Ophiobolin A Induces an Apoptotic Transcriptional Signature and Modulates Redox Homeostasis in T98G and U118MG Glioblastoma Cells: A Machine Learning Approach Authors: Paweł Woźnicki, Dorota Hudy, Oliwia Trzaskoś, Paul Avijit, Marvin Xavierselvan, Jacek Tabarkiewicz, Joanna Katarzyna Strzelczyk, David Aebisher Published: 2026 ABSTRACT Gliomas are the most common group of primary brain tumors, among which glioblastoma multiforme (GBM) is characterized by a particularly poor prognosis and the limited effectiveness of available treatments. Ophiobolin A (OP-A), a natural sesterterpenoid, exhibits promising anticancer properties, including the ability to cross the blood-brain barrier and induce paraptosis-like cell death. However, the molecular mechanisms underlying its action, especially in the early phase of the cellular response, remain not fully understood. The aim of this study was to analyze early changes in the expression of genes associated with apoptosis, ferroptosis, and antioxidant mechanisms in T98G and U118MG glioma cells exposed to OP-A. Gene expression was assessed by RT-qPCR, apoptosis was evaluated using Annexin V/PI staining and flow cytometry, and treatment-induced morphological changes were documented by brightfield microscopy. Statistical analysis was performed using the Mann–Whitney U test. Descriptive Annexin V/PI analysis showed a lower proportion of viable cells and a higher proportion of early apoptotic cells in the analyzed OP-A-treated T98G and U118MG samples compared with the corresponding vehicle-control samples. These preliminary observations were based on technical replicates from a single biological experiment and require confirmation in independent biological replicates. Transcriptional profiling revealed a shift toward a pro-apoptotic phenotype, characterized by increased BAX and FAS expression together with a trend toward reduced BCL2 expression, whereas ferroptosis-associated genes remained largely unchanged. Notably, SLC7A11 upregulation suggested activation of compensatory antioxidant mechanisms in response to OP-A treatment. In T98G cells, OP-A induced a distinct and reproducible transcriptional signature that enabled accurate discrimination from control conditions (AUC = 0.833). Feature importance and SHAP analyses identified BAX as the most informative predictor, followed by SLC7A11 and FAS, with bootstrap validation confirming BAX as a stable marker. Pathway analysis demonstrated selective activation of apoptosis- and cysteine metabolism-related pathways, while hierarchical clustering revealed that OP-A generated a transcriptional profile distinct from oxidative stress-inducing agents. The predictive performance of this molecular signature was cell-line dependent, showing weaker discrimination in U118MG cells.Short-term OP-A exposure in T98G and U118MG cells was associated with exploratory trends in apoptosis- and redox-related gene expression and a higher proportion of Annexin V-positive cells. The machine-learning analyses identified candidate discriminatory features within this limited dataset but should be regarded as hypothesis-generating. Larger studies with independent biological replication, additional GBM models, different exposure conditions, and functional validation are required to confirm these observations and clarify the mechanism of OP-A action. 1. Introduction Gliomas are the most common group of primary brain tumors, accounting for approximately 81% of all malignant tumors of the central nervous system [1]. The World Health Organization classification distinguishes four grades of malignancy—Grade I and II gliomas are classified as low-grade tumors, while Grade III and IV gliomas are classified as high-grade tumors [2]. A higher grade of malignancy typically correlates with a poorer clinical prognosis [3]. The most diagnosed tumor among grade IV gliomas is glioblastoma multiforme (GBM), for which the five-year survival rate is only about 6% [4]. The current standard of care, which includes maximal tumor resection, followed by radiotherapy targeting the tumor bed and concurrent chemotherapy using temozolomide in combination with a specific radiation dose [5,6], remains insufficient. For this reason, there is an intensive search for new therapeutic strategies in the treatment of gliomas [7]. In this context, an important area of research remains the identification of naturally occurring bioactive compounds that exhibit anticancer properties and may serve as a basis for the development of new therapies [8]. One such compound