Abstract Background Lower-grade glioma (LGG) is a highly heterogeneous disease that presents challenges in accurately predicting patient prognosis. Mitochondria play a central role in the energy metabolism of eukaryotic cells and can influence cell death mechanisms, which are critical in tumorigenesis and progression. However, the prognostic significance of the interplay between mitochondrial function and cell death in LGG requires further investigation. Methods We employed a robust computational framework to investigate the relationship between mitochondrial function and 18 cell death patterns in a cohort of 1467 LGG patients from six multicenter cohorts worldwide. A total of 10 commonly used machine learning algorithms were collected and subsequently combined into 101 unique combinations. Ultimately, we devised the mitochondria-associated programmed cell death index (mtPCDI) using machine learning models that exhibited optimal performance. Results The mtPCDI, generated by combining 18 highly influential genes, demonstrated strong predictive performance for prognosis in LGG patients. Biologically, mtPCDI exhibited a significant correlation with immune and metabolic signatures. The high mtPCDI group exhibited enriched metabolic pathways and a heightened immune activity profile. Of particular importance, our mtPCDI maintains its status as the most potent prognostic indicator even following adjustment for potential confounding factors, surpassing established clinical models in predictive strength. Conclusion Our utilization of a robust machine learning framework highlights the significant potential of mtPCDI in providing personalized risk assessment and tailored recommendations for metabolic and immunotherapy interventions for individuals diagnosed with LGG. Of particular significance, the signature features highly influential genes that present further prospects for future investigations into the role of PCD within mitochondrial function. Supplementary Information The online version contains supplementary material available at 10.1186/s12967-023-04468-x. Keywords: Machine learning, Precision oncology, Lower-grade glioma, Programmed cell death, Mitochondrial function Introduction Gliomas are the majority prevalent primary tumors that develop of the nervous system in the brain [[41]1]. The classification of gliomas by the WHO is based on histological differences, resulting in four grades, with WHO II and III being classified as lower-grade gliomas (LGG). LGG have a more favorable prognosis compared to glioblastoma (GBM) [[42]2]. While surgery is the recommended treatment for LGG, the invasive nature or close proximity of these tumors to critical tissues can make complete removal challenging. The current standard treatment for LGG involves surgery followed by radiotherapy, with chemotherapy serving as a promising alternative therapy [[43]3]. Unfortunately, the findings suggest that all LGG survivors, regardless of their treatment approach (surgical only management or no treatment), are at risk of experiencing long-term cognitive impairments in various domains [[44]4]. Despite the current standard treatment, which yields a median survival time of 5–10 years for LGG patients [[45]5]. Hence, there is a pressing need to unravel the underlying molecular mechanisms and develop a dependable molecular classification model that can effectively evaluate prognosis and guide personalized treatment strategies for individuals diagnosed with LGG. Programmed cell death (PCD) is a crucial physiological process that plays a pivotal role in maintaining tissue homeostasis and eliminating damaged or unwanted cells. PCD can occur through various mechanisms, including apoptosis, anoikis, autophagy, alkaliptosis, cuproptosis, entosis, entotic cell death, immunogenic cell death, ferroptosis, lysosome-dependent cell death, methuosis, necroptosis, netoticcelldeath, NETosis, oxeiptosis, pyroptosis, parthanatos, and paraptosis [[46]6]. Imagine PCD as a housekeeping system within the body. Just like we clean our houses to maintain cleanliness and order, PCD acts as an internal cleaning mechanism that removes damaged or unnecessary cells to keep the tissues healthy. Among these PCD mechanisms, mitochondrial dysfunction has been implicated in several of them [[47]7]. Mitochondria play a pivotal role in supplying energy for cellular functions, regulating cellular signaling pathways, and governing PCD. Studies have shown that mitochondrial dysfunction, characterized by changes in mitochondrial structure, function, and dynamics, is associated with decreased mitochondrial respiration, altered mitochondrial morphology, and impaired mitochondrial quality control in LGG [[48]8–[49]10]. Apoptosis is a widely recognized mechanism of PCD, which serves an essential function in preserving tissue homeostasis and eliminating damaged or unnecessary cells. Apoptosis is typified by a sequence of biochemical and morphological alterations [[50]11, [51]12]. Pyroptosis is a form of PCD that occurs following inflammasome