Metab8D: a metabolic regulome network from multiomics and machine learning Authors: Ryan Schildcrout, Kirk Smith, Rupa Bhowmick, Yuntao Lu, Suraj Menon, Minali Kapadia, Emily Kurtz, Anya Coffeen-Vandeven, Srikar Nelakuditi, Sriram Chandrasekaran Published: 2026 ABSTRACT AbstractTo explore multiomic regulation of the metabolome, we used machine learning to predict metabolomic variation across ~1000 different cancer cell lines with matched omics data from eight biomolecular classes: genomic copy number variation, mutations, DNA methylation, histone post-translational modifications (PTMs), transcriptomics and RNA splice variants, non-coding transcriptomics (miRNA and lncRNA), proteomics, and phosphoproteomics. Overall, the metabolome is tightly associated with the transcriptome, with coding and non-coding RNAs emerging as top predictors. Peripheral metabolites are predictable via levels of corresponding enzymes, while those in central metabolism require combinatorial predictors in signaling and redox pathways, and may not reflect corresponding pathway expression. We reconstruct multiomic interaction subnetworks for highly predictable metabolites, and YAP1 signaling emerged as a top global predictor across four omic layers. We prioritize predictive multiomic features for single-cell and spatial metabolomics assays. Top predictors were enriched for synthetic-lethal interactions and synergistic combination therapies that target compensatory metabolic modulators. Subject terms: Gene regulatory networks, Metabolic pathways Introduction Variation in metabolite levels between cells may arise from differences in regulation at the transcriptional, post-transcriptional, translational, and post-translational levels1,2. Understanding metabolic regulation holistically remains challenging due to the complex, interconnected nature of metabolic networks, comprising thousands of reactions, that are controlled by an equally complex web of regulatory interactions3. In addition to thermodynamics and nutrient availability, intracellular metabolite levels are driven by enzyme activity, which is the basis for the metabolome’s relationship with proteogenomic networks. In turn, metabolites themselves modulate gene expression, creating complex feedback between metabolic and regulatory networks. Because of their wide availability, transcriptomics or proteomics have been used to predict the metabolome with limited success4–7. However, these studies focus on a limited set of omics data and do not examine the relative influence and predictive power of each omics class on the metabolome8–18. Further, it is unclear if certain metabolites are more predictable than others. Here, we leverage natural metabolic variation across hundreds of cancer cell lines to infer a multiomic predictor network of the metabolome (Metab8D) using machine learning (ML). Our approach utilizes ten feature sets from eight omics classes from the Cancer Cell Line Encyclopedia (CCLE), allowing for direct comparison of each class on predicting metabolite variation. We hypothesized that the extent of predictability is indicative of the strength of relationship with the metabolome. Similar assumptions have been used in numerous studies to assess relationships between transcription factors and target genes19–22. Although many metabolic regulatory relationships have been documented, they represent a small fraction of the number of interactions that likely exist in biological systems. ML algorithms can fill this gap by learning relationships from experimental data. This study comprises three major sections. The first section highlights the ML model’s ability to predict the metabolome from a diverse set of omics inputs, quantify model robustness, and recapitulate known regulatory relationships. Second, previously uncharacterized regulators of metabolism and molecular interactions are identified via the reconstructed Metab8D network, and systems-level insights are explored. Finally, we highlight Metab8D model application to drug combination discovery. From these analyses, we put forth a set of regulatory relationships that may be prioritized for further study, especially considering omics whose relationship with the metabolome is poorly understood, such as non-coding RNAs. The Metab8D network identified significant multiomic predictors of metabolomic variation that could serve as markers for either individual metabolites or the overall metabolome for in vivo, single-cell, or spatial omics studies that require limited measurements23–26. The identified synthetic lethal interactions among top metabolite predictors can help design combination therapies by targeting compensatory mechanisms. Results Metab8D model construction using machine learning Analogous to gene regulatory inference with ML, wherein the expression value of each gene is taken as the learning objective, we use metabolite level variation as the prediction task in a regression problem using multiomic data as input21,27. Each feature set is used individually to predict the metabolome, enabling rational comparisons of predictive capabilities across omics. The output is an “influence” network model that implicitly captures regulatory events at the transcriptomic, proteomic or post-translational levels. Our approach used omics from the CCLE, which include biomolecular assays from up to ~1000 cancer cell types28, to predict the corresponding metabolome. We used ten feature sets from eight biomolecular classes: 1. genomic copy number variation (CNV), 2. genomic mutations, 3. epigenomic DNA methylation, 4. epigenomic histone PTMs, transcriptomics – 5a. coding transcripts and 5b. RNA splicing, non-coding transcriptomics – 6a. miRNA and 6b. lncRNA, 7. proteomics, and 8. phosphoproteomics to train ML models for predicting relative levels of each metabolite across matched cell lines (Graphical Abstract, Supplementary Table 1). Random forest regression ML algorithms were used to generate metabolomic predictions, as they generally outperformed other ML methods based on cross-validation in our benchmarking (Supplementary Fig. 1). Random forests have also outperformed other methods in the inference of gene regulatory networks in prior studies21,27. We further make use of inherent characteristics of random forests to obtain the importance of each feature in prediction tasks. Higher feature importances indicate a greater possibility of a regulatory relationship between a gene or protein and the target metabolite. Such regulatory relationships could reflect either a feature’s impact on the metabolite or the metabolite’s impact on the feature. Separate random forest models were trained iteratively for each of 225 metabolites in the CCLE metabolome using nine feature sets individually (excluding mutations due to their binary structure, discussed further in the “ML Feature Importance Analysis reveals global and metabolite-specific predictors,” section) to create a total of 2025 models (Fig. 1A, B). Model performance was evaluated using cross-validation, and significantly predictable metabolites were identified based on a Bonferroni-corrected P value of 0.05 to account for false positives (Fig. 1C). We aimed to avoid bias in our models by employing a wide array of omics without incorporating information from previously known molecular interactors, as some omics classes are studied far more in depth than others. This framework enables rational comparisons across feature sets regardless of previous knowledge of their regulatory influence. Random forest model performance indicates transcriptomics, non-coding transcriptomics, and DNA methylation as accurate predictors of metabolomic variation. A Representative examples of 1-methylnicotinamide predictions using transcriptomic, miRNA, and proteomic inputs. Each feature set is used individually to predict relative 1-methylnicotinamide levels, which yields a predicted metabolite variation array which can be compared to the true values. Scatter plots of predicted vs true values in cross-validation help visualize the accuracy of each model. Finally, feature importance analysis identifies the most influential features on model output. In this case, the model identified NNMT, the enzyme involved in 1-methylnicotinamide synthesis, as the top feature predictive of 1-methylnicotinamide. B Number of features in each input dataset. C Number of significantly predicted (Bonferroni-corrected p-value < 0.05/225) metabolites per input feature set for each of 225 metabolites measured in the CCLE panel, also represented as a percentage out of 225. D Box plots showing the distribution of Pearson’s R values for predicted vs true values from each model type over 225 metabolite models. E Clustered heatmap showing correlation values for all non-lipid metabolites (columns) with select metabolites labeled. ML model performance indicates metabolic variation reflects transcriptional differences between cell lines Across all ML algorithms and validation procedures, transcriptomics consistently generated the most accurate predictions (Fig. 1C–E and Supplementary Figs. 1–3). While histone PTMs and phosphoproteomics were not as accurate as other omics classes, they had fewer features by an order of magnitude, showcasing that small feature sets can still impart meaningful predictions. The ability of each omics class to make accurate predictions independent of feature set size instills confidence in the hypothesis that the ML models are picking up mechanistic relationships and are not overfitting to the data. For example, with