Comprehensive Analysis of Metabolic Isozyme Targets in Cancer

Abstract

Metabolic reprogramming is a hallmark of malignant transformation, and loss of isozyme diversity (LID) contributes to this process. Isozymes are distinct proteins that catalyze the same enzymatic reaction but can have different kinetic characteristics, subcellular localization, and tissue specificity. Cancer-dominant isozymes that catalyze rate-limiting reactions in critical metabolic processes represent potential therapeutic targets. Here, we examined the isozyme expression patterns of 1,319 enzymatic reactions in 14 cancer types and their matching normal tissues using The Cancer Genome Atlas mRNA expression data to identify isozymes that become cancer-dominant. Of the reactions analyzed, 357 demonstrated LID in at least one cancer type. Assessmentof the expression patterns in over 600 cell lines in the Cancer Cell Line Encyclopedia showed that these reactions reflect cellular changes instead of differences in tissue composition; 50% of the LID-affected isozymes showed cancer-dominant expression in the corresponding cell lines. The functional importance of the cancer-dominant isozymes was assessed in genome-wide CRISPR and RNAi loss-of-function screens: 17% were critical for cell proliferation, indicating their potential as therapeutic targets. Lists of prioritized novel metabolic targets were developed for 14 cancer types; the most broadly shared and functionally validated target was acetyl-CoA carboxylase 1 (ACC1). Small molecule inhibition of ACC reduced breast cancer viability in vitro and suppressed tumor growth in cell line– and patient-derived xenografts in vivo. Evaluation of the effects of drug treatment revealed significant metabolic and transcriptional perturbations. Overall, this systematic analysis of isozyme expression patterns elucidates an important aspect of cancer metabolic plasticity and reveals putative metabolic vulnerabilities.

Introduction

Cancer cells undergo metabolic adaptation that enables rapid cell proliferation and survival in various tissue environments (1). Metabolic adaptation is driven by altered regulation of enzymatic functions by mutated oncogenes (e.g., PI3K, KRAS), changes in enzyme expression levels due to DNA copy-number or epigenetic changes, and mutations, which can alter substrate specificities (2–5). An important feature of metabolic rewiring is altered isozyme composition in cancer cells relative to normal cells (4, 6). Isozymes are distinct proteins encoded by different genes that catalyze the same enzymatic reaction, but often with distinct intracellular localization, and different kinetics and substrate affinity. Most normal cells express diverse sets of isozymes, which contributes to the redundancy and robustness of critical biochemical processes, cancer cells often lose that diversity and express a single cancer-dominant isozyme. For example, of the four isozymes (PKL, PKR, PKM1, PKM2) of pyruvate kinase (PK) that catalyzes the final step of glycolysis, the L and R isoforms are expressed predominantly in the liver and red blood cells, the M1 isoform is common to most adult tissues, but in tumor tissues the M2 isoform (PKM2) is dominant (7, 8).

It has not been systematically examined how many metabolic enzymes are affected by cancer specific loss of isozyme diversity (LID), defined as one isoform becoming the dominant isoform in cancer while corresponding normal tissues express a broad range of isoforms. LID could represent a potential therapeutic vulnerability of cancer cells if it affects a rate-limiting catalytic step in a cell survival critical biological process. We hypothesize that cancer cells may become dependent on a single isozyme to catalyze a critical metabolic step and that selective inhibition of such isozyme could lead to a metabolic bottleneck and cell death (4, 9–11).

In this study, we examined isozyme expression in matched tumor and normal tissues from The Cancer Genome Atlas (TCGA) to identify enzymes that demonstrate LID. Next, the expression patterns of the shortlisted enzymes were assessed in the Cancer Cell Line Encyclopedia (CCLE) to ascertain that the cancer-dominant expression reflects changes in cancer cells rather than differences in stromal cell composition between cancer and normal tissue. The functional importance, and therefore therapeutic potential, of the cancer-dominant isoforms was subsequently assessed in genome-wide clustered regularly interspaced short palindromic repeats (CRISPR) and RNAi loss-of-function screens. We further characterized in vitro and in vivo anticancer effect of a small molecule inhibitor, PF-05175157, of the top ranked metabolic targets, acetyl-CoA carboxylase 1 (ACC1/ACACA). ACC1/ACACA catalyzes the initial rate-limiting step of de novo fatty acid synthesis and is one of the most frequently affected enzymes by LID across all 14 cancers analyzed in this study. PF-05175157 is dual inhibitor of ACC1 and ACC2 and has been tested in Phase I/II clinical trials in healthy volunteers and patients with type 2 diabetes (NCT01819922, NCT01792635, NCT01396161).

Materials and Methods

Pan-cancer human gene expression data

RNA sequencing (RNA-seq) expression data were obtained from the TCGA RNAseqDB database (12, 13). Only cancer types with >10 paired tumor-normal samples were included in the analysis. Expression data was normalized using upper quantile normalization and log₂-transformed [log₂(x + 1)]. For each cancer, only genes expressed in >25% of tumor-normal pairs or showed a ≥ 40% expression change in tumor versus normal were considered. Methods of assignment of breast cancer subtypes are described in Supplementary Methods.

