Feature
IAN C. HAYDON/UW INSTITUTE FOR PROTEIN DESIGN
Two protein assemblies (right) were developed using an artificial-intelligence tool called RFdiffusion.
‘TRANSFORMATIVE’ AI DESIGNS
CUSTOM PROTEINS ON DEMAND
Computer-devised biomolecules could form the basis of new
vaccines or medicines. By Ewen Callaway
“O
K. Here we go.” David Juergens, Gevorg Grigoryan, the co-founder and chief perform as the software suggests.
a computational chemist at the technical officer of Generate Biomedicines The tools have revolutionized the process of
University of Washington (UW) in Somerville, Massachusetts, a biotechnol- designing proteins in the past year, research-
in Seattle, is about to design a ogy company applying protein design to drug ers say. “It is an explosion in capabilities,” says
protein that, in 3-billion-plus development. Mohammed AlQuraishi, a computational biol-
years of tinkering, evolution The tools are inspired by AI software that ogist at Columbia University in New York City,
has never produced. synthesizes realistic images, such as the whose team has developed one such tool for
On a video call, Juergens opens a cloud- Midjourney software that, this year, was protein design. “You can now create designs
based version of an artificial intelligence (AI) famously used to produce a viral image of that have sought-after qualities.”
tool he helped to develop, called RFdiffusion. Pope Francis wearing a designer white puffer “You’re building a protein structure
This neural network, and others like it, are jacket. A similar conceptual approach, customized for a problem,” says David Baker,
helping to bring the creation of custom pro- researchers have found, can churn out realistic a computational biophysicist at UW whose
teins — until recently a highly technical and protein shapes to criteria that designers spec- group, which includes Juergens, developed
often unsuccessful pursuit — to mainstream ify — meaning, for instance, that it’s possible RFdiffusion. The team released the software
science. to speedily draw up new proteins that should in March 2023, and a paper describing the
These proteins could form the basis for vac- bind tightly to another biomolecule. And early neural network appears this week in Nature1.
cines, therapeutics and biomaterials. “It’s been experiments show that when researchers man- (A preprint version was released in late 2022,
a completely transformative moment,” says ufacture these proteins, a useful fraction do at around the same time that several other
236 | Nature | Vol 619 | 13 July 2023
©
2
0
2
3
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
teams, including AlQuraishi’s and Grigoryan’s, a theoretical protein structure, it wasn’t able to The tool has gained widespread use in
reported similar neural networks2,3). come up with diverse solutions to a problem Baker’s laboratory. “The design process is
For the first time, protein designers now that would increase the odds of success. almost unrecognizable compared to a year
have the kinds of reproducible and robust That is where RFdiffusion and similar ago,” he says. The neural network has excelled
tools around which a new industry can be protein-designing AIs, released in recent in design challenges that have been inefficient,
created, Grigoryan adds. “The next challenge months, come in. They are based on the same difficult or impossible using other approaches.
becomes, what do you do with it?” principles as neural networks that generate In one analysis reported in their study1,
realistic images, such as Stable Diffusion, the researchers started with a snippet from
Grand designs DALL-E and Midjourney. These ‘diffusion’ another protein, such as a portion of a viral
Juergens inputs a few specifications for the networks are trained on data, be they images protein recognized by immune cells, and
protein he wants into a web form resembling or protein structures, which are then made tasked AI-based tools with churning out
an online tax calculator. It must be 100 amino progressively noisier, eventually bearing no 100 different new proteins, to see how many
acids long and form a symmetrical two-protein resemblance to the starting image or struc- would incorporate the desired motif. The
complex called a homodimer. Many cell recep- ture. The network then learns to ‘denoise’ the team carried out this challenge for 25 different
tors adopt this configuration, and a new data, performing the task in reverse. initial shapes. The results didn’t always incor-
homodimer could be a synthetic cell-signalling Networks such as RFdiffusion are trained on porate the starting snippet, but RFdiffusion
molecule, chimes in Joe Watson, a UW com- tens of thousands of real protein structures produced at least one protein that did for 23 of
putational biochemist who co-developed stored in a repository called the Protein Data the motifs, compared with 15 for hallucination
RFdiffusion, and is also on the video call. But Bank (PDB). When the network makes a new and 12 for inpainting.
this morning’s design isn’t meant to do any- protein, it begins with total noise: a random RFdiffusion has also proved adept at mak-
thing except resemble a realistic protein. assortment of amino acids. “You’re asking ing proteins that self-assemble into complex
Researchers have struggled for decades to what is the protein that gave rise to the noise,” nanoparticles that might be able to deliver
build new proteins. At first, they tried to cobble explains Watson. After rounds of denoising, it drugs or vaccine components. Previous AI
together useful parts of existing proteins, such produces something resembling a real — but approaches5 can also make these kinds of pro-
as a pocket of an enzyme in which a chemical new — protein. tein, but Watson says RFdiffusion’s designs are
reaction is catalysed. This approach relied on When Baker’s team tested RFdiffusion with- much more sophisticated.