is ophiobolin A (OP-A), which belongs to the group of sesterterpenoids produced by pathogenic fungi of the genus Bipolaris [9]. This compound exhibits significant anticancer properties [10], and its potential in GBM therapy has been confirmed in preclinical studies. In an orthotopic mouse glioma model, OP-A led to tumor growth inhibition, prolonged survival, and demonstrated the ability to cross the blood-brain barrier [10], which is one of the main limitations of many anticancer therapies targeting the central nervous system [11]. Furthermore, a key aspect of OP-A’s mechanism of action is its relative selectivity toward cancer cells—it has been shown that exposure to this compound did not significantly affect the viability of normal astrocytes, and its effect on liver cells was minimal [10]. To date, studies have shown that OP-A induces paraptosis-like cell death in GBM [12,13], independent of classical apoptotic pathways and mechanisms dependent on reactive oxygen species [12]. This suggests the potential for using OP-A to overcome GBM cells’ resistance to therapy, which often results from abnormalities in apoptotic pathways [14]. Despite a growing body of evidence pointing to the anticancer potential of OP-A, the molecular mechanisms underlying its action in glioma cells remain not fully understood. Previous studies have focused primarily on analyzing cell morphology, the types of induced cell death, and the pathways leading to its occurrence, while changes in the expression of genes associated with apoptosis and redox homeostasis have not been clearly characterized. Ferroptosis is an iron-dependent form of regulated cell death driven by excessive lipid peroxidation and the disruption of cellular redox homeostasis. Natural small molecules may promote ferroptosis in tumor cells by increasing ROS production and weakening antioxidant defenses [15]. Understanding these early changes is crucial for elucidating the mechanisms of OP-A action and may provide information useful for optimizing therapeutic strategies in GBM. Therefore, the aim of this study was to analyze early changes in the expression of genes associated with apoptosis, ferroptosis, and antioxidant mechanisms in T98G and U118MG glioma cells exposed to OP-A. 2. Results 2.1. The Effect of OP-A on the Expression of Genes Associated with Cell Death Pathways and Redox Homeostasis 2.1.1. The Effect of OP-A on the Expression of Apoptosis-Related Genes Analysis of the expression of apoptosis-related genes revealed a similar direction of change for some of the markers studied in both cell lines. In both lines, a trend toward increased BAX expression was observed (T98G: log2FC = 1.3006; U118MG: log2FC = 1.8764), accompanied by a tendency toward decreased BCL2 expression (T98G: log2FC = −0.1458; U118MG: log2FC = −0.7446). Furthermore, a trend toward increased FAS expression was observed in both cell lines (T98G: log2FC = 1.5164; U118MG: log2FC = 1.0099). A different pattern of changes was observed for CASP3, whose expression showed an upward trend in T98G cells (log2FC = 0.6436) and a downward trend in U118MG cells (log2FC = −1.3614). Despite the observed directional changes, none of the analyzed differences reached statistical significance (Figure 1). 2.1.2. The Effect of OP-A on the Expression of Genes Associated with Ferroptosis and Antioxidant Mechanisms An analysis of gene expression related to redox homeostasis and the mechanisms regulating ferroptosis revealed a varied pattern of changes in response to OP-A in the studied cell lines. In both lines, a trend toward increased SLC7A11 expression was observed (T98G: log2FC = 2.9314; U118MG: log2FC = 1.0214), indicating a marked transcriptional induction of this gene. Furthermore, GPX4 expression showed an upward trend in both T98G cells (log2FC = 0.5666) and U118MG cells (log2FC = 0.8856). Opposite trends were observed for ACSL4 and GCH1. In T98G cells, both genes showed a tendency toward increased expression (ACSL4: log2FC = 0.6321; GCH1: log2FC = 0.2724), whereas in U118MG cells, a downward trend was observed (ACSL4: log2FC = −0.7139; GCH1: log2FC = −0.8911). Despite the observed directional changes, none of the analyzed differences reached statistical significance (Figure 2). 