Independent breast cancer gene expression datasets

The E-GEOD-70951 data includes background-corrected, quantile normalized, and log₂-transformed microarray results generated on two different platforms (cohort 1 and cohort 2) from 195 paired breast cancer and normal breast tissues (14). We detected a large batch effect between the two platforms and several isozymes of interest were not measured in cohort 1, therefore we restricted our validation analysis to cohort 2 of E-GEOD-70951 only, that included n = 108 HR⁺/HER2⁻, n = 10 HER2⁺, n = 30 triple-negative breast cancer (TNBC) samples and matching normal tissues. The E-GEOD-76250 data includes normalized and log₂-transformed gene expression data from 33 paired TNBC and normal tissues (15). Both datasets are available through ArrayExpress (https://www.ebi.ac.uk/arrayexpress/).

Enzyme annotation

Information about enzymatic reactions and the isozymes were retrieved from the ENZYME nomenclature database, release October 2018 (https://enzyme.expasy.org), that contained 7,524 enzymatic reactions for which an enzyme commission (EC) number exists based on experimental data, and lists all isozymes that catalyze these reactions (16). We only considered the 1,341 enzymatic reactions that involved human proteins. Methods and databases used for detailed characterization of enzymatic reactions and isozymes are provided in Supplementary Methods.

Definition of LID and ranking of candidate isozyme targets

To detect enzymatic reactions with LID and to identify the most promising potential therapeutic targets, we followed a three-step selection process. First, we selected enzymatic reactions catalyzed by only 2 to 5 isozymes. Second, we identified reactions for which only one isozyme was significantly upregulated, or unchanged, in cancer compared with corresponding normal tissue while all other isozymes catalyzing the same reaction showed lower expression in cancer. Third, we ranked isozymes on the basis of how closely their expression characteristics approached an ideal LID expression profile using a target selection score that reflected five separate expression features, M1 through M5. M1 represents the expression of the target isozyme in cancer. M2 represents the expression of the complementary isozymes that catalyze the same reaction in cancer. M3 represents the difference between M1 and M2 (M3 = M1–M2 in cancer) and quantifies the “cancer-dominance” of an isozyme. M4 quantifies the expression difference between the cancer-dominant target isozyme and the complementary isozyme in corresponding normal samples [M4 = (target isozyme) – (complementary isozyme in normal tissue)], we assume that the smaller M4 is, the less potential for toxic effects from a target isozyme-specific inhibitor due to the higher expression of the complementary enzymes that would not be inhibited. M5 measures the difference between target isozyme dominance in cancer versus its dominance in normal tissues (M5 = M3–M4), and we assume that the larger the difference the greater the therapeutic index. We used these five measures to calculate a score that was the average of the sum of ranks of M1 through M5 for each candidate target isozyme (M2 and M4 were ranked in increasing order, while M1, M3, and M5 in decreasing order) presented in percentage scale and calculated separately in each cancer type.

Cell line gene expression data and gene dependency datasets

We obtained cell line RNA-seq data from the (CCLE; https://portals.broadinstitute.org/ccle, release of 8-NOV-2017; ref. 17). The data was normalized using upper quantile method and log₂ transformed [log₂(x + 1)]. In vitro functional importance of isozymes was estimated based on loss-of-function screens using RNAi (DEME-TER2; ref. 18) and CRISPR (DepMap Public 19Q3; ref. 19). DEME-TER2 contains screens for 17,309 genes in 712 cell lines, DepMap contains results for 17,634 genes in 563 cell lines. Both databases can be obtained through DepMap portal (https://depmap.org/portal). Results for both screens are expressed as a gene dependency score (GD), which is defined as the effect of single-gene knockdown on cell viability after normalization and scaling to allow comparisons across difference cell lines and genes. A GD of zero means no effect on cell viability, and negative values indicate decreased viability. Details about statistical analysis of gene dependency data are provided in Supplementary Methods.

In vitro cell viability assay

All cell lines analyzed in the study, BT20, BT474, BT483, BT549, HCC1143, HCC1187, HCC1395, HCC1500, HCC38, HCC70, HMEC, MDA-MB-231, MDA-MB-468, T47D, and ZR75–1, were purchased from ATCC and were cultured at 37°C, in 5% CO2 in RPMI1640 media (Gibco), except BT20 cells that were grown in EMEM media (Sigma), supplemented with 10% FBS (Gibco). Three thousand cells were plated in each well of a 96-well plate and 24 hours later treated with various concentrations of PF-05175157 (Pfizer) or corresponding DMSO diluent. Cell viability was determined using CellTiter-Glo Luminescent Cell Viability Assay (Promega) after 24, 48, and 72 hours of drug exposure. The assay reagent was added in equal amounts to the wells and cell lysis was performed at room temperature for 20 minutes. Luminescence was measured in a BioTek Synergy2 plate reader (BioTek). The luminescence from the treated cells was compared with their respective untreated cells that served as a control for normalization. Dose–response data were analyzed using drc R package (20) by fitting symmetric log-logistic model. The main outputs of the model were: a lower limit of the response when the dose approaches infinity, the upper limit when the dose approaches 0, and EC₅₀, the dose required to reduce the viability half-way between the upper and lower limit.

Cell-cycle analysis

Cell lines were plated in duplicate in 25-mm tissue culture dishes or in 6-well plates and treated 24 hours later with the indicated concentrations of PF-05175157, docetaxel (10 nmol/L; as a positive control) or with a DMSO control. At 24 and 72 hours posttreatment, cells were harvested and fixed overnight in 70% ethanol. Cells were then washed in PBS, stained with 50 ug/mL propidium iodide (Abcam) containing 550 U/mL RNase A, and assayed for DNA content on an LSR Fortessa Cytometer (BD Biosciences). Analysis was carried out using FlowJo software (BD Biosciences).