the seventh-most number of features out of nine feature sets, miRNA inputs produced the fourth-most significant predictions. In contrast, genomics (CNV) had the most features, yet only the sixth-highest predictive performance. We further validated this assumption with benchmarking models trained on the top 40 global features per dataset identified using holdout validation with a training set to test set ratio of 75 to 25 (Supplementary Fig. 1). The transcriptomic and miRNA data achieved high accuracy (97% and 96% of metabolites predicted significantly, respectively, with Bonferroni-corrected p-value < 0.05/225) while only using 40 features. The top 40 features can thus potentially serve as a reduced set of molecular markers that can be measured for reconstructing the metabolome (Supplementary Table 2). Box plots and clustered heatmaps of Pearson’s correlations (model predictions versus true relative metabolite levels) illustrate the differences in predictive power across omics inputs (Fig. 1D, E). While box plots help visualize predictability across the metabolome, the heatmap of individual metabolite models is useful for identifying metabolites that are differentially predictable with certain feature sets. While some metabolites are highly predictable across omics layers, such as 1-methylnicotinamide and phosphocreatine, others have stark differences in predictability. For example, anthranilic acid, 2-deoxyadenosine, and alpha-glycerophosphocholine predictions were more accurate with transcriptomic and proteomic inputs. This suggests that individual metabolites can be influenced, or impart influence, preferentially at particular omics levels. Conversely, some metabolites, such as pipecolic acid, do not have high correlations with any model, suggesting that their regulatory relationships might not be sufficiently captured in these data. The ability to quantify associations with the metabolome across omics classes is vital for prioritizing putative regulators. Globally, model results indicate that non-coding RNAs are highly associated with metabolomic variation across cell lines as compared to other omics classes, highlighting their importance for further study in context with metabolic regulatory networks. Control and validation experiments show consistency and robustness in predictive capability Several control experiments were conducted to account for confounding factors that may drive spurious correlations, such as tissue type, culture conditions, and tumor sample origin. We trained models using cell line subsets separated by growth type (adherent and suspension), cancer lineage (lung and hematopoietic/lymphoid tissue, as well as all lineages except lung or hematopoietic/lymphoid tissue), and by tumor sample origin (primary and metastatic) (Supplementary Fig. 2). Out of the eight control experiments conducted, confidence scores were calculated for each feature in the top 20 (0–8) depending on the number of controls where the feature also appeared in the top 20 (see “Methods”). This system assesses the model’s ability to identify regulator-metabolite relationships rather than assess overall accuracy. Top features discussed throughout are shown with corresponding confidence scores (CS: 0–8), and those discussed in the text have a CS of at least 2, meaning that they were observed in the overall model and at least two other control conditions. A CS cutoff of 2 for passing the control experiments was informed by the average CS for the top 20 features across all feature sets (2.31) (Supplementary Table 3). While the raw accuracies of the models in the control experiments are lower due to loss of observations (cell lines) per dataset, thereby limiting training data, similar trends in the predictive power of each omic subset are observed across controls. Models controlled for tumor origin show a high level of concordance in accuracy and feature importance. Overall, transcriptomics had the highest predictive accuracy across all omics models in all control experiments. Additionally, to account for data leakage and assess model accuracy when only trained on a subset of cancer lineages, models trained on lung or hematopoietic/lymphoid lineages, which contained the largest number of cell lines, were tested on the remaining cell lines. These experiments showed similar accuracies to those from cross-validation overall, with some notable differences in proteomics and histone PTMs trained on lung cancer cell lines (higher accuracy when tested on the unseen cell lines) and transcriptomics and lncRNAs trained on hematopoietic/lymphoid cell lines (lower accuracy when tested on the unseen cell lines) (Supplementary Fig. 3). To assess the robustness of the top features in independent datasets, validation experiments were conducted where top features from CCLE-trained models were compared to corresponding relationships between features and metabolites in the NCI60 cancer cell line panel29. Random forest models were trained on NCI60 genomics (CNV), transcriptomics, miRNA, and proteomics data, and top features from the CCLE models were compared to the top features from the NCI60 models (see “Methods”, Supplementary Fig. 4). CCLE-identified features showed significantly higher feature importance scores in the NCI60 data as compared to the whole feature set (Supplementary Fig. 4). To assess the generalizability of this framework to single-cell datasets, wherein the metabolome is challenging to measure, we performed the same feature robustness analysis with single-cell transcriptomics of the cell cycle and matched metabolome from synchronized cells30,31. CCLE-identified features from metabolic gene-only models (filtered for genes present in Recon3D32) show significantly higher feature importance scores in metabolic gene-only models trained on single-cell transcriptomics data than expected out of random chance (Supplementary Fig. 4). Finally, we repeated these analyses with an additional independent set of metabolomics measurements taken in controlled growth conditions for a subset of CCLE cell lines33. Although fewer cell lines were used for this study, feature importances were consistent with models trained on the original CCLE metabolomics (Supplementary Fig. 4). These experiments demonstrate generalizability of the feature-metabolite associations from CCLE-trained models across cell line panels. Further, to ensure feature-metabolite relationships were not driven by data preprocessing methods, experiments were conducted using K-nearest neighbors (KNN) imputation of missing values rather than zero imputation (Methods), which showed nearly identical accuracies (Supplementary Fig. 5). The same feature robustness analysis was performed relative to the original models, which showed that the top features were highly conserved for every feature set (Supplementary Fig. 4). ML feature importance analysis reveals global and metabolite-specific predictors For each metabolite, top predictors from each feature set were identified based on random forest feature importance scores (see “Methods”). Shapley feature importance analysis was also conducted for the top three predicted metabolites per feature set (lowest P value) (Supplementary Fig. 6)34. The transcriptomics models not only had high predictive power, but also recalled several known mechanistic metabolite-gene relationships. For some metabolites, the mRNA transcript of the corresponding enzyme that synthesizes the metabolite was the top predictor. These include the choline O-acetyltransferase gene (CS: 5) for the metabolite acetylcholine, choline dehydrogenase (CS: 4) for betaine, cystathionine beta-synthase (CS: 7) for cystathionine, and glycerol 3-phosphate dehydrogenase (CS: 8) for alpha-glycerophosphate. For other metabolites, consumption or catabolic degradation was more predictive of their levels. For example, cytidine deaminase (CS: 6) for cytidine, and monoamine oxidase A (CS: 4) for serotonin. Several metabolites’ corresponding transporters also featured as top predictors, including SLC6A6 (CS: 7) for taurine and SLC6A8 (CS: 8) for phosphocreatine. Nicotinamide and tryptophan metabolism shared several top predictors, suggesting shared regulation. Nicotinamide phosphoribosyltransferase (NAMPT) (CS: 5) was a top predictor of NAD. Indoleamine 2,3-dioxygenase (IDO) (CS: 6) and tryptophan 2,3-dioxygenase (CS: 5) were most predictive of kynurenine. Anthranilic acid was predicted by kynureninase (CS: 6), and both methylnicotinamide and niacinamide were predicted by nicotinamide N-methyltransferase (NNMT) (CS: 8 and 4, respectively). The proteomics model also recapitulates several known relationships between metabolites and specific enzymes that impact their levels. For example, the top predictor in the proteomics model for cystathionine was cystathionine beta-synthase (CBS) (CS: 5), and the top predictor for phosphoenol pyruvate was pyruvate kinase (PKM) (CS: 3). Thymine and uracil’s top predictor was dihydropyrimidine dehydrogenase (DPYD) (CS: 2 and 5, respectively)—the initial and rate-limiting enzyme in uracil and thymidine catabolism. Similarly, malic enzyme (ME2) (CS: 4), part of the malate-aspartate shuttle, was a top predictor of aspartate levels, while glutaminase (GLS) (CS: 3) was the top predictor of glutamine. In addition to examining top features for each metabolite individually, we systematically assessed the proximity of top features from genomic, DNA methylation, transcriptomic, and proteomic models to reactions involving the predicted metabolite using the Recon3D metabolic model, discussed in the “Top predictors are close topologically to the corresponding metabolites,” section below (Fig. 2A). Global insights across all 225 ML models reveal distinct patterns of predictability. A Workflow for topologically mapping features to Recon3D. Top one, two, and five features are identified from ML, and are assessed for overlap with genes present in the reactions for its predicted metabolite (both the enzyme target and the first-degree neighbor). B Pie chart showing the percentage (of 225) of top features from transcriptomic inputs. YAP1, GPX8, and CCN1 account for 14.2%, 5.3%, and 4.4% of top 1 features, respectively. No other top feature was present for more than three metabolite models. C Percentage of top 1 features from biologically relevant feature subsets using transcriptomics inputs. 