducible results. Moreover, the findings offer valuable insights into the relationship between cell density and FCCP concentration, facilitating the adjustment of experimental parameters for future metabolic and physiological studies. Fig. 7Comparative analysis of the optimal cell density and FCCP dosage on experimental results using the proposed artificial neural network (ANN) model. Values are expressed as mean ± SD. Different letters denote statistically significant differences. The groups (G1-12) were designed with respective cell density and FCCP concentration. OCR: Oxygen consumption rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone. Comparative analysis of the optimal cell density and FCCP dosage on experimental results using the proposed artificial neural network (ANN) model. Values are expressed as mean ± SD. Different letters denote statistically significant differences. The groups (G1-12) were designed with respective cell density and FCCP concentration. OCR: Oxygen consumption rate; FCCP: Carbonyl cyanide 4-(trifluoromethoxy)phenylhydrazone. Identification of bioenergetic phenotypes The optimal clustering analysis identified three distinct metabolic phenotypes as the most reliable partitioning of the data. The elbow method revealed a pronounced reduction in total WSS from K = 1 to K = 3, with diminishing returns at higher K values (Fig. 8a). Consistently, the average silhouette width peaked at K = 3 (0.36), supporting the selection of three cohesive and well-separated clusters (Fig. 8b). The adjusted Rand index (ARI), which assessed agreement across clustering algorithms, showed excellent concordance between K-means and Spectral clustering (ARI = 0.862), moderate agreement with Gaussian Mixture Models (ARI = 0.743), and weak agreement with DBSCAN (ARI = 0.312) (Fig. 8c). Fig. 8Identification and validation of three distinct cellular metabolic phenotypes. (a) Elbow plot of total within-cluster sum of squares (WSS) for k-values 1–6, indicating k = 3 as the optimal cluster number. (b) Silhouette analysis for k = 2–6, with k = 3 showing the highest average silhouette score, reflecting strong cluster cohesion and separation. (c) Adjusted Rand Index (ARI) comparing clustering consistency across algorithms (K-means, Spectral Clustering, Gaussian Mixture Model (GMM), and DBSCAN). K-means and Spectral clustering at k = 3 achieved the highest agreement (ARI = 0.862), confirming a robust solution. (d) Pie chart illustrating the distribution of cells across the three identified phenotypes. (e) Heatmap of z-score–normalized Seahorse metabolic parameters, with rows representing individual cells and columns representing the seven measured variables. Phenotype assignments (Energetic: red, Quiescent: blue, Intermediate: green) highlight distinct metabolic patterns. (f) Pairwise scatter plot matrix of all seven parameters, with cells colored by phenotype and convex hulls encircling each group. Upper panels display correlation coefficients, while diagonal panels show parameter density distributions. Together, these analyses validate that the three clusters represent distinct and separable metabolic states. BR: Basal respiration; MR: Maximum respiration; ATP: Adenosine triphosphate production; bOCR: Basal oxygen consumption rate; bECAR: Basal extracellular acidification rate; sOCR: Stressed oxygen consumption rate; sECAR: Stressed extracellular acidification rate; Identification and validation of three distinct cellular metabolic phenotypes. (a) Elbow plot of total within-cluster sum of squares (WSS) for k-values 1–6, indicating k = 3 as the optimal cluster number. (b) Silhouette analysis for k = 2–6, with k = 3 showing the highest average silhouette score, reflecting strong cluster cohesion and separation. (c) Adjusted Rand Index (ARI) comparing clustering consistency across algorithms (K-means, Spectral Clustering, Gaussian Mixture Model (GMM), and DBSCAN). K-means and Spectral clustering at k = 3 achieved the highest agreement (ARI = 0.862), confirming a robust solution. (d) Pie chart illustrating the distribution of cells across the three identified phenotypes. (e) Heatmap of z-score–normalized Seahorse metabolic parameters, with rows representing individual cells and columns representing the seven measured variables. Phenotype assignments (Energetic: red, Quiescent: blue, Intermediate: green) highlight distinct metabolic patterns. (f) Pairwise scatter plot matrix of all seven parameters, with cells colored by phenotype and convex hulls encircling each group. Upper panels display correlation coefficients, while diagonal panels show parameter density distributions. Together, these analyses validate that the three clusters represent distinct and separable metabolic states. BR: Basal respiration; MR: Maximum respiration; ATP: Adenosine triphosphate production; bOCR: Basal oxygen consumption rate; bECAR: Basal extracellular acidification rate; sOCR: Stressed oxygen consumption rate; sECAR: Stressed extracellular