<s>
AlphaFold	O
is	O
an	O
artificial	B-Application
intelligence	I-Application
(	O
AI	B-Application
)	O
program	O
developed	O
by	O
DeepMind	B-Application
,	O
a	O
subsidiary	O
of	O
Alphabet	O
,	O
which	O
performs	O
predictions	O
of	O
protein	O
structure	O
.	O
</s>
<s>
The	O
program	O
is	O
designed	O
as	O
a	O
deep	B-Algorithm
learning	I-Algorithm
system	O
.	O
</s>
<s>
AlphaFold	O
AI	B-Application
software	I-Application
has	O
had	O
two	O
major	O
versions	O
.	O
</s>
<s>
A	O
team	O
that	O
used	O
AlphaFold	B-Application
2	I-Application
(	O
2020	O
)	O
repeated	O
the	O
placement	O
in	O
the	O
CASP	O
competition	O
in	O
November	O
2020	O
.	O
</s>
<s>
AlphaFold	B-Application
2	I-Application
's	O
results	O
at	O
CASP	O
were	O
described	O
as	O
"	O
astounding	O
"	O
and	O
"	O
transformational.	O
"	O
</s>
<s>
On	O
15	O
July	O
2021	O
the	O
AlphaFold	B-Application
2	I-Application
paper	O
was	O
published	O
at	O
Nature	O
as	O
an	O
advance	O
access	O
publication	O
alongside	O
open	B-Application
source	I-Application
software	I-Application
and	O
a	O
searchable	O
database	O
of	O
species	O
proteomes	O
.	O
</s>
<s>
AlphaFold	O
started	O
competing	O
in	O
the	O
2018	O
CASP	O
using	O
an	O
artificial	B-Application
intelligence	I-Application
(	O
AI	B-Application
)	O
deep	B-Algorithm
learning	I-Algorithm
technique	O
.	O
</s>
<s>
DeepMind	B-Application
is	O
known	O
to	O
have	O
trained	O
the	O
program	O
on	O
over	O
170,000	O
proteins	O
from	O
a	O
public	O
repository	O
of	O
protein	O
sequences	O
and	O
structures	O
.	O
</s>
<s>
The	O
program	O
uses	O
a	O
form	O
of	O
attention	B-General_Concept
network	I-General_Concept
,	O
a	O
deep	B-Algorithm
learning	I-Algorithm
technique	O
that	O
focuses	O
on	O
having	O
the	O
AI	B-Application
identify	O
parts	O
of	O
a	O
larger	O
problem	O
,	O
then	O
piece	O
it	O
together	O
to	O
obtain	O
the	O
overall	O
solution	O
.	O
</s>
<s>
The	O
overall	O
training	O
was	O
conducted	O
on	O
processing	O
power	O
between	O
100	O
and	O
200	O
GPUs	B-Architecture
.	O
</s>
<s>
Combining	O
a	O
statistical	O
potential	O
based	O
on	O
this	O
probability	O
distribution	O
with	O
the	O
calculated	O
local	O
free-energy	O
of	O
the	O
configuration	O
,	O
the	O
team	O
was	O
then	O
able	O
to	O
use	O
gradient	B-Algorithm
descent	I-Algorithm
to	O
a	O
solution	O
that	O
best	O
fitted	O
both	O
.	O
</s>
<s>
Central	O
to	O
AlphaFold	O
is	O
a	O
distance	O
map	O
predictor	O
implemented	O
as	O
a	O
very	O
deep	O
residual	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
with	O
220	O
residual	B-Algorithm
blocks	I-Algorithm
processing	O
a	O
representation	O
of	O
dimensionality	O
64×64×128	O
–	O
corresponding	O
to	O
input	O
features	O
calculated	O
from	O
two	O
64	O
amino	O
acid	O
fragments	O
.	O
</s>
<s>
Each	O
residual	B-Algorithm
block	I-Algorithm
has	O
three	O
layers	O
including	O
a	O
3×3	O
dilated	O
convolutional	O
layer	O
–	O
the	O
blocks	O
cycle	O
through	O
