Important Biology Problem Solved By AI

Protein Folding Predictions: A Science Transformation For Medicine

Development of AlfaFold

Proteins are the building blocks of all life and understanding their structure and function is foundational to understanding the causes and treatment of physical disorders.  Scientists have been studying protein structure since the 1960’s, but it is complex and each protein can take months to years to identify the 3-dimensional shape that determines its specific function in the body.  Predicting the 3-D shape that a protein will adopt based solely on its amino acid sequence has been an important open research problem for more than 50 years.  In 2021 and 2022, Deep Mind, a subsidiary of Google and Alphabet, was successful in predicting the 3-D structure of almost every known protein, or about 200 million proteins, using an AI computational program called AlphaFold.  Demis Hassabis is the software develop, co-founder, and CEO  of Deep Mind Technologies.   He and senior staff research scientist John Jumper were jointly awarded a $3 million Breakthrough Prize in Life Sciences for the achievement in 2022. 

Deep Mind began to develop AlphaFold soon after its AlphaGo AI software gained worldwide attention by beating the world champion, Lee Sedol, at the Chinese game of Go.  Hassabis considered gaming accomplishments as a learning step in the process for his end goal of solving important problems in science.   Deep Mind has made the structures of proteins from nearly every species with known amino acid sequences freely available as open source in a public database.  This facilitates the development of multiple applications ranging from molecular biology to drug development. 

Deep Mind’s mission was to develop general-purpose algorithms that can be applied across many problems.  It took 5 or 6 years to create the first version and compete at the CASP13 convention.  It was the first time machine learning had been used as the core component of a system to solve a problem.  From there, they moved on to the second version AlfaFold 2 in the search for more atomic accuracy.  It has a complicated architecture with 32 different component algorithms from biology, chemistry and biophysics disciplines.

As is true for AI in other applications, the software developers do not fully know the exact mechanisms of how the AI solved the problem nor its rules for protein folding.  This is called the Black Box of AI.   By definition, AI is able to learn in a manner similar to the human mind.  AI is presented with examples of known data and problem solutions, then given the task of solving problems where the answer is unknown.  For example, AlfaFold was given data about chemical bond angles and amino acid sequencing that added in some constraints to narrow the search space of possible protein structures, which AlfaFold then reverse engineered.  Confidence measures were added to each amino acid so users could know which parts are the most reliable.

Protein Physiology and Folding Problem

Proteins consist of long chains of amino acids which spontaneously fold into a 3D structure in a process called protein folding.  The 3D structure is specific to the function of that protein with distinct surface characteristics that determine what other molecules can bond or interact with it.     Understanding how the amino acid sequence determines the 3D structure is very challenging and called the protein folding problem.

 Before AlphaFold, techniques including X-ray crystallography, cryo-electron microscopy and nuclear magnetic resonance were used but were very expense and time consuming.  Those efforts identified the structures of 170,000 proteins over 60 years, out of the 200 million known proteins across all life forms.  AlfaFold is the first accurate computational method for protein structure predictions.

Except for water, protein is the most common substance in the body, forming muscles, bones, organs, skin, connective tissue and nails.    They are used to make hormones, enzymes, antibodies, and immune globulins.  Proteins are responsible for nearly every cellular task including growth, repair and maintenance of tissues, metabolic functions, maintenance of fluid balance and pH, transportation and storage of nutrients in and out of cells, structure of organs and connective tissue, and storage of energy. 

Failure of a polypeptide chain to fold properly produces either an inactive or toxic protein that malfunctions causing a number of diseases.    Misfolds can result in allergies and neurodegenerative disorders, including Alzheimer’s, Parkinson’s, Huntington’s diseases and Cystic Fibrosis.

Demis Hassabis reported that almost every pharmaceutical company is using AlfaFold.  It has been used to address antibiotic resistance, tackling plastic pollution and alfalfa crop sustainability.  It’s important to understand that AI is already actively transforming science and medicine.  It has been called an astounding game changer for technological advancement