understanding how proteins fold up and work, out providing any guidance except the length Neural networks such as RFdiffusion seem
as well as intuition and a lot of trial and error. of the protein, the network generated diverse, to really shine when tasked with designing
Scientists sometimes screened thousands of proteins that can stick to another specified
designs to identify one that worked as hoped. “The design process protein. Baker’s team has used the network to
A light-bulb moment came with AlphaFold create proteins that bind strongly to proteins
(developed by the London-based AI firm
is almost unrecognizable implicated in cancers, autoimmune diseases
DeepMind, now Google DeepMind) and compared to a year ago.” and other conditions. One as-yet unpublished
other AI-based models that could accurately success, he says, was to design strong bind-
predict protein structures from amino-acid ers for a hard-to-target immune-signalling
sequences, says Baker. Designers realized that realistic-looking proteins, different from any- molecule called the tumour necrosis factor
these neural networks, trained on real protein thing it had been trained on in the PDB. receptor — the target for antibody drugs that
sequences and structures, could also help to But the researchers are also able to direct generate billions of dollars in revenue each
create proteins from scratch. the program to make proteins according to year. “It is broadening the space of proteins
In the past few years, Baker’s team and specific design constraints during the denois- we can make binders to and make meaningful
others in the field have released a slew of ing process, a process called conditioning. therapies” for, Watson says.
AI-based protein-design tools (Nature 609, For instance, Baker’s team conditioned
661–662; 2022). One approach these tools RFdiffusion to make proteins that include a Real-world testing
use, called hallucination, involves creating a specific fold, or that can nestle against the Baker’s team is cranking out so many designs
random string of amino acids that is then opti- surface of another molecule (an interaction that testing whether they work as intended
mized by AlphaFold, or a similar tool called that underlies binding). Grigoryan’s team has become a serious bottleneck. “One
RoseTTAFold, until it resembles something even developed a diffusion network called machine-learning person can generate
that the neural network suggests is likely to Chroma and then conditioned it to make pro- enough designs to keep 100 biologists busy
fold into a specific structure. Another, called teins shaped to resemble the 26 capital letters for months,” says Kevin Yang, a biomedical
inpainting, takes a specified snippet of a pro- used in English, as well the Arabic numerals3. machine-learning researcher at Microsoft
tein sequence or structure and builds the rest Research in Cambridge, Massachusetts whose
of the molecule around it using RoseTTAFold. Signal from noise team has developed its own diffusion-based
But these tools are far from perfect. Juergens’ computer screen initially shows protein design tool6.
Experiments tended to show that structures noise, the random assortment of amino acids But early signs suggest that RFdiffusion’s
designed by hallucination methods didn’t that the AI system starts with. They are repre- creations are the real deal. In another challenge
always form well-folded proteins when they sented as red, smudgy squiggles that resemble described in their study, Baker’s team tasked
were made in the laboratory, and ended up as a toddler’s fingerpainting. They morph, frame the tool with designing proteins containing a
gunk at the bottom of a test tube, for instance. by frame, into ever-more-complex shapes, key stretch of p53, a signalling molecule that is
Hallucination methods also struggled to make with protein-like features such as tight spirals overactive in many cancers (and a sought-after
anything but small proteins (although other known as α-helices and ribbony shapes that drug target). When the researchers made
researchers showed, in a February preprint, double back on themselves, called β-sheets. 95 of the software’s designs (by engineer-
how the technique could be used to design “It’s a nice mixed alpha–beta topology,” says ing bacteria to express the proteins), more
longer molecules4). Inpainting also did a poor Juergens, smiling as he admires a creation than half maintained p53’s ability to bind to
job of forming proteins when given shorter that took only a few minutes to make. “This its natural target, MDM2. The best designs
snippets. Even when the approach did produce is looking good.” did so around 1,000 times more strongly
Nature | Vol 619 | 13 July 2023 | 237
©
2
0
2
3
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Feature
room to grow,” Yang says.
IAN C. HAYDON/UW INSTITUTE FOR PROTEIN DESIGN
The latest protein-design tools have proved
to be extremely powerful at creating pro-
teins that can do a particular task — so long
as that function can be described in terms of
a shape, such as the surface of a protein to
bind to, says AlQuraishi. But, he adds, tools
such as RFdiffusion aren’t yet able to handle
other kinds of specifications, such as making
a protein that can carry out a particular reac-
tion regardless of its shape — when “you know
what you want but you don’t know what the
RFdiffusion generated a protein that binds to the parathyroid hormone, shown in pink. geometry is”.