dilation	O
of	O
values	O
1	O
,	O
2	O
,	O
4	O
,	O
and	O
8	O
.	O
</s>
<s>
Alongside	O
a	O
distance	O
map	O
in	O
the	O
form	O
of	O
a	O
very	O
finely-grained	O
histogram	O
of	O
distances	O
,	O
AlphaFold	O
predicts	O
Φ	B-Application
and	I-Application
Ψ	I-Application
angles	I-Application
for	O
each	O
residue	O
which	O
are	O
used	O
to	O
create	O
the	O
initial	O
predicted	O
3D	O
structure	O
.	O
</s>
<s>
The	O
AlphaFold	O
authors	O
concluded	O
that	O
the	O
depth	O
of	O
the	O
model	O
,	O
its	O
large	O
crop	O
size	O
,	O
the	O
large	O
training	O
set	O
of	O
roughly	O
29,000	O
proteins	O
,	O
modern	O
Deep	B-Algorithm
Learning	I-Algorithm
techniques	O
,	O
and	O
the	O
richness	O
of	O
information	O
from	O
the	O
predicted	O
histogram	O
of	O
distances	O
helped	O
AlphaFold	O
achieve	O
a	O
high	O
contact	O
map	O
prediction	O
precision	O
.	O
</s>
<s>
The	O
2020	O
version	O
of	O
the	O
program	O
(	O
AlphaFold	B-Application
2	I-Application
,	O
2020	O
)	O
is	O
significantly	O
different	O
from	O
the	O
original	O
version	O
that	O
won	O
CASP	O
13	O
in	O
2018	O
,	O
according	O
to	O
the	O
team	O
at	O
DeepMind	B-Application
.	O
</s>
<s>
The	O
DeepMind	B-Application
team	O
had	O
identified	O
that	O
its	O
previous	O
approach	O
,	O
combining	O
local	O
physics	O
with	O
a	O
guide	O
potential	O
derived	O
from	O
pattern	O
recognition	O
,	O
had	O
a	O
tendency	O
to	O
over-account	O
for	O
interactions	O
between	O
residues	O
that	O
were	O
nearby	O
in	O
the	O
sequence	O
compared	O
to	O
interactions	O
between	O
residues	O
further	O
apart	O
along	O
the	O
chain	O
.	O
</s>
<s>
As	O
a	O
result	O
,	O
AlphaFold	O
1	O
had	O
a	O
tendency	O
to	O
prefer	O
models	O
with	O
slightly	O
more	O
secondary	O
structure	O
(	O
alpha	O
helices	O
and	O
beta	O
sheets	O
)	O
than	O
was	O
the	O
case	O
in	O
reality	O
(	O
a	O
form	O
of	O
overfitting	B-Error_Name
)	O
.	O
</s>
<s>
AlphaFold	B-Application
2	I-Application
replaced	O
this	O
with	O
a	O
system	O
of	O
sub-networks	O
coupled	O
together	O
into	O
a	O
single	O
differentiable	O
end-to-end	O
model	O
,	O
based	O
entirely	O
on	O
pattern	O
recognition	O
,	O
which	O
was	O
trained	O
in	O
an	O
integrated	O
way	O
as	O
a	O
single	O
integrated	O
structure	O
.	O
</s>
<s>
Local	O
physics	O
,	O
in	O
the	O
form	O
of	O
energy	O
refinement	O
based	O
on	O
the	O
AMBER	B-Device
model	O
,	O
is	O
applied	O
only	O
as	O
a	O
final	O
refinement	O
step	O
once	O
the	O
neural	O
network	O
prediction	O
has	O
converged	O
,	O
and	O
only	O
slightly	O
adjusts	O
the	O
predicted	O
structure	O
.	O
</s>
<s>
A	O
key	O
part	O
of	O
the	O
2020	O
system	O
are	O
two	O
modules	O
,	O
believed	O
to	O
be	O
based	O
on	O
a	O
transformer	B-Algorithm
design	O
,	O
which	O
are	O
used	O
to	O
progressively	O
refine	O
a	O
vector	B-General_Concept
of	I-General_Concept
information	I-General_Concept
for	O
each	O
relationship	O
(	O
or	O
"	O
edge	O
"	O
in	O
graph-theory	O
terminology	O
)	O
