Deciphering and targeting cancer metabolism using executable models

The progression of cancer is driven by mutations and errors in the cell that promote the "hallmarks of cancer". These are a set of distinct behaviours that cells in tumours possess, enabling the survival and growth of the cancer. A longstanding observation in cancer cells is the Warburg effect; the switch away from oxidative phosphorylation to anaerobic glycolysis. Recent evidence has shown that metabolic reprogramming- large scale modification of how the cells process energy- is achieved by mutations in oncogenes including KRAS. This enables cancer cell survival in the tumour. Furthermore mutations in individual metabolic enzymes, such as fumarate hydratase, can affect a transition from epithelial to a mesenchymal morphology. Both the links between individual metabolites and cell behaviour, and the role of the metabolic network in cancer development however remains unclear. Knowledge of both of these are necessary to interpret metabolic changes in cancer and to identify new drug targets that are robust to network effects.

Computational modelling offers the opportunity to formalise the relationships between elements in the network and to address this issue. However, conventional approaches based on ordinary differential equations (ODEs) and flux balance analysis (FBA) are not suitable. ODE models require precise physico-chemical parameters that are not available for human metabolism. FBA does not have this requirement but is not capable of modelling the accumulation of metabolites that can occur in response to mutation.

I propose to use an "executable" modelling approach to construct models of the metabolic network and link them to cellular behaviour. Executable modelling approaches do not require detailed physical parameters, and can model accumulation events. They have the further advantage that they are amenable to a class of model analysis known as formal verification. Here mathematical proofs are used to offer guarantees of cell properties that are encoded in formal logic. These guarantees can apply over all possible states of the system and so offer a uniquely powerful way to test model correctness. This is particularly useful for highly robust systems such as the metabolic network, but also in taking account of rare events. This latter issue is particularly important when considering how rare events determine the development of cancer. Model building will be guided by the use of machine learning approaches (t-SNE, decision forests) to address publicly available expression data and metabolomic datasets shared with us by AstraZeneca. Finally, these models will be used to make novel predictions about the metabolism can change cell phenotype. The development of these models will allow us to understand how metabolic networks drive cancer progression and help us identify novel oncogenes and drug targets.

The progression of cancer is associated with extensive metabolic reprogramming. A key feature of cancer is the Warburg effect, where cells switch from relying on oxidative phosphorylation to anaerobic glycolysis. Similarly, KRAS and BRAF mutations drive widescale reprogramming of metabolism, whilst the mutation of individual metabolic enzymes such as fumarate hydratase can promote cancer progression.

Computational models of cancer metabolism could elucidate the links between metabolism and cellular phenotype. Here we propose the construction of an executable model of cancer metabolism, guided by the use of machine learning applied to high throughput expression and metabolomic datasets and the existing literature on cancer metabolism. Executable models do not require the detailed physical parameterisation necessary for continuous models, such as those based on ordinary differential equations. They also allow more complex model behaviours to be described than flux balance analysis, including metabolite accumulation. Finally, they are amenable to analyses based on formal verification, where mathematical proofs are used to offer powerful guarantees of model behaviour by effectively searching all possible states of the model. The models and the new predictions that arise from the model construction will be validated through experiments performed with the Frezza group.

This project will lead to demonstrable contributions within academia, industrial research, and wider society. Complex models of metabolism will be constructed using existing publicly available and privately shared high throughput datasets analysed using machine learning approaches, alongside the literature on cancer metabolism. These models will be tested using formal verification based analyses. The work will give rise to valuable insights into how metabolism can determine the cancer phenotype, and will be published in high impact papers and through conferences.

Within the academic community, the interdisciplinary nature of the work will support biologists working in metabolism, and computer scientists working on formal verification approaches applied to large models. The models that are generated will encompass large numbers of observations that have been made and as such can be used to test and validate experimental approaches before they are taken into the laboratory. Similarly, model building in the life sciences has generated new problems for existing approaches and algorithms, requiring the development of new techniques for analysis. As such, the construction of new, complex models will support algorithm developments, at a minimum by extending the test sets with large, realistic models, or by requiring new research.

Through working with high throughput datasets, supplied in part by AstraZeneca, we will be able to apply our models to relevant data on the metabolic behaviours of different cell lines. This will give new insights into a large body of experimental data covering multiple cell lines under constant conditions. This in turn will be used to inform drug development within the company, and will lead to future collaborations based around the research.

In addition to impact from the direct results of the project, the work will have additional impacts important to the wider UK. One post-doctoral research associate will be trained over the course of the fellowship in the building and analysis of executable models of biological processes, and the development of formal specifications describing biological processes. Current models of metabolism have been used in outreach activities (including demonstrations for school parties) and will be used to demonstrate to the wider public how computational methods aid scientific understanding of cancer and wider biology. This includes the development of worksheets for A-level students currently under discussion with the OCR.