Networks are pretty amazing. If you have a bunch of nodes that sit there and do very simple tasks that you can train a dog to do, and then you connect them in the right way, you can create a programme that can get through a level of Mario! In neural networks, all the nodes are doing is getting an input, then deciding what works best based on that input, and sending that information somewhere else. It’s like teaching your dog to get help when you’re in trouble. You say “Help!”, it makes a decision where to go to find help, and then someone else is alerted to your situation. Now you link those inputs and outputs, and suddenly you can solve some very interesting problems. Like making your dog get another dog to do something, who tells another dog… okay… at some point the analogy may break down. What I want to get across, is that networks can take very simple things and create complexity by just adding a few connections between them.
Getting back to the biological theme, proteins aren’t really all that simple. From what you’ll have read on other posts in this blog there’s a lot of effort that goes into modelling them. The position of even a small peptide loop can have a big effect on how a protein functions. So if you have a network of 15 – 25 thousand proteins you can imagine that the complexity of that system is quite incredible. That system is the human body.
Looking at systems from the perspective of interactions between components creating complex behaviour, it’s really not that surprising that human diseases often function by disrupting the network of protein interactions as found by this paper recently published in Cell. Assuming there’s a human disease which has survived for generations and generations, it is likely to be quite effective in how it functions. And if you have a complex system like the human body, the easiest way to disrupt that arising behaviour called “a healthy life”, is to alter a few of the interactions. Showing this for a large group of disease-associated missense mutations is however pretty impressive and time-consuming, which is probably why there are about 50 people from over 20 different institutions in the author list.
So what exactly did they show? They showed what happens when there is a mutation in a gene that causes a disease. Such a mutation replaces an amino acid in the protein sequence encoded by the gene. The resulting protein is then somehow different. The group around Marc Vidal and Susan Lindquist showed that rather than affecting the stability of the protein, the system tends to be targeted via the interactions of the protein. What is more, they showed that different mutations of the same gene can affect different sets of interactions. Using their “edgotyping” approach it is possible to characterize the type of effect a mutation will have and possibly predict the change in phenotype (i.e. the disease state).
Now if the possibility of predicting how a single mutation (1 in ~300 amino acids), in a single protein (1 in ~ 25,000 proteins) affects the human body doesn’t convince you that networks are pretty powerful, I really don’t know what will. But now back to the important stuff…. how do you make a real life neural network with communicating dogs? And could you make them reach level two on Mario? Or… Super Mario Dog?