Did AI Just Solve a 50 Year Old Biology Problem?

Biology has had a difficult problem for 
around 50 years the protein folding problem   did ai just solve this perplexing problem you may remember google's deep mind from 
their ai system alphago which beat the world   champion go player in 2016. this was a major 
achievement for ai that arrived sooner than   predicted just recently deepmind has achieved 
significant success in accurately predicting   protein structures using its system alpha fold 
this is exciting news it illustrates how ai may   indeed be moving at exponential speeds and may 
help science in more immediate and unexpected ways   proteins are essential to life tiny machines 
that support nearly all its functions made up   of chains of amino acids they are large complex 
molecules a protein's function mostly depends on   its unique 3d structure some diseases arise from 
errors in protein folding so a full understanding   of proteins is monumental in addition it 
could allow more precision and drug creation   in his acceptance speech for the 1972 nobel 
prize in chemistry christian athenson famously   postulated that in theory a protein's amino acid 
sequence should fully determine its structure   and so began a five-decade quest to solve this 
problem this isn't easy a protein can fold in   an astronomical number of ways add to this a 
huge amount of amino acid sequence possibilities   and the job is incredibly challenging dr 
cyrus leventhal estimated that a protein has   10 to 300 possible conformations scientists have 
determined actual protein structures by using   techniques such as nuclear magnetic 
resonance and x-ray crystallography   however these methods require the use of 
multi-million dollar equipment and can take years   of rigorous and elaborate work for each protein 
we know of over 200 million proteins yet we've   only mapped around 170 000 of them using these 
experimentally determined methods as early as the   1980s scientists tried computational solutions but 
with poor results starting in 1994 professors john   molt and christoph fidelis founded critical 
assessment of protein structure prediction   casp to accelerate this approach occurring 
every other year the casp challenge has   prompted various teams to advance methods of 
identifying protein structure from sequence   casp chooses recently determined protein 
structures that are not published in advance   therefore teams must go in blind to predict 
the structure of previously unknown proteins   and verify accuracy against the real model no 
cheating alpha fold competed in 2018 and did very   well but at casp 14 it effectively crushed it by 
scoring over twice as high as the next best team   alpha fold was trained on all 170 000 publicly 
available protein structures the team expanded   on their prior model using new deep learning 
architectures which enabled unparalleled accuracy   a full paper will be published that will describe 
the specific deep learning methods used for this   recent model alpha fold structure predictions were 
in some cases nearly identical to the originals   made using experimental methods how does casp 
compare the accuracy of the predicted proteins   to the experimentally obtained models a method 
known as global distance test gdt ranks results   on a scale from 1 to 100 gdt can be thought of 
as the percentage of amino acid residues beads   in the protein chain within a threshold distance 
from the correct position a score of around 90 gdt   is informally considered to be on par with 
results obtained from experimental methods   the alpha fault system achieved a median score of 
92.4 gdt overall across all targets one scientist   used an alpha fold prediction to determine the 
structure of a bacterial protein in half an hour   he had been trying for a decade to get a solution 
by other methods on a ground level the remarkable   breakthrough of alpha fold will allow scientists 
understand diseases more quickly and develop more   effective drugs to fight them it could also have 
impact outside the medical fields such as building   enzymes to reduce plastic waste or even removing 
carbon from the atmosphere deepmind cautions that   there is still much to learn about how multiple 
proteins form complexes and how they interact with   dna rna or small molecules with that caveat alpha 
fold is a promising example of how ai is becoming   one of humanity's most useful tools expanding 
our scientific knowledge at an ever faster pace   to learn more about protein folding and deep 
learning head to the links below the video   like follow subscribe and catch us next time to 
see how you plus science can help save the world you