2.2. Evaluation of OP-A-Induced Apoptosis in T98G and U118MG Cells Using Annexin V/PI Staining Apoptosis was evaluated by Annexin V/PI flow cytometry after 1 h of exposure to 47.8 µM OP-A. The corresponding vehicle-control samples were exposed to 0.4% (v/v) DMSO. The descriptive results for T98G and U118MG cells are presented in Table 1 and Table 2, respectively. The reported values were obtained from technical replicates of a single biological sample for each experimental condition. Therefore, they reflect technical measurement variability rather than biological variability and should be interpreted as preliminary observations. No inferential statistical comparisons were performed, and confirmation in independent biological experiments is required (Figure 3, Table 1 and Table 2). Bright-field microscopy images illustrating morphological changes in glioblastoma multiforme cell lines after prolonged exposure to ophiobolin A (OP-A) is presented in Figure 4. 2.3. Exploratory Machine Learning and Descriptive Analyses 2.3.1. Classification Performance (OP-A vs. Control) Using an exploratory Random Forest classifier with LOOCV, we obtained an AUC of 0.833 (accuracy: 83.3%) for T98G cells and an AUC of 0.667 (accuracy: 66.7%) for U118MG cells (Table 3). Within this limited dataset, OP-A-treated and control samples were more consistently distinguished in T98G cells than in U118MG cells. However, because only three biological replicates were included per condition, these performance estimates are highly sensitive to individual observations and should be interpreted descriptively. They do not establish a robust or generalizable transcriptional signature or confirm cell-line-specific predictive performance. Furthermore, the absence of functional comparative experiments precludes conclusions regarding differential sensitivity to OP-A or the involvement of p53 status (Table 3). 2.3.2. Feature Importance Feature importance analysis identified BAX as the most discriminative variable, followed by SLC7A11, FAS, and CASP3 (Table 4). SHAP summary plots (refer to Figure 5a) confirmed that higher values of these features push the prediction toward OP-A (BAX, SLC7A11, FAS, and CASP3 are the top contributors to the model’s decisions) (Figure 5). 2.3.3. Bootstrap Fold Change for BAX (Effect Size) Exploratory bootstrap analysis yielded an estimated 5.57-fold increase in BAX expression in OP-A-treated T98G cells (95% internally bootstrapped CI: 2.41–16.12). However, because this estimate was derived from only three biological replicates per condition, it should be interpreted descriptively and does not provide confirmatory evidence of robust activation of the mitochondrial apoptosis pathway. 2.3.4. Feature Selection Stability To explore the internal consistency of feature selection, we repeatedly resampled the dataset and recorded how frequently individual features were selected. In this internal bootstrap resampling analysis, BAX was the most frequently selected feature (82.2%), followed by FAS (58.7%) and SLC7A11 (39.9%). Ferroptosis-related genes, including GPX4, ACSL4, and GCH1, were selected less frequently (<10%) (Table 5). Given the limited number of independent biological replicates, these results identify candidate features for further investigation rather than validated or robust discriminators. 2.3.5. Pathway Activity Scoring Descriptive pathway scoring yielded higher apoptosis- and cysteine metabolism-related scores in OP-A-treated cells than in control cells, whereas the ferroptosis score showed only a limited difference (Table 6). Given the small sample size and the absence of inferential testing, these findings should be interpreted as exploratory trends and do not demonstrate activation or inhibition of the respective pathways. 2.3.6. Gene Correlation Network Exploratory correlation analysis showed apparent positive associations among BAX, FAS, and SLC7A11 and negative associations between BCL2 and selected pro-death-related genes (Figure 4B). Given the very small number of observations, these patterns are descriptive and should not be interpreted as evidence of coordinated gene regulation. 3. Discussion Studies conducted to date have shown that OP-A-induced GBM cell death occurs via a paraptosis-like mechanism [12,13]. Two mechanisms for this phenomenon have been proposed. Kim et al. demonstrated that OP-A induces paraptosis-like cell death through covalent binding to the thiol groups (–SH) of cysteine and glutathione and the formation of Michael adducts, which disrupts cellular thiol homeostasis, leads to abnormal disulfide bonds, endoplasmic reticulum stress, and its expansion, independent of mechanisms involving reactive oxygen species [12]. In turn, Bury et al. demonstrated that OP-A acts in glioma cells by blocking BKCa channels in the cell membrane, mitochondria, and