Work / Technology & tools

ILLUSTRATION: THE PROJECT TWINS

SEVEN TECHNOLOGIES
TO WATCH IN 2024
Advances in artificial intelligence are at the heart of many of this year’s most
exciting areas of technological innovation. By Michael Eisenstein

F
rom protein engineering and 3D print- protein-based vaccines and vehicles for drug an algorithm called ProtGPT2 that consist-
ing to detection of deepfake media, delivery. “Things that were impossible a year ently comes up with synthetic proteins that
here are seven areas of technology and a half ago — now you just do it.” fold stably when produced in the laboratory1.
that Nature will be watching in the Much of that progress comes down to Another tool co-developed by Ferruz, called
year ahead. increasingly massive data sets that link pro- ZymCTRL, draws on sequence and functional
tein sequence to structure. But sophisticated data to design members of naturally occurring
Deep learning for protein design methods of deep learning, a form of artificial enzyme families2.
Two decades ago, David Baker at the University intelligence (AI), have also been essential. Sequence-based approaches can build on
of Washington in Seattle and his colleagues ‘Sequence based’ strategies use the large lan- and adapt existing protein features to form
achieved a landmark feat: they used computa- guage models (LLMs) that power tools such as new frameworks, but they’re less effective for
tional tools to design an entirely new protein the chatbot ChatGPT (see ‘ChatGPT? Maybe the bespoke design of structural elements or
from scratch. ‘Top7’ folded as predicted, but it next year’). By treating protein sequences like features, such as the ability to bind specific tar-
was inert: it performed no meaningful biolog- documents comprising polypeptide ‘words’, gets in a predictable fashion. ‘Structure based’
ical functions. Today, de novo protein design these algorithms can discern the patterns that approaches are better for this, and 2023 saw
has matured into a practical tool for gener- underlie the architectural playbook of real- notable progress in this type of protein-design
ating made-to-order enzymes and other pro- world proteins. “They really learn the hidden algorithm, too. Some of the most sophisti-
teins. “It’s hugely empowering,” says Neil King, grammar,” says Noelia Ferruz, a protein bio- cated of these use ‘diffusion’ models, which
a biochemist at the University of Washington chemist at the Molecular Biology Institute of also underlie image-generating tools such as
who collaborates with Baker’s team to design Barcelona, Spain. In 2022, her team developed DALL-E. These algorithms are initially trained to

