Reconstruction of 1056 cancer patient-specific GEMs across 25 cancer types

To develop the computational workflow for the prediction of MGPs across multiple cancer types, cancer patient-specific GEMs were first reconstructed using the PCAWG and TCGA RNA-seq data. A previously developed generic human GEM “Recon 2M.2” [12] was integrated with the PCAWG and TCGA RNA-seq data, which attempted to reconstruct GEMs for 1056 cancer patients that represent 25 cancer types (Fig. 2a). Here, in this study, samples from CNS-GBM and CNS-Oligo were combined to obtain a greater number of IDH1 mutants from both gliomas that have been relatively well studied for the IDH1 mutation-associated oncometabolites [19, 20]. The reconstructed GEMs for 7 Eso-AdenoCA and 6 TCGA-LAML samples were discarded in this study because they did not complete up to 24 out of 182 metabolic tasks (“Methods” and Additional file 1: Fig. S1), whereas patient-specific GEMs from other cancer types successfully completed all the metabolic tasks. All the reconstructed GEMs were further evaluated using MEMOTE (metabolic model tests) [21], which, as a result, showed a high level of consistency: average scores of 95% for “Mass Balance” (i.e., equal masses of reactants and products), 93% for “Charge Balance” (i.e., equal net charges of reactants and products), and 98% for “Metabolite Connectivity” (i.e., each metabolite being part of at least one reaction). The resulting 1043 patient-specific GEMs across the 24 cancer types contained, on average, information on 72 metabolic pathways, 3829 reactions, and 1214 unique metabolites (Fig. 2b, Additional file 2: Table S1 and Additional file 3: Table S2). Liver-HCC GEMs appeared to have the greatest average number of reactions (i.e., 3964 reactions on average), while TCGA-LAML GEMs showed the smallest average number of reactions (i.e., 3448 reactions on average); this difference in the model size is likely attributed to unique metabolic activities associated with each cancer type (Fig. 2b). To further confirm that the cancer type-specific GEMs reflect different tissues of origin, reaction contents of the 1043 reconstructed GEMs were subjected to t-distributed stochastic neighbor embedding (t-SNE) [22], which clearly showed that the GEMs from the same cancer type tend to be better clustered than those from different cancer types (Fig. 2c). Distinct clustering of cancer type-specific GEMs was further substantiated by using the Jaccard index (Additional file 1: Fig. S2). This result partly demonstrates the biological quality of the patient-specific GEMs reconstructed in this study. Further analysis of these GEMs is available in Additional file 1: Fig. S3.

Fig. 2

Overview of reconstructing cancer patient-specific genome-scale metabolic models (GEMs). a Reconstruction of 1056 patient-specific GEMs for 25 cancer types by using the PCAWG and TCGA RNA-seq data. The number of samples for each cancer type is presented in a parenthesis next to the cancer type abbreviations. The presented cancer types are as follows: Biliary-AdenoCA, biliary adenocarcinoma; Bladder-TCC, bladder transitional cell carcinoma; Breast-AdenoCA, breast adenocarcinoma; Breast-LobularCA, breast lobular carcinoma; Cervix-AdenoCA, cervix adenocarcinoma; Cervix-SCC, cervix squamous cell carcinoma; CNS-GBM/Oligo, central nervous system glioblastoma or oligodenroglioma; ColoRect-AdenoCA, colorectal adenocarcinoma; Eso-AdenoCA, esophagus adenocarcinoma; Head-SCC, head-and-neck squamous cell carcinoma; Kidney-ChRCC, kidney chromophobe renal cell carcinoma; Kidney-RCC, kidney renal cell carcinoma; Liver-HCC, liver hepatocellular carcinoma; Lung-AdenoCA, lung adenocarcinoma; Lung-SCC, lung squamous cell carcinoma; Lymph-BNHL, lymphoid mature B-cell lymphoma; Ovary-AdenoCA, ovary adenocarcinoma; Prost-AdenoCA, prostate adenocarcinoma; Skin-Melanoma, skin melanoma; SoftTissue-Leiomyo, leiomyosarcoma of soft tissue; SoftTissue-Liposarc, liposarcoma of soft tissue; Stomach-AdenoCA, stomach adenocarcinoma; TCGA-LAML, acute myeloid leukemia; Thy-AdenoCA, thyroid low-grade adenocarcinoma; and Uterus-AdenoCA, uterus adenocarcinoma. b Number of reactions (pink) and metabolites (purple) across the 1043 GEMs. The model statistics for 1056 GEMs, including the discarded GEMs not presented herein, are available in Additional file 1: Fig. S1. c t-SNE plot of the reaction contents of the 1043 cancer patient-specific GEMs. Same colors are used as presented in a