Apoptosis assay

Cell lines were plated in duplicate in 25-mm tissue culture dishes or in 6-well tissue culture plates and treated 24 hours later with the indicated concentrations of PF-05175157, docetaxel (10 nmol/L, as a positive control) or with a DMSO control. At 24 and 72 hours posttreatment, cells were harvested and stained with Annexin V-FITC according to the manufacturer’s instructions (ANNEX300F, BioRad). Immediately prior to analysis by flow cytometry, propidium iodide (1 μg/mL) was added to the cells and assayed for apoptosis on an LSR Fortessa Cytometer (BD Biosciences). Analysis was carried out using FlowJo software (BD Biosciences). Annexin V single positive and Annexin V/PI double positive cells were considered apoptotic.

Mouse patient-derived xenograft models

The patient-derived xenograft (PDX) experiments were performed by the In Vivo Services at The Jackson Laboratory Sacramento facility, according to an Institutional Animal Care and Use Committee–approved protocol and in compliance with the Guide for the Care and Use of Laboratory Animals (National Research Council, 2011). The J000102184 PDX model of TNBC was tested in female NSG mice, NOD.Cg-Prkdc^scid Il2rg^tm1Wjl/SzJ (NSGTM: JAX 005557). Mice were housed in individually ventilated polysulfone cages with HEPA filtered air at a density of up to 5 mice per cage. The animal room was lighted entirely with artificial fluorescent lighting, with a controlled 12-hour light/dark cycle (6 a.m. to 6 p.m. light). The normal temperature and relative humidity ranges in the animal rooms were 22°C to 26°C and 30% to 70%, respectively. The animal rooms were set to have up to 15 air exchanges per hour. Filtered tap water, acidified to a pH of 2.5 to 3.0, and standard rodent chow was provided ad libitum. J000102184 tumor cells were injected into 7-week-old NSGTM mice at a volume of 40 μl per mouse. Mice were enrolled when tumor volume reached between 100 to 200 mm³. Body weights, clinical observations, and digital caliper measurements were recorded two times weekly post enrollment. Animals that reached a body condition score of ≤2, a body weight loss of ≥20%, ulcerated tumors, or a tumor volume >2,000 mm3 were to be euthanized before study terminus.

On study day 0, dosing of mice began with PF-05175157. Group 1 was dosed with the vehicle comprised of 0.5% methycellulose in water. Group 2 was dosed with the 20 mg/kg of PF-05175157. Both groups were dosed twice daily by oral gavage for 4 weeks. On study day 43, tumor volumes were measured, and all animals were euthanized via CO₂ asphyxiation. At take-down, tumors from all mice were harvested for FFPE.

Mouse cell line–derived xenograft models

Ten million MDA-MB-468 (ATCC HTB132-) cells were implanted subcutaneously into the right flank of immune-deficient Rag2/IL2RG double knockout mice (Envigo) in the presence of Matrigel (Corning) at the Yale Center for Precision Cancer Modeling under an Institutional Review Board—approved animal protocol. All mice scratched informed consent. Twelve mice with palpable, similar sized tumors on postimplantation day 7 were randomized to two experimental groups: vehicle and PF-05175157 treatment (20 mg/kg). A 25-mg/mL stock of PF-05175157 was prepared by reconstitution in DMSO. The stock was diluted to 5 mg/mL in sterile PBS and administrated daily for five times a week by intraperitoneal injection in a 100 μL volume until the end of the experiment. Vehicle was prepared by diluting equal volume of DMSO in PBS. Tumor volumes were recorded by caliper measurements at 3-day intervals and tumor volumes calculated using the formula $\frac{1}{2}\times \mathrm{length}\times {\mathrm{width}}^2$.

Mouse tumor growth curve analysis

Both, cell line–derived xenograft (CDX) and PDX data were analyzed with mixed-effect models using lme4 R package (21). There were two fixed effects in the model: day of study and type of drug used (Vehicle or PF-05175157), and a random effect allowing for random intercept for each animal to include individual variability in tumor volume. The interaction term between two fixed effects was added to model the effects of drug on the rate of tumor growth.

RNA-seq data preparation and analysis

RNA-seq libraries were prepared from 1 μg of RNA using PolyA selection with oligo-dT beads, followed by random priming using the Illumina TruSeq Stranded Total RNA kit. Samples were sequenced with a target coverage of 50 million reads, paired-end, using the Illumina NovaSeq 6000 S4 platform. RNA-seq was performed at the Yale Center for Genome Analysis. Three biological replicates were created in each scenario (details about data preprocessing are provided in Supplementary Methods). Gene was called significant for FDR < 0.05 and absolute value of log₂ fold change higher than 1. Gene was called validated for P value < 0.05 in independent cohort and the same direction of expression change (details are provided in Supplementary Methods).

Metabolic data preparation and analysis

Cells in each well were quenched by 150 μL of ice-cold quenching buffer. Quenching buffer recipe was as follows: 20% MEOH into 0.1% formic acid, 3 mmol/L NaF, and 5.5 μg/mL D8-Phenylalanine (used as an internal standard). Then, the material was transferred to a LC/MS/MS V-bottom plate on dry ice, stored in −80°C freezer until liquid is completely frozen and lyophilized overnight. The lyophilized powder was reconstituted in 50 μL/well D4-taurine solution (another internal standard). Six biological replicates were created in each scenario. Each sample was run with two columns and two MS modes per column to get the best coverage of metabolites: reverse phase column with positive and negative MS mode; hypercarb column with positive and negative MS mode. The targeted group of metabolites was based on our in-house IROA libraries (~600 metabolites in total) and the untargeted curation was based on the Kyoto Encyclopedia of Genes and Genomes (KEGG) database (~2,700 metabolites). To call a metabolite significant, the same rules were applied as in the gene expression data analysis (details are provided in Supplementary Methods).