844 | Nature | Vol 625 | 25 January 2024

 remove computer-generated noise from large And then there’s the challenge of implemen- Other methods exploit a CRISPR-based
numbers of real structures; by learning to dis- tation. The US Defense Advanced Research method called prime editing to introduce short
criminate realistic structural elements from Projects Agency’s Semantic Forensics (Sema- ‘landing pad’ sequences that selectively recruit
noise, they gain the ability to form biologically For) programme has developed a useful tool- enzymes that in turn can precisely splice large
plausible, user-defined structures. box for deepfake analysis, but, as reported in DNA fragments into the genome. In 2022, for
RFdiffusion software3 developed by Baker’s Nature (see Nature 621, 676–679; 2023) major instance, genome engineers Omar Abudayyeh
lab and the Chroma tool by Generate Biomed- social-media sites are not routinely employing and Jonathan Gootenberg at the Massachusetts
icines in Somerville, Massachusetts4, exploit it. Broadening the access to such tools could Institute of Technology, Cambridge and their
this strategy to remarkable effect. For exam- help to fuel uptake, and to this end Lyu’s team colleagues first described programmable
ple, Baker’s team is using RFdiffusion to has developed the DeepFake-O-Meter7, a cen- addition through site-specific targeting ele-
engineer novel proteins that can form snug tralized public repository of algorithms that ments (PASTE), a method that can precisely
interfaces with targets of interest, yielding can analyse video content from different insert up to 36 kilobases of DNA8. PASTE is
designs that “just conform perfectly to the sur- angles to sniff out deepfake content. Such especially promising for ex vivo modification
face,” Baker says. A newer ‘all atom’ iteration of resources will be helpful, but it is likely that of cultured, patient-derived cells, says Cong,
RFdiffusion5 allows designers to computation- the battle against AI-generated misinforma- and the underlying prime-editing technology
ally shape proteins around non-protein targets tion will persist for years to come. is already on track for clinical studies. But for in
such as DNA, small molecules and even metal vivo modification of human cells, SSAP might
ions. The resulting versatility opens new hori- Large-fragment DNA insertion offer a more compact solution: the bulkier
zons for engineered enzymes, transcriptional In late 2023, US and UK regulators approved PASTE machinery requires three separate viral
regulators, functional biomaterials and more. the first-ever CRISPR-based gene-edit- vectors for delivery, which could undermine
ing therapy for sickle-cell disease and editing efficiency relative to the two-compo-
Deepfake detection transfusion-dependent β-thalassaemia — a nent SSAP system. That said, even relatively
The explosion of publicly available generative major win for genome editing as a clinical tool. inefficient gene-replacement strategies could
AI algorithms has made it simple to synthe- CRISPR and its derivatives use a short be sufficient to mitigate the effects of many
size convincing, but entirely artificial images, programmable RNA to direct a DNA-cutting genetic diseases.
audio and video. The results can offer amus- enzyme such as Cas9 to a specific genomic And such methods are not just relevant
ing distractions, but with multiple ongoing site. They are routinely used in the lab to to human health. Researchers led by Caixia
geopolitical conflicts and a US presidential disable defective genes and introduce small Gao at the Chinese Academy of Sciences in
election on the horizon, opportunities for sequence changes. The precise and program- Beijing developed PrimeRoot, a method that
weaponized media manipulation are rife. mable insertion of larger DNA sequences span- uses prime editing to introduce specific tar-
Siwei Lyu, a computer scientist at the Uni- ning thousands of nucleotides is difficult, but get sites that enzymes can use to insert up to
versity at Buffalo in New York, says he’s seen emerging solutions could allow scientists to 20 kilobases of DNA in both rice and maize9. Gao
numerous AI-generated ‘deepfake’ images and replace crucial segments of defective genes thinks that the technique could be broadly use-
audio related to the Israel–Hamas conflict, or insert fully functional gene sequences. Le ful for endowing crops with disease and patho-
for instance. This is just the latest round in a Cong, a molecular geneticist at Stanford Uni- gen resistance, continuing a wave of innovation
high-stakes game of cat-and-mouse in which versity in California and his colleagues are in CRISPR-based plant genome engineering. “I
AI users produce deceptive content and Lyu exploring single-stranded annealing proteins believe that this technology can be applied in
and other media-forensics specialists work (SSAPs) — virus-derived molecules that mediate any plant species,” she says.
to detect and intercept it. DNA recombination. When combined with a
One solution is for generative-AI develop- CRISPR–Cas system in which the DNA-slicing Brain–computer interfaces
ers to embed hidden signals in the models’ function of Cas9 has been disabled, these Pat Bennett has slower than average speech,
output, producing watermarks of AI-gener- SSAPs allow precisely targeted insertion of up and can sometimes use the wrong word. But
ated content. Other strategies focus on the to 2 kilobases of DNA into the human genome. given that motor neuron disease, also known
content itself. Some manipulated videos, for
instance, replace the facial features of one
public figure with those of another, and new
algorithms can recognize artefacts at the
boundaries of the substituted features, says
Lyu. The distinctive folds of a person’s outer
ear can also reveal mismatches between a
face and a head, whereas irregularities in
the teeth can reveal edited lip-sync videos in
which a person’s mouth was digitally manip-
ulated to say something that the subject
didn’t say. AI-generated photos also present
a thorny challenge — and a moving target. In
STEVE FISCH/STANFORD MEDICINE

2019, Luisa Verdoliva, a media-forensics spe-

cialist at University Federico II of Naples, Italy,
helped to develop FaceForensics++, a tool for
spotting faces manipulated by several widely
used software packages6. But image-forensic
methods are subject- and software-specific,
and generalization is a challenge. “You cannot
have one single universal detector — it’s very Brain–computer interface technology has allowed Pat Bennett (seated) to regain her speech.
difficult,” she says.

Nature | Vol 625 | 25 January 2024 | 845

Work / Technology & tools

ChatGPT?
Maybe next year

MAX IGLESIAS, MAX PLANCK INST. OF BIOCHEMISTRY

Why chatbots didn’t make the cut
this time.