49% (110) of top features (out of 225 metabolite models, metabolic gene-only) are redox enzymes, and 42% (94) of top features (out of 225 metabolite models, whole-data) are signaling enzymes. 22% (21) of top features (out of 95 metabolites that overlapped between CCLE and Recon3D) were involved in the primary reaction for their respective metabolite, and 85% (81) of top features (out of 95 metabolites) were involved within the first-degree neighboring reaction for their respective metabolite. 0% (0/10) top features from the central carbon metabolites and 25% (21/85) of peripheral metabolites (from 95 total metabolites that overlapped between CCLE and Recon3D) were involved in the enzymatically associated reaction for their respective metabolites. Lastly, out of 136 non-lipid metabolite models, 19% (3/16) of top features were redox enzymes for central carbon metabolites, and 9% (11/120) of top features were redox enzymes for peripheral metabolites (whole-data models). D Contrasting Shapley importance features for citrate and taurine from transcriptomics models. While the taurine transporter SLC6A6 is clearly shown to be the most important feature in predicting relative levels of taurine, citrate does not have one clear top predictor. In this way, feature importance analysis of citrate shows high order, complex influence by multiple features. E Heatmap of log2-scaled feature importance scores for 136 non-lipid metabolites. The top five features for each feature set are shown for every metabolite in order of transcriptomics feature importance (descending). This shows a set of metabolites that have one strong predictor (top of heatmap) and those metabolites, primarily related to central metabolism, that have multiple predictors (bottom of the heatmap). Analysis of genomic mutations revealed several well-known associations between genetic regulators and metabolite variation in cancers. Random forests yielded few accurate metabolite predictions using binary mutation input data, most likely due to underlying decision trees overfitting to weak relationships. To more clearly assess the relationship between mutations and the metabolome, support vector machines (SVMs) were used, as their underlying architecture is more suited for working with sparse data and is less likely to overfit in this scenario (Supplementary Fig. 1). While relatively less accurate than other omics across the metabolome, the mutation ML models nevertheless recapitulated known associations between important cancer markers and the metabolome. Out of the 107 metabolites predicted significantly with mutation data, mutual information calculations between mutation features and metabolites revealed a handful of critical cancer-related genes consistently appearing in the top 20 features across metabolites (Supplementary Table 4). For example, TP53 appeared in the top 20 features 20 times, APC 19 times, CDKN2A 17 times, PTEN 16 times, RB1 14 times, and MUC16 12 times. TP53 was a top feature for several nucleotides and APC was associated with several amino acids and lipids, consistent with their known effects on cancer metabolism (S. Table 5). The ML models also have the ability to identify previously uncharacterized associations between features and metabolites. While the top feature for five out of ten of the most accurately predicted metabolites from transcriptomics was found to be mechanistically connected to its respective metabolite, the remaining five metabolites either had YAP1 or CCN1 as their top predictive feature (oxalate (YAP1, CS: 3), xanthine (YAP1, CS: 2), UMP (CCN1, CS: 3), AMP (CCN1, CS: 4), and kynurenic acid (YAP1, CS: 2)). Calcium oxalate has been found to be associated with YAP135, and xanthine has shown an indirect relationship with YAP136, but the association between kynurenic acid and YAP1 has yet to be established. Similarly, AMP and CCN1 have an indirect relationship through AMPK37, but the relationship between UMP and CCN1 has yet to be established. Comparison of predictive capabilities between omics is particularly useful in the case of lncRNAs, where very few features have known functions, yet show tight associations with the metabolome. Overall, lncRNAs produced highly accurate metabolome predictions, generally second-best behind transcriptomics of protein coding genes. Interesting patterns emerged in the lncRNA data, such as MIR3936HG being the top feature for carnitine, acetyl-, propionyl-, malonyl-, butyryl-, and valeryl-carnitine (CS: 7, 7, 6, 6, 4, and 8, respectively). Top lncRNA features with high confidence scores (>5) prioritized for further study in context with their predicted metabolites are provided in the supplementary materials (S. Table 6). The ML models also revealed long range metabolite-enzyme interactions. Top features in the proteomics model for glutathione, a key redox metabolite, included SLC7A11 (CS: 0) and ALDH3A1 (CS: 2). Although it has a low confidence score, SLC7A11 is a cysteine and glutamate transporter (xCT), and its activity has been shown to influence glutathione levels38. Aldehyde dehydrogenase protects cells from oxidative stress and consumes redox cofactors. Interestingly, SLC7A11 and ALDH3A1 also form a synthetic lethal pair, suggesting that they may compensate for each other’s activity in maintaining redox balance39. Interestingly, the most common top predictors of the metabolome across various omics data involved either signaling proteins like YAP1 or redox enzymes such as GPX8 (signaling and redox enzymes were identified using the KEGG database40) (Fig. 2B, C). YAP1 was among the top 2 predictors in the transcriptomics model for 42 metabolites including numerous amino acids and nucleotides, while GPX8 was among the top 2 predictors for 22 metabolites including various lipids. This suggests that the combination of growth signaling and redox homeostasis influences a significant part of the metabolome. Additionally, CCN1, which YAP1 is known to regulate41, was the top feature for ten metabolite models from transcriptomics, and the third-most-common top feature overall (Fig. 2B). While many metabolites were predicted preferentially by a single top feature, others were predicted by multiple features, suggesting that they might be involved in multiple pathways and may be regulated combinatorially. Such metabolites will be harder to predict using transcriptomic variation of a single gene. For example, citrate is an intermediate in multiple core metabolic pathways, and its concentration remains low in cancer due to its rapid conversion into acetyl-CoA and oxaloacetate42. As opposed to metabolites such as taurine where there is one clear regulator, citrate’s feature importances show a more balanced regulatory influence among several top features (Fig. 2D). The top predictor for taurine was SLC6A6 (CS: 7), the taurine transporter, and was shown by our models to have an extremely high influence in predicting relative taurine levels as compared to even the second most important feature by Shapley analysis (Fig. 2D). However, citrate’s feature importance plot does not show one clear top predictor, but rather an array of features that appear to impact relative citrate levels similarly. In this way, ML can uncover complex metabolic control by multiple inputs in addition to strong associations between single features and metabolites. For each metabolite model, the distribution of feature importances for the top five features shows differences between metabolites that have one clear top feature and those that require a combination of features to predict, such as those in central carbon metabolism (Fig. 2E). Interestingly, metabolites that had a distinct single predictor like glucuronate, lactose, and adenosine were also more generalizable to NCI60 (S. Figure 4). This may also be related to the limited ability of ML to tease out complex combinatorial relationships using small data such as the NCI60 panel. Histone PTMs are increasingly being recognized for their role in sensing and regulating metabolism43–45. The CCLE histone PTM data measure total PTM levels and lack gene-level resolution. Total levels of H3K9 trimethylation and H3K27 di- & trimethylation were the top features in histone PTMs; these are well known repressive marks, and importantly, are also sensitive to the metabolic state of the cell. Histone methylation is sensitive to the level of one-carbon metabolites like SAM and one-carbon sink metabolites such as methyl nicotinamide44. Global feature importances from control experiments also identified features integral to specific cellular states. For example, the ML models built using primary or metastatic cell lines revealed distinct miRNA predictors that are known drivers