endoplasmic reticulum, leading to an increase in intracellular K+ concentration. In response, water influx, cell swelling, and vacuolization occur, partially compensated by exocytosis. Membrane depolarization activates calcium channels, causing an increase in Ca2+ concentration, which is further enhanced by the release of ions from mitochondria and the endoplasmic reticulum (ER) via IP3 and ryanodine receptors [13]. It is worth noting that both of the described mechanisms of OP-A action induce the same cell death phenotype—paraptosis-like death. The literature describes that OP-A-induced paraptosis-like cell death is associated with severe endoplasmic reticulum stress, in which the CHOP transcription factor plays a significant role [12]. Activation of CHOP in response to endoplasmic reticulum stress increases the expression of proapoptotic genes, such as BAX [16], while simultaneously reducing the expression of the anti-apoptotic BCL2 [17]. Furthermore, CHOP regulates the expression of death receptors, such as DR4 and DR5, which activates the extrinsic apoptotic pathway [18]. To date, however, no direct evidence has been found that CHOP mechanistically induces FAS expression. In our study, analysis of the expression of apoptosis-related genes showed that exposure to OP-A is associated with a tendency toward activation of the stress response pathway in both cell lines studied. In both cell lines, a trend toward increased expression of the proapoptotic genes BAX and FAS was observed, accompanied by a tendency toward decreased expression of the anti-apoptotic gene BCL2. This profile of changes indicates a shift in the molecular balance toward signals promoting cell death, involving both components of the mitochondrial and extrinsic apoptotic pathways. At the same time, CASP3 gene expression exhibited a distinct pattern depending on the cell line—an upward trend was observed in T98G cells, whereas a downward trend was seen in U118MG cells. Despite directional changes in the expression of all analyzed markers, none of the observed differences reached statistical significance; therefore, these results should be interpreted as potential activation of early components of the apoptotic response rather than as evidence of full execution of the apoptotic program. The observed trends toward increased BAX and FAS expression, in the absence of clear transcriptional induction of CASP3, may reflect a CHOP-dependent stress response triggered by the disruption of thiol homeostasis rather than full activation of the classical apoptotic program, which is consistent with previous reports of OP-A-induced paraptosis-like cell death. Although the directional changes in apoptosis-related gene expression were broadly consistent with a pro-death response, their limited magnitude and lack of statistical significance contrasted with the pronounced increase in Annexin V-positive cells. In both cell lines, exposure to OP-A increased the proportion of Annexin V-positive cells, predominantly within the early apoptotic population, whereas only a small percentage of cells exhibited features of late apoptosis or cell death. This discrepancy may indicate that the early phenotypic response to OP-A involves post-transcriptional or post-translational mechanisms rather than substantial transcriptional regulation of the analyzed genes. However, because protein abundance, activation or cleavage status, and caspase activity were not assessed, this interpretation remains hypothetical and requires further investigation. Overall, these findings suggest that OP-A induces early apoptosis-associated cellular changes but do not provide evidence of full execution of the classical apoptotic cascade. Additionally, both the T98G and U118MG lines, similar to the U373-MG line analyzed by Kim et al. [12] and Bury et al. [13], carry mutations in the TP53 tumor suppressor gene [19]. These mutations may impair canonical p53-dependent signaling and limit the activation of classical apoptotic pathways [20,21], which may partially explain the observed lack of full activation of the caspase-3 axis despite a tendency toward increased expression of genes associated with the proapoptotic response. At the same time, an analysis of genes associated with redox homeostasis and the regulation of ferroptosis revealed a varied response pattern of GBM cells to OP-A. The most pronounced change was a trend toward increased expression of SLC7A11 in both cell lines studied, particularly in T98G cells, which may indicate the activation of compensatory mechanisms