AlphaFold 🧬

An AI system revolutionizing biological research.

By: Evan Luksay

TL;DR

Google’s artificial intelligence subsidiary, DeepMind, recently developed a deep learning system named AlphaFold, which effectively solves the 50-year-old protein folding problem.

The Breakdown

• Proteins are large complex molecules composed of chains of amino acids. The function of a protein depends on its unique 3D structure
• As one of the fundamental building blocks of life, proteins are involved in nearly all cellular processes
• Previously, the structure of proteins was determined using nuclear magnetic resonance and X-ray crystallography machines; however, these methods are timely, error-prone, and costly
• Theoretically, a protein’s 1D amino acid sequence should fully determine its 3D structure
• Since the 1970s, scientists have hoped that computational methods could one day determine how a protein folds (and thus derive its structure) by analyzing its amino acid sequence

The Tech

• AlphaFold makes use of a deep learning attention network (a type of neural network) to determine the “distances between pairs of amino acids and the angles between chemical bonds that connect those amino acids” from 1D amino acid sequences. These datasets are then used to predict the protein’s 3D structure
• Next, one compares these predicted structures to known protein structures and then further refines the models by using a gradient descent algorithm
• Block structure of AlphaFold’s neural network:

• AlphaFold determines a Global distance test score for the prediction, which rates how close the prediction of the protein’s structure is to reality. The most recent AlphaFold system received an average test score of 92.4 (with 100 being a perfect match) — a level of accuracy similar to that determined by the most technically advanced traditional methods used today

The Significance

• AlphaFold has demonstrated the powers of AI in solving difficult scientific problems in remarkable timeframes
• The AI system is poised to improve further and help researchers access more cost-efficient and accurate methods to determine protein structure, which is essential in many medical and industrial applications
• Computational methods used in AlphaFold’s neural network are likely to revolutionize drug development and disease research areas. Additionally, it will likely be utilized in CRISPR techniques to carry out futuristic gene editing
• AlphaFold has already begun to prove its utility by determining spike proteins’ structure for the SARS-CoV-2 virus

Additional Info

AlphaFold Version One Code

AlphaFold a solution to a 50-year-old grand challenge in biology

‘It will change everything’: DeepMind’s AI makes a gigantic leap in solving protein structures

Full Source List