Google DeepMind – Genetic origami

A big step forward but still needs more work.

• DeepMind has convincingly won an academic competition that tests researchers’ ability to predict the structure of proteins from their genetic sequence which opens a door to improved understanding and treatment of genetic disorders.
• Proteins are the exclusive product of DNA and as such are responsible for 100% of all functions and features that result from an organism’s genetic make-up.
• Proteins are chains of amino acids which, after they have been synthesized from DNA, are folded by the electrostatic interactions between the amino acids to make the shape which enables their function.
• Protein structure is crucial because it is in effect the execution of instructions that are encoded on the DNA strand.
• However, because there are 20 different amino acids that can be used and because proteins can be many hundreds of amino acids long, the number of possible structures that can be formed from one chain is practically infinite.
• Hence, calculating each possibility (as a regular computer would do (brute force)) is impossible as Cyrus Levinthal calculated in 1969 that a protein with 100 amino acids has 3¹⁹⁸ possible combinations (see here).
• Hence, in order to create an algorithm that has any possibility to be any good at predicting protein structure, the machine needs to be able to work out where to look and what possibilities to exclude without having to calculate them.
• This is exactly the technique that allowed DeepMind to crack the game of Go before anyone else (see here) which has now been taken and applied to a much more practical problem.
• I am sure that this is why the algorithm is called AlphaFold as the way that it works is very similar to AlphaGo.
• Like AlphaGo, AlphaFold has multiple neural networks that have been trained to work out various features of the problem in hand that are then combined to the narrow down the number of possibilities that need to be calculated or searched to arrive at a viable result.
• The results are pretty startling as in an algorithmic competition to predict protein structure, AlphaFold placed 1^st out of 98 entrants with the most accurate prediction of 25 out of 43 protein structures.
• The second-place algorithm managed to be most accurate in just 3 of 43 of the 43 proteins under test.
• The 43 protein structures had been derived using the normal, time-consuming, manual and very expensive methods.
• It is important to note that DeepMind managed to be the most accurate which implies that it did not get any of the structures completely right.
• In the world of protein biology, this is important as tiny changes in a protein’s structure can radically alter its biological properties meaning for this to be useful, there remains a very long way to go.
• Hence, while this is a big step forward by DeepMind, much to the chagrin of its competitors in academia, I think it is still very far from having any real practical application.
• DeepMind may get a lot less attention than other Google acquisitions like Nest, but it is on track to more than justify the enormous $500m price tag paid in 2014.
• Once again, it is Google and Google owned entities that are driving forward AI leaving Apple, Facebook and everyone else fairly far behind.