involved in maintaining thiol balance. SLC7A11 encodes the light subunit of the Xc− transporter system, which is responsible for the exchange of cystine and glutamate [22], and its increased expression may reflect an attempt to increase the availability of cysteine, which is essential for glutathione synthesis. This is particularly relevant in the context of OP-A activity, which leads to the inactivation of proteins and glutathione through covalent binding to their –SH groups [12]. The remaining genes analyzed exhibited a more complex regulatory profile. The expression of GPX4, a key enzyme responsible for the reduction of lipid peroxides and the protection of cells against ferroptosis [23], showed an upward trend in both cell lines, whereas changes in the expression of ACSL4 and GCH1 depended on the cell type—in the T98G line, an upward trend was observed, while in the U118MG line, the direction of change was opposite. Due to the lack of statistical significance, the observed differences should be interpreted as potential adaptive trends rather than as confirmed activation or inhibition of ferroptosis-related pathways. The lack of clear induction of genes key to the regulation of ferroptosis [24], coupled with a tendency toward increased SLC7A11 expression, is consistent with earlier observations by Kim et al., according to whom the cytotoxic effect of OP-A in GBM cells does not result from the classical mechanism of ferroptosis, but is primarily associated with a disruption of thiol homeostasis and the induction of endoplasmic reticulum stress leading to paraptosis-like cell death [12]. Machine learning analysis complemented the experimental findings by identifying BAX as the most robust feature discriminating OP-A-treated from control cells. Although individual gene expression changes did not reach statistical significance, the consistent ranking of BAX across bootstrap resampling suggests that modulation of stress-responsive apoptotic signaling represents a reproducible component of the cellular response to OP-A. Similarly, the high contribution of SLC7A11 supports the involvement of adaptive thiol homeostasis rather than activation of canonical ferroptosis, consistent with the proposed mechanism of OP-A-induced paraptosis-like cell death. The lower classification performance observed in U118MG cells likely reflects intrinsic biological differences between glioblastoma cell lines and indicates that the transcriptional response to OP-A is context-dependent. It should be emphasized that our study involved brief exposure of T98G and U118MG cells to a high concentration of OP-A (8 µL, 0.012 mol/L, 1 h), which allowed us to assess early changes in gene expression associated with endoplasmic reticulum stress pathways and antioxidant mechanisms. However, the limitations of this approach should be noted: the exposure time does not allow for a full assessment of the morphological and functional effects of paraptosis, such as vacuolization or loss of ionic homeostasis. Importantly, only three biological replicates per condition were included, and the study was limited to two cell lines. The involvement of ferroptosis could not be conclusively assessed because the analysis was limited to a small panel of ferroptosis-related transcripts. Functional measurements, such as lipid ROS accumulation, labile iron levels, rescue with ferroptosis inhibitors, and GPX4 protein expression or activity, will be required to determine whether ferroptosis contributes to the effects of OP-A. The present study should therefore be considered preliminary and hypothesis-generating, as the small sample size limits the statistical power and generalizability of the findings. Consequently, the classification performance estimates, feature rankings, and pathway activity scores should be interpreted as descriptive and exploratory rather than as evidence of robust transcriptional signatures, confirmed pathway activation, or cell-line-specific predictive performance. Although LOOCV and bootstrap resampling were used for internal exploratory analysis, these approaches cannot compensate for limited independent biological replication or provide external validation. Further studies involving larger independent datasets, additional GBM cell lines, different OP-A concentrations and exposure times, and functional validation are therefore required to confirm these preliminary observations and better characterize the mechanism of action of OP-A. Considering the results obtained, OP-A may be a particularly interesting