of tumor phenotypes in each condition. miR-142-5p, which has been shown to limit cell invasion and migration46, was the most important global feature from the miRNA models and the second-most-important feature in the primary tumor controls. miR-142 was also a top global feature in the DNA methylation data (fourth in whole-data models and third in primary controls) but is not a top predictor in the metastatic ML models. In the metastatic samples, miR-142-5p was ranked 232nd out of 734 miRNAs, showing that its relation to metabolite level variation is blunted in the metastatic state. Additionally, miR-181a, which is known to promote proliferation and invasion47, is the 43rd most important feature in the primary tumor control models versus 16th in the metastatic controls, highlighting its association with metabolism in metastatic cells in particular. Top predictors are close topologically to the corresponding metabolites To further substantiate the ability of ML to identify metabolic regulatory interactions, we mapped features from genomic, DNA methylation, transcriptomic, and proteomic models to reactions present in the Recon3D model (Fig. 2A, see “Methods”). Recon3D is a metabolic network model built on biochemical reactions curated from literature32. It is a comprehensive collection of known metabolic reactions in human cells, making it a prime candidate for systematically validating ML models and uncovering topological relationships between top features and their respective predicted metabolites. To ensure overlap with Recon3D, ML models were re-trained on the Recon3D metabolic gene subset, which contained between 1300 and 1900 genes per feature set (Supplementary Fig. 7A). The resulting metabolic subset ML models showed similar accuracy to whole-data models (Supplementary Fig. 7B, C). Feature importances from metabolic gene-only models and whole-data models were also conserved (Supplementary Fig. 7D, E). Models were also trained on a random set of 2000 genes, which happened to produce similar accuracies, but did not conserve mechanistic relationships in transcriptomic top features (Supplementary Table 7). Furthermore, the average feature importance score of the most important transcriptomic feature across all models was higher in the metabolic gene-only models (0.11) as compared to the random set (0.07). The top 5 features for each metabolite from all the metabolic gene-only ML models displayed a substantial number of overlaps (92–96%) with either the corresponding reactions directly associated with each metabolite or their first-degree neighbors (Supplementary Table 8). This suggests an arms-length or one-step influence strategy wherein the top predictors for ~95% of the metabolites are within one step of the reactions involving those metabolites. While the top features were on average topologically closer to the corresponding metabolite, consistent with prior studies6,48, we found distinctions for metabolites in peripheral or central metabolism (Fig. 2C). For example, in the transcriptomics ML models, none of the top features for the metabolites in central carbon metabolism overlapped with the corresponding reaction. In contrast, the top features for 25% of the metabolites in peripheral metabolism were involved in their corresponding reactions, suggesting that peripheral metabolites are more influenced by direct enzymatic regulation as compared to those in central carbon metabolism (Fig. 2C). To further explore if redox enzymes are a larger player in central carbon metabolism than peripheral metabolism, we determined if top features in each category were redox enzymes for all non-lipid metabolites. We found that 19% of metabolites in central carbon metabolism had redox enzymes as the top feature compared to only 9% for peripheral metabolites (Fig. 2C). These findings suggest that central carbon metabolism is influenced to a greater extent by redox enzymes than peripheral metabolism. YAP1 and GPX8 are globally important features across the metabolome As mentioned previously, certain features, such as YAP1, GPX8, and CCN1 in the transcriptomics models, showed up as the top feature in numerous metabolite models, highlighting their importance as global regulators. After identifying the most influential features for individual metabolite models, we next pinpointed features that were globally important across the metabolome using the average rank of the features for all 225 metabolite models (see “Methods”, Supplementary Table 9). These results were then scaled by Z-score and top features for each input omics class were identified. Important features of interest were checked for overlap between whole-data models and controls for robustness. Well-known metabolic regulators such as mTOR are featured in the top global regulators. mTOR phosphorylation was a top ten feature for several metabolites in central and nucleotide metabolism, including oxalate, succinate, xanthine, and phosphogluconate; similarly, phosphorylation of Rictor, which interacts with mTOR, was a top feature for amino acids, including arginine, consistent with their role in amino acid sensing and regulation49. Interestingly, top global predictors showed consistency across omics classes. The most intriguing example of this was YAP1, which was found to be the most important global feature for phosphoproteomics, the sixth most important for transcriptomics, and the seventh for DNA methylation (Supplementary Table 9). YAP1 is a coactivator in Hippo signaling, which is essential for growth signaling, oncogenic adaptations, and the epithelial mesenchymal transition (EMT)50. YAP1 was the most important feature for 32 metabolite models using transcriptomics, 2 using proteomics, 5 using DNA methylation, and 134 using phosphoproteomics. Other examples of cross-omic consistencies for top features were miR142 (methylation and miRNA), RB (mutations and phosphoproteomics), and ERBB2 (DNA methylation and phosphoproteomics). Not only do different omics display conserved global regulators, but globally important features have established relationships. For example, the top global feature from RNA splicing, EPHA2, has been shown to induce accumulation and activation of YAP1, subsequently promoting glutamine metabolism51. To validate the association between YAP1 transcripts and the 32 metabolites for which it was the top predictor, we used the DepMap Data Explorer to quantify the relationships between two isoforms of YAP1 protein and each metabolite with linear regression (Supplementary Table 10)52,53. We found that most associations were statistically significant (Bonferroni-corrected p-value < 0.05/225), thus increasing confidence that YAP1 acts as an important predictor for a wide array of metabolites. Construction of multiomic metabolite regulome networks for top predictive metabolites The Metab8D network, containing the top 20 predictors (and associated confidence scores) identified from random forest feature importance scores from each of the omic layers for all 225 metabolites, is available in S. Data 1 and on the GitHub repository54. To understand how top omics features interact with one another as well as with the metabolite of interest, metabolite regulome networks were constructed using top features identified by Shapley importance analysis (see “Methods”) (Fig. 3A). Shapley importance analysis for the top 3 most accurately predicted metabolites for each omics class was conducted to assess the directionality and relative magnitude of feature-metabolite associations (Supplementary Fig. 6). Integrating protein-protein, miRNA-gene and metabolite-gene interactions with Metab8D interactions reveals high consistency between feature importance and topological distribution. A Proposed metabolic regulome network from top eight features from histone PTM, DNA methylation, miRNA, transcriptomic, and proteomic-trained models for predicting 1-methylnicotinamide (highest mean Shap value) as determined by Shapley analysis. B This figure presents a comparative analysis of relationships between top features derived from Metab8D multiomics, including transcriptomics, proteomics, phosphoproteomics, genomics (CNV), methylation, miRNA, splicing, and epigenomics, and their corresponding metabolites. Top global features from each omics dataset are organized in distinct, color-coded circular layouts surrounding a central circle of metabolites. The size of each node reflects the number of metabolites associated with that feature. Edges connecting top features to metabolites illustrate the consistency in feature importance and topological distribution across the datasets. Edges between top feature factors represent correlations in their metabolite-consistency profiles, with edge width corresponding to the strength of the correlation. The node selection is based on the top features across the eight feature sets. For example, YAP1 shows high consistency with 11 metabolites based on distance of YAP1 to these metabolites in the PPI/miRNA/Recon3D networks and their feature importance in Metab8D, and shares similar metabolite interaction patterns as miR-320b. Met and Stat3 also share similar metabolite interaction patterns with each other and with WIZ (in DNA methylation). We further focused on 1-methylnicotinamide and phosphocreatine, which were the most highly predictable metabolites (lowest P values and largest correlations) across omics classes (Fig. 3A, Supplementary Fig. 8). Relationships between miRNA and RNA were identified using TarBase55, and between DNA methylation, RNA, and