Data and source code availability

All data associated with this study are available in the paper or the Supplementary Materials, or publicly available. RNA-seq gene expression data for TCGA samples were downloaded from the RNAseqDB database (https://github.com/mskcc/RNAseqDB). Matched breast cancer and normal gene expression data sets were downloaded from ArrayExpress (https://www.ebi.ac.uk/arrayexpress/) under the accession numbers E-GEOD-70951 and E-GEOD-76250. Breast cancer cell lines RNA-seq gene expression data we downloaded from the CCLE (https://portals.broadinstitute.org/ccle, release of 8-NOV-2017). RNA-seq expression data and metabolite abundances data for BT474 and MDA-MB-468 cell lines have been deposited in the Sequence Read Archive under project accession number PRJNA757722.

Bioinformatic pipeline (LIDopener) to identify isozymes that show LID and rank them as potential therapeutic targets is implemented in MATLAB software (R2017b, Mathworks, Inc) and is freely available for download from https://github.com/mmarczyk363/LIDopener.

Results

Isozyme expression patterns in cancer and in corresponding normal tissues

We identified 2,331 unique human isozymes catalyzing 1,319 different enzymatic reactions in the ENZYME nomenclature database and obtained their mRNA expression data in 14 different cancer types and corresponding normal tissues from TCGA RNASeqDB database (Fig. 1A). The number of available paired tumor-normal samples ranged from 11 in esophageal cancer to 105 in breast cancer (Fig. 1B). The overall number of expressed isozymes ranged from 2,200 to 2,261 across cancer types, 93% of enzymatic reactions and 90% of individual isozymes were detected in all cancer types, consistent with their fundamental roles in cell biology (Fig. 1C and D; Supplementary Fig. S1A). Fifteen isozymes showed cancer type specific expression (Fig. 1C; Supplementary Table S1).

Figure 1.

Characterization of isozyme mRNA expression in 14 cancer types. A, Data sources and sample sizes for human enzymatic reactions (ENZYME database) and isozyme expression (TCGA data downloaded from RNAseqDB). B, Number of paired tumor/normal samples, enzymatic reactions, and isozymes included in our analysis in each cancer type. C, Venn diagram of shared and private enzymatic reactions between cancer types. D, Top (red heat map), median Pearson correlation coefficients of isozyme expression levels between patients within and across cancer types. Bottom (blue heat map), Jaccard index of differentially expressed isozymes between tumor and corresponding normal tissues (defined as absolute log-fold change >2 and FDR <0.05). E, T-Distributed stochastic neighbor embedding (t-SNE) plot based on expression levels of all isozymes (n = 2,331) in all tumor and normal samples. F, Proportion of differentially expressed isozymes between tumor and normal tissues. Each dot represents a patient; medians are indicated by the red vertical lines.

Within each cancer type, the sample-to-sample median correlation coefficients for isozyme expression ranged from 0.85 to 0.94, indicating some between-cancer variation among histologically similar cancers. Differences in isozyme expression patterns were smaller between cancer and matching normal tissues than between different cancer types (Fig. 1E), indicating that most tissue-specific metabolic features persist in cancer cells.

Kidney and lung cancers showed the greatest difference from their corresponding normal tissues (Fig. 1E). Kidney chromophobe tumors also formed a cluster distinct from the other two types of kidney cancers and lung adenocarcinomas also separated from lung squamous cell carcinomas that clustered more closely with head and neck squamous cell cancers in the isozyme expression space (Fig. 1E). These observations are consistent with distinct cell types of origin for these cancer types. When we examined differentially expressed isozymes between cancer and corresponding normal tissue, we found that most differentially expressed isozymes showed lower expression in cancer compared with normal tissues (Supplementary Figs. S1B and S2A). Analyzing tumor-normal pairs at patient level, we observed substantial variation in the proportion of differentially expressed isozymes from cancer to cancer in all cancer types (Fig. 1F).

LID in cancer

To identify isozymes affected by LID, we applied the following selection criteria (Fig. 2A): (i) the isozyme family includes 2 to 5 members that catalyze the same enzymatic reaction; (ii) one of the isozymes shows similar or elevated expression in cancer compared with normal (i.e., the cancer-dominant isoform) while the other isoforms (i.e., complementary isozymes) show lower expression in cancer compared with normal; and (iii) all isozyme members of the family show similar expression levels in normal tissue.

Figure 2.

LID in cancer and identification of potential therapeutic isozyme targets. A, Schematic representation of the strategy for finding an ideal therapeutic target isozyme showing LID. B, Venn diagram of private and shared enzymatic reactions showing LID across cancers. C, Left, total number of enzymatic reactions affected by LID in each cancer type. Right, Jaccard index of enzymes affected by LID across cancer types. D, Linear regression line of average selection scores and their SD across cancer types. Each point represents an enzymatic reaction; the color scale (yellow to red) indicates the number of cancers in which it shows LID. Gray shading around the regression line indicates the 95% confidence intervals. E, Over-representation analysis of LID enzymes in KEGG pathways. Color scale (yellow to red) reflects P values from the enrichment test (*, adjusted P < 0.01; **, adjusted P < 0.001). Number of pathways affected in different cancer types are shown on sidebars.

