4 ms·

Protein folding seems like the kind of problem that people were better at than computers 10 years ago, before deep learning

by tamana 10y ago

dnautics 10y ago
exactly which deep learning technique do you propose unleashing on protein folding? Deep learning is not 'magic', it is a set of tools good at a certain subset of machine learning tasks, and the impressive thing is how many machine learning tasks can be shoehorned and contorted into the deep learning rubric, i.e. 'fuzzy classification via scoring'. It should be obvious how image recognition and voice recognition fit this rubric. "Playing Go", which does not fit this rubric obviously, I think, is conquered by deep learning because there is a discrete matrix of 'moves' which looks awfully lot like an image recognition matrix, and so a move which optimizes score ('classification') can be selected. Fundamentally one of the problems is that we don't know how to score proteins. We also don't necessarily know how to score go, but it's trivial to generate solutions by running hypothetical matches because the game outcome is deterministic, solvable, and definite. While there is a decent subset of proteins whose known structures could be used to seed a database for scoring, it is very small in comparison to the universe of possibilities (20^n, where n is the length of the amino acid chain). In short: Given an arbitrary "end-board layout" I can with our existing human knowledgebase tell you which go player has won. Given an arbitrary amino acid sequence and a proposed structure, I cannot with our existing human knowledgebase tell you if the structure is correct. Because of this limitation in creating a scoring system, while I won't say that deep learning won't be used to improve protein folding (maybe it will, indirectly) it almost certainly won't be used directly, because it is not in the same class of problem.

Protein folding seems like the kind of problem that people were better at than computers 10 years ago, before deep learning

by tamana 10y ago

Protein folding seems like the kind of problem that people were better at than computers 10 years ago, before deep learning

dnautics 10y ago

exactly which deep learning technique do you propose unleashing on protein folding? Deep learning is not 'magic', it is a set of tools good at a certain subset of machine learning tasks, and the impressive thing is how many machine learning tasks can be shoehorned and contorted into the deep learning rubric, i.e. 'fuzzy classification via scoring'. It should be obvious how image recognition and voice recognition fit this rubric. "Playing Go", which does not fit this rubric obviously, I think, is conquered by deep learning because there is a discrete matrix of 'moves' which looks awfully lot like an image recognition matrix, and so a move which optimizes score ('classification') can be selected. Fundamentally one of the problems is that we don't know how to score proteins. We also don't necessarily know how to score go, but it's trivial to generate solutions by running hypothetical matches because the game outcome is deterministic, solvable, and definite. While there is a decent subset of proteins whose known structures could be used to seed a database for scoring, it is very small in comparison to the universe of possibilities (20^n, where n is the length of the amino acid chain). In short: Given an arbitrary "end-board layout" I can with our existing human knowledgebase tell you which go player has won. Given an arbitrary amino acid sequence and a proposed structure, I cannot with our existing human knowledgebase tell you if the structure is correct. Because of this limitation in creating a scoring system, while I won't say that deep learning won't be used to improve protein folding (maybe it will, indirectly) it almost certainly won't be used directly, because it is not in the same class of problem.