proteins using BioGrid56. Shapley feature importance analysis was used to identify network features for each metabolite. The interconnectedness of these networks highlights the model’s ability to identify rational biochemical relationships. For example, nicotinamide N-methyltransferase (NNMT), the enzyme involved in the methylation of nicotinamide, was the top predictor of 1-methylnicotinamide levels from both the transcriptomic (CS: 8) and proteomic (CS: 6) models. NNMT has also been shown to modulate histone methylation57,58. Shapley analysis also identifies the directionality of those relationships. For example, NNMT is positively correlated with 1-methylnicotinamide concentrations, and hsa-miR-200c, which was identified as a regulator of NNMT translation via TarBase, is negatively correlated with 1-methylnicotinamide levels (Fig. 3A). Such networks may reflect the regulatory effect of the metabolite on gene expression as well as the feature’s effect on the metabolite. Similarly, solute carrier family 6 member 8 (SLC6A8), a creatine transporter, was the top predictor of phosphocreatine in the transcriptomics model (CS: 8) and the second most important predictor in the DNA methylation model (CS: 5). Creatine kinase (CKMT1A), the enzyme responsible for phosphorylating creatine to create phosphocreatine, was the third-most-important phosphocreatine predictor in the proteomics model (CS: 3) and the second-most-important predictor in the transcriptomics model (CS: 4) (Supplementary Fig. 8). CKMT1B was also the third most important feature in the transcriptomics model for phosphocreatine (CS: 3). In addition to known interactions, these networks incorporate proposed feature-metabolite relationships identified by Metab8D. The interconnectedness of the top predictive features across models suggests that the relative levels of these metabolites involve high-order interactions between each of these omics layers. Physical interactions are frequent among top metabolite predictors It is difficult to tease out metabolic regulation by individual biomolecules given buffering and combinatorial regulation. By integrating physical and genetic interactions with multiomic associations, we hypothesized that we could understand how groups of regulators complement each other and maintain metabolic homeostasis. We overlaid protein-protein interaction (PPI) data from the STRING database, miRNA-gene interaction data from the TarBase database and metabolite-protein interactions from Recon3D with Metab8D interactions (Fig. 3B; Supplementary Fig. 9; see “Methods”)32,55,59. We further analyzed the consistency between the top features in Metab8D with highly connected nodes in the PPI, Recon3D, and miRNA networks. We computed a distance matrix between the nodes in the multiple networks and compared it with ML feature importances. Interactions that show high consistency between Metab8D and other interaction networks were visualized (Fig. 3B). This cross-network analysis highlighted several shared interactions between omic layers, especially between various signaling regulators in proteomics and phosphoproteomics, with transcription factors in transcriptomics and DNA methylation. Chek1 in phosphoproteomics shared common patterns of interactions with the transcription factor HOXA5 (DNA methylation); MET and STAT3 (phosphoproteomics) shared similar interactions with each other and with the transcriptional regulators WIZ (in DNA methylation) and the microRNA miR-1246. Similarly, enzymes in fatty acid metabolism shared cross-omic similarity with transcriptional regulators. Lipoic Acid Synthetase (LIAS) enzyme shared similar targets to H3K36 trimethylation, and fatty acid synthase (FASN) with miR-142-3p. Acyl-CoA Synthetase (ACSF3; DNA methylation) had similar targets to aldo/keto reductase AKR7A2 (splicing) with ferritin (FTL), which are all related to redox homeostasis. Genomic copy number changes in Neuroblastoma Breakpoint Family Member 1 (NBPF1) had similar interactions as the GPX8 gene methylation and the histone demethylase KDM3B proteomic levels (Supplementary Data 2). These results suggest that Metab8D predicted interactions are consistent with known molecular interactions, and our cross-omic analysis highlights potential shared regulation between omic layers. Synthetic lethal interactions reveal compensatory control mechanisms Next, inspired by the synthetic lethal (SL) interaction centered around glutathione highlighted earlier, we systematically assessed for SL interactions across all top metabolite predictors in Metab8D. Since metabolite levels are controlled by multiple factors, we hypothesized that the SL interaction analysis can identify compensatory regulators for each metabolite. We compiled curated experimental human SL interactions from SynLethDB and tested if these interactions are enriched among the top ten predictors of each metabolite (S. Data 3)60. The transcriptome model had the most striking enrichment for SL interactions among top predictors for each metabolite, with 52 interactions involving 45 metabolites (27-fold higher than random chance, p-value = 1 × 10−34) (Fig. 4A). Top metabolite predictors are enriched for synthetic lethality (SL) and drug synergy. A Frequency of SL gene pairs among top ten influencers using genomic, transcriptomic, proteomic, phosphoproteomic and multiomic (combination of top ten influencers from genomic, transcriptomic, proteomic, phosphoproteomic ML models). This is compared with frequency of SL pairs obtained from random sampling using the corresponding omics data. SL pairs were obtained from SynLethDB 2.0 which contains 35,943 human SLs. B High-confidence genetic and physical interactions among the top ten global predictors of phosphoproteomics ML model from the STRING database. C High-confidence genetic and physical interactions among the top ten global predictors of transcriptomics ML model from the STRING database. D High-confidence genetic and physical interactions among the top ten predictors of the metabolite glutathione. E Average Bliss drug interaction score (positive values implies synergy) for pairs of drugs that target top predictors of a given metabolite derived from various omics. Top metabolites associated with synergy or antagonism are highlighted. Drug interaction data was derived from DrugCombDB. Several SL interactions involve a combination of signaling factors, suggesting that combinations of signaling pathways influence variation in metabolite levels, and each gene in the SL pair may compensate for the other’s absence. For example, the top predictors of nucleotides and nucleotide-precursors showed several SL interactions involving growth signaling factors like AXL, YAP1, CCN1, PTPN14, CSK, and HRAS. Of note, YAP1 and PTPN14 are both top predictors of hypoxanthine, and they both form an SL pair and also physically interact61. SL interactions among top predictors also highlight potential long range regulatory effects. Several interactions involved both a growth signaling and a redox-related gene and were connected to metabolites in central metabolism (Fig. 4B–D). For example, the top predictive network for lactate contained an SL interaction between the redox enzyme GPX8 and CAV2, which is involved in cell signaling, growth, and lipid metabolism. Similarly, the top predictors for glyceraldehyde-3-phosphate and dihydroxyacetone phosphate, intermediates in glycolysis, also had an SL interaction between GPX8 and the tyrosine kinase AXL. Synergistic drug combinations converge on nucleotide metabolism pathways Synthetic lethality is directly related to drug synergy, wherein drugs targeting proteins in compensatory pathways show enhanced potency when combined. The abundance of SL interactions among top predictors in Metab8D suggests that drugs that target predictors of specific metabolites are likely to be synergistic. To test this hypothesis, we utilized curated drug-drug interaction data in various cancer cell lines from DrugCombDB, which includes 448,555 drug combinations covering 2887 unique drugs and 124 human cancer cell lines62. The drugs were mapped onto known targets using data from DrugBank63. We tested if synergistic interactions were enriched among drugs that target the top ten predictors of each metabolite. The transcriptome and phosphoproteome models had the strongest enrichment for synergistic interactions among top predictors for each metabolite (Fig. 4E). Drug combinations that target top predictors in the transcriptomic model were highly likely to be synergistic with an average bliss interaction score greater than +4, while randomly selected combinations were likely to be antagonistic (bliss score < −5) (Fig. 4E). Similar to observations with SL interactions, several synergistic interactions involved a combination of signaling pathway inhibitors. We next analyzed metabolites that were associated with synergistic interactions. Notably, nucleotides and nucleotide precursors showed numerous synergistic interactions across a wide range of drugs targeting diverse signaling and metabolic pathways. Guanosine, inosine, pantothenate, thymine, xanthine and hypoxanthine were among the top metabolites associated with most synergistic drug combinations (Fig. 4E and Supplementary Data 4). Out of 104 drug combinations associated with these metabolites, 100 were synergistic (Supplementary Data 4). Drugs that target diverse signaling pathways may lead to similar downstream effects on nucleotide metabolism. For example, several tyrosine kinase inhibitors including sorafenib