We identified 357 distinct enzymatic reactions that met our selection criteria and showed LID in at least 1 cancer type. Fifty-five (15%) of these were observed only in a single cancer type (Fig. 2B and C; Supplementary Table S2). The number of LID-affected enzymatic reactions ranged from 47 in esophageal to 151 in kidney renal clear cell cancers (Fig. 2C; Supplementary Fig. S2B). Across all cancer types, the most frequently LID-affected enzymatic reaction was EC 6.3.4.14/EC 6.4.1.2, catalyzed by ACACA, also known as ACC1, which was the cancer-dominant isoform in 12 cancer types (Supplementary Table S1).

To rank the cancer-dominant LID-affected isozymes as potential therapeutic targets, we next applied a target selection score that used five different features (M1 through M5; see Materials and Methods) of their expression pattern that were combined into a single score that ranged between 0 and 100. The 15 isozymes with the highest target selection scores in 8 or more cancer types are listed in Supplementary Table S1. Several isozymes showed consistently high selection scores across all cancer types (Fig. 2D) and three, including ACACA, l-dopachrome isomerase (MIF), and Beta-glucuronidase (GUSB), had uniformly high scores in more than 10 cancer types.

The 357 LID-affected metabolic reactions represent a broad range of biochemical pathways (Fig. 2E; Supplementary Fig. S2C and S2D). The top three KEGG pathways enriched for LID in 8 or more cancer types were all related to energy metabolism and biomolecule synthesis including pyruvate-, butanoate-, and propanoate metabolism.

Expression of candidate isozyme targets in cancer cell lines and in vitro cell viability dependency from genome-wide knockout screens

To assess the functional importance of the 357 cancer-dominant isozymes, we ranked their relevance for cell survival using GDs derived from in vitro whole-genome CRISPR and RNAi knockout datasets collated in DepMap and extracted isozyme expression levels for each cell line from the CCLE. To prioritize the isozymes for functional relevance, we applied the following three selection criteria: (i) target isozyme expression is greater than the expression of the complementary isozymes in CCLE cancer cell lines (i.e., mimics expression pattern observed in corresponding tissues); (ii) target knockdown impairs tumor cell survival, as reflected by a negative GD; and (iii) target isozyme GD is lower than that of the complementary isozyme dependency score, implying a greater impact on cell survival. In this analysis, we only included cancer types for which 10 or more cell lines were available in the CCLE and DepMap databases (Fig. 3A and B).

Figure 3.

Prioritization of LID-based target isozymes by functional importance. A, Number of LID-based target isozymes that satisfied the three criteria for the functional assessment by cancer type. Each row corresponds to different criteria. B, Selection score distributions of isozymes that passed all three functional validation criteria and those that failed. C, Association between number of cancer types in which isozymes showed LID (x-axis) and number of corresponding cancer cell line models in which the isozymes were functionally validated (y-axis). Color scale shows percent functional validation rate, and the size of dots corresponds to the number of reactions.

Approximately 50% of the LID-affected isozymes showed cancer-dominant expression in the corresponding cell lines, but many were considered nonessential for cell viability based on dependency scores. The number of isozymes that satisfied all three functional prioritization criteria ranged from 9 in esophageal cancer to 26 in kidney clear cell carcinoma, resulting in an average functional validation rate of 17% (range 10%–21%) across cancer types (Supplementary Table S2). Overall, 33 isozymes met our functional prioritization criteria in the cancer types in which they showed LID in the TCGA (Fig. 3C; Supplementary Table S2). The most broadly shared (i.e., showing LID in 9 cancer types) and functionally validated (in 7 cancer types in the CCLE/DepMap data) therapeutic target isozyme was ACACA.

Isozyme targets in breast cancer subtypes

The largest single cancer type in our data set was breast cancer. Because modern classification divides breast cancer into at least three clinically and molecularly distinct subtypes including hormone receptor positive/human epidermal growth factor receptor 2 negative (HR⁺/HER2⁻), HER2-positive (HER2⁺), and TNBCs, we applied our strategy to discover potential isozyme targets for each of the distinct breast cancer subtypes. We identified 127, 102, and 93 isozymes showing LID in HR⁺/HER2⁻, HER2⁺, and TNBC subtypes, respectively (Supplementary Fig. S3A; Supplementary Table S3). Forty-nine of the identified isozymes were cancer-dominant in all three subtypes, resulting in 178 unique potential metabolic targets (Fig. 4A). At the pathway level, carbohydrate metabolism pathways were the most enriched in therapeutic targets (Fig. 4B).

Figure 4.

Identification and prioritization of potential therapeutic isozyme targets in breast cancer subtypes. A, Venn diagrams of shared and private LID-based isozyme targets in breast cancer subtypes. JI, Jaccard index comparing similarity between subtypes. B, Overrepresentation analysis of LID enzymes in KEGG pathways. Color scale (yellow to red) reflects P values from the enrichment test (*, adjusted P < 0.01; **, adjusted P < 0.001). Number of pathways affected in different subtypes are shown on sidebars. C, Number of LID-based target isozymes that met four steps of the functional assessment by breast cancer subtype. D, Normalized log₂ expression of BDH1 isozyme validated in all breast cancer subtypes as target. Black lines show average values for each isozyme.

