A multimodal atlas of tumour metabolism reveals the architecture of gene–metabolite covariation

Abstract

Tumour metabolism is controlled by coordinated changes in metabolite abundance and gene expression, but simultaneous quantification of metabolites and transcripts in primary tissue is rare. To overcome this limitation and to study gene–metabolite covariation in cancer, we assemble the Cancer Atlas of Metabolic Profiles of metabolomic and transcriptomic data from 988 tumour and control specimens spanning 11 cancer types in published and newly generated datasets. Meta-analysis of the Cancer Atlas of Metabolic Profiles reveals two classes of gene–metabolite covariation that transcend cancer types. The first corresponds to gene–metabolite pairs engaged in direct enzyme–substrate interactions, identifying putative genes controlling metabolite pool sizes. A second class of gene–metabolite covariation represents a small number of hub metabolites, including quinolinate and nicotinamide adenine dinucleotide, which correlate to many genes specifically expressed in immune cell populations. These results provide evidence that gene–metabolite covariation in cellularly heterogeneous tissue arises, in part, from both mechanistic interactions between genes and metabolites, and from remodelling of the bulk metabolome in specific immune microenvironments.

Main

Coordinated changes of genetically encoded metabolic enzymes and transporters, and the metabolites they act on, underpin diverse cancer-associated phenomena, including tumorigenesis¹, pluripotency^2,3, the onset of drug resistance^4,5,6 and the modulation of immune responses^7,8,9,10,11. However, despite the high value of joint profiling of metabolites and gene expression/protein levels, previous large-scale studies of tumour metabolism have overwhelmingly focused on the analysis of gene expression data¹². Conversely, the few instances of multimodal metabolomic and transcriptomic profiling of human tumour specimens have largely been performed in disparate, unrelated studies by a multitude of research teams^{13,14,15,16,17,18,19,20,21,22}. Integration of metabolomic datasets produced in different patient cohorts is challenging due to technical batch effects and the semi-quantitative nature of untargeted metabolomic data (reported in arbitrary units of relative abundance). Thus, both the scarcity of multimodal metabolomic/transcriptomic data from tissue specimens and the challenges of harmonizing available datasets fundamentally impede the discovery of recurrent, coordinated changes in metabolic gene expression and metabolite abundance across cancers.

Tumours from diverse cancer types differ in their cell-type composition, vascularization and other factors ultimately influencing metabolism. Yet, they share a convergent set of metabolic alterations^{23,24,25,26,27}. For example, several meta-analyses of the tumour metabolic transcriptome have identified recurrent upregulation of genes in one-carbon metabolism and oxidative phosphorylation across cancer types^26,28,29. Analogously, meta-analyses of metabolomics data have demonstrated that numerous central carbon metabolites (for example, lactate) and effector metabolites (for example, kynurenine) are at higher abundance in tumour tissue compared to normal tissue across many cancer types^23,30. These studies have illustrated the power of meta-analysis for distilling highly recurrent metabolic phenotypes from heterogeneous data but have left unresolved the question of how metabolic gene expression and metabolite abundance are coordinated and ultimately shape tumour physiology.

To systematically investigate gene–metabolite covariation in cancer, we assembled, harmonized and integratively analysed metabolomics and transcriptomics profiles from 988 primary tumour and matched adjacent normal tissue collected in 15 independent studies covering 11 cancer types. The preprocessed and harmonized data constitute the Cancer Atlas of Metabolic Profiles (CAMP), representing what is, to our knowledge, the largest harmonized dataset of multimodal metabolomic and transcriptomic data on primary tumour specimens. The CAMP is publicly available for download on Zenodo (https://doi.org/10.5281/zenodo.7150252) and can be interactively explored at https://rezniklab.shinyapps.io/CAMP-shiny-app/. Leveraging the diversity of diseases in our dataset, we designed a concordance-based statistical meta-analysis approach to discover instances of gene–metabolite interactions (GMIs) that transcended cancer type. This revealed two distinct classes of GMIs: First, we identified a small number of strong interactions between enzymes and metabolites involved in the same or subsequent reactions (‘proximal’ GMIs), suggesting that these enzymes are the primary determinants of their respective metabolite pool sizes. A second group of GMIs consisted of a small number of metabolites broadly correlated to large numbers of genes. Interestingly, this second class of GMIs was enriched for genes specifically expressed in immune cells, and for metabolites related to nicotinamide adenine dinucleotide (NAD⁺), a pleiotropic metabolite that acts both as a central cofactor in metabolism³¹ and as a signalling molecule influencing cell identity³². Taken together, these findings suggest that gene–metabolite covariation in tumours emerges, in part, from two complementary phenomena: the expression of enzymes with strong control over metabolite pool size, and the presence of specific cell populations in the tumour microenvironment with characteristic metabolomic profiles.

Results

The Cancer Atlas of Metabolic Profiles

Since metabolomic profiling has so far been excluded from large multimodal tumour profiling projects (for example, the TCGA³³), there is no unified resource of metabolomic/transcriptomic data in the cancer research field. However, several groups have independently produced and released matched metabolomic/transcriptomic data in diverse cancer types^{13,14,15,16,17,18,19,20,21,22}. We combined these datasets with several in-house studies to create a comprehensive collection of 988 samples (764 tumour samples and 224 adjacent normal samples) across 11 different cancer types, covering 15 datasets, which we called the CAMP (Table 1 and Fig. 1a). The overall collection includes a total of more than 40,000 unique transcripts and almost 2,500 unique metabolites. To maximize comparability across these heterogeneous studies, we applied a unified workflow to process RNA expression from microarray and RNA-sequencing (RNA-seq) data, harmonize metabolite names and annotations, and standardize data normalization and preprocessing (Methods). The CAMP represents an unprecedented resource to investigate the covariation of metabolite levels and gene expression at scale across diverse lineages of human cancers and normal tissues. To ensure high quality of the data, we evaluated several measures of quality control (QC). These included confirming that changes in metabolite abundance and gene expression between tumour and normal samples recapitulated those from prior work (Extended Data Fig. 1a,b), and demonstrating that covariation between pairs of metabolites was strongest between proximal metabolites in the metabolic network (that is, metabolite pairs acted upon by a common enzyme; Extended Data Fig. 1c)…