This AI will revolutionize biology and medicine as it unfolds the 3D shape of almost all proteins and speeds up drug development.
DeepMind’s AI program, AlphaFold, has solved one of biology’s oldest problems – How protein folds-up. The initial version of AlphaFold, according to CASP – short for Critical Assessment of Structure Prediction, has achieved highest accuracy in determining the structure of proteins. “This will change medicine. It will change research. It will change bioengineering. It will change everything,” says Andrei Lupas, an evolutionary biologist at the Max Planck Institute for Developmental Biology in Tübingen, Germany, who assessed the performance of different teams in CASP.
Proteins support every biological process in every living thing. Proteins are responsible for almost everything happening in the body of a biological organism where its function is defined by its 3D shape which tends to form without any help. The shape of protein is necessary for the biological processes and is to know to understand the binding of proteins with other molecules. This problem of ‘protein folding’ has been lying unanswered since decades.
AlphaFold is trained on the sequences and structures of over a hundred thousand of proteins mapped out by hundreds of scientists around the world. The initial version of the AlphaFold was developed by applying the concept of deep learning to the structural and genetic data to predict the distance between the amino acids in a protein. Later this data is used to develop a consensus model of what the protein should look like. This approach did not meet the requirement and was improvised by creating an AI network that incorporated additional information about the physical and geometric constraints that determine how a protein folds. Which means that the system predicts the final structure of the protein instead of the relationship between the amino acids.
The results at CASP show that two-thirds of the predictions were comparable to the experimental structure of the protein. AlphaFold scored almost 90 out of 100 in predicting the structure and once this value extends over 90 it will be considered to be approximately similar to the experimental value.
Following these results the applications and solutions to other biological problems may appear at a very fast pace. This could possibly help us understand the biological world to a greater extent and maybe even the chemical world.