therapeutic candidate for the treatment of gliomas. GBM cells are characterized by an increased ability to neutralize oxidative stress, which limits the effectiveness of therapies based on the generation of reactive oxygen species: both radiation therapy [25] and chemotherapy [26]. In this context, it is important to note that OP-A’s mechanism of action is not based on inducing oxidative stress, but rather on directly disrupting thiol homeostasis by interacting with the –SH groups of proteins and glutathione. Furthermore, the observed increase in SLC7A11 expression suggests an attempt at compensatory replenishment of cysteine and glutathione pools, which indirectly confirms their depletion in response to the compound’s action. This approach sensitizes cells to therapies utilizing oxidative stress in combination models [26,27,28]. It is worth noting, however, that elevated expression of the Xc−system in GBM cells promotes resistance to oxidative stress and simultaneously reduces sensitivity to temozolomide [29]. Therefore, it is essential to select an OP-A dose that effectively induces GBM cell death without exacerbating their resistance to therapy. Furthermore, the correlation between apoptosis defects in GBM cells and poor prognosis [14] highlights the significant role of apoptosis resistance and underscores the need to develop therapeutic strategies that induce alternative forms of cell death. In this context, OP-A’s ability to induce paraptosis-like cell death, independent of classical apoptotic mechanisms, makes it a particularly promising compound for the treatment of gliomas. Taken together, short-term exposure to 47.8 µM OP-A in T98G and U118MG cells revealed exploratory trends toward altered expression of apoptosis- and redox-related genes and a higher proportion of Annexin V-positive cells compared with the corresponding vehicle controls. These preliminary findings require confirmation in larger studies incorporating independent biological replication, additional GBM models, different OP-A concentrations and exposure times, and functional validation before definitive conclusions regarding the mechanism of OP-A action can be drawn. 4. Materials and Methods 4.1. Reagents and Chemicals Ophiobolin A (OP-A) (Sigma Aldrich, Warsaw, Poland) was used to prepare a 12 mM stock solution of OP-A was prepared in dimethyl sulfoxide (DMSO) from Sigma Aldrich, Warsaw, Poland. The addition of 8 µL of the stock solution to 2 mL of culture medium resulted in a final OP-A concentration of 47.8 µM and a final DMSO concentration of approximately 0.4% (v/v). Vehicle-control cells received an equivalent volume of DMSO without OP-A. Eagle’s Minimum Essential Medium (EMEM) and fetal bovine serum (FBS) were purchased from ATCC (Manassas, VA, USA). Penicillin, streptomycin, and neomycin were purchased from Sigma-Aldrich (Warsaw, Poland). 4.2. Cell Culture The human glioblastoma cell lines T98G and U118MG were obtained from ATCC (Manassas, VA, USA) and cultured according to the supplier’s recommendations. Cells were maintained in EMEM supplemented with 10% FBS and 1% antibiotic solution at 37 °C in a humidified atmosphere containing 5% CO2. Cells were passaged at appropriate intervals according to ATCC recommendations. 4.3. Viability of Cells Cell viability was assessed using the Muse® Cell Count & Viability Kit (Luminex, Austin, TX, USA). To 20 µL of cell suspension, 380 µL of reagent was added, incubated for 5 min at room temperature, and then analyzed using the Guava® MUSE® Cell Analyzer (Cytek Biosciences B.V., Amsterdam, The Netherlands). 4.4. Treatment of T98G and U118MG Cells with OP-A T98G and U118MG cells were seeded in 6-well plates at a density of 2 × 105 cells per well in 2 mL of culture medium and incubated for 24 h under standard culture conditions. Cells were subsequently treated by adding 8 µL of a 12 mM OP-A stock solution prepared in DMSO to each well. This resulted in a final OP-A concentration of 47.8 µM and a final DMSO concentration of approximately 0.4% (v/v). Vehicle-control cells received 8 µL of DMSO without OP-A. After 1 h of exposure, cells intended for molecular analyses were harvested, pelleted, and stored at −80 °C until RNA isolation. The molecular analyses included three independent biological replicates per experimental condition. 4.5. RNA Isolation T98G and U118MG cell pellets were lysed in RLT buffer, and total RNA was isolated using the RNeasy Mini Kit (Qiagen, Hilden, Germany) according to the manufacturer’s protocol. On-column DNase digestion was performed using the RNase-Free DNase Set (Qiagen, Hilden, Germany). RNA concentration and purity were assessed using a NanoPhotometer Pearl spectrophotometer (Implen, Munich, Germany). Isolated RNA was stored at −80 °C until further analysis. 