between	O
an	O
amino	O
acid	O
residue	O
of	O
the	O
protein	O
and	O
another	O
amino	O
acid	O
residue	O
(	O
these	O
relationships	O
are	O
represented	O
by	O
the	O
array	O
shown	O
in	O
green	O
)	O
;	O
and	O
between	O
each	O
amino	O
acid	O
position	O
and	O
each	O
different	O
sequences	O
in	O
the	O
input	O
sequence	B-Algorithm
alignment	I-Algorithm
(	O
these	O
relationships	O
are	O
represented	O
by	O
the	O
array	O
shown	O
in	O
red	O
)	O
.	O
</s>
<s>
Internally	O
these	O
refinement	O
transformations	O
contain	O
layers	O
that	O
have	O
the	O
effect	O
of	O
bringing	O
relevant	O
data	O
together	O
and	O
filtering	O
out	O
irrelevant	O
data	O
(	O
the	O
"	O
attention	B-General_Concept
mechanism	I-General_Concept
"	O
)	O
for	O
these	O
relationships	O
,	O
in	O
a	O
context-dependent	O
way	O
,	O
learnt	O
from	O
training	O
data	O
.	O
</s>
<s>
The	O
output	O
of	O
these	O
iterations	O
then	O
informs	O
the	O
final	O
structure	O
prediction	O
module	O
,	O
which	O
also	O
uses	O
transformers	B-Algorithm
,	O
and	O
is	O
itself	O
then	O
iterated	O
.	O
</s>
<s>
In	O
an	O
example	O
presented	O
by	O
DeepMind	B-Application
,	O
the	O
structure	O
prediction	O
module	O
achieved	O
a	O
correct	O
topology	O
for	O
the	O
target	O
protein	O
on	O
its	O
first	O
iteration	O
,	O
scored	O
as	O
having	O
a	O
GDT_TS	O
of	O
78	O
,	O
but	O
with	O
a	O
large	O
number	O
(	O
90%	O
)	O
of	O
stereochemical	O
violations	O
–	O
i.e.	O
</s>
<s>
DeepMind	B-Application
stated	O
this	O
update	O
succeeded	O
about	O
70%	O
of	O
the	O
time	O
at	O
accurately	O
predicting	O
protein-protein	O
interactions	O
.	O
</s>
<s>
In	O
December	O
2018	O
,	O
DeepMind	B-Application
's	O
AlphaFold	O
placed	O
first	O
in	O
the	O
overall	O
rankings	O
of	O
the	O
13th	O
Critical	O
Assessment	O
of	O
Techniques	O
for	O
Protein	O
Structure	O
Prediction	O
(	O
CASP	O
)	O
.	O
</s>
<s>
AlphaFold	O
gave	O
the	O
best	O
prediction	O
for	O
25	O
out	O
of	O
43	O
protein	O
targets	O
in	O
this	O
class	O
,	O
achieving	O
a	O
median	O
score	O
of	O
58.9	O
on	O
the	O
CASP	O
's	O
global	O
distance	O
test	O
(	O
GDT	O
)	O
score	O
,	O
ahead	O
of	O
52.5	O
and	O
52.4	O
by	O
the	O
two	O
next	O
best-placed	O
teams	O
,	O
who	O
were	O
also	O
using	O
deep	B-Algorithm
learning	I-Algorithm
to	O
estimate	O
contact	O
distances	O
.	O
</s>
<s>
In	O
January	O
2020	O
,	O
implementations	O
and	O
illustrative	O
code	O
of	O
AlphaFold	O
1	O
was	O
released	O
open-source	B-Application
on	O
GitHub	B-Application
.	O
</s>
<s>
The	O
feature	O
generation	O
code	O
is	O
tightly	O
coupled	O
to	O
our	O
internal	O
infrastructure	O
as	O
well	O
as	O
external	O
tools	O
,	O
hence	O
we	O
are	O
unable	O
to	O
open-source	B-Application
it.	O
"	O
</s>
<s>
In	O
November	O
2020	O
,	O
DeepMind	B-Application
's	O
new	O
version	O
,	O
AlphaFold	B-Application
2	I-Application
,	O
won	O
CASP14	O
.	O
</s>
<s>
Overall	O
,	O
AlphaFold	B-Application
2	I-Application
made	O
the	O
best	O
