3 ms·

You are right. I did not mean that sequencing shouldn't be done, what I meant was that the simple idea of sequence -> nearest match -> most appropriate treatmen

by Dn_Ab 15y ago

You are right. I did not mean that sequencing shouldn't be done, what I meant was that the simple idea of sequence -> nearest match -> most appropriate treatment won't work due to the massive amount of variance involved. So the sequencing will be more useful at higher level. So more understanding patterns and devising strategies than personalizing cocktails.

epistasis 15y ago
Ah, I see what you're saying. Though even going with the best treatment in the nearest match is a huge leap forward over what we can currently do. I see two ways forward, using the current class of chemotherapeutics + small molecule/antibody targetted therapies. As straw men, they are: 1) come up with a totally complete model of what everything in the cell does, how the mutations affect those interactions amongst gene products, and change the cell, and then designing some treatment to selectively kill that tumor population and to also block off obvious avenues to evolve resistance. 2) the "dumb" approach, figure out classes of disease, and try to use induction from the results of previous patients in that class of disease to figure out the best treatment for the current patient. Right now both avenues are being pursued as best as possible. The first straw man is far from being approachable; we know so little about molecular biology, and when physics can't even figure out how to fold a protein, how are we going to be able to figure out changes to protein-protein binding dynamics, etc? Not that it can't be done eventually, but it's a very tough road. The second approach is probably the best we can do now, and corresponds very well to the type of machine learning that is being researched these days. Even without full understanding of the biological system and the effects of all possible mutations, we can combine what we currently know with machine learning to construct a black-box for the parts of molecular biology that we don't yet know. I think that this is probably the most reasonable path forward for the time being, though we clearly need some intelligent way to guide studies into the combinatorial effect of targeted therapies.
- Dn_Ab 15y ago
  I think Moore's law can help to an extent with approach 1, especially the protein folding bottleneck. We also need better computational techniques, taking more from dynamical systems maths for example. I think approach 2 can also inform approach 1. Especially if grey and clearbox models are used - probabilistic graphical models, markov logic networks even genetic programming. So instead of using them for just prediction or a distance metric, we study the patterns in the models they produce. For example, instead of merely fitting a bunch of weights in some sum product model, one could use genetic programming where the primitives represented actual computations from a simplified model of cell dynamics. So while not guidance it still offers some way of embedding some intelligence in the search.

by Dn_Ab 15y ago

epistasis 15y ago

Ah, I see what you're saying. Though even going with the best treatment in the nearest match is a huge leap forward over what we can currently do. I see two ways forward, using the current class of chemotherapeutics + small molecule/antibody targetted therapies. As straw men, they are: 1) come up with a totally complete model of what everything in the cell does, how the mutations affect those interactions amongst gene products, and change the cell, and then designing some treatment to selectively kill that tumor population and to also block off obvious avenues to evolve resistance. 2) the "dumb" approach, figure out classes of disease, and try to use induction from the results of previous patients in that class of disease to figure out the best treatment for the current patient. Right now both avenues are being pursued as best as possible. The first straw man is far from being approachable; we know so little about molecular biology, and when physics can't even figure out how to fold a protein, how are we going to be able to figure out changes to protein-protein binding dynamics, etc? Not that it can't be done eventually, but it's a very tough road. The second approach is probably the best we can do now, and corresponds very well to the type of machine learning that is being researched these days. Even without full understanding of the biological system and the effects of all possible mutations, we can combine what we currently know with machine learning to construct a black-box for the parts of molecular biology that we don't yet know. I think that this is probably the most reasonable path forward for the time being, though we clearly need some intelligent way to guide studies into the combinatorial effect of targeted therapies.

Dn_Ab 15y ago
I think Moore's law can help to an extent with approach 1, especially the protein folding bottleneck. We also need better computational techniques, taking more from dynamical systems maths for example. I think approach 2 can also inform approach 1. Especially if grey and clearbox models are used - probabilistic graphical models, markov logic networks even genetic programming. So instead of using them for just prediction or a distance metric, we study the patterns in the models they produce. For example, instead of merely fitting a bunch of weights in some sum product model, one could use genetic programming where the primitives represented actual computations from a simplified model of cell dynamics. So while not guidance it still offers some way of embedding some intelligence in the search.