4.6. cDNA Synthesis A total of 10 ng of RNA from each sample was reverse-transcribed into cDNA using the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems, Foster City, CA, USA) according to the manufacturer’s protocol. Reverse transcription was performed using a Mastercycler Personal Thermal Cycler (Eppendorf, Hamburg, Germany) under the following conditions: 25 °C for 10 min, 37 °C for 120 min, and 85 °C for 5 min, followed by a hold at 4 °C. 4.7. Gene Expression Analysis by RT-qPCR Expression of chosen genes was assessed with RT-qPCR using TaqMan® Gene Expression Assays (Assay IDs: GAPDH-Hs99999905_m1; ACSL4-Hs00244871_m1; GPX4-Hs00157812_m1; GCH1-Hs00609198_m1; SLC7A11-Hs00921938_m1; BAX-Hs00180269_m1; FAS-Hs00236330_m1; CASP3-Hs00234387_m1; BCL2-Hs00608023_m1) and TaqMan® Fast Advanced Master Mix for qPCR buffer (Applied Biosystems, Foster City, CA, USA). The reactions were done on QuantStudio 5 Real-Time PCR System (Applied Biosystems, Foster City, CA, USA). Results were calculated with 2−ΔΔCt method with GAPDH as reference gene. Expression was evaluated in three biologicals repeats, and every reaction was done in triplicates. 4.8. Microscopic Observation Bright-field images of T98G and U118MG cells were acquired immediately before OP-A exposure (0 h) and after 72 h of treatment with 47.8 µM OP-A using a Zeiss Primovert inverted bright-field microscope (Oberkochen, Germany). This analysis was performed to document morphological effects associated with prolonged OP-A exposure and was conducted separately from the molecular and Annexin V/PI analyses performed after 1 h of treatment. 4.9. Annexin Assay T98G and U118MG cells were treated with OP-A by adding 8 µL of a 12 mM stock solution prepared in DMSO to 2 mL of culture medium, resulting in a final OP-A concentration of 47.8 µM and a final DMSO concentration of approximately 0.4% (v/v). Vehicle-control cells received an equivalent volume of DMSO without OP-A. After 1 h of exposure, adherent cells were detached using trypsin/EDTA, collected, and resuspended in serum-supplemented medium at a concentration of 1 × 105 to 5 × 105 cells/mL. For staining, 100 µL of the cell suspension was mixed with 100 µL of Muse® Annexin V & Dead Cell Reagent. Samples were incubated for 20 min at room temperature in the dark and subsequently analyzed using a Guava® Muse® Cell Analyzer. Cells were classified as viable, early apoptotic, or late apoptotic/dead based on Annexin V and dead-cell marker fluorescence. The same analysis settings and gating thresholds were applied to all samples. The Annexin V assay was performed using one biological sample per experimental condition, with each sample analyzed in 5 technical replicates. Technical replicates represented repeated measurements of the same biological sample and were not treated as independent biological observations. Therefore, the results were analyzed descriptively, and no inferential statistical comparisons were performed. Annexin V/PI flow cytometry was performed using one biological sample for each experimental condition. Each sample was measured in three technical replicates, which consisted of repeated flow cytometric analyses of the same prepared cell sample. Therefore, the technical replicates were used only to assess measurement repeatability and were not considered independent biological replicates. 4.10. Machine Learning Methodology Machine learning and deep learning analyses were performed in Python 3.14.6 (Boston, MA, USA) (scikit-learn, shap, TensorFlow). RT-qPCR expression data (three biological replicates) were preprocessed by adding a pseudo count (1 × 10−8) to avoid division by zero, then normalized per cell line using MinMax scaling to [0, 1]. Four pathway-derived features were engineered: BAX/BCL2 ratio, apoptosis-ferroptosis crosstalk, thiol-disruption index, and combined stress response. For each cell line and treatment, we computed the following pathway-derived features:BAX/BCL2 ratio—mitochondrial apoptosis propensity.Apoptosis–ferroptosis crosstalk—(BAX + FAS)/(GPX4 + ε).Thiol-disruption index—SLC7A11/(GPX4 + ε).Combined stress response—(BAX + SLC7A11)/(BCL2 + GPX4 + ε). BAX/BCL2 ratio—mitochondrial apoptosis propensity. BAX/BCL2 ratio—mitochondrial apoptosis propensity. Apoptosis–ferroptosis crosstalk—(BAX + FAS)/(GPX4 + ε). Apoptosis–ferroptosis