prediction	O
for	O
88	O
out	O
of	O
the	O
97	O
targets	O
.	O
</s>
<s>
On	O
the	O
group	O
of	O
targets	O
classed	O
as	O
the	O
most	O
difficult	O
,	O
AlphaFold	B-Application
2	I-Application
achieved	O
a	O
median	O
score	O
of	O
87	O
.	O
</s>
<s>
Measured	O
by	O
the	O
root-mean-square	B-General_Concept
deviation	I-General_Concept
(	O
RMS-D	O
)	O
of	O
the	O
placement	O
of	O
the	O
alpha-carbon	O
atoms	O
of	O
the	O
protein	O
backbone	O
chain	O
,	O
which	O
tends	O
to	O
be	O
dominated	O
by	O
the	O
performance	O
of	O
the	O
worst-fitted	O
outliers	O
,	O
88%	O
of	O
AlphaFold	B-Application
2	I-Application
's	O
predictions	O
had	O
an	O
RMS	O
deviation	O
of	O
less	O
than	O
4	O
Å	O
for	O
the	O
set	O
of	O
overlapped	O
C-alpha	O
atoms	O
.	O
</s>
<s>
AlphaFold	B-Application
2	I-Application
also	O
achieved	O
an	O
accuracy	O
in	O
modelling	O
surface	O
side	O
chains	O
described	O
as	O
"	O
really	O
really	O
extraordinary	O
"	O
.	O
</s>
<s>
In	O
all	O
four	O
cases	O
the	O
three-dimensional	O
models	O
produced	O
by	O
AlphaFold	B-Application
2	I-Application
were	O
sufficiently	O
accurate	O
to	O
determine	O
structures	O
of	O
these	O
proteins	O
by	O
molecular	O
replacement	O
.	O
</s>
<s>
Of	O
the	O
three	O
structures	O
that	O
AlphaFold	B-Application
2	I-Application
had	O
the	O
least	O
success	O
in	O
predicting	O
,	O
two	O
had	O
been	O
obtained	O
by	O
protein	O
NMR	O
methods	O
,	O
which	O
define	O
protein	O
structure	O
directly	O
in	O
aqueous	O
solution	O
,	O
whereas	O
AlphaFold	O
was	O
mostly	O
trained	O
on	O
protein	O
structures	O
in	O
crystals	O
.	O
</s>
<s>
For	O
all	O
targets	O
with	O
a	O
single	O
domain	O
,	O
excluding	O
only	O
one	O
very	O
large	O
protein	O
and	O
the	O
two	O
structures	O
determined	O
by	O
NMR	O
,	O
AlphaFold	B-Application
2	I-Application
achieved	O
a	O
GDT_TS	O
score	O
of	O
over	O
80	O
.	O
</s>
<s>
In	O
2022	O
DeepMind	B-Application
did	O
not	O
enter	O
CASP15	O
,	O
but	O
most	O
of	O
the	O
entrants	O
used	O
AlphaFold	O
or	O
tools	O
incorporating	O
AlphaFold	O
.	O
</s>
<s>
AlphaFold	B-Application
2	I-Application
scoring	O
more	O
than	O
90	O
in	O
CASP	O
's	O
global	O
distance	O
test	O
(	O
GDT	O
)	O
is	O
considered	O
a	O
significant	O
achievement	O
in	O
computational	O
biology	O
and	O
great	O
progress	O
towards	O
a	O
decades-old	O
grand	O
challenge	O
of	O
biology	O
.	O
</s>
<s>
Propelled	O
by	O
press	O
releases	O
from	O
CASP	O
and	O
DeepMind	B-Application
,	O
AlphaFold	B-Application
2	I-Application
's	O
success	O
received	O
wide	O
media	O
attention	O
.	O
</s>
<s>
Writing	O
about	O
the	O
event	O
,	O
the	O
MIT	O
Technology	O
Review	O
noted	O
that	O
the	O
AI	B-Application
had	O
"	O
solved	O
a	O
fifty-year	O
old	O
grand	O
challenge	O
of	O
biology.	O
"	O
</s>
<s>
The	O
same	O
article	O
went	O
on	O
to	O
note	O
that	O
the	O
AI	B-Application
algorithm	O
could	O
"	O
predict	O
the	O
shape	O
of	O
proteins	O
to	O
within	O
the	O
width	O
