Thoughts Regarding AlphaFold

Over the weekend Google’s Artificial Intelligence team based out of London (DeepMind) made a historic announcement that they had effectively solved protein folding problem. This announcement left me flabbergasted because I had always felt that the protein folding problem wouldn’t be solved until it could be inserted into a quantum computer. I felt this way because it made logical sense to assume that a protein folding prediction would only work if it mimicked the folding process of the natural world. This means that we would need a quantum computer to simulate the environment a string of amino acids would be exposed to within a cell. However in the end this really wasn’t necessary since crystallographers had already painstakingly determined the structure of 150k+ proteins. This provided the DeepMind team with enough raw data to train their AlphaFold program accurately enough to identify the final structure of almost all the amino acid chains presented at CASP2020.

What’s very surprising about these results is that AlphaFold doesn’t actually understand the concept of an amino acid, it simply knows that amino acid sequence X is statistically likely to fold in a certain way. Despite the hype this problem isn’t completely solved, since AlphaFold struggled with some amino acid sequences. However I am sure that most crystallographers will tell you that having a fairly accurate prediction greatly aids in helping them decode what an X-ray image of the protein is showing them. This will certainly result in a positive feedback loop where AlphaFold will speed up the discovery of protein structures and in turn the additional confirmed protein structures will improve AlphaFold. I am excited for the possibilities but I am also a little apprehensive since AI is starting to hit a lot closer to home thus making me wonder what else it can do.

rss facebook twitter github youtube mail spotify instagram linkedin google pinterest medium vimeo