For breast cancer, in addition to the functional relevance criteria derived from the in vitro gene dependency analysis, and confirmed cancer-dominant expression in the CCLE, we introduced an additional gene expression–based validation step, which required confirmation of cancer-dominant expression of the target isozyme in two independent human breast cancer tissue mRNA expression datasets (E-GEOD-70951 and E-GEOD-76250; Supplementary Fig. S3B). Nine isozymes met all criteria for high priority metabolic target in HR⁺/HER2⁻ cancers, 4 for HER2⁺ cancers, and 11 for TNBC (Fig. 4C), corresponding to 17 unique isozymes (Supplementary Table S3). Only one of these isozymes, 3-hydroxybutyrate dehydrogenase 1 (BDH1), fulfilled all validation criteria in all three cancer subtypes (Fig. 4D). Fructose-bisphosphate aldolase, 3-hydroxyacyl-CoA dehydrogenase, Diamine N-acetyltransferase, Aconitate hydratase, and Choline kinase passed the four-step prioritization in at least two subtypes, ACACA met all criteria in TNBC only.

To explore potential mechanisms that could explain LID expression pattern of the 17 most promising isozyme targets in breast cancer, we examined somatic mutation, copy-number and methylation status of these genes in the TCGA breast cancer subset that was used for the LID analysis (n = 105 cases with matched normal and cancer tissue; Supplementary Fig. S4). Only 4 of the 17 genes had nonsynonymous potentially deleterious mutations affecting 3 patients, and 8 of the 21 complementary isozymes carried mutations in 4 patients. Gene amplifications affected 6 target isozymes, most frequently observed in ADAR in 12 patients followed by ALDOA and CHKA in 7 patients each. Deletions affected 8 complementary isozymes but were rare, affecting only 1 or 2 patients each. When we compared isozyme gene methylation status on a continuous scale (Supplementary Table S3), 6 complementary isozymes (ACACB, CHKB, PHGDH, ADARB1, POMT2, ELAC1) showed statistically significant hypermethylation compared with the corresponding target isozymes in all breast cancer subtypes. These findings suggest that methylation mediated gene silencing, rather than mutations or copy-number alterations, are the most common cause of LID in breast cancer.

Effect of small molecule inhibitor of ACACA/ACACB on breast cancer viability in vitro and in vivo

We selected ACACA for functional validation because of its direct translational relevance. ACACA is a biotin-dependent enzyme that catalyzes the carboxylation of acetyl-CoA to produce malonyl-CoA, which is the rate-limiting step in de novo long-chain fatty acid synthesis. Pfizer has developed a drug, PF-05175157, that inhibits ACACA and ACACB with in vitro IC₅₀ concentrations of 27 nmol/L and 33 nmol/L, respectively (22). This drug has been tested in Phase I/II clinical trials in healthy volunteers and patients with type 2 diabetes (NCT01819922, NCT01792635, NCT01396161). These trials demonstrated significant inhibitory effect on lipid synthesis in patients and showed good pharmacokinetic properties with plasma C_max between 20 to 40 μg/mL (23, 24).

We exposed breast cancer cells and normal breast epithelial (HMEC) cells to 0.1, 1, 10, and 25 μg/mL of PF-05175157 (courtesy of Pfizer) or corresponding dilutions of DMSO for 24, 48, and 72 hours in vitro and observed statistically significant time- and dose-dependent inhibition of cell viability (Supplementary Fig. S5; Supplementary Tables S4 and S5). A total of 15 breast cancer cell lines were tested (Supplementary Tables S4) and PF-05175157 at 10 μg/mL concentration significantly inhibited cell viability in all but 1 of the cell lines (Fig. 5A; Supplementary Table S5). Log-logistic model fitted to dose–response curves gave different slopes, EC₅₀, and minimum viability for each cell line, but all measures of the inhibitory effect were highly significant. The estimated EC₅₀ ranged from 0.96 μg/mL in T47D to 66.9 μg/mL in BT549 cell line (Fig. 5B). The most resistant cells were HMEC with EC₅₀ of 117 μg/mL. Next, we examined cell-cycle progression and apoptosis at 24 and 72 hours after exposure to PF-05175157 on the highly sensitive MDA-MB-468 and BT474 cells. There was a time dependent increase in cells at G₂–M phase (Fig. 5C; Supplementary Table S6) followed by significant increase in apoptosis in both cell lines at 72 hours (Fig. 5D). The effect of ACACA inhibition may be rescued by lipid uptake from the surrounding tissues therefore we also assessed the efficacy of PF-05175157 in MDA-MB-468 tumor xenografts (CDX; n = 12) and a triple-negative PDX (J000102184, PDX; n = 10) model (Fig. 5E; Supplementary Fig. S6; Supplementary Table S7). In both models, PF-05175157 suppressed tumor growth compared with control (Fig. 5E).

Figure 5.

Inhibition of breast cancer cell viability and tumor growth by PF-05175157. A, Dose–response curve with 72 hours of exposure in 16 cell lines. Red dotted lines, no effect of treatment. Black frame highlights the normal epithelial cells. B, Association between predicted maximum cell viability inhibition and log EC₅₀. Black frame highlights normal epithelial cell line. C, Proportion of cells in different cell-cycle phases measured in 2 cell lines (rows) and 2 time points (columns). D, Proportion of apoptotic cells measured in 2 cell lines (rows) and 2 time points (columns). Significant differences between different doses, *, P < 0.05. E, Tumor growth curves in mouse MDA-MB-468 xenograft and patient-derived TNBC xenografts with and without (Vehicle) drug treatment. Lines indicate prediction of mixed-effect model. Error bars show 95% confidence intervals around mean values of the data.