crosstalk—(BAX + FAS)/(GPX4 + ε). Thiol-disruption index—SLC7A11/(GPX4 + ε). Thiol-disruption index—SLC7A11/(GPX4 + ε). Combined stress response—(BAX + SLC7A11)/(BCL2 + GPX4 + ε). Combined stress response—(BAX + SLC7A11)/(BCL2 + GPX4 + ε). Binary classification (OP-A vs. control) used a Random Forest classifier (100 trees, max depth = 2 and random state = 42) with LOOCV. Feature selection (SelectKBest, k = 3) was applied inside each LOOCV fold. Because the sample size was small (n = 6 per cell line: 3 control + 3 OP-A), we employed LOOCV to obtain an almost unbiased estimate of generalization performance. For each LOOCV fold, the training set was first reduced to the top k = 3 features using SelectKBest with the ANOVA F-test to mitigate overfitting. The classifier was then trained on the selected features and evaluated on the left-out sample. Performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, and the confusion matrix. Global feature importance was derived from the Random Forest model’s Gini impurity reduction, averaged over all LOOCV folds. To provide local interpretability, we used SHAP (SHapley Additive exPlanations). A Random Forest classifier was retrained on the full T98G dataset (OP-A vs. control) using the top 4 features selected by the same ANOVA criterion. SHAP values were computed with a TreeExplainer, and for the binary classification, the SHAP values for the positive class (OP-A) were extracted. To evaluate the robustness of feature selection, we performed bootstrap stability analysis with 1000 resamples. In each iteration, we drew a bootstrap sample (same size as the original T98G OP-A/control dataset) with replacement, applied the same SelectKBest procedure (k = 3), and recorded which features were selected. The stability of a feature was defined as the proportion of bootstrap samples in which it was chosen. Features with a stability > 0.5 were considered robust. For key genes (BAX, FAS, SLC7A11), we calculated the geometric mean fold change (OP-A/control) together with 95% bootstrap confidence intervals. From the original three control and three OP-A replicates, we generated 10,000 bootstrap samples (sampling with replacement, keeping group sizes equal). The fold change was computed for each bootstrap sample as the ratio of the geometric means. The 2.5th and 97.5th percentiles of the bootstrap distribution defined the 95% confidence interval. We defined three custom gene sets based on the biological pathways investigated:Apoptosis: {BAX, BCL2, CASP3, FAS}.Ferroptosis: {GPX4, ACSL4, GCH1}.Cysteine metabolism: {SLC7A11}. Apoptosis: {BAX, BCL2, CASP3, FAS}. Apoptosis: {BAX, BCL2, CASP3, FAS}. Ferroptosis: {GPX4, ACSL4, GCH1}. Ferroptosis: {GPX4, ACSL4, GCH1}. Cysteine metabolism: {SLC7A11}. Cysteine metabolism: {SLC7A11}. For each treatment group and cell line, we computed a single-sample pathway activity score as the average of the scaled expression values (MinMax-scaled per cell line) of the genes belonging to the set. This approach provides a simplified, interpretable summary of pathway-level responses. To visualize the overall relationship among different treatments, we performed hierarchical clustering on the T98G treatment centroids (mean expression of all 12 features per treatment). The distance matrix was calculated using Pearson correlation distance (1 − Pearson correlation), and the dendrogram was constructed using the Ward linkage method. All machine learning results are reported with LOOCV to minimize overfitting. Bootstrapping (10,000 iterations) was used to obtain confidence intervals for fold changes and feature selection stability. No p-values are reported for classification performance; instead, we provide AUC with the full confusion matrices. Pathway activity scores are presented as descriptive summaries without inferential testing, given the exploratory nature of the analysis. 4.11. Statistical Analysis Differences between the control group and the treatment group were assessed using the nonparametric Mann–Whitney U test due to the small sample size in each group (n = 3). Additionally, the log2 fold change (log2FC) was calculated as the base-2 logarithm of the ratio of the arithmetic mean expression values in the treatment group to those in the control group (treatment/control). Analyses were performed using GraphPad Prism version 11 (GraphPad Software, Boston, MA, USA), and p-values < 0.05 were considered statistically significant. Data visualization was performed using GraphPad Prism version 11 (GraphPad Software, USA).