acidification rate; The three identified phenotypes—quiescent, intermediate, and energetic—were statistically validated by one-way ANOVA, which confirmed highly significant differences across all metabolic parameters (p < 0.001) (Table S3). Post-hoc Tukey’s HSD further supported pairwise distinctions among the groups. Parameters such as ATP production (F = 44.02) and maximal respiration (F = 115.34) showed particularly strong differences, indicating that between-phenotype variation far exceeded within-group variability. The clusters were evenly distributed, comprising 26.7% energetic, 40% intermediate, and 33.3% quiescent phenotypes (Fig. 8d). Each phenotype displayed a unique metabolic profile: energetic cells exhibited elevated activity across energy-producing pathways, quiescent cells showed broadly reduced metabolic function, and intermediate cells occupied a moderate state with mixed characteristics. Heatmap visualization reinforced these distinctions, highlighting high maximal respiration and glycolytic capacity in energetic phenotypes, markedly reduced activity in quiescent cells, and intermediate patterns with partial activity across pathways (Fig. 8e). Correlation analyses revealed distinct regulatory regimes within each phenotype (Fig. 8f). Energetic cells demonstrated strong positive coupling between maximal respiration and glycolysis (stressed ECAR, r = 0.63) as well as between ATP production and baseline OCR (r = 0.85), reflecting synergistic energy production. In contrast, quiescent phenotypes displayed negative correlations, such as between ATP production and baseline ECAR (r=−0.60) and between ATP production and basal respiration (r=−0.38), suggesting inefficient or dysregulated metabolism. The intermediate phenotype exhibited more complex and often inverse correlations, for example between basal and maximal respiration (r=−0.44), supporting its role as a transitional metabolic state. These internal correlation structures provide strong evidence that the phenotypes represent biologically distinct and functionally coherent cellular states. The ANN achieved exceptional classification accuracy, perfectly distinguishing the three phenotypes—energetic, intermediate, and quiescent (Fig. 9a). The model architecture included seven input neurons (metabolic parameters), a hidden layer of ten neurons, and a three-neuron output layer corresponding to the phenotypes. All datasets were classified with 100% accuracy, with precise identification of 12 energetic, 18 intermediate, and 15 quiescent samples. Training (n = 31) preserved a balanced phenotype distribution (25.8% energetic, 41.9% intermediate, 32.2% quiescent), while the test (n = 7) and validation (n = 4) sets confirmed generalizability with distinct phenotype representation. Fig. 9An Artificial Neural Network validates the distinctness of metabolic phenotypes. (a) Confusion matrices showing classification performance of the ANN model for the full dataset as well as training, test, and validation subsets. The matrix presents the frequency of actual class labels (rows) versus predicted class labels (columns). The numerical class labels correspond to the following categories: 1 = Energetic, 2 = Intermediate, and 3 = Quiescent. The values along the main diagonal represent correct classifications, while off-diagonal values indicate errors. The model achieved 100% accuracy, perfectly classifying all cells. The adjacent table details the number and percentage of each phenotype within each subset. (b) Heatmap of weights connecting the seven input metabolic parameters to the ten hidden-layer neurons, illustrating the complex non-linear feature combinations learned for phenotype discrimination. (c) PCA plot of the dataset colored by ANN-assigned phenotype labels, demonstrating clear separation among energetic, intermediate, and quiescent groups. (d) Schematic of the ANN architecture, including the seven-neuron input layer, a ten-neuron hidden layer, and a three-neuron output layer representing the three phenotypes (energetic, intermediate, quiescent). BR: Basal respiration; MR: Maximum respiration; ATP: Adenosine triphosphate production; bOCR: Basal oxygen consumption rate; bECAR: Basal extracellular acidification rate; sOCR: Stressed oxygen consumption rate; sECAR: Stressed extracellular acidification rate; In confusion matrix, An Artificial Neural Network validates the distinctness of metabolic phenotypes. (a) Confusion matrices showing classification performance of the ANN model for the full dataset as well as training, test, and validation subsets. The matrix presents the frequency of actual class labels (rows) versus predicted class labels (columns). The numerical class labels correspond to the following categories: 1 = Energetic, 2 = Intermediate, and 3 = Quiescent. The values along the main diagonal represent correct classifications, while off-diagonal values indicate errors. The model achieved 100% accuracy, perfectly classifying all cells. The adjacent