Identification of anticancer drugs for hepatocellular carcinoma through personalized genome-scale metabolic modeling

Abstract

Genome-scale metabolic models (GEMs) have proven useful as scaffolds for the integration of omics data for understanding the genotype-phenotype relationship in a mechanistic manner. Here, we evaluated the presence/absence of proteins encoded by 15,841 genes in 27 hepatocellular carcinoma (HCC) patients using immunohistochemistry. We used this information to reconstruct personalized GEMs for six HCC patients based on the proteomics data, HMR 2.0, and a task-driven model reconstruction algorithm (tINIT). The personalized GEMs were employed to identify anticancer drugs using the concept of antimetabolites; i.e., drugs that are structural analogs to metabolites. The toxicity of each antimetabolite was predicted by assessing the in silico functionality of 83 healthy cell type-specific GEMs, which were also reconstructed with the tINIT algorithm. We predicted 101 antimetabolites that could be effective in preventing tumor growth in all HCC patients, and 46 antimetabolites which were specific to individual patients. Twenty-two of the 101 predicted antimetabolites have already been used in different cancer treatment strategies, while the remaining antimetabolites represent new potential drugs. Finally, one of the identified targets was validated experimentally, and it was confirmed to attenuate growth of the HepG2 cell line.

Figure 1. General pipeline for the identification of antimetabolites

1. The presence/absence of 15,841 proteins in tumors obtained from 27 HCC patients was evaluated using immunohistochemistry. tINIT algorithm was developed and used for reconstruction of personalized GEMs for six HCC patients and GEMs for 83 healthy cell types based on proteomics data and HMR 2.0. A method identifying potential antimetabolites for the treatment of the HCC patients was developed, and the toxicity of each antimetabolite was predicted using GEMs for healthy cells types.

2. tINIT was used for reconstructing GEMs which are in agreement with omics data and which could perform a set of predefined tasks. In this illustration, the model should perform two simple tasks; production of “D” from “A” and “E” from “B.” The resulting model should contain as many of the green reactions and as few of the red ones as possible. In the first step, all reactions were identified which, if removed from the network, cause any of the tasks to fail. These reactions were marked blue. In the second step, the INIT algorithm was used to find the network with the maximal number of green reactions compared to red, with the additional constraints that the model must be functional and that all blue reactions must be included. This would result in the dotted reactions being removed from the network. At this stage, the first task would be possible, but not the second one (since uptake of “C” makes the production of “E” possible without using any red reactions). In the final step, each task was tested and a gap‐filling algorithm was used to reinsert the reactions which were required for all tasks to work. This would result in the inclusion of the lower‐most red reaction.

3. The effect of antimetabolites can be predicted in silico by using metabolic network and potential use of antimetabolites is illustrated.

Figure 2. Proteomics data for 27 HCC patients

1. Clustering of the generated proteomics data between 27 different HCC patients showed notable differences. The color indicates the protein expression differences between tissue samples.

2. Due to the coverage of the proteomics data, we focused on the reconstruction of the personalized models for six HCC patients. The number of the evaluated proteins in each HCC patients varies between 9,312 and 14,561.

3. The expression level of 4,936 proteins measured in all six HCC patients and the proteomics data showed notable differences between the HCC patients and hepatocytes.

Figure 3. Comparison of the personalized GEMs for HCC patients

1. A, B

   The pairwise differences and similarities of the reactions (A) and genes (B) between personalized HCC models and the generic HCC model that is reconstructed based on the average protein expression level of 27 HCC patients.

Figure 4. Prediction of antimetabolites for HCC patients

1. 147 antimetabolites are predicted as potential anticancer drugs through personalized HCC models and 101 of these antimetabolites are effective for inhibiting HCC tumor growth in all six HCC patients. Antimetabolites are also predicted through the use of a generic HCC model that is reconstructed based on the average protein expression data in HCC patients, and 127 potential antimetabolites are identified. Twenty‐six of the antimetabolites predicted based on the generic HCC models are not effective in all six HCC patients.

2. Distribution of the antimetabolites that are predicted to be effective in number of the personalized HCC models.

3. 46 of the antimetabolites identified through the use of personalized models cannot be used for inhibition of the HCC tumor in all six HCC patients. The differences between the 46 predicted antimetabolites are shown through the use of personalized and generic HCC models.

Figure 5. Evidence levels of the predicted antimetabolites for HCC patients

Antimetabolites are a type of drugs, which acts by inhibiting the use of a normal metabolite, most often by being structurally similar to the metabolite in question. The 101 predicted antimetabolites were categorized based on their known use as antimetabolites and/or their connection to HCC (see Supplementary Dataset S11 for literature evidence). In summary, 22 of the predicted antimetabolites are currently in use as anticancer drugs, and 9 are used as drugs against other diseases. For 39 of them, the corresponding enzymes are currently targets for drugs but not with antimetabolites. The remaining 31 have not been investigated as drugs or drug targets, but all of them show a strong correlation with cancer progression. These results speak strongly for the validity of the in silico predictions.

Figure 6. Usage of l‐carnitine antimetabolites for the treatment of HCC

l‐Carnitine and metabolites in the l‐carnitine biosynthetic pathway, as well as two essential amino acids, lysine and methionine, necessary for the synthesis of l‐carnitine were identified through our modeling approach. The analogs of l‐carnitine were proposed as antimetabolites for the treatment of HCC patients and the predicted consequence of the use of an l‐carnitine antimetabolite is presented. l‐Carnitine antimetabolites may result in reduced β‐oxidation, de novo synthesis of fatty acids, and eventually may suppress or kill the growth of the HCC tumor. The abbreviations and the detailed explanations for the metabolites as well as the associated genes for each reaction are presented in HMR 2.0.

Figure 7. Inhibitory effect of perhexiline on the proliferation of the HepG2 cell line

1. A, B

   Perhexiline was used to mimic the effect of the l‐carnitine analog on the proliferation of the HepG2 cell line. The number of viable cells was determined after treatment with (A) perhexiline (2, 4, 8, and 20 μM) and (B) sorafenib (2 and 4 μM) for 24 and 48 h. Both perhexiline and sorafenib were dissolved in DMSO, and for each concentration of compounds analyzed, controls with corresponding concentration of DMSO were analyzed. Each bar represents the results from eight replicate samples, and mean ± s.d. values are presented. Student's t‐test versus untreated cells: *P‐values < 0.001.

2. C

   Example images for the HepG2 cell line after 24 h of the treatment with 20 μM perhexiline and corresponding concentration of DMSO.