of	O
an	O
atom.	O
"	O
</s>
<s>
Although	O
a	O
30-minute	O
presentation	O
about	O
AlphaFold	B-Application
2	I-Application
was	O
given	O
on	O
the	O
second	O
day	O
of	O
the	O
CASP	O
conference	O
(	O
December	O
1	O
)	O
by	O
project	O
leader	O
John	O
Jumper	O
,	O
it	O
has	O
been	O
described	O
as	O
"	O
exceedingly	O
high-level	O
,	O
heavy	O
on	O
ideas	O
and	O
insinuations	O
,	O
but	O
almost	O
entirely	O
devoid	O
of	O
detail	O
"	O
.	O
</s>
<s>
Unlike	O
other	O
research	O
groups	O
presenting	O
at	O
CASP14	O
,	O
DeepMind	B-Application
's	O
presentation	O
was	O
not	O
recorded	O
and	O
is	O
not	O
publicly	O
available	O
.	O
</s>
<s>
DeepMind	B-Application
is	O
expected	O
to	O
publish	O
a	O
scientific	O
paper	O
giving	O
an	O
account	O
of	O
AlphaFold	B-Application
2	I-Application
in	O
the	O
proceedings	O
volume	O
of	O
the	O
CASP	O
conference	O
;	O
but	O
it	O
is	O
not	O
known	O
whether	O
it	O
will	O
go	O
beyond	O
what	O
was	O
said	O
in	O
the	O
presentation	O
.	O
</s>
<s>
Nevertheless	O
,	O
as	O
much	O
as	O
Google	O
and	O
DeepMind	B-Application
do	O
release	O
may	O
help	O
other	O
teams	O
develop	O
similar	O
AI	B-Application
systems	O
,	O
an	O
"	O
indirect	O
"	O
benefit	O
.	O
</s>
<s>
In	O
late	O
2019	O
DeepMind	B-Application
released	O
much	O
of	O
the	O
code	O
of	O
the	O
first	O
version	O
of	O
AlphaFold	O
as	O
open	O
source	O
;	O
but	O
only	O
when	O
work	O
was	O
well	O
underway	O
on	O
the	O
much	O
more	O
radical	O
AlphaFold	B-Application
2	I-Application
.	O
</s>
<s>
Another	O
option	O
it	O
could	O
take	O
might	O
be	O
to	O
make	O
AlphaFold	B-Application
2	I-Application
structure	O
prediction	O
available	O
as	O
an	O
online	O
black-box	O
subscription	O
service	O
.	O
</s>
<s>
Convergence	O
for	O
a	O
single	O
sequence	O
has	O
been	O
estimated	O
to	O
require	O
on	O
the	O
order	O
of	O
$	O
10,000	O
worth	O
of	O
wholesale	B-Device
compute	I-Device
time	I-Device
.	O
</s>
<s>
But	O
this	O
would	O
deny	O
researchers	O
access	O
to	O
the	O
internal	O
states	O
of	O
the	O
system	O
,	O
the	O
chance	O
to	O
learn	O
more	O
qualitatively	O
what	O
gives	O
rise	O
to	O
AlphaFold	B-Application
2	I-Application
's	O
success	O
,	O
and	O
the	O
potential	O
for	O
new	O
algorithms	O
that	O
could	O
be	O
lighter	O
and	O
more	O
efficient	O
yet	O
still	O
achieve	O
such	O
results	O
.	O
</s>
<s>
Fears	O
of	O
potential	O
for	O
a	O
lack	O
of	O
transparency	O
by	O
DeepMind	B-Application
have	O
been	O
contrasted	O
with	O
five	O
decades	O
of	O
heavy	O
public	O
investment	O
into	O
the	O
open	O
Protein	O
Data	O
Bank	O
and	O
then	O
also	O
into	O
open	O
DNA	O
sequence	O
repositories	O
,	O
without	O
which	O
the	O
data	O
to	O
train	O
AlphaFold	B-Application
2	I-Application
would	O
not	O
have	O
existed	O
.	O
</s>
<s>
We	O
’ve	O
been	O
heads	O
down	O
working	O
flat	O
out	O
on	O
our	O
full	O
methods	O
paper	O
(	O
currently	O
under	O
review	O
)	O
with	O
accompanying	O
open	B-Application