Computational workflow for predicting metabolite-gene-pathway sets (MGPs)

Using the 1043 patient-specific GEMs and the mutation data from the PCAWG whole genome sequencing (WGS) data and TCGA whole exome sequencing (WES) data for 24 cancer types, MGPs were predicted using a computational workflow that consists of four steps (Fig. 1). This workflow is applied to a metabolite and generates MGPs as an output. Therefore, this workflow is repeated for entire metabolites of each patient-specific GEM across the 24 cancer types, except for currency metabolites (e.g., ATP and H+; Additional file 4: Table S3). This workflow begins with the calculation of so-called flux-sum value [23] of each metabolite (step 1 in Fig. 1). Flux-sum of a metabolite is defined to be the total sum of all the fluxes necessary for the generation or consumption of that metabolite, essentially representing its turnover rate under a pseudo-steady state condition. Biologically, a metabolite with a higher turnover rate is in high demand by serving as key intermediate or essential end product for cellular function. Therefore, the flux-sum can be seen as a measure, which quantifies the intracellular importance of that metabolite. Flux-sum approach was used to examine the robustness of bacterial metabolism [23], predict antibacterial targets [24, 25], and redesign bacterial metabolism for the enhanced chemical production [26]. Beyond bacteria, this approach has been used to reveal metabolic reprogramming of rice under salinity stress [27], predict Warburg-like effects in mouse hepatocyte deficient in a microRNA called miR-122a [28], and predict oncogenes in head-and-neck squamous cell carcinoma [29]. Also, a variant of the flux-sum approach, using an artificial sink reaction to calculate a metabolite’s production rate, was used for cancer studies [13, 30]. To justify the use of the flux-sum approach to predict mutation-associated metabolites, we examined the capability of flux-sum values to distinguish metabolic differences between normal and cancer samples by using reported metabolome data [31]. The metabolome data represent 5 distinct cancer types, and the flux-sum values were mostly successful in distinguishing the two groups (Additional file 1: Fig. S4).

In the second step, a metabolite was paired with a gene if flux-sum distributions of the metabolite appeared to be significantly different upon mutation of the gene (step 2 in Fig. 1). For convenience, a metabolite and a mutated gene involved in MGP candidates are referred to as “target metabolite” and “target gene” hereafter, respectively. For this metabolite-gene (MG) pairing, PCAWG and TCGA mutation data were prepared, which covered a total of 930 samples, each having 0–586 mutated genes and representing 18 cancer types (“Methods”; Additional file 1: Fig. S5 and Additional file 5: Table S4). At this stage, as a result, a unique list of 31,521 MG pairs was generated across the 18 cancer types. In this unique list, a compartment for a metabolite was not considered; for example, two pairs, IDH1 mutant with akg_c and akg_m (α-ketoglutarate in cytoplasm and mitochondria, respectively), were counted as one.

Next, MG pairs predicted from the previous step were connected with metabolic pathways that biosynthesize a target metabolite if these pathways show significantly different “target flux-sum” values upon mutation of a target gene (step 3 in Fig. 1). Here, the target flux-sum value refers to the summation of all the fluxes from a metabolic pathway that contributes to the biosynthesis of a target metabolite. Also, a contributing metabolic pathway considered in MGP candidates is referred to as a “target pathway.” MG pairs from the previous step were removed if target pathways were not found. Information from this step was thought to help understand the mechanism behind the association between a target metabolite and a target gene. Indeed, among the MG pairs predicted from the second step, there were pairs that showed a statistical significance (P value < 0.05; “Methods”) between a target metabolite and a target gene at a genome-scale level, but with no such significance at a pathway level. For example, 2-oxoglutarate was predicted to be significantly affected by COL6A3 mutation in CNS-GBM/Oligo at a genome scale (in the step 2), but no such significance was observed between 2-oxoglutarate and COL6A3 mutation at individual 2-oxoglutarate biosynthetic pathways, including alanine and aspartate metabolism; citric acid cycle; glutamate metabolism; glycine, serine, alanine and threonine metabolism; urea cycle; transport reactions; and additional unassigned reactions. Therefore, “2-oxoglutarate-COL6A3” pair was not selected for an MGP from this workflow. Additionally, MG pairs associated with exchange/demand reactions, transport reactions (except for those associated with essential amino acids), or unassigned reactions were not considered because they provide limited information on explaining the biological link between a target metabolite and a target gene (Additional file 3: Table S2); in GEMs, transport reactions are usually annotated with genes at a lower confidence than typical metabolic genes. As a result, 17,656 MGP candidates were generated from this step.