Readers might detect a theme in this

year’s technologies to watch: the outsized
impact of deep-learning methods. But
one such tool did not make the final cut:
the much-hyped artificial-intelligence
(AI)-powered chatbots. ChatGPT and its
ilk seem poised to become part of many
A form of imaging called RESI could allow the imaging of individual base pairs in DNA. researchers’ daily routines and were feted
as part of the 2023 Nature’s 10 round-up
as amyotrophic lateral sclerosis, had previ- Edward Chang at the University of Califor- (see go.nature.com/3trp7rg). Respondents
ously left her unable to express herself ver- nia, San Francisco. In that work, a BCI neuro­ to a Nature survey in September (see
bally, that is a remarkable achievement. prosthesis allowed a woman who was unable to go.nature.com/45232vd) cited ChatGPT
Bennett’s recovery comes courtesy of a speak as a result of a stroke to communicate at as the most useful AI-based tool and were
sophisticated brain–computer interface (BCI) 78 words per minute — roughly half the average enthusiastic about its potential for coding,
device developed by Stanford University neu- speed of English, but more than five times faster literature reviews and administrative tasks.
roscientist Francis Willett and his colleagues at than the woman’s previous speech-assistance Such tools are also proving valuable
the US-based BrainGate consortium10. Willett device. The field is seeing progress in other from an equity perspective, helping those
and his colleagues implanted electrodes in areas as well. In 2021, Collinger and biomed- for whom English isn’t their first language
Bennett’s brain to track neuronal activity and ical engineer Robert Gaunt at the University to refine their prose and thereby ease their
then trained deep-learning algorithms to trans- of Pittsburgh implanted electrodes into the paths to publication and career growth.
late those signals into speech. After a few weeks motor and somatosensory cortex of an indi- However, many of these applications
of training, Bennett was able to say as many vidual who was paralysed in all four limbs represent labour-saving gains rather than
as 62 words per minute from a vocabulary of to provide rapid and precise control over a transformations of the research process.
125,000 words — more than twice the vocabu- robotic arm along with tactile sensory feed- Furthermore, ChatGPT’s persistent
lary of the average English speaker. “It’s really back12. Also under way are independent clini- issuing of either misleading or fabricated
truly impressive, the rates at which they’re cal studies from BrainGate and researchers at responses was the leading concern of more
communicating,” says bioengineer Jennifer UMC Utrecht in the Netherlands, as well as a than two-thirds of survey respondents.
Collinger, who develops BCI technologies at trial from BCI firm Synchron in Brooklyn, New Although worth monitoring, these tools
the University of Pittsburgh in Pennsylvania. York, to test a system that allows people who need time to mature and to establish their
BrainGate’s trial is just one of several studies are paralysed to control a computer — the first broader role in the scientific world.
from the past few years demonstrating how industry-sponsored trial of a BCI apparatus.
BCI technology can help people with severe As an intensive-care specialist, Hochberg really trying to close the gap from super-reso-
neurological damage to regain lost skills and is eager to deliver these technologies to his lution microscopy to structural-biology tech-
achieve greater independence. Some of that patients with the most severe disabilities. But niques like cryo-electron microscopy,” says
progress stems from the steady accumu- as BCI capabilities evolve, he sees potential to Ralf Jungmann, a nanotechnology researcher
lation of knowledge about functional neu- treat more-moderate cognitive impairments at the Max Planck Institute of Biochemistry
roanatomy in the brains of individuals with as well as mental-health conditions, such as in Planegg, Germany, referring to a method
various neurological conditions, says Leigh mood disorders. “Closed-loop neuromodu- that can reconstruct protein structures with
Hochberg, a neurologist at Brown University lation systems informed by brain–computer atomic-scale resolution.
in Providence, Rhode Island, and director of interfaces could be of tremendous help to a Researchers led by Hell and his team at the
the BrainGate consortium. But that knowl- lot of people,” he says. Max Planck Institute for Multidisciplinary
edge has been greatly amplified, he adds, by Sciences in Göttingen made an initial foray
machine-learning-driven analytical methods Super-duper resolution into this realm in late 2022 with a method
that are revealing how to better place elec- Stefan Hell, Eric Betzig and William Moerner called MINSTED that can resolve individual
trodes and decrypt the signals that they pick were awarded the 2014 Nobel Prize in Chem- fluorescent labels with 2.3-ångström precision
up. istry for shattering the ‘diffraction limit’ that — roughly one-quarter of a nanometre — using
Researchers are also applying AI-based lan- constrained the spatial resolution of light a specialized optical microscope13.
guage models to speed up the interpretation microscopy. The resulting level of detail — in Newer methods provide comparable res-
of what patients are trying to communicate — the order of tens of nanometres — opened a olution using conventional microscopes.
essentially, ‘autocomplete’ for the brain. This wide range of molecular-scale imaging exper- Jungmann and his team, for instance,
was a core component of Willett’s study, as well iments. Still, some researchers yearn for better described a strategy in 2023 in which indi-
as another11 from a team led by neurosurgeon — and they are making swift progress. “We’re vidual molecules are labelled with distinct

