Another (big) advance for DeepMind...

[doc75]

Amazing stuff continues to unfold at DeepMind (pardon the pun):

 

https://www.theguardian.com/technology/2020/nov/30/deepmind-ai-cracks-50-year-old-problem-of-biology-research

[patience_and_focus]

Amazing stuff continues to unfold at DeepMind (pardon the pun):

 

https://www.theguardian.com/technology/2020/nov/30/deepmind-ai-cracks-50-year-old-problem-of-biology-research

 

This is actually quite impressive. This is (err, was) a very hard problem.  I had the fortune of being part of a group of some very bright minds when they were  working on this protein folding problem back in the day and participating in CASP (the competition that AlphaFold won this year by a 2.5X margin to its nearest competitor).

 

It is a crazy and stunning advance, the effect of which cannot be overstated.  This could be as significant as human genome sequencing. Because if sequencing the genome was equivalent to translating a book of "unknown" language to english then this advance is equivalent to creating a play out of the characters of the book, understanding them and see in real life how they potentially function and play their part in the human body.

[thowed]

I would be very grateful if anyone has any sense of what types of companies are likely to benefit from these amazing developments.  General biotech, or more specific areas?  Though presumably this will be priced in pretty quickly.

 

Many thanks.

[clutch]

I would be very grateful if anyone has any sense of what types of companies are likely to benefit from these amazing developments.  General biotech, or more specific areas?  Though presumably this will be priced in pretty quickly.

 

Many thanks.

 

Google (or Alphabet).

 

The expertise of DeepMind is currently used to solve and optimize many problems in Google.

[Investor20]

Protein folding is very important in drug development.

 

However, what is important is how these protein structures interact with other proteins or molecules.

https://www.nature.com/articles/s41573-020-0078-4

 

For example, if we can find a molecule that binds specifically to a protein, it can change how the protein functions and could be a potential medicine.

 

I am not sure if being able to predict protein folding itself can do that job. As I understand Alphafold worked by using database of known protein structures to predict the protein structures.  But is there a data base for the algorithms to learn from different molecules binding to the protein?

 

An important part of protein research is protein misfolding that is supposed to be cause of cataract and many degenerative diseases.

https://www.nature.com/scitable/topicpage/protein-misfolding-and-degenerative-diseases-14434929/

 

Please read about levinthals paradox.  Essentially if you take a protein of a chain with 100 amino acids that tries different ways it can fold randomly, till it reaches the most stable configuration, that would take millions of years.  But proteins fold within seconds.

 

So the idea is they fold not by randomly trying different structures but they fold in an order...a pathway...to reach a structure. Thus this may not be the most thermodynamically stable structure, but the structure it has reached through a path of folding steps.

 

However,  if they dont fold properly, then these misfolded proteins stick to each other and precipitate.  Once misfolded, they neither know how to go back or go forward..its like getting lost without a map.  For example when you boil egg, the proteins in there are precipitating.  They are misfolded proteins stuck to each other. But if you put it in a cup of water, they are not going to dissolve back.  This is unlike the sugar that might be precipitated from boiling water.  But you put the sugar in more water, it will dissolve rightaway. You could try putting a boiled egg in a big tank of water for very long time and it wont go back into water.

 

I am not sure this DeepMind will help solve the protein interaction problem.  Definitely very interesting, but it did require a database of already solved protein structures and can it be extrapolated to how proteins bind to other molecules or other proteins?

 

 

 

 

[thowed]

Very interesting, Investor20, thanks for this.

 

It's always good to hear an independent view - it's easy for everyone just to say, 'this is great' without thinking.

[clutch]

 

"Why is DeepMind's AlphaFold a major breakthrough? When will it begin impacting drug discovery? Dangers of AI at scale"

[cameronfen]

So just looking at Twitter, I don’t think the protein folding problem is solved as a lot of the hyperbole is due to the good marketing teams at google.  Could this win a Nobel prize? Potentially.  However it didn’t outperformamong all the CASP subcategories (according to a Twitter person) so other models which are still mediocre outperformed in some cases. 

 

 

Protein folding is very important in drug development.

 

However, what is important is how these protein structures interact with other proteins or molecules.

https://www.nature.com/articles/s41573-020-0078-4

 

For example, if we can find a molecule that binds specifically to a protein, it can change how the protein functions and could be a potential medicine.

 

I am not sure if being able to predict protein folding itself can do that job. As I understand Alphafold worked by using database of known protein structures to predict the protein structures.  But is there a data base for the algorithms to learn from different molecules binding to the protein?

 

An important part of protein research is protein misfolding that is supposed to be cause of cataract and many degenerative diseases.

https://www.nature.com/scitable/topicpage/protein-misfolding-and-degenerative-diseases-14434929/

 

Please read about levinthals paradox.  Essentially if you take a protein of a chain with 100 amino acids that tries different ways it can fold randomly, till it reaches the most stable configuration, that would take millions of years.  But proteins fold within seconds.

 

So the idea is they fold not by randomly trying different structures but they fold in an order...a pathway...to reach a structure. Thus this may not be the most thermodynamically stable structure, but the structure it has reached through a path of folding steps.

 

However,  if they dont fold properly, then these misfolded proteins stick to each other and precipitate.  Once misfolded, they neither know how to go back or go forward..its like getting lost without a map.  For example when you boil egg, the proteins in there are precipitating.  They are misfolded proteins stuck to each other. But if you put it in a cup of water, they are not going to dissolve back.  This is unlike the sugar that might be precipitated from boiling water.  But you put the sugar in more water, it will dissolve rightaway. You could try putting a boiled egg in a big tank of water for very long time and it wont go back into water.

 

I am not sure this DeepMind will help solve the protein interaction problem.  Definitely very interesting, but it did require a database of already solved protein structures and can it be extrapolated to how proteins bind to other molecules or other proteins?

 

As far as this, afaik if you have a folded protein finding protein protein interaction is easier because it’s more of a geometric problem rather than a physics thermodynamics energy problem.  If you know protein A binds with something and you can protein fold, you just build peptide chains until you get a peptide chain that looks like the shape of A in the activation site.  We often know the shape of A and what A binds to so afaik this stuff is hugely useful in designing molecules.  Again keep in mind, alphafold is trained on the shape of protein in biological systems, so even if they would fold differently in different systems, how they fold in biological systems are what medical researchers care about.  Not an expert here but this is what I gathered from reading about it.