Finally, MGPs were selected by identifying target genes in each target pathway that show corresponding target flux-sum values significantly different from target flux-sum values of other target genes in the same pathway (step 4 in Fig. 1). For this, for each target gene in a target pathway, the mean of its target flux-sum values was calculated, and converted to the modified Z-score (“Methods”). The resulting modified Z-scores would subsequently reveal target genes that show atypical target flux-sum values despite being in the same target pathway for MGP candidates. For example, 42 MGP candidates involving 42 target genes, all predicted to be associated with 5,10-methenyltetrahydrofolate in Lymph-BNHL, were collected for folate metabolism. Despite their involvement in folate metabolism, only two target genes, BTK and EP300, encoding Bruton’s tyrosine kinase and histone acetyltransferase p300, respectively, appeared to have the mean flux-sum values significantly different from the other 40 target genes according to the modified Z-scores. Therefore, BTK and EP300 were selected to be final target genes for the target metabolite “5,10-methenyltetrahydrofolate” and folate metabolism in Lymph-BNHL. If fewer than three MGP candidates are available for a target pathway, all the MGP candidates are considered to be significant. From this step, 4335 MGPs were generated as final sets for the 18 cancer types (Additional file 6: Dataset S1).

Evaluation of the MGP-predicting computational workflow using AML and renal cell carcinoma samples

The computational workflow predicting MGPs was first evaluated using multi-omics data from the 17 AML and 21 renal cell carcinoma (RCC) samples. Here, the multi-omics data include mutation data (from either targeted gene sequences or WES data), transcriptome (RNA-seq), and metabolome data; they were experimentally obtained in this study for the AML samples (“Methods” and Additional file 7: Dataset S2 and Additional file 8: Dataset S3) and the RCC samples (“Methods” and Additional file 9: Dataset S4 and Additional file 10: Dataset S5). This evaluation was in particular focused on whether the computational workflow would generate biologically meaningful MG pairs included in the final MGPs predicted from the 17 AML samples and the 21 RCC samples (Fig. 1). As with the PCAWG data, 17 AML patient-specific GEMs and 21 RCC patient-specific GEMs were first reconstructed using the corresponding RNA-seq data (Additional file 1: Fig. S6a,b). It should be noted that one AML patient-specific GEM was discarded in this study because it did not satisfy all the metabolic tasks (i.e., the incapacity to use L-lysine in mitochondria), and among the 21 RCC samples initially collected, RNA-seq data was not properly generated for the sample “P28” due to the too low RNA sample purity, and therefore, 20 RCC patient-specific GEMs were generated as a result. Subsequently, the reconstructed 16 AML GEMs and 20 RCC GEMs were subjected to the computational workflow (Fig. 1). In this evaluation, seven mutated somatic genes were considered for the 16 AML samples (Fig. 3a), and another six mutated somatic genes were considered for the 20 RCC samples (Fig. 3e), based on bioinformatic analysis of DNA sequencing data (“Bioinformatics analysis of DNA sequences” in Methods) and consideration of additional criteria that were also applied to the PCAWG and TCGA data (“Preparation of mutation data from PCAWG WGS data and TCGA WES data” in “Methods”).