and sunitinib are predicted to be associated with these nucleotides (Supplementary Data 4). These drugs are known to affect levels of nucleotides including guanine, adenine, adenosine, guanosine, hypoxanthine, and inosine64. To further validate these associations, we used the DepMap Data Explorer to quantify the relationships between cellular sensitivity to these signaling drugs and levels of nucleotides, and found several statistically significant associations (Supplementary Data 5)53. These results confirm the connection predicted by our analysis. In contrast, antagonistic drug interactions were surprisingly associated with metabolites in the TCA cycle. Citrate, isocitrate, and fumarate were the top metabolites associated with most antagonistic drug combinations. Several drug combinations associated with these TCA cycle metabolites involve drugs such as sirolimus, everolimus and tetracycline that are known to reduce the activity of the TCA cycle (Supplementary Data 4)65,66. Drugs that perturb TCA cycle activity may lead to antagonism with agents that require TCA cycle activity. Thus, Metab8D can provide insights into understanding combinatorial metabolic regulation, synthetic lethality, and drug synergy. Discussion Cellular metabolism is controlled by multiple layers of regulatory networks. Here, we quantified the influence of potential regulatory interactors on metabolomic variation from matched omics across eight biomolecular classes. The metabolome was surprisingly highly predictable from transcriptomic data. While the absolute levels of metabolites may depend on thermodynamics, variation may result from differences in transcriptional regulation. Further, the relationship between transcripts and metabolites is nonlinear, and a combination of transcripts influences metabolite levels. In many cases, there is no direct correspondence between metabolites and enzymes they are involved with. Signaling-related transcripts, proteins, and phosphoproteins were top predictors across omics layers. Redox metabolism emerged as a second common global predictor, exemplified by GPX enzymes, that influence glutathione levels and overall redox state of the cell. Levels of histone PTMs, miRNA, and phosphoproteins were also highly predictive of the metabolome despite having the fewest features by an order of magnitude. Analysis of feature importance levels of individual metabolites revealed that metabolites in peripheral metabolism, like taurine, were most predictable via levels of features related to those metabolites, while metabolites in central metabolism were more predictable by a combination of regulators, including redox enzymes. Because central carbon metabolism has been more extensively studied than peripheral metabolism, we would expect to find strong, clear associations with central carbon metabolism. However, we find the opposite, that central carbon metabolites often do not have a clear top predictor, suggesting that these metabolites are more prone to combinatorial regulation. We found these patterns to hold true in independent datasets from NCI60 and single-cell omics. Our analysis thus uncovered key systems-level insights on metabolome variation between cells, the generalizability of ML models, and predictability of individual metabolites. Although we make use of ML to predict the metabolome, such models are not directly transferrable to predicting metabolomics from other sources due to several factors, including differences between in vitro and in vivo conditions, differences between cell line panels, and limited feature overlap within and across omics. Furthermore, one of the key observations from this study is that specific metabolites, such as those in central metabolism, may have distinct patterns in predictability, which can affect the generalizability of ML models that aim to predict metabolomes. While many metabolites were predicted in our Metab8D model to be preferentially influenced by a single top feature, others were predicted by multiple features, suggesting that they may be regulated combinatorially. Such metabolites will be harder to predict using a single omics feature. Hence, the focus of Metab8D is not the prediction of absolute metabolite levels, but rather the associations between features and metabolites, as reflected in our validation experiments. A limitation of our study is that our analysis was done in cancer cell lines in culture and may not reflect metabolism in vivo; however, patterns observed here related to transcriptional variation and topological relatedness of gene-metabolite associations have also been observed in vivo in humans and model systems48,67,68. Further, the ‘guilt by association’ strategy used by network inference methods may not identify causative mechanisms. We address this through integration of multiple layers of orthogonal omics information that we found to be consistent with each other69. Further, these interactions are robust across hundreds of cell lines in independent datasets spanning multiple tissue lineages and growth conditions. Our model also correctly recalls several known interactions reported in literature. The above suggests that the statistically inferred associations may represent real interactions. Another limitation is that network associations driven by systems-level data are descriptive, and findings are not necessarily mechanistic, even among high confidence features. However, along with extensive documentation of Metab8D associations reflected in known molecular interactors, these findings suggest strong mathematical relationships between the metabolome and other omics classes. Experimentally validating proposed interactions from this study, even for a small percentage of interactions, would require an immense effort. We hope that these prioritized associations may inform rational exploration of previously uncharacterized metabolic regulatory interactors. Additionally, some high-throughput measurements do not provide a comprehensive panel of molecules that exist within a cell. For example, the metabolite panel used for this study consisted of 225 metabolites, a fraction of those that exist within a cell across distinct cellular compartments. Furthermore, there exist more proteins in human cells than used in this study, and the 90 phosphoproteins present were measured using a targeted approach, and therefore are not representative of the entire cellular array of phosphoproteins. To make a strong inference as to the relative influence of these classes on the metabolome, these analyses would need to be conducted on more comprehensive, untargeted data. Similarly, differences in sample quantities can influence accuracies, where omics with too few samples (cell lines) may not perform as well (Supplementary Fig. 1 and Supplementary Table 1). Differences in feature size across omics motivated the decision to keep feature sets separate when building the metabolite prediction models, as models trained on combined feature sets are imbalanced and therefore are challenging to interpret. Additionally, models that use combined feature sets display lower overall accuracy, further obscuring meaningful insights. By training models on each feature set independently, we can consistently compare accuracies while maintaining interpretability. This was substantiated by models trained on the top 40 global features per dataset, which maintain relatively high accuracies (except in the case of lncRNA) while controlling for feature size. Overall, our work can guide the design of rational multiomic data integration methods for metabolome prediction. The feature importance heatmap (Fig. 2E) identifies multiomic feature subsets that could work in tandem to predict particular metabolites, and could inform algorithm design that accounts for combinatorial regulation. Beyond the relative predictability of the metabolome from omics inputs, we find that top predictors from ML are mechanistically related to the associated metabolites. For example, in the transcriptomics ML model using genes present in Recon3D, the metabolites’ top predictors were enzymes closely associated, either directly (22%) or first-degree neighbor (85%), with those metabolites. The top predictors for each metabolite across different omics layers were highly enriched for genetic and pharmacological interactions. This further supports the ability of our approach to uncover mechanistic interactions and highlights potential multiomic compensatory buffering between top metabolomic predictors. Finally, we find that drugs that target top predictors in nucleotide metabolism are likely to be more synergistic. This builds upon our observation of increased synthetic lethality among the top predictors of each metabolite. Notably, a wide range of signaling drugs intersect with nucleotide metabolism and are predominantly synergistic when used in combination (mean bliss score > 10, Supplementary Data 5, Fig. 4E). Our analysis suggests that these drugs, which target signaling pathways, may also impact nucleotide homeostasis. This is consistent with a prior study that found combinations of Mek and PI3 Kinase inhibitors show synergy via nucleotide metabolism70. In contrast, drug antagonism was associated with metabolites in the TCA cycle. A similar observation was made in bacteria wherein antibiotics that have opposing effects on TCA cycle and bacterial respiration are antagonistic71. Several drugs require active growth and metabolism to be potent. Inhibition of TCA cycle may result in growth arrest which may lead