Future protein-design tools will also need
than did natural p53. When the researchers binding proteins such as antibodies, or the the capacity to churn out proteins to numer-
attempted this task with hallucination, the protein receptors used by T cells (a type of ous different criteria, says Grigoryan. A
designs — although predicted to work — did immune cell). These proteins have flexible potential therapeutic protein must not only
not pan out in the test tube, says Watson. loops that interlock with their targets, as bind to its target, but also not bind to others
Overall, Baker says his team has found that opposed to the sandwich-like, flat interfaces and should possess properties that make it
10–20% of RFdiffusion’s designs bind to their that RFdiffusion has excelled at so far. Baker easy to mass-produce.
intended target strongly enough to be use- says they are making progress with antibodies. One direction that researchers are exploring
ful, compared with less than 1% for earlier, Ovchinnikov and others say it’s challeng- is whether proteins could be designed using
pre-AI methods. (Previous machine-learning ing, in general, to design biomolecules whose plain language text descriptions, similar to the
approaches were not able to reliably design function depends on floppy regions that give prompts fed to image-generation tools such as
binders, Watson says). Biochemist Matthias them the ability to adopt many different Midjourney. “You can really imagine we will be
Gloegl, a colleague at UW, says that lately he shapes. These are features that have proved able to write descriptions of a protein and have
has been hitting success rates approaching difficult to model using AI. “If the problem them synthesized and tested,” says Watson.
50%, which means it can take just a week or two is, can we bind to something else and inhibit Grigoryan and his colleagues have taken
to come up with working designs, as opposed it,” says Ovchinnikov, “I think that problem is a step towards this goal. In their December
to months. “It’s really insane,” he says. going to be solved with these methods. But in 2022 preprint3, they trained Chroma to attach
The cloud-based version of RFdiffusion had order to do something more complex, more descriptions to its designs and spit out designs
around 100 users each day by late June, accord- like what nature does, you need to introduce to text-based specifications, including ‘protein
ing to Sergey Ovchinnikov, an evolutionary some flexibility.” with a CHAD domain’ (a protein shape incor-
biologist at Harvard University in Cambridge, Tanja Kortemme, a computational biolo- porating multiple helices) or ‘crystal structure
Massachusetts. Joel Mackay, a biochemist at gist at the University of California, San Fran- of aminotransferases’ (enzymes involved in
the University of Sydney in Australia, has been cisco, is using RFdiffusion to design proteins making and breaking down proteins).
dabbling with RFdiffusion to design proteins that can be used as sensors or as switches to The protein Juergens created in a few
capable of binding to other proteins that his minutes this morning is only a model of a
lab studies, which include molecules called “You can really imagine protein’s 3D structure. Juergens then uses
transcription factors that control gene activity another AI tool to come up with sequences of
in cells. He found the design process simple,
we will be able to write amino acids that should fold up into that struc-
and used computer modelling to validate that, descriptions of a protein ture. As a final check, he plugs the sequences
in theory, the proteins should bind to the tran- and have them synthesized.” into AlphaFold to see whether the software
scription factors. predicts folded structures that match the
Mackay is now testing whether the proteins design. They’re spot on, with the AlphaFold
can alter gene expression as intended when control cells. She says that if a protein’s active predictions differing from the design by an
they are produced in cells. He has his fingers site depends on the placement of a few amino average of just 1 ångström (the width of a
crossed, because such a finding would amount acids, the AI network does well, but it strug- hydrogen atom).
to a simple way to switch specific transcription gles to design proteins with more-complex “This is at the accuracy that we would class
factors on and off within cells, instead of using active sites, requiring many more key amino as a design success,” says Watson. The only
drugs that can take years to identify, if they acids to be in place — a challenge she and her thing left to do, he says, is to see how the
can be discovered at all. “If this method works colleagues are trying to tackle. protein performs in real life.
reliably for our types of proteins, it would be a Another limitation of the latest diffusion
total game-changer,” he says. methods is their inability to create proteins Ewen Callaway is a senior reporter for Nature
that are vastly different from natural proteins, in Bristol, UK.
Future improvements says Yang. That is because the AI systems have
The latest models such as RFdiffusion are been trained only on existing proteins that
1. Watson, J. L. et al. Nature https://doi.org/10.1038/s41586-
a “step change” says Charlotte Deane, an scientists have characterized, he says, and 023-06415-8 (2023).
immune informatician at the University of tend to create proteins that resemble those. 2. Lin, Y. & AlQuraishi, M. Preprint at https://arxiv.org/
Oxford, UK. But key challenges remain. “What Generating more-alien-looking proteins might abs/2301.12485 (2023).
3. Ingraham, J. et al. Preprint at bioRxiv https://doi.
it will do is inspire people to see how far we require a better understanding of the physics org/10.1101/2022.12.01.518682 (2022).
can push these diffusion methods,” she says. that imbues proteins with their function. 4. Frank, C. et al. Preprint at bioRxiv https://doi.
One application that she and other scientists That could make it easier to design org/10.1101/2023.02.24.529906 (2023).
5. Wicky, B. I. M. et al. Science 378, 56–61 (2022).
and biotechnology companies are particu- proteins to carry out tasks no natural protein 6. Wu, K. E. Preprint at https://arxiv.org/abs/2209.15611
larly interested in is designing more complex has ever evolved to do. “There’s still a lot of (2022).
238 | Nature | Vol 619 | 13 July 2023
©
2
0
2
3
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