Fig. 3

Analysis of metabolite-gene (MG) pairs from metabolite-gene-pathway sets (MGPs) predicted for the 16 AML samples and 20 RCC samples. a Number of MGPs predicted for the seven mutated genes from the 16 AML samples. It should be noted that samples having IDH1 or IDH2 mutation were also considered together, presented as “IDH,” in order to examine the overall effects of mutations in IDH1 and IDH2. b Classification of the detected peaks from relative quantification of metabolites from the 17 AML samples. c AUC values of target metabolites from the final six MGPs, which were predicted from the computational workflow and supported with the AML metabolome data (Fig. 1). AUC values of target metabolites were predicted using MetaboAnalyst [33]. The black dashed line indicates the AUC value of 0.7. d AUC values for target metabolites from the final six MGPs (red dots) and 60 metabolites from the AML metabolome data; these 60 metabolites include those not predicted as a target metabolite for MGPs and are paired with each of the presented target genes (box plots). These 60 metabolites correspond to the peaks in the metabolome data that are annotated, and also present in the GEMs in b. e–h Same analyses (a–d) conducted for the 21 RCC samples. In e, MGPs were predicted for the six mutated genes from the 20 RCC samples. Samples having NOTCH1 or NOTCH2 mutation were considered together as “NOTCH,” and samples having ERBB2, ERBB3, or ERBB4 mutation were considered together as “ERBB” in order to collect the sufficient number of samples to generate AUC values. In h, AUC values for target metabolites from the final 15 MGPs (red dots) and 104 metabolites from the RCC metabolome data are presented. These 104 metabolites include those not predicted as a target metabolite for MGPs and are paired with each of the presented target genes (box plots)

With 355 unique metabolites with flux-sum values from the 16 AML GEMs, 60 MGPs involving 59 MG pairs were predicted from the computational workflow (Fig. 3a). Five target metabolites (i.e., citrate, L-lysine, L-phenylalanine, phosphoenolpyruvate and L-threonine) that belong to six MG pairs out of the final 59 MG pairs were detected in the 17 AML metabolome data (Fig. 3b, c). Next, biological significance of the target metabolites detected in the AML metabolome data was examined whether these target metabolites would show significantly different concentrations, depending on mutation of a target gene across the AML samples. The significance of a metabolite is presented in terms of the area under the receiver operating characteristic (ROC) curve (AUC), a metric often used as a discriminating power for biomarkers [32], by using MetaboAnalyst [33] (Methods). Among the final six MG pairs supported with the metabolome data, target metabolites paired with DNMT3A, IDH, IDH2, or NRAS showed AUC values greater than 0.7 [34, 35] (Fig. 3c). Here, it should be noted that the samples having the IDH1 or IDH2 mutation were also considered together, presented as “IDH,” in order to examine the overall effects of mutations in both IDH1 and IDH2. AUC values of the target metabolites in these four MG pairs also appeared to be mostly higher than AUC values of 60 metabolites detected in the metabolome data (Fig. 3d). Evaluation of the empirical statistical significance suggests that the probability of four out of six MG pairs receiving AUC > 0.7 is extremely low (empirical P value = 0.044; Additional file 1: Fig. S6c). These results revealed that the computational workflow played a role in selecting biologically more meaningful MG pairs in the final MGPs.

Similar conclusion was also derived from evaluation of the computational workflow using the RCC samples. By applying the computational workflow to the 20 RCC GEMs, 70 MGPs including 69 MG pairs were initially predicted (Fig. 3e); 14 target metabolites involved in 15 out of the final 69 MG pairs were detected in the 21 RCC metabolome data, which allowed the same evaluation as the MG pairs from the AML samples (Fig. 3f, g). As a result, eight out of the 15 MG pairs showed AUC values greater than 0.7 (Fig. 3g). As with the AML samples, empirical statistical significance was observed for eight out of the 15 MG pairs, which showed AUC > 0.7 (empirical P value = 0.018; Additional file 1: Fig. S6d). Also, the target metabolites in these eight MG pairs mostly showed greater AUC values than 104 metabolites detected in the metabolome data (Fig. 3h).

It should be noted that no significant difference was observed in AUC values between metabolites from the metabolome data, which were available in the cancer patient-specific GEMs, and those not available in the GEMs (Additional file 1: Fig. S6e, f); this suggests that metabolites in the GEMs are not necessarily more significantly associated with mutations than metabolites absent in the GEMs. Finally, the computational workflow also generated biologically valid MG pairs with empirical statistical significance (empirical P value = 0.042) from transcriptome and metabolome data generated for 67 breast cancer samples [36] (Additional file 1: Fig. S7). The predicted MG pairs include mevalonate pathway-associated metabolites (i.e., (R)-mevalonate, (R)-5-phosphomevalonate, and isopentenyl diphosphate), which were supported by the literature [37, 38].