Transcriptomic and metabolic profiling after ACACA/ACACB inhibition in breast cancer cell lines

We next examined transcriptomic and metabolomic changes after 6 and 24 hours of exposure to 10 μg/mL PF-05175157 in BT474 cells and after 24 hours exposure in MDA-MB-468 cells. In BT474 cells, at 6 hours, 148 genes were upregulated and 210 downregulated, and at 24 hours transcriptomic changes increased to 534 upregulated and 499 downregulated genes (Fig. 6A; Supplementary Table S8). The gene expression changes were highly correlated between the two time points (r_S = 0.66, Supplementary Fig. S7A). Two hundred thirteen significantly upregulated and 251 downregulated genes at the 24-hour time point in the BT474 cells were also significantly upregulated in MDA-MB-468 cells, indicating a shared transcriptional response (Fig. 6B). At pathway level, expression of many genes involved in metabolic processes was increased, suggesting a large-scale metabolic compensatory response. Amino acid synthesis, glycolysis, lysosomal degradation, mitochondrial respiration, and pentose phosphate pathways showed a significant rapid upregulation at 6 hours (Fig. 6C). We also observed a significant decrease in DNA repair and cell-cycle–related genes by 24 hours in the MDA-MB-468 cells. A closer look at lipid metabolism showed significant upregulation of enzymes involved in fatty acid release from lipid stores (ATGL, DAGL, MAGL), fatty acid uptake from the extracellular space (LDLR) and citrate availability (IDH, SLC25A1; Fig. 6D). In contrast, an important regulator of fatty acid beta oxidation CPT1 decreased. These transcriptional changes suggest activation of processes that could provide alternate sources of fatty acids when de novo synthesis is inhibited (Fig 6D). Metabolic profiling also showed extensive changes in metabolite levels across many metabolic pathways that tended to increase from 6 to 24 hours (Fig. 7A; Supplementary Table S9). Metabolite changes correlated over time, indicating a consistent and sustained metabolic response (r_S = 0.71; Supplementary Fig. S7B). Two hundred seventy-eight significant metabolites upregulated in BT474 cells, and 66 downregulated metabolites showed the same changes in MDA-MB-468 cells at the corresponding time points (Fig. 7B). At the metabolite abundance level, alterations in ether lipid, glycerolipid, and glycerophospholipid metabolism were significant in both time points in BT474 cells, and similar changes were seen in MDA-MB-468 cells (Fig. 7C).

Figure 6.

In vitro transcriptomic changes induced by PF-05175157. A, Volcano plots of differentially expressed genes at different time points in BT474 cells. Red dots, significantly affected genes (FDR < 0.05 and log₂FC > 1). In the lower corners of each plot are the number of genes up- or downregulated after treatment. B, Validation of significantly altered genes at 24 hours in BT474 cells using data from MDA-MB-468 cells treated with PF-05175157 for 24 hours. Red dots, validated genes (i.e., P < 0.05 in MDA-MB-468 cells and the same direction of change in both cell lines); blue dots, not validated genes. C, Gene set enrichment analysis using the NanoString metabolic pathways. Red, higher expression of gene set after treatment; blue, higher expression in control (*, adjusted P < 0.05; **, adjusted P < 0.01). D, mRNA expression changes in key fatty acid metabolism enzymes mapped onto a schema of fatty acid metabolism (modified after Montesdeoca and colleagues; ref. 27).

Figure 7.

In vitro metabolomic changes induced by PF-05175157. A, Volcano plots show increased and decreased metabolites after PF-05175157 treatment of BT474 cells. Red dots, significantly affected metabolites (FDR < 0.05 and log₂FC > 1). In the lower corners of each plot are the number of metabolites that increased or decreased after treatment. Metabolites are coded by KEGG compound ID (https://www.genome.jp/kegg/compound/) and are listed in Supplementary Table S9. B, Validation of metabolites significantly altered in BT474 cells using MDA-MB-468 cells. Red dots, validated metabolites (P < 0.05 in MDA-MB-468 and the same direction of expression change in both cell lines); blue dots, not validated metabolites. C, Enrichment analysis of metabolites on KEGG metabolic pathways. Red, enrichment of a metabolic pathway after treatments (*, adjusted P < 0.05; **, adjusted P < 0.01).

Discussion

In this study, we demonstrated that while most metabolic processes remain preserved in tissues after malignant transformation, about 5% to 30% of isozymes are differentially expressed in cancer compared with corresponding normal tissues. This provides rationale to pursue metabolic enzymes as potential therapeutic targets (4, 25). However, how to prioritize promising candidates from the relatively large pool of metabolic enzyme targets remains a challenge because most enzymatic reactions are catalyzed by multiple distinct isozymes that might substitute for one another, not all enzymatic steps are rate-limiting in a metabolic process, and not all metabolic processes are equally critical for cell survival. In this analysis, we hypothesized that the “ideal” target isozyme expression profile includes high expression of the putative target and low expression of its complementary isozymes in cancer, while all isoforms are expressed at similar levels in corresponding normal tissues. We call this expression pattern LID. We found 357 distinct enzymatic reactions, representing a broad range of metabolic processes, which showed LID in at least 1 of the 14 cancer types that we examined. There was no single reaction that was universally affected by LID in all cancer types, the most frequently affected pathways involved carbohydrate, energy and fatty acid metabolism. We applied a therapeutic target score to further rank the 357 reactions and corresponding cancer-dominant isozymes as potential therapeutic targets based on their expression patterns in cancer and corresponding normal tissues and provide the top ranked isozyme targets for all 14 cancer types (Supplementary Table S2).