activation and caspase-1 cleavage. Cellular enlargement, membrane rupture, and the production of pro-inflammatory cytokines are its defining features [[52]13]. Ferroptosis is a recently identified type of PCD defined by iron-dependent cellular demise and lipid peroxidation [[53]14]. Autophagy is a cellular mechanism that is essential for preserving cellular equilibrium by breaking down impaired proteins and organelles. Autophagy can serve as a mechanism for either promoting cell survival or inducing cell death, depending on the specific context in which it occurs [[54]15]. Necroptosis is a form of PCD that is characterized by necrosis-like cell death and is triggered by the activation of RIPK1 and RIPK3 [[55]6]. Cuproptosis is a form of PCD that is triggered by copper overload and is characterized by lipid peroxidation and mitochondrial dysfunction. Entotic cell death occurs exclusively in viable cells and their adjacent regions. Unlike the traditional apoptotic pathway, entotic cell death does not require the activation of apoptotic executioner pathways [[56]16]. Netotic cell death is an additional type of PCD that occurs due to the discharge of neutrophil extracellular traps (NETs), commonly observed in response to infections or injuries [[57]17]. Parthanatos is a tightly controlled type of cell death triggered by excessive activation of the nuclease PARP-1 [[58]18]. The process of lysosome-mediated cell death involves the action of hydrolases that enter the cytosol through membrane permeabilization [[59]19]. Additionally, alkaliptosis, an emerging type of programmed cell death, is controlled by the process of intracellular alkalinization [[60]20]. Oxeiptosis, which utilizes the reactive oxygen sensing capabilities of KEAP1, is a recently discovered cellular pathway that is likely to operate in conjunction with other cell death pathways [[61]21]. In the context of LGG, tumor cells can evade PCD mechanisms through various strategies, similar to concealing garbage in hidden corners of a room. They may change their shape or activate specific signaling pathways to escape elimination. The increased understanding of PCD mechanisms has led to the development of numerous drugs that target these pathways. For instance, the FDA approved a BCL-2 inhibitor that regulates cell apoptosis, which is effective in treating lymphoma [[62]22]. GSDME-induced pyroptosis, a distinct form of programmed cell death, has shown potential as an anti-tumor immunotherapy [[63]23]. Furthermore, research has demonstrated that obstructing ferroptosis can trigger cellular resistance to anti-PD-1/PD-L1 treatment [[64]24]. These findings demonstrate the importance of PCD research in advancing our understanding of LGG and developing new therapies to combat them. Disrupted mitochondrial morphology, such as changes in shape, size, or cristae organization, can disrupt normal mitochondrial function and trigger PCD [[65]25–[66]27]. Structural abnormalities may affect the release of pro-apoptotic factors from the mitochondria, leading to caspase activation and subsequent apoptosis [[67]27]. Mitochondrial function is also closely linked to PCD mechanisms. Dysfunctional mitochondria with impaired oxidative phosphorylation and ATP production can induce cellular stress and initiate PCD pathways [[68]25]. To better link mitochondrial function to PCD as described above, a series of related biomarkers were screened in this study. Nowadays, several molecular markers have been identified as clinically significant in both the diagnosis and prognosis of LGG [[69]28–[70]31]. Markers such as IDH1 mutation and MGMT promoter methylation play a critical role in determining the optimal postsurgical treatment, including adjuvant chemotherapy and radiotherapy, and are strongly associated with the prognosis of LGG [[71]32]. In addition to genetic factors, various environmental and lifestyle factors can contribute to the development of LGG [[72]33]. The carcinogens present in tobacco smoke have the potential to inflict DNA damage on brain cells, thus facilitating tumor formation [[73]34]. There have been studies examining the potential correlation between exposure to electromagnetic fields (EMF), such as those emitted by power lines or electronic devices, and the risk of brain tumors, including LGG [[74]35]. In the process of screening and proving numerous biomarkers, sources of bias such as sample selection bias, tumor heterogeneity, analytical bias, and publication bias can impact the accuracy and generalizability of the results [[75]36]. Hence, the identification of survival-associated genes through transcriptome-based databases is necessary for prognostic prediction and targeted treatment selection. For decades, researchers have demonstrated that mitochondrial dysfunction and PCD mechanisms are essential for the development and spread of malignant neoplasms. In order for malignant cells to progress, they must overcome various forms of cell death and mitochondrial dysfunction. Nevertheless, there is still a lack of comprehensive