MGPs predicted for the 18 cancer types

A total of 4335 MGPs from the computational workflow across the 18 cancer types were next evaluated. Overall, Lymph-BNHL generated the greatest number of MGPs (534 MGPs), followed by Liver-HCC (368 MGPs), Breast-AdenoCA (364 MGPs), and Lung-SCC (356 MGPs) (Fig. 4a). These cancer types also had the greatest number of mutated genes among the 18 cancer types except for Breast-AdenoCA: 244, 231, and 221 mutated genes for Lymph-BNHL, Liver-HCC, and Lung-SCC, respectively (Additional file 5: Table S4). There were also cancer types that had a relatively high number of mutated genes despite a small number of samples (e.g., Lung-SCC and LungAdenoCA in Fig. 4a), and the opposite (i.e., greater number of samples than mutated genes; e.g., Ovary-AdenoCA, Kidney-RCC, CNS-GBM/Oligo and TCGA-LAML in Fig. 4a) was also observed. Regarding the number of MGPs predicted, Lymph-BNHL showed a substantially greater number than Liver-HCC although these two cancer types had similar numbers of samples and mutated genes (Additional file 5: Table S4). Moreover, Breast-AdenoCA showed a similar number of MGPs as Liver-HCC although Liver-HCC had almost twice the number of mutated genes than Breast-AdenoCA. These statistics suggest that the resulting MGPs were not necessarily biased by the number of samples and mutated genes. Interestingly, oncogenes and tumor suppressor genes appeared to be slightly more associated with the MGPs than other target genes across the 18 cancer types (Additional file 1: Fig. S8).

Fig. 4

Overview of the predicted MGPs across 18 cancer types. a Number of samples and mutated genes considered in this study, and the number of predicted MGPs for each cancer type. b Percentages of submetabolisms (on the basis of KEGG pathways) where MGPs were predicted for each cancer type. Colors in bar graphs indicate submetabolisms that are presented in d. c, d Ten target genes associated with the greatest number of MGPs where c the number of cancer types and d the number of submetabolisms are presented for each target gene. e Distribution of target metabolites associated with the MGPs predicted for the 18 cancer types across the genome-scale human metabolic pathways. Target metabolites and submetabolisms related to target pathways in the MGPs are presented in the metabolic map without target genes. Frequency, shown with different colors between blue and red, refers to the number of cancer types where a target metabolite appeared

Next, the predicted 4335 MGPs were categorized into eight different submetabolisms according to the target pathways to gain better insights into these MGPs. As a result, in each cancer type, MGPs were mostly shown to belong to amino acid metabolism (38.5% of MGPs on average for the 18 cancer types), followed by carbohydrate metabolism (19.1%) and lipid metabolism (18.9%) (Fig. 4b). The results are overall consistent with the knowledge of cancer metabolism: for example, increased intracellular concentration of L-leucine associated with KRAS mutation in amino acid metabolism [39], and generation of D-2-hydroxyglutarate (carbohydrate metabolism) [19] and altered cholesterol homeostasis (lipid metabolism) [40] as a result of the IDH1 mutation. Interestingly, the percentage of the predicted MGPs associated with lipid metabolism was remarkably different between two sarcomas, SoftTissue-Liposarc and SoftTissue-Leiomyo, and this different metabolic composition appeared to be consistent with their biology [41]; SoftTissue-Leiomyo (leiomyosarcoma) occurs in smooth muscle [42], whereas SoftTissue-Liposarc (liposarcoma) appears in adipocytes [43]. Cell growth of the liposarcoma is highly affected by fatty acid biosynthesis, which has been suggested as a therapeutic target [44].

A closer look into the target genes involved in the predicted MGPs across the 18 cancer types further showed that seven out of the top ten target genes were cancer driver genes: TP53, IDH1, BRAF, PBRM1, PIK3CA, CREBBP, and FAT1 [45] (Fig. 4c). The MGPs associated with these target genes appeared to be involved in multiple cancer types with the exception of BRAF-associated MGPs. BRAF-associated MGPs were predicted to occur solely in Thy-AdenoCA, and also, IDH1-associated MGPs mostly appeared in CNS-GBM/Oligo. As expected, these driver genes all appeared to be associated with multiple submetabolisms through MGPs with the three most representative submetabolisms being amino acid metabolism, carbohydrate metabolism, and lipid metabolism (Fig. 4d). Representative target metabolites from these three submetabolisms (Fig. 4e) were 4-aminobutanal and 2-oxoglutarate (predicted in 15 out of 18 cancer types), and 4-aminobutanoate, L-lysine, putrescine, (3R,5S)-1-pyrroline-3-hydroxy-5-carboxylate, and trans-4-hydroxy-L-proline (14 out of 18 cancer types) from amino acid metabolism; D-fructose 6-phosphate and 6-phospho-D-gluconate (13 out of 18 cancer types), and acetyl-CoA, glyceraldehyde 3-phosphate, and 2-oxoglutarate (12 out of 18 cancer types) from carbohydrate metabolism; and decanoyl-CoA, dodecanoyl-CoA, and octanoyl-CoA (12 out of 18 cancer types) from lipid metabolism. Taken together, MGPs predicted from the 18 cancer types overall appeared to be in good agreement with the existing knowledge of cancer metabolism.

