AI-Powered AlphaFold 2: Revolutionizing Protein Folding

[BWV]

This is cool - AI is the future of biology, humans just arent smart enough

https://deepmind.com/research/case-studies/alphafold

https://predictioncenter.org/casp14/doc/presentations/2020_12_01_TS_predictor_AlphaFold2.pdf

Deepmind trained on 100K proteins and is now in the second iteration of the algorithm

 

[.Scott]

Saying that humans aren't smart enough to predict the folding of a protein molecule is like saying that we don't have the arm strength to fly.

We concoct planes to fly and AI to solve the folding problem.

 

[BWV]

Saying that humans aren't smart enough to predict the folding of a protein molecule is like saying that we don't have the arm strength to fly.

We concoct planes to fly and AI to solve the folding problem.

Sure, but the difference is we understand how planes fly, but can't fully grasp the details on how & why these algorithms work for protein folding, Go, chess or whatever. You get a statistical certainty, but no absolute proof.

 

[Tom.G]

Sure, but the difference is we understand how planes fly, but can't fully grasp the details on how & why these algorithms work for protein folding...

Maybe we will understand when we fold proteins for as long as we've been flying airplanes.

On the other hand, we may bump into the conundrum of whether anything can totally understand itself, whether machine or animal.

Anyhow, food for endless thought, ...or nightmares!

 

[.Scott]

If I use a computer to calculate pi to 1000 digits, is it because I am not smart enough to calculate it myself?

 

[BWV]

If I use a computer to calculate pi to 1000 digits, is it because I am not smart enough to calculate it myself?

Not the same, as you know how to do the calculation and explicitly provide these instructions to the computer.

We don't really know how complex ANNs arrive at their answers - can certainly study and learn from them, but if we really understood AlphaZero, for example, then we could presumably create Stockfish or Deepblue type programs to beat it. protein folding is a larger combinatorial space than Go or Chess and unlike the games, there may be ‘rules‘. we are not aware of

 

[BWV]

More from Nature, this may be the biggest science story of the year
https://www.nature.com/articles/d41586-020-03348-4

This is a big deal,” says John Moult, a computational biologist at the University of Maryland in College Park, who co-founded CASP in 1994 to improve computational methods for accurately predicting protein structures. “In some sense the problem is solved.”
...

nearly two-thirds were comparable in quality to experimental structures. In some cases, says Moult, it was not clear whether the discrepancy between AlphaFold’s predictions and the experimental result was a prediction error or an artefact of the experiment.

 

[.Scott]

Not the same, as you know how to do the calculation and explicitly provide these instructions to the computer.

If I know the AI algorithm - and code it myself, then is it like long division?

Calling an algorithm "Artificial Intelligence" is like calling a driver-assist tool "Autopilot". The name alone tempts you to think of the application as independently smart and "person-like".

 

[BWV]

Calling an algorithm "Artificial Intelligence" ...tempts you to think of the application as independently smart and "person-like".

I did not make up the term and am not tempted to think about ANNs as either ’smart’ or ‘person-like’

 

[.Scott]

I did not make up the term and am not tempted to think about ANNs as either ’smart’ or ‘person-like’

But coding up something that uses AI techniques is different than coding up something that doesn't rely on statistics or pattern discovery?

 

[BWV]

But coding up something that uses AI techniques is different than coding up something that doesn't rely on statistics or pattern discovery?

it is if you don't understand exactly the algorithm is doing. We don't understand exactly how or why AlphaZero is better at chess than Stockfish. One could imagine, as an analogy, some dumbass feeding a bunch of data into a NN of measurements of force, acceleration and mass then relying on the algorithm without understanding F=MA