understanding regarding the interplay between mitochondrial dysfunction and PCD in LGG, and there are limited detailed functional studies of these processes in LGG. To fill these knowledge gaps, we introduced a novel metric called the mitochondrial programmed cell death index (mtPCDI) to forecast the efficacy and prognosis of therapeutic interventions in LGG. Through our investigation, we discovered the heterogeneity among LGG patients and evaluated their clinical outlook. While our findings rely on the hypothesis of an interaction between mitochondrial dysfunction and PCD, this provides valuable guidance for selecting the best treatment options. Materials and methods Data collection We obtained clinical details and transcriptome data of individuals suffering from LGG from four databases: the Cancer Genome Atlas (TCGA, [76]https://portal.gdc.cancer.gov), Chinese Glioma Genome Atlas (CGGA, [77]http://www.cgga.org.cn/), Gene Expression Omnibus (GEO, [78]http://www.ncbi.nlm.nih.gov/geo), and ArrayExpress ([79]https://www.ebi.ac.uk/biostudies/arrayexpress). The total analysis included 1467 samples, with 506 samples from TCGA-LGG, 172 samples from CGGA-325, 420 samples from CGGA-693, 121 samples from Rembrandt, 121 samples from [80]GSE16011, and 142 samples from E-MTAB-3892. To improve comparability across datasets, all RNA-seq data were converted to transcripts per million (TPM) format and corrected for batch effects using the “combat” function of the “sva” package. Prior to analysis, all data were log-transformed. In data collection for differential expression analysis, we acquired RNAseq data in TPM (Transcripts Per Million) format from two distinct sources: the TCGA and the Genotype-Tissue Expression (GTEx) project. Specifically, we retrieved 506 LGG samples (WHO grade II and III) from TCGA and 105 normal cerebral cortex samples from GTEx. For consistency and standardized processing, all the data underwent uniform processing using the Toil process [[81]37] from UCSC XENA ([82]https://xenabrowser.net/datapages/). Harmonized data processing ensures that the datasets are compatible and reduces any potential bias arising from variations in data preprocessing methods. Identification of prognostic mitochondria-related genes and PCD-related genes We conducted a literature search [[83]38] and gathered 18 patterns of PCD and key regulatory genes, which included 580 genes related to apoptosis, 367 genes related to autophagy, 7 genes related to alkaliptosis, 338 genes related to anoikis, 19 genes related to cuproptosis, 15 genes related to enteric cell death, 87 genes related to ferroptosis, 34 genes related to immunogenic cell death, 220 genes related to lysosome-dependent cell death, 101 genes related to necroptosis, 8 genes related to netotic cell death, 24 genes related to NETosis, 5 genes related to oxeiptosis, 52 genes related to pyroptosis, 9 genes related to parthanatos, and 66 genes related to paraptosis. Additionally, there were 8 Methuosis genes and 23 Entosis genes, resulting in a total of 1964 PCD-related genes. We removed 416 duplicates, resulting in 1548 PCD-related cluster genes for our analysis (Additional file [84]9: Table S1). From MitoCarta 3.0 [[85]39], we extracted 1136 mitochondria-related genes (Additional file [86]10: Table S2). We employed the “limma” package to identify genes with differential expression in LGG and their associated normal tissues. To determine differential expression, we set thresholds of log2 fold change (log2FC) greater than 2 and a false discovery rate (FDR) less than 0.05. Subsequently, we utilized the “VennDiagram” package to show visual representations in differentially expressed genes (DEGs) related to mitochondrial function and programmed cell death. Additionally, we conducted Pearson correlation analysis on the RNA-seq data of TCGA-LGG samples to identify mtPCD (mitochondrial programmed cell death) co-expressed genes that exhibit a correlation coefficient (R) greater than 0.6 and a p-value (P) less than 0.001. Development of prognostic model We integrated ten diverse machine learning algorithms and evaluating 101 algorithmic combinations [[87]40, [88]41]. These machine learning algorithms included Support Vector Machine (SVM), Least Absolute Shrinkage and Selection Operator (Lasso), Gradient Boosting Machine (GBM), Random Forest, Elastic Net, Stepwise Cox, Ridge, CoxBoost, Super Partial Correlation (SuperPC), and Partial Least Squares with Cox regression (plsRcox). We followed a sequential approach [[89]40] that involved identifying prognostic variables using univariate Cox regression modeling, developing prediction models on the TCGA-LGG cohort, validating these models on five external and independent datasets (CGGA-325, CGGA-693, Rembrandt, [90]GSE16011, and E-MTAB-3892), and calculating the Harrell Consistency Index (C-index) for model selection. We defined the model with the highest average C-index in all cohorts as the optimal model. Based on previous descriptions in the references [[91]40, [92]42], we categorized LGG