Solving the protein structure prediction problem by AlphaFold2(AF2) from sequence seems at its core a game-changing breakthrough. Major areas of biology and biophysics will thrive and others may become diminished or even obsolete.
Will AF2 hurt or benefit experimental methods for determination of protein structure? Structures of large macromolecular machines should be enabled by having accurate computational structures for subunits and components. X-ray structure value may become more specialized. One thing that seems likely is that the already great value of sequence data (which is doubling every 8 months) is likely to become far greater by being more directly connected to spatial information. At its core solving the protein folding problem will enhance sequence impact and thereby increase the overall the pace of biology and biophysical advances by improving the ability of structural biology to better harness the flood of sequence data. How much will medical areas such as cancer biology and cancer drug discovery benefit? How about research areas such as protein design? What areas are likely to be most powerfully advanced and which most negatively impacted? What do you think?
How should senior and early stage researchers position themselves to ride the wave of growing positive impacts and reduce research wipeouts when this new mega-wave adds constructively or destructively with current research waves of systems biology and single molecule and single cell advances? What should we change in training for graduate students and fellows to insure they are correctly positioned for science with reliable protein structure prediction?
John,
You have asked an interesting question. For a senior member of structural biology community it shows the yearning for practical solutions as well as naivety of the real difficulties in even defining the problem. The protein folding problem is an ill-defined and most likely unsolvable problem in its classic sense of axiomatic reductionist sciences. If the old Anfinsen view was true about the minimum of global free energy there never would be life on Earth or anywhere else. The new paradigm that I partially formulated is that every protein (as a matter of fact every macromolecule of life) has its own specific recipe how to combine the structural with dynamical features to accomplish the desired function (PNAS 106(2009)10505). It means that there is a unique proportion of conditionally stable structure to conditionally unstable elements to perform given function. This ill defined condition is not solvable by classic axioms but it just might be partially solvable by fuzzy methods such as AI. Why AI is so good? Because the protein folding problem is a practical not a theoretical problem. Depending on conditions and their possible changes the protein suppose to perform a certain function. This is what physics call "self organized criticality". This is far from equivalency of having a certain structure nor a possible unique binding site nor anything that remotely can be called a solution. AI just simply captures better the set of heuristic rules than other methods particularly dynamical coupling with solvent. There are millions of examples to support this view. If anybody wants to engage I will be happy to, but not now. There is no place here to mention even a single one.
So the only hope is that we found an effective heuristic tool that gives us better approximation to reality. To be properly tested the method needs to be exposed to finding bordering cases when it breaks down. What conditions produce divergencies. What percentages of certain structures are folded and are not. For instance only around 30% of genomes of higher eukaryotes are properly folded. Around 30% is conditionally folded (on the target, in the particular compartment) since the misfolding diseases. And finally around 30% is almost never folded with residual presence of folding nuclei that precipitate the function. So all these classes of proteins must be tested against this new AI approach to see when the rules break down and how to enhance them.
Therefore, it is not a breakthrough that John is so excited about. But it is definitely progress. But, I would not see anytime in the future, abolishing or significant diminished importance of any particular field of science, in exactly the same way as radio did not get extinct by TV.
Bog
I have a question about this AI based prediction. Is it going to predict exact (very likely) structure or will give an ensemble of structures? If the later is so how do we know which one is correct? Does this mean we should do experiment to prove it?
Very important questions John! I believe that AlphaFold 2's prediction of protein structures within experimental error is the culmination of years of training enabled by experimental results and is definitely a positive development. Jeff and, subsequently, Zhang's Tasser had been winning CASP for years but has never achieved the level of precision that AF2 has shown, especially for monomeric proteins. This development ensures that students and postdocs do not end up spending months and, sometimes, years on struggling to express, crystallize and phase with MAD or MIR-based techniques which, as you are aware, can sometimes be excruciatingly time-consuming. This enables researchers to be researchers rather than technicians and also, hopefully, will end the tradition of rewarding, in terms of high-impact publications, technical feats rather than scientific ones. Having said that, we still need to wait and watch the evolution of AF2 vis-a-vis any improvements in its ability to predict protein-protein interactions or, as Yasin has pointed out, its ability to predict dynamics. My two-pence to the discussion! :)
Yasin Pourfarjam - I'm not sure it is yet able to do conformations or ensembles - however given complete structures there are computational approaches to model these.
John Tainer Thank you very much for your answer. I am very curios to know how the prediction is performed. Is it based on known crystal structures deposited to PDB? Given the size of predicted human proteome I think there will be a long long way to fully understand the mechanism of protein folding.
John,
You have asked an interesting question. For a senior member of structural biology community it shows the yearning for practical solutions as well as naivety of the real difficulties in even defining the problem. The protein folding problem is an ill-defined and most likely unsolvable problem in its classic sense of axiomatic reductionist sciences. If the old Anfinsen view was true about the minimum of global free energy there never would be life on Earth or anywhere else. The new paradigm that I partially formulated is that every protein (as a matter of fact every macromolecule of life) has its own specific recipe how to combine the structural with dynamical features to accomplish the desired function (PNAS 106(2009)10505). It means that there is a unique proportion of conditionally stable structure to conditionally unstable elements to perform given function. This ill defined condition is not solvable by classic axioms but it just might be partially solvable by fuzzy methods such as AI. Why AI is so good? Because the protein folding problem is a practical not a theoretical problem. Depending on conditions and their possible changes the protein suppose to perform a certain function. This is what physics call "self organized criticality". This is far from equivalency of having a certain structure nor a possible unique binding site nor anything that remotely can be called a solution. AI just simply captures better the set of heuristic rules than other methods particularly dynamical coupling with solvent. There are millions of examples to support this view. If anybody wants to engage I will be happy to, but not now. There is no place here to mention even a single one.
So the only hope is that we found an effective heuristic tool that gives us better approximation to reality. To be properly tested the method needs to be exposed to finding bordering cases when it breaks down. What conditions produce divergencies. What percentages of certain structures are folded and are not. For instance only around 30% of genomes of higher eukaryotes are properly folded. Around 30% is conditionally folded (on the target, in the particular compartment) since the misfolding diseases. And finally around 30% is almost never folded with residual presence of folding nuclei that precipitate the function. So all these classes of proteins must be tested against this new AI approach to see when the rules break down and how to enhance them.
Therefore, it is not a breakthrough that John is so excited about. But it is definitely progress. But, I would not see anytime in the future, abolishing or significant diminished importance of any particular field of science, in exactly the same way as radio did not get extinct by TV.
Bog
My response was not really correct - AF2 does provide a set of conformations. However as the current network is trained on crystal structures, the resulting conformations may not appropriately represent the solution ensemble. So this is an area that merits further development.
The reliability of the structural model predicted using AlphaFold2 depends heavily on the existing experimental structures, but there are always some proteins with unknown structures that cannot efficiently obtain available structural information from the existing structures. Moreover, some proteins with similar sequences have significant functional differences due to several detailed differences in structure. No matter how accurate the structure prediction is, there is no guarantee that the given structure is exactly the same as the real structure. Why is it called prediction? Because it predicts the structure state that the protein structure will present when it is actually measured in the future. I think AlphaFold2 improves the accuracy of structure prediction, but it is not enough to be a reliable basis for structure determination and drug design. Ultimately, we still need to verify the predicted model with experimental methods.
- Badges
- Science topic
- Similar topics
- Physical Sciences
- Waves