to reduced potency of growth-dependent drugs. In contrast, cells with nucleotide imbalances continue to grow72. This may result in errors in DNA replication and transcription leading to cell death. Overall, Metab8D may provide a rational framework for metabolic engineering and designing synergistic combination therapies. Methods Data acquisition and preprocessing Omics from the Cancer Cell Line Encyclopedia (CCLE) were obtained from the Broad Institute (Cambridge, MA) Dependency Map Portal (including associated studies for proteomics and phosphoproteomics, which were obtained from Nusinow et al. and Li et al., respectively)28,73,74. These include genomic copy number variation, mutations, epigenomic bulk histone H3 modifications, epigenomic DNA methylation, transcriptomics, RNA splicing, miRNA, lncRNA, proteomics, phosphoproteomics, and metabolomics. Omics data acquisition methods can be found in CCLE-associated publications28,73–75. The number of cell lines used for training can be found in Supplementary Table 1. Data preprocessing involved replacing missing values (NaN) with zeros prior to scaling, taking the median value of repeated observations (if required), standardizing cell line or feature names across datasets, and removing features with zero variance. To ensure that zero imputation did not bias results, models were trained on data with missing values replaced using KNN imputation within each cross-validation fold (and after scaling). Such models showed little difference in accuracy and had robust feature conservation (Supplementary Figs. 4 and 5). Only lncRNA, RNA splicing, and mutation data had features with zero variance (38, 1131 and 146 features, respectively). Once removed, there were 3767, 2957 and 17,203 features used for model training, respectively. RNA splicing ensemble IDs were converted to common names using BioTools.fr76. Each data type was merged independently with the metabolomics dataset by common cell line, where each matrix was organized by cell lines (rows) and features (columns). Input datasets for the models containing metabolically relevant genes were filtered for genes present in the Recon3D genome-scale metabolic model77. The same data preprocessing steps and scaling methods were used for validation experiments with NCI60 data, which were obtained from the National Cancer Institute CellMiner portal29,78. Machine learning model training ML models that predict metabolomic variation were trained on individual omic data from the CCLE. A separate model was trained and tested for each metabolite using 5-fold cross-validation, where data were normalized by Z-score within each CV fold. Statistically significant predictions were determined using a Bonferroni-corrected P value (0.05/225) to account for false positives. Because the number of features was much greater than the number of observations in these data, it was important to select algorithms that penalize large coefficients and limit overfitting. Random forests were primarily used for this study, although several algorithms were tested for benchmarking, including Ridge Regression, Lasso Regression, XGBoost, and SVMs79–83. Each of these algorithms is easily interpretable (provides feature importance scores or coefficients) and characterizes different mathematical relationships. Non-linear SVMs are more difficult to interpret, and therefore mutual information calculations were required to provide surrogate feature importance scores. The Python script and random forest results for these algorithms can be found at the Metab8D GitHub repository54 and in the Zenodo record84, which also includes the CCLE data used for model training and the trained random forest models. Random forests are comprised of many decision trees, and are classified as an ensemble learning method that aggregates many iterations of decisions. This algorithm works by randomly selecting a training set from the input data (bootstrapping) and using that training set to dictate a unique decision tree85. These predictions are averaged over all the decision trees, which is what allows for more accurate predictions from the overall model than any individual tree could provide85. Ultimately, we chose random forests for the bulk of our analyses for their balanced and interpretable framework as well as their overall high accuracy in comparison to other algorithms. More complex models can be challenging to interpret, and simpler models may not identify non-linear or combinatorial relationships. Hyperparameters Random forests were trained using Sci-kit Learn’s Random Forest Regressor algorithm with the following parameters: (random_state = 0, n_estimators = 100, n_jobs = −1, max_depth = 10). These parameters were selected after considering computation time while maintaining sufficient depth. While random forests are robust to changes in hyperparameters, we elected to not tune such parameters to avoid overfitting. A similar approach was taken for XGBoost and SVMs. For ridge and lasso regression, an array of alphas was used in conjunction with cross validation to determine the appropriate regularization term for each model. We used an array containing alpha values of 1e-2, 0.1, 1, 10, 100, and 1000. The best alphas were saved for each cross-validated model and applied to final model training using the whole data. Statistics and reproducibility Performance of each model was assessed using Pearson’s correlation between predicted and true metabolite concentrations. The statistical significance of each model prediction was assessed using the Pearson’s correlation between predicted and true metabolite concentrations along with its associated t-test P value. To be considered statistically significant, model results must have a positive R value (as negative R values would indicate poor model performance) and a P value lower than the Bonferroni cutoff (0.05/225). Cross-validation was used to assess overall accuracy of each data type, and a final model trained on the entire data set was used to examine feature importances. Feature importance scores were generated from each random forest model and were analyzed locally (for each metabolite) and globally (averaged across all metabolite models for a given feature set). Global importances were sorted by median rank and scaled by Z-score (Supplementary Table 9). Final models trained on the whole data and associated feature importances are reproducible using the code available in the GitHub repository54. The number of cell lines (samples) used for training may be found in Supplementary Table 1. Model interpretation Mutation feature importances for SVMs were calculated using mutual information—a model-agnostic method for assessing the dependency between a feature and a metabolite. This was done in lieu of directly interpretable model characteristics of non-linear SVMs. Mutated genes that appeared in the top 20 features for each metabolite were tallied (similar to control experiment confidence scores) to assess the impact of each mutation in predicting the metabolome (Supplementary Table 4). Feature importances for select predicted metabolites were also quantified using Shapley analysis, a game-theoretic approach that calculates feature importances by adding each feature into the model sequentially, and repeating that process for all orderings of features, thereby assessing the influence of each feature on the model as a whole86,87. The three metabolites with the lowest P values were selected for Shapley analysis for each data type (Supplementary Fig. 6). Each dot on the Shapley beeswarm plot corresponds to a cell line observation. Red data points indicate high feature values and blue points indicate low feature values. High SHAP values correspond with higher metabolic concentration outputs from the model, and low SHAP values correspond with lower outputs. Therefore, a gene with a high feature value and SHAP value (red and to the right) illustrates a positive correlation, and a high feature value and a low SHAP value (red and to the left) illustrates a negative correlation. Consistencies across control experiments were quantified with a confidence scoring system, where the top 20 features were extracted from whole-data and control models. We then tallied the occurrence of each of the top 20 features conserved in the top 20 features of each control experiment, out of 8 experiments, leading to a confidence scoring system ranging from 0 to 8 for each feature. The 8 experiments included models generated on the following subsets: lung cancers, hematopoietic and lymphoid cancers, all cancers except lung, all cancers except hematopoietic and lymphoid, suspension cell culture, adherent cell culture, primary tumor samples, and metastatic samples. Comparison with Recon3D The top one, two, and five features (quantified by random forest feature importance scores) from genomics (CNV), DNA methylation, transcriptomics and proteomics models (for metabolic gene-only models) were mapped back to the Recon3D metabolic model reactions. This was accomplished by matching metabolites from the CCLE panel to metabolites present in the Recon3D metabolic model. Common metabolic cofactors were removed to avoid overrepresenting neighboring reactions, including ATP, ADP, water, hydrogen, phosphate, and diphosphate, which may be found in the recon_mapping section of the GitHub repository54. Functions from the COBRA toolbox allowed for the identification of all reactions associated with the relevant metabolites (both primary and first-degree neighbors), as well as the