4
$\begingroup$

In the news, DeepMind's AlphaFold is said to have solved the protein folding problem using neural networks, but isn't this a problem only optimised quantum computers can solve?

To my limited understating, the issue is that there are too many variables (atomic forces) to consider when simulating how an amino acid chain would fold, in which case only a quantum computer can be used to simulate it.

Is the neural network just making a very good estimate, or is it simulating the actual protein structure?

Improve this question
$\endgroup$
1
  • 1
    $\begingroup$ the CASP competition entered by deepmind is rigorous in its scoring & is unbiased wrt prediction technique. increasingly there is not really a technical difference between "estimation" vs "simulation" by ML. yes QM computers are thought to have an advantage here but actually applying them to real world problems is still relatively far off. the computational complexity of protein folding is very high but its a case where the "actual complexity" is not really known/ proven. so deepmind work is a demonstration/ "proof of principle" its within reach of existing systems but only via excellent engr. $\endgroup$
    – vzn
    Dec 14, 2020 at 21:19

2 Answers 2

Reset to default
8
$\begingroup$

AlphaFold (version 1 and 2) predicts (so estimates) the 3D shape of the protein from the sequence of amino acids. AlphaFold's performance is measured with the global distance test (GDT), which is a measure of similarity between two protein structures (the prediction and the ground-truth) that ranges from 0 to 100.

There is a short video and a longer one (both by DeepMind) that summarise the issue of protein folding, how it is important, how well AlphaFold approximately solves it (in the competition Critical Assessment of protein Structure Prediction (CASP)), i.e. AlphaFold 2 achieves a median GDT score of 92.4 (and 87 on the hardest proteins), which is a lot higher than AlphaFold 1's GDT score of 58 (which was the highest achieved score at the time), where, according to John Moult (president of CASP), a score around 90 is considered a satisfactory solution to the protein folding problem. You can find more details about AlphaFold 2 in this DeepMind blog post and about AlphaFold 1 in this other blog post or the associated paper published in Nature this year. You can find the code for AlphaFold 1 here, but there are other community/open-source implementations.

Despite the importance of the problem and achievement, there is clearly a lot of hype about this breakthrough (given also that it was achieved by DeepMind). This is also discussed in this video by Lex Fridman.

Improve this answer
$\endgroup$
2
  • $\begingroup$ what is the difference between "hype" vs "celebration"? its a thin line sometimes. think most of the official announcement + much media coverage is not really "hype" because its a real breakthrough, there is strong scientific consensus on that, many authorities quoted endorsing that its a breakthru, and the worldwide multi-decade CASP competition format also underlines that. $\endgroup$
    – vzn
    Dec 14, 2020 at 21:16
  • $\begingroup$ @vzn This is an important milestone for the machine learning community, but I don't know exactly how useful it will really be for healthcare, medicine, or related areas. We will see in the next years. There's a lot of hype (i.e. as you put it, celebration), but that's not necessarily bad. It will probably help DeepMind get even more funds, which is probably not bad, given their achievements in the last 5-7 years (considering also that it has been funded in 2010). $\endgroup$
    – nbro
    Dec 14, 2020 at 21:28
0
$\begingroup$

the issue is that there are too many variables (atomic forces) to consider when simulating how an amino acid chain would fold, in which case only a quantum computer can be used to simulate it.

These many variables, taking as an example the ones you mentioned, the atomic forces, are somehow grouped in order to facilitate the calculations; thus, it is not necessary for absolutely everything to be simulated simultaneously.

A notable example of this is when the K computer (one of the most powerful supercomputers in Japan, but far from being an optimized quantum computer) calculated the force needed to untie the DNA strands of histones (without taking into account the interactions between each nucleotide).

Multi Scale Modeling of Chromatin and Nucleosomes

1:37 -

By treating multiple atoms as one single particle we can increase the number of phenomena in our simulation

Basically, they start from the following question: "What is the maximum number of elements and interactions that we can exclude from an analysis object in a simulation in order to still preserve its properties without too much loss and with good accuracy?"

Another example of how this is done is when they apply the Finite Element Method to study hemodynamics in the heart.

They simply transform the heart into a set of regular tetrahedrons, yet still manage to simulate a very wide range of phenomena - such as the rate of consumption of ATP in each part of the heart and the variation in thickness of cardiac muscles during functioning.

Multi-scale Multi-physics Heart Simulator UT-Heart

In other words, they are far from wanting to simulate the heart in all its tissue details. On the contrary, they simplify as much as possible so that simulations are workable on our current supercomputers.

Improve this answer
$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .