What happens when you change more than one thing in a cell? Will discuss models from experiments (used cancer drugs in combinations to derive a model of the breast cancer cell survival phenotype), and models from patients (much data from human tumors and so you can generate a human model, and then test predictions in cells), and the biology of combinatorial perturbations.
A defining feature of cancer cells is a breakdown in regulation by aquired GOF (gain of function) or LOF (loss of function) mutations. One of the advances in cancer pharmacology are compounds that target the function of the proteins in these pathways. Can be receptor antibodies or tyrosine kinase inhibitors. Even though cancers can often be a single phenotype, the underlying geneotypes may be complex and plural.
Approaches to CP (combinatorial perturbations) include metabolic flux models, and what we need is an equivalent for the cancer cells. View the regulation as a system with inputs and outputs, and the drugs as inputs,and the phenotypic and molecular response as the output. Search for modesl that are suitable. The model designed by them concerns MCF7 breast cancer cells. Each perturbation pattern consists of 1, 2 or 3 perturbations in sequence. This pattern is converted into a phosphoprotein and phenotype pattern. They use a linear differentiation equation model, buffered with sigmoid transfer functions. Start with good tradeoffs between model fit and simplicity. They have an error term that measures fit and another variable that measures simplicity. They use a MC simulation, which results in the probabilities of functional interaction.
The top-scoring model can be viewed as a graph, and consists of: mapk cascade with a negative feedback loop, and more (in essense, it looks pretty good). This model predicts the effect of unseen drug pairs by leaving them out of the training set and seeing how the model performs.
So, what about models from patients? Three different types of data can be divided into genotype and phenotype measurements. They adapted CoPIA to cancer genomics data, and the main issue was speed. End up with matrices of various types, including ones of regulatory dependencies. Their method is derived from the S-System (Savageau 1969 J Theor Biol), and they developed 3 different ways to find S. They tried CoPIA-like (best, but only effective up to 20 genes), Bolasso and others that weren’t as suitable (Allyson: not sure if I got this section right).
The glioblastoma CoPIA model makes a number of testable predictions. Is NDN a previously-overlooked glioblastoma tumor suppressor? They did a simple wet-lab experiment to check. The growth (inhibition of that growth upon expression that was dose-dependent) and response of the target genes were measured. Part of the summary: they propose to view cancer tumors as CP experiments for in vivo reverse engineering.
Please note that this post is merely my notes on the presentation. They are not guaranteed to be correct, and unless explicitly stated are not my opinions. They do not reflect the opinions of my employers. Any errors you can happily assume to be mine and no-one else’s. I’m happy to correct any errors you may spot – just let me know!