3 ms·

It's not a semantic difference. That's the problem. There is a difference between knowledge of protein folding (which is what the article is talking about), i.e

It's not a semantic difference. That's the problem. There is a difference between knowledge of protein folding (which is what the article is talking about), i.e. understanding proteins well enough to predict how they will fold, and modeling a folded structure (which is what AlphaFold is doing, based off of other pre-existing experimental structures). AlphaFold isn't simulating folding at all. It's modeling what the final structure should look like based off other data, not knowledge of how a protein folds. That's what is frustrating me so much with this topic. It's being sold as "we can predict how proteins will fold" as opposed to "we can model a folded protein based on its sequence". There's a huge difference in our understanding of protein folding, between those two points.

epgui 5y ago
It does sound like a semantic difference. To me, and I believe to most in the field, the protein folding problem was always pretty much about figuring out the final 3D structure (the final state), and not necessarily about the act of folding (ie.: the gerundive) itself. There is a lot of value in learning more about folding (in the gerundive sense), but the focus has always been on finding the folded structure, because as you very well know, structure is function.
- t_serpico 5y ago
  The “protein folding problem” consists of three closely related puzzles: (a) What is the folding code? (b) What is the folding mechanism? (c) Can we predict the native structure of a protein from its amino acid sequence? (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2443096/ https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2443096/). It seems the purist definition of solving protein folding means specifying the sequence of actions to arrive at the native structure for a given input sequence. This implies your solution is grounded in a physical reality. Such approaches (i.e MD simulations) have not been very successful to my knowledge and are computationally intractable for most proteins. Even if you are able to do this, how do you verify that your folding mechanism is in fact correct? We don't have ground truth experimental data at the temporal resolution one would want to verify such a prediction. For now, the only practical framing of the protein folding problem is to predict a native structure given an input sequence. This is simply because we have the requisite data...
  - YeGoblynQueenne 5y ago
    We have a lot of data we don't understand and we train a model to represent that data. The model represents the data very well, but we don't understand the model. So now we have a lot of data we don't understand and a model we don't understand. What new knowledge have we acquired? The OP is right to complain that there is no scientific knowledge gained by the mere act of modelling a poorly understood process, when the modelling is performed automatically and its results are also poorly understood. The goal of science is to gain knowledge of the processes in nature, not merely to predict their results. That's how science has advanced in the last few thousands of years and it's startling to see Nature and Science standing behind such a different idea of "science".

epgui 5y ago

It does sound like a semantic difference. To me, and I believe to most in the field, the protein folding problem was always pretty much about figuring out the final 3D structure (the final state), and not necessarily about the act of folding (ie.: the gerundive) itself. There is a lot of value in learning more about folding (in the gerundive sense), but the focus has always been on finding the folded structure, because as you very well know, structure is function.

t_serpico 5y ago
The “protein folding problem” consists of three closely related puzzles: (a) What is the folding code? (b) What is the folding mechanism? (c) Can we predict the native structure of a protein from its amino acid sequence? (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2443096/ https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2443096/). It seems the purist definition of solving protein folding means specifying the sequence of actions to arrive at the native structure for a given input sequence. This implies your solution is grounded in a physical reality. Such approaches (i.e MD simulations) have not been very successful to my knowledge and are computationally intractable for most proteins. Even if you are able to do this, how do you verify that your folding mechanism is in fact correct? We don't have ground truth experimental data at the temporal resolution one would want to verify such a prediction. For now, the only practical framing of the protein folding problem is to predict a native structure given an input sequence. This is simply because we have the requisite data...
- YeGoblynQueenne 5y ago
  We have a lot of data we don't understand and we train a model to represent that data. The model represents the data very well, but we don't understand the model. So now we have a lot of data we don't understand and a model we don't understand. What new knowledge have we acquired? The OP is right to complain that there is no scientific knowledge gained by the mere act of modelling a poorly understood process, when the modelling is performed automatically and its results are also poorly understood. The goal of science is to gain knowledge of the processes in nature, not merely to predict their results. That's how science has advanced in the last few thousands of years and it's startling to see Nature and Science standing behind such a different idea of "science".