genes associated with each reaction88. Once identified, gene names were converted to BiggIDs and matched to each set of reactions. The number of metabolite models that had at least one top feature present in the relevant set of reactions (out of the 95 metabolites in both the CCLE panel and Recon3D) were compiled (Supplementary Table 8). We performed this same analysis on a set of random transcripts, which show fewer overlaps (Supplementary Table 8). A random set of 5 transcripts was taken for each metabolite and overlaps were counted. This process was repeated 50 times, and the median value of overlaps is represented in the table (Supplementary Table 8). Model validation in independent datasets The top features from CCLE-trained models were isolated and used to analyze relationships between transcripts and metabolites in models trained on NCI60 data, single-cell data, additional CCLE metabolomics from Cherkaoui et al.33 and models trained using KNN-imputed features (Supplementary Fig. 4). Random forest feature importance scores from models trained on validation datasets were used in comparison to CCLE outputs, where top features as identified by CCLE models were isolated in validation data to examine if they were larger than by random chance. Z-score-scaled feature importance ranks were used for this analysis. Student’s t tests and Kolmogrov–Smirnov tests compared isolated feature importances to the entire feature set. Data were stacked to create one distribution among all metabolites (pooled). Aggregated importances showed significant differences between the top features and overall feature set. These analyses were conducted on sets of 1, 3, 5, 10, 50, and 100 top features from each metabolite model (except for datasets with fewer features than top feature subsets). Effect of feature size, missing features, and metabolic features We explored how well metabolic genes alone could predict the metabolome. Genomic, DNA methylation, transcriptomic, and proteomic datasets were filtered for genes present in the Recon3D metabolic model77, and new models were trained on the metabolic gene subset. The difference in feature size between omics classes can affect their predictive power, and filtering for genes present in Recon3D standardized these data to provide mechanistic information within the context of known metabolic networks. Each filtered dataset had a similar number of features, which provided an avenue through which to assess the impact of feature size on model accuracy. We first filtered the feature sets for metabolically relevant genes from Recon3D (Supplementary Fig. 7A). With fewer than 2000 genes, transcriptomics inputs predicted over 99% of metabolites statistically significantly, and other feature sets conferred similar accuracy to whole-data models in predicting each of the 225 metabolites (Supplementary Fig. 7B). The median correlation for each omics class was slightly lower in the metabolic gene-only models as compared to the whole-data models, but the distributions were similar overall (Supplementary Fig. 7C). Similar accuracies from the metabolic gene-only models and the whole-data models illustrate the lack of impact of feature size on model predictions. We further explored this by training models on a random set of 2000 genes, which also had similar accuracies. While models built on randomly selected features might not provide mechanistic insights (Supplementary Table 7), they show that the metabolome is highly predictable within a reasonable range of input feature sizes. Feature importances from metabolic gene-only models and whole-data models were also conserved significantly (Supplementary Fig. 7D, E). Both the number of training samples (cell lines) and features can impact accuracy. Proteomics and phosphoproteomics had fewer cell lines for training than the rest of the omics (Supplementary Table 1), which could explain lower accuracies comparatively. Additionally, we hypothesize that the reduced feature size in the proteomic data may account for the reduced accuracy in these models compared to transcriptomics. We interrogated this by comparing the transcriptomic and proteomic metabolic gene-only datasets to see if there were discrepancies in the feature set that might influence accuracy. Among the 471 genes that were present in the transcriptomics feature set but were absent from proteomics, 19 unique genes were the top 1 predictor in 36 metabolite models. These missing features explain the discrepancy in accuracy (~16%) between transcriptomic and proteomic inputs, suggesting that access to more comprehensive proteomic data might be more predictive of relative metabolite levels. Statistical analysis for synthetic lethality and drug combination analysis All P values were corrected using Bonferroni multiple hypothesis correction. Statistical significance of the overlaps with synthetic lethal interactions from SynLethDB 2.0 and drug interactions from DrugCombDB was assessed by creating 50 sets of random top features for each metabolite in all omics models and quantifying the overlap. This was then compared with the actual overlap using a t-test. Multiomics topological analysis based on metabolic, miRNA and PPI networks The metabolic network data was obtained from the Recon3D genome-scale metabolic model. Protein-protein interaction (PPI) data was sourced from the STRING database (https://string-db.org/), and miRNA-gene interaction data was sourced from the TarBase database (https://dianalab.e-ce.uth.gr/tarbasev9). A metabolic network linking metabolites to genes was constructed based on the metabolite-reaction-gene relationships provided by the Recon3D genome-scale metabolic model. The PPI network and miRNA-gene network were constructed directly using the pairwise relationships extracted from the STRING and TarBase databases, respectively. The Construction of Distance Matrices and Rank Difference Analysis involves two steps. First, the top features (top 20 for epigenomics, top 40 for phosphoproteomics, top 100 for other omics) were selected by feature importance ranking, mapped to the PPI network, and used to compute an n×l distance matrix (n: number of top genes in the PPI network; l: number of shared genes/proteins between the metabolic and PPI networks). Next, an l×m metabolic distance matrix was generated from the positions of genes in the Recon3D metabolic network (m: metabolites common to the metabolic network and CCLE metabolomics data). The final n×m omics-metabolite distance matrix was obtained by multiplying these matrices. For miRNA, a three-step process was used: first, a distance matrix based on miRNA-gene interactions was generated, followed by multiplication with the previous two matrices to yield the final omics-metabolite distance matrix. To compare the ranking differences of top features between the distance and feature importance matrices, both were converted into ranks, and the absolute differences were normalized to a range of 0 to 1. For each metabolite, the top 10% and bottom 10% of feature-metabolite relationships were classified as high and low consistency, respectively. Cross-omic edges were determined based on similarity of metabolite interaction patterns based on Pearson’s correlation coefficient, and a threshold of +0.4 was used to visualize significant cross-omic similarities. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Material Description of Additional Supplementary files Supplementary Data 1 Supplementary Data 2-5 Reporting Summary Acknowledgements We thank Deepak Nagrath, Maciek Antoniewicz, Analisa Difeo, and Yatrik Shah for suggestions and critical feedback on the project. The graphical abstract (https://BioRender.com/9i1s5kh), Fig. 3A (https://BioRender.com/yii2yml), Supplementary Fig. 3A (https://BioRender.com/fqyzj6q), Supplementary Fig. 4A (https://BioRender.com/x01g473), and Supplementary Fig. 8 (https://BioRender.com/pn2ro51) were created using BioRender (Schildcrout, R. (2026)). Author contributions S.C. and R.S. designed and performed research, and wrote the manuscript; K.S., R.B., Y.L., S.M., A.C.V., E.K., M.K., and S.N. performed research. Peer review Peer review information Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary handling editor: Tobias Goris. Funding This work was supported by faculty start-up funds from the University of Michigan (UM), Camille and Henry Dreyfus Foundation, the Rogel Cancer Center at UM, R35 GM13779501 from NIH to S.C., the Advanced Proteogenomics of Cancer (T32 CA140044) Program to R.S., and the Graduate Assistance in Areas of National Need (GAANN) Fellowship to R.S. Data availability All datasets are provided in figures, the supplementary materials, the GitHub repository (https://github.com/sriram-lab/Metab8D/tree/main)54, and the Zenodo record (https://doi.org/10.5281/zenodo.20100480)84. Code availability Code used to generate all models is provided in the GitHub repository (https://github.com/sriram-lab/Metab8D/tree/main)54 and the Zenodo record (https://doi.org/10.5281/zenodo.20100480)84. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s42003-026-10409-3. References Associated Data Data Availability Statement All datasets are provided in figures, the supplementary materials, the GitHub repository (https://github.com/sriram-lab/Metab8D/tree/main)54, and the Zenodo record (https://doi.org/10.5281/zenodo.20100480)84. Code used to generate all models is provided in the GitHub repository (https://github.com/sriram-lab/Metab8D/tree/main)54 and the Zenodo record (https://doi.org/10.5281/zenodo.20100480)84.