It has been known that the same gene mutation can show different metabolic effects in different cancer types [46]. To examine this idea, the MGPs predicted to be associated with PBRM1, PIK3CA, CREBBP, or FAT1 were further examined (Additional file 1: Fig. S9). Indeed, the different metabolic effects of the same gene mutation were observed, depending on a cancer type, for these four target genes. For example, PBRM1-associated MGPs predicted for Kidney-RCC and Liver-HCC showed that histidine metabolism and fatty acid biosynthesis in Kidney-RCC appeared to be affected by PBRM1 mutation in contrast to pentose phosphate pathway for Liver-HCC (Additional file 1: Fig. S9a). Some of these predicted MGPs were supported by previously reported experimental evidences, including deregulation of histidine metabolism in Kidney-RCC with PBRM1 mutation [47], and decreased availability of cholesterol upon PIK3CA mutation in human breast epithelial line (MCF10A) [48]. Thus, it is expected that the MGPs predicted herein can serve as a reference for further examining the different metabolic effects of gene mutations that have not been experimentally validated.

The novel MGPs predicted across the multiple cancer types may also have a therapeutic potential as supported by following examples. First, a MGP “L-leucine-BRCA1-transport, extracellular” was predicted for Ovarian-AdenoCA. L-Leucine activates mTOR pathway [49], which has been suggested as a therapeutic target for BRCA1-deficient cancer [50]. Thus, L-leucine restriction in the diet may help treat ovarian cancer with BRCA1 mutation by less activating mTOR pathway. Next, two MGPs, “phosphoenolpyruvate-PIK3CA-glycolysis/gluconeogenesis” and “fumarate-PIK3CA-citric acid cycle,” were predicted for Breast-AdenoCA, and may provide hypotheses for overcoming trastuzumab resistance in breast cancer with PIK3CA mutation [51]. One study showed that trastuzumab resistance might be treated by targeting altered glucose metabolism [52], and, in accordance with the two MGPs, both phosphoenolpyruvate and fumarate were reported to be more available in trastuzumab-resistant gastric cancer [53]. Thus, controlling the availability of phosphoenolpyruvate and/or fumarate may contribute to treat trastuzumab-resistant breast cancer with PIK3CA mutation. Another three MGPs paired with VHL, a cancer driver gene frequently mutated in RCC [54], support the inhibition of indoleamine 2,3-dioxygenase 1 (IDO1) as a drug target, which was previously attempted [55]. IDO1 converts L-tryptophan to N-formyl-L-kynurenine in tryptophan metabolism, and the three predicted MGPs are: “N-formyl-L-kynurenine-VHL-tryptophan metabolism” and “anthranilate-VHL-tryptophan metabolism” for the RCC samples collected in this study, and “N-formylanthranilate-VHL-tryptophan metabolism” for Kidney-RCC. IDO1 inhibition can stabilize tryptophan metabolism that is often upregulated in RCC, and causes immunosuppression [54]. Finally, two MGPs, “reduced glutathione-KEAP1-glutamate metabolism” from Lung-AdenoCA and “L-leucine-KRAS-transport, extracellular” from ColoRect-AdenoCA, are well aligned with previous drug target suggestions: inhibition of glutaminase in lung adenocarcinoma with KEAP1 mutation [56], and inhibition of LAT1 (or SLC7A5) encoding “solute carrier family 7 member 5” in colorectal cancer with KRAS mutation [57], respectively. These evidences suggest that the predicted MGPs are not only consistent with the knowledge of cancer metabolism, but also provide reasonable treatment strategies, especially drug targets.

MGPs predicted for CNS-GBM/Oligo

Finally, MGPs predicted for CNS-GBM/Oligo were further analyzed in comparison with the reported studies on these cancers. This comparative analysis would reveal specific MGPs that agree with previous findings as well as novel MGPs that can be validated in future. First, generation of D-2-hydroxyglutarate as a result of the IDH1 mutation has been well studied in gliomas [