table details the number and percentage of each phenotype within each subset. (b) Heatmap of weights connecting the seven input metabolic parameters to the ten hidden-layer neurons, illustrating the complex non-linear feature combinations learned for phenotype discrimination. (c) PCA plot of the dataset colored by ANN-assigned phenotype labels, demonstrating clear separation among energetic, intermediate, and quiescent groups. (d) Schematic of the ANN architecture, including the seven-neuron input layer, a ten-neuron hidden layer, and a three-neuron output layer representing the three phenotypes (energetic, intermediate, quiescent). BR: Basal respiration; MR: Maximum respiration; ATP: Adenosine triphosphate production; bOCR: Basal oxygen consumption rate; bECAR: Basal extracellular acidification rate; sOCR: Stressed oxygen consumption rate; sECAR: Stressed extracellular acidification rate; In confusion matrix, Analysis of hidden-layer weights revealed how the ANN learned multivariate signatures unique to each phenotype (Fig. 9b). For example, neuron H1 assigned strong negative weights to ATP production (−1.04) and stressed OCR (−1.35) but high positive weights to basal respiration (+ 0.92) and stressed ECAR (+ 0.93). Conversely, neuron H3 strongly emphasized ATP production (+ 0.94), stressed OCR (+ 1.45), and maximal respiration (+ 1.75). These patterns indicate that the network extracted complex, non-linear relationships to discriminate phenotypes. Principal component analysis (PCA) of ANN outputs further confirmed the clear separation of the three groups (Fig. 9c). The final ANN structure is depicted in Fig. 9d, consisting of a seven-neuron input layer, a ten-neuron hidden layer, and a three-neuron softmax output layer. Together, the perfect classification performance and distinct weight patterns provide compelling computational evidence that energetic, intermediate, and quiescent phenotypes represent genuine and discrete metabolic states. Conclusion and future directions In this paper, an ANN was used to model and predict OCR and ECAR values measured by the Seahorse Mito Stress Test on hNPCs, aiming to optimize cell density and FCCP concentration. The proposed ANN model is closely aligned with experimental data, exhibiting minimal errors. The results demonstrate that ANN can effectively model and predict OCR values obtained from Seahorse assays, marking a significant advancement in optimizing the Seahorse metabolic flux assay. A major contribution of this study is using ANN and machine learning algorithms to enhance Seahorse metabolic flux assessments. Our model functions as an in-silico experimental design tool, enabling researchers to optimize protocols by predicting metabolic outcomes prior to wet-lab validation, thereby greatly enhancing screening efficiency. The first ANN model successfully established predictive relationships between experimental parameters and metabolic outputs (OCR/ECAR), yet its architecture imposed inherent limitations on biological interpretability and broader applicability. To address this, we developed a new ANN framework designed to classify metabolic profiles into fundamental phenotypic states. Using K-means clustering on the entire dataset, we first identified natural, data-driven phenotypes independent of treatment conditions, which then served as biologically meaningful labels for supervised classification. This approach streamlines metabolic phenotyping by reducing experimental workload, minimizing variability, and improving reproducibility. Looking ahead, several promising directions remain. Expanding the training dataset to include diverse cell types and disease-specific models will be critical for developing a universal platform for metabolic phenotyping. Integrating multi-omics layers, such as transcriptomics and proteomics, may uncover mechanistic links between genetic markers and metabolic states. A longer-term goal is adapting this paradigm for clinical applications—for example, applying patient-derived cells in in-silico drug screening to predict therapeutic efficacy and potential side effects. Finally, to democratize AI-powered experimental design for the broader biological community, future development should focus on transforming this framework into an accessible and intuitive software interface. Accurate forecasting of Seahorse assay results using ANN holds substantial potential for advancing bioenergetics research, particularly regarding human nucleus pulposus cells, where mitochondrial function is crucial for understanding cellular health and disease progression. Integrating artificial neural networks and machine learning into Seahorse test improvements offers an efficient, reliable, and cost-effective approach for advancing metabolic flux analysis. This methodology is applicable across various cell types and experimental settings, providing researchers with a powerful tool for advancing metabolic research. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1