While cancer type specific isozyme expression patterns exist, we also observed substantial patient to patient variation in what isozymes showed LID within any given cancer type, suggesting flexibility in the metabolic adaptation processes. This variability may also reflect clonal heterogeneity; different neoplastic cell subpopulations may have distinct metabolic needs that can also evolve and change over the course of the disease (26).

We used GDs from the DepMap project to assess the in vitro functional importance and therapeutic target potential of cancer-dominant isozymes. Overall, about 17% of the isozymes showing LID in human cancers demonstrated importance in cancer cell survival in the in vitro CRISPR and RNAi screens. We identified 33 isozymes with in vitro functional validation in the DepMap data that we consider highly promising metabolic targets in cancer. The number of candidate therapeutic targets ranged from 9 in esophageal cancer to 26 in kidney clear cell carcinoma. Notably, most of these isozymes have not previously been implicated in cancer biology and represent novel targets. Using breast cancer as an example, we also showed that subtype-specific metabolic differences exist within a given cancer type. We identified 17 LID-based targets in breast cancer that differed by subtype with only one potential target, BDH1 that is involved in fatty acid catabolism, shared by all three subtypes. ACC1/ACACA had the broadest cancer-dominant expression pattern across cancer types and met all other functional prioritization criteria. Increased de novo fatty acid synthesis is a hallmark of cancer cell metabolism, and ACACA catalyzes the rate-limiting initial step in this process (27). Several other investigators have also identified this enzyme as a promising therapeutic target. Treatment of xenograft and genetically engineered mouse models of non–small cell lung cancer with an allosteric inhibitor of ACACA/ACC1 and ACACB/ACC2 inhibited tumor growth in vitro and in vivo (28). In breast cancer cell lines, silencing of ACC1 with shRNA or inhibition with a nonspecific small molecule inhibitor resulted in cell death (29, 30). Similar results were seen in prostate cancer (31) and glioblastoma multiforme cell lines as well, where a dual ACACA/ACACB inhibitor or siRNA knockdown of these genes caused increased cell death (32). In head and neck cancer, cetuximab plus TOFA [5-(tetradecyloxy)-2-furancarboxylic acid], an allosteric inhibitor of ACC, also showed significant growth inhibition in cetuximab-resistant tumor xenografts (33).

Our target discovery strategy is based on the premise of developing isoform-specific enzyme inhibitors. This remains a medicinal chemistry challenge and we recognize that isoform specificity is relative and is often dose dependent. However, there is a growing number of success stories including the development of clinically useful isoform-specific phosphoinositide 3-kinase and histone deacetylase inhibitors.

Elevated fatty acid synthesis and impaired fatty acid oxidation are also hallmarks of obesity, type 2 diabetes, and nonalcoholic fatty liver disease. Pfizer Inc has developed a small molecule dual ACACA/ACACB inhibitor, PF-05175157, that has completed Phase I clinical testing in healthy volunteers and Phase II studies in patients with type 2 diabetes (12–14). Because of the direct clinical translational relevance, we tested the anticancer activity of single agent PF-05175157 in breast cancer cell lines and mouse xenograft and PDX model. We observed a broad single agent growth inhibitory effect in vitro at concentrations that are achievable in human plasma. The drug also slowed tumor growth in xenograft and PDX models in vivo. Transcriptional and metabolic profiling of BT474 and MDA-MB-468 cells showed large scale metabolic and transcriptomic response by 24 hours. The transcriptomic changes suggested activation of compensatory mechanisms including fatty acid release from lipid stores, increased fatty acid uptake from the extracellular space, increased citrate production and upregulation of fatty acid synthase. However, despite this initial metabolic activity, there was substantial and significant increase in apoptosis by 72 hours in all 15 cancer cell lines. The in vivo growth inhibitory effects were also significant, but PF-05175157 did not completely abolish tumor growth, suggesting that for maximal growth inhibition combination therapies may be required. An unexpected observation in the transcriptomic data was decreased expression of many DNA repair enzymes after ACC inhibition that raises the possibility of therapeutic synergy with DNA-damaging agents.

Finally, to enable other investigators to explore LID in cancer, we generated a publicly available tool, LID-opener, that can be applied to any data in which metabolic isozyme expression differences between biological conditions are examined (e.g., different tumor subtypes within the same cancer, primary versus metastatic lesions or pre- and post-therapy samples) (https://github.com/mmarczyk363/LIDopener). Our approach can also be scaled to study metabolic adaptation at the single cell level.

In summary, we performed a systematic analysis of isozyme expression patterns across 14 cancer types and identified several hundred cancer-dominant isozymes. We further prioritized the cancer-dominant isozymes as potential therapeutic targets with in vitro functional importance and experimentally validated ACC1/ACACA as a cancer drug target using a clinically relevant small molecule inhibitor. We believe that this work elucidates an important but overlooked aspect of metabolic plasticity inherent in cancer biology and could lead to further studies to fully characterize the putative metabolic vulnerabilities suggested by our analysis.

Significance:

This study exploits the loss of metabolic isozyme diversity common in cancer and reveals a rich pool of potential therapeutic targets that will allow the repurposing of existing inhibitors for anticancer therapy.