What is so impressive with Google and DeepMind’s AlphaFold Technology

mediumThis post was originally published by Emil Walleser at Medium [AI]

DeepMind surpasses expectations with protein folding predictions and here is why it is such an impressive feat.

Google-owned DeepMind has pushed the limits of Artificial intelligence. One of the first introductions most people had to DeepMind was through AlphaZero [1]. “AlphaZero achieved within 24 hours a superhuman level of play in the games of chess and shogi (Japanese chess) as well as Go, and convincingly defeated a world-champion program in each case.” AlphaZero achieved all of this through a process called reinforcement learning, basically playing repeated games against itself until it identified winning strategies. This lack of human input data creates a scenario where AlphaZero was uninhibited by the ideas of what humans think is a good chess strategy and created its own gameplay strategy. The topped ranked chess player in the world Magnus Carlsen stated that AlphaZero has influenced his own chess strategy “In essence, I have become a very different player in terms of style than I was a bit earlier, and it has been a great ride.” [2] While the scope of games such as Chess and Go are limited, the idea that AI can not only outperform humans but do so in ways we haven’t comprehended previously is impressive.

The next big focus for DeepMind… protein folding. Protein sequencing (determining the components of an actual protein itself) was introduced around the 1950s [3]. However, that is just one part of the puzzle. Protein takes on an amazing three-dimensional shape consisting of swirling alpha helices and beta sheets. The chains themselves interact with one another form a maze of sorts — kind of this mystical ball of tangled yarn [4]. An example of this can be seen in the image below.

What is so special about this shape? If we already know the sequence how complicated can the rest be? Clearly, there must be some rule and the folding can’t be completely random. Levinthal’s Paradox expounds on this idea that such a complex shape can occur so quickly and appear uncontrolled in nature.

“Estimates have suggested that it could take all the computers in the world longer than the age of the universe to compute the structure of a single 200-residue protein.” [5]

The sheer complexity of such a “simple” problem is astounding. Clearly, a brute force approach isn’t the answer. The Critical Assessment of protein Structure Prediction (CASP) is the organization dedicated to advancing knowledge and understanding of how proteins fold and promoting the best prediction models for them [6]. Every two years CASP announces the best performing prediction model for protein folding. Judges compare a set of predicted proteins to a known protein sequence calculating a measurement to calculate the best score. the detail the best performing protein prediction models produced and compared to known protein folding by experts. AlphaFold took home top honors in 20 and significantly advanced the prediction of protein folding structure [7]. Using complex deep convolutional neural networks and looking at covariations of related proteins, AlphaFold significantly outperformed other competitors. Not to be outdone by anyone but themselves, DeepMind improved over the original design of AlphaFold for CASP14. They harnessed the power of attention networks and created an incredibly reliable model [8]. From the DeepMind blog post [9]:

This means that our predictions have an average error (RMSD) of approximately 1.6 Angstroms, which is comparable to the width of an atom (or 0.1 of a nanometer).

DeepMind and AlphaFold have demonstrated the incredible power of artificial intelligence and deep learning have to solve problems previously stumping humans. In the case of chess, AlphaZero entered mastery level in under a single day. While the protein folding question took longer to train, errors within the width of an atom shows just how powerful this product is and how artificial intelligence will keep pushing the world forward.

[1] Silver, David, et al. “Mastering chess and shogi by self-play with a general reinforcement learning algorithm.” arXiv preprint arXiv:1712.01815 (2017).
[2] When Magnus met AlphaZero, Newinchess.com
[3] B. Foltman, Protein Sequencing: Past and Present (1980)
[4] Protein Folding, the Good, the Bad, and the Ugly (2010), Harvard University
[5] Feldman et al., Probabilistic sampling of protein conformations: new hope for brute force? (2002), Proteins
[6] https://predictioncenter.org/
[7]Senior, A. W., Evans, R., Jumper, J., Kirkpatrick, J., Sifre, L., Green, T., Qin, C., et al., Improved Protein Structure Prediction Using Potentials from Deep Learning, Nature, vol. 577, no. 7792, pp. 706–10, 2020. DOI: 10.1038/s41586–019–1923–7
[8]Jumper, J., Evans, R., Pritzel, A. et al. Highly accurate protein structure prediction with AlphaFold. Nature (2021). https://doi.org/10.1038/s41586-021-03819-2
[9] The AlphaFold Team, AlphaFold: a solution to a 50 year old grand challenge in biology (2020), deepmind.com

Spread the word

This post was originally published by Emil Walleser at Medium [AI]

Related posts