source	I-Application
code	I-Application
and	O
on	O
providing	O
broad	O
free	O
access	O
to	O
AlphaFold	O
for	O
the	O
scientific	O
community	O
.	O
</s>
<s>
However	O
it	O
is	O
not	O
yet	O
clear	O
to	O
what	O
extent	O
structure	O
predictions	O
made	O
by	O
AlphaFold	B-Application
2	I-Application
will	O
hold	O
up	O
for	O
proteins	O
bound	O
into	O
complexes	O
with	O
other	O
proteins	O
and	O
other	O
molecules	O
.	O
</s>
<s>
Where	O
structures	O
that	O
AlphaFold	B-Application
2	I-Application
did	O
predict	O
were	O
for	O
proteins	O
that	O
had	O
strong	O
interactions	O
either	O
with	O
other	O
copies	O
of	O
themselves	O
,	O
or	O
with	O
other	O
structures	O
,	O
these	O
were	O
the	O
cases	O
where	O
AlphaFold	B-Application
2	I-Application
's	O
predictions	O
tended	O
to	O
be	O
least	O
refined	O
and	O
least	O
reliable	O
.	O
</s>
<s>
With	O
so	O
little	O
yet	O
known	O
about	O
the	O
internal	O
patterns	O
that	O
AlphaFold	B-Application
2	I-Application
learns	O
to	O
make	O
its	O
predictions	O
,	O
it	O
is	O
not	O
yet	O
clear	O
to	O
what	O
extent	O
the	O
program	O
may	O
be	O
impaired	O
in	O
its	O
ability	O
to	O
identify	O
novel	O
folds	O
,	O
if	O
such	O
folds	O
are	O
not	O
well	O
represented	O
in	O
the	O
existing	O
protein	O
structures	O
known	O
in	O
structure	O
databases	O
.	O
</s>
<s>
AlphaFold	B-Application
2	I-Application
's	O
difficulties	O
with	O
structures	O
obtained	O
by	O
protein	O
NMR	O
methods	O
may	O
not	O
be	O
a	O
good	O
sign	O
.	O
</s>
<s>
On	O
its	O
potential	O
as	O
a	O
tool	O
for	O
drug	O
discovery	O
,	O
Stephen	O
Curry	O
notes	O
that	O
while	O
the	O
resolution	O
of	O
AlphaFold	B-Application
2	I-Application
's	O
structures	O
may	O
be	O
very	O
good	O
,	O
the	O
accuracy	O
with	O
which	O
binding	O
sites	O
are	O
modelled	O
needs	O
to	O
be	O
even	O
higher	O
:	O
typically	O
molecular	O
docking	O
studies	O
require	O
the	O
atomic	O
positions	O
to	O
be	O
accurate	O
within	O
a	O
0.3	O
Å	O
margin	O
,	O
but	O
the	O
predicted	O
protein	O
structure	O
only	O
have	O
at	O
best	O
an	O
RMSD	B-General_Concept
of	O
0.9	O
Å	O
for	O
all	O
atoms	O
.	O
</s>
<s>
So	O
AlphaFold	B-Application
2	I-Application
's	O
structures	O
may	O
only	O
be	O
a	O
limited	O
help	O
in	O
such	O
contexts	O
.	O
</s>
<s>
But	O
even	O
with	O
such	O
caveats	O
,	O
AlphaFold	B-Application
2	I-Application
was	O
described	O
as	O
a	O
huge	O
technical	O
step	O
forward	O
and	O
intellectual	O
achievement	O
.	O
</s>
<s>
The	O
AlphaFold	O
Multimer	O
model	O
is	O
published	O
separately	O
as	O
open-source	B-Application
,	O
but	O
pre-run	O
models	O
are	O
not	O
available	O
.	O
</s>
<s>
Specifically	O
,	O
AlphaFold	B-Application
2	I-Application
's	O
prediction	O
of	O
the	O
structure	O
of	O
the	O
ORF3a	O
protein	O
was	O
very	O
similar	O
to	O
the	O
structure	O
determined	O
by	O
researchers	O
at	O
University	O
of	O
California	O
,	O
Berkeley	O
using	O
cryo-electron	O
microscopy	O
.	O
</s>
