Within our cells, proteins perform countless functions that are essential for life. They build tissues, transport chemicals, fight infection, and much more. But what gives each protein its unique capabilities is its three-dimensional structure. Like intricately folded origami, a protein is assembled from a long chain of smaller molecules called amino acids that bend and twist into complex shapes. A protein’s shape determines what it can interact with, enabling its specific function. This entire process by which a protein folds into a functional form is known as protein folding.
For over 50 years, biologists have been trying to solve the protein folding problem – predicting a protein’s 3D structure from its amino acid sequence. But the process is mind-bogglingly complex, with interactions occurring at the scale of billionths of a meter and trillionths of a second. Using traditional physics-based simulation techniques, it would take powerful supercomputers months or years to computationally fold a single protein. Devising new proteins with shapes tailored for medical and industrial uses seemed an insurmountable challenge.
Enter DeepMind’s AlphaFold, an artificial intelligence system designed specifically to crack the protein folding problem. Inspired by the neural networks of the human brain, AlphaFold uses deep learning, a form of machine learning, to analyze huge datasets of known protein structures. By discovering patterns in these massive amounts of data, AlphaFold learned to model the complex molecular forces that guide protein folding. In effect, AlphaFold derived its own ‘physics engine’ just by observing nature’s handiwork through protein folding examples.
Equipped with this deep learning model, AlphaFold could then take the amino acid sequence of a never-before-seen protein and predict its 3D structure with stunning accuracy in minutes rather than months. This breakthrough opened up revolutionary possibilities for computational biology and drug discovery research. Scientists could now elucidate the shapes of important proteins involved in diseases and gain new insights into their functions. Rather than costly lab experiments, they could scan protein structural databases for drug candidates. The potential was enormous.
Recently, AlphaFold’s full power was demonstrated when DeepMind used it to predict structures for nearly every protein catalogued by science – over 200 million in total! This massive Protein Almanac provides a new lens for exploring biology, revealing key details for proteins behind diseases like COVID-19, malaria, HIV and cancer. Researchers everywhere now have an incredible resource for their work and a head start on solving countless medical mysteries.
Of course, AlphaFold is still an AI system with limitations to improve. But it provides an unbelievable proof of concept – given enough data, AI can learn rules well enough to simulate immensely complex natural processes. Similar deep learning techniques are now being applied to other biology puzzles, from mapping neural networks to designing enzyme pathways. AlphaFold has ushered in a new era of AI-accelerated bioengineering, where digital models of life provide a limitless canvas for discovery.