846 | Nature | Vol 625 | 25 January 2024

 DNA strands14. These molecules are then disorders such as Parkinson’s disease from from 10,000 donors. But HCA is also part of
detected with dye-tagged complementary blood samples, says Rizzoli. Jungmann is a broader ecosystem of intersecting cellular
DNA strands that bind to their corresponding similarly enthusiastic about the potential for and molecular atlas efforts. These include the
targets transiently but repeatedly, making it RESI to document reorganization of individ- Human BioMolecular Atlas Program (HuBMAP)
possible to discriminate individual fluores- ual proteins in disease or in response to drug and the Brain Research through Advancing
cent ‘blinking’ points that would blur into a treatments. It might even be possible to zoom Innovative Neurotechnologies (BRAIN) Initia-
single blob if imaged simultaneously. This in more tightly. “Maybe it’s not the end for the tive Cell Census Network (BICCN), both funded
resolution enhancement by sequential imag- spatial resolution limits,” Jungmann says. “It by the US National Institutes of Health, as well
ing (RESI) approach could resolve individual might get better.” as the Allen Brain Cell Atlas, funded by the Allen
base pairs on a DNA strand, demonstrating Institute in Seattle, Washington.
ångström-scale resolution with a standard Cell atlases According to Michael Snyder, a genomicist
fluorescence microscope. If you’re looking for a convenient cafe, Google at Stanford University and former co-chair of
The one-step nanoscale expansion (ONE) Maps can find nearby options and tell you how the HuBMAP steering committee, these efforts
microscopy method, developed by a team led to get there. There’s no equivalent for navi- have been driven in part by the development
by neuroscientists Ali Shaib and Silvio Rizzoli gating the much more complex landscape of and rapid commercialization of analytical
at University Medical Center Göttingen, the human body, but ongoing progress from tools that can decode molecular contents at
Germany, doesn’t quite achieve this level of various cell-atlas initiatives — powered by the single-cell level. For example, Snyder’s team
resolution. However, ONE microscopy offers advances in single-cell analysis and ‘spatial routinely uses the Xenium platform from 10X
an unprecedented opportunity to directly omics’ methods — could soon deliver the Genomics in Pleasanton, California, for its
image fine structural details of individual tissue-wide cellular maps that biologists crave. spatial transcriptomics analyses. The platform
proteins and multiprotein complexes, both makes it possible to survey the expression of
in isolation and in cells15. “Maybe it’s not the roughly 400 genes at once in 4 tissue samples
ONE is an expansion-microscopy-based every week. Multiplexed antibody-based meth-
approach that involves chemically coupling
end for the spatial ods such as the PhenoCycler platform by Akoya
proteins in the sample to a hydrogel matrix, resolution limits. Biosciences in Marlborough, Massachusetts,
breaking the proteins apart, and then allowing It might get better.” allow the team to track large numbers of pro-
the hydrogel to expand 1,000-fold in volume. teins with single-cell resolution in a format that
The fragments expand evenly in all directions, enables 3D tissue reconstruction. Other ‘multi-
preserving the protein structure and enabling The largest — and perhaps the most ambi- omics’ methods allow scientists to profile mul-
users to resolve features separated by a few tious — of these initiatives is the Human Cell tiple molecular classes in the same cell at once,
nanometres with a standard confocal micro- Atlas (HCA). The consortium was launched including the expression of RNA, the structure
scope. “We took antibodies, put them in the in 2016 by cell biologist Sarah Teichmann at of chromatin and the distribution of protein.
gel, labelled them after expansion, and were the Wellcome Sanger Institute in Hinxton, Last year saw dozens of studies showcas-
like, “Oh — we see Y shapes!” says Rizzoli, refer- UK, and Aviv Regev, now head of research ing progress in the generation of organ-spe-
ring to the characteristic shape of the proteins. and early development at biotechnology cific atlases using these techniques. In June,
ONE microscopy could provide insights into firm Genentech in South San Francisco, Cali- for example, the HCA released an integrated
conformationally dynamic biomolecules or fornia. It encompasses some 3,000 scientists analysis of 49 data sets from the human lung16.
enable visual diagnosis of protein-misfolding in nearly 100 countries, working with tissues “Having that very clear map of the lung informs
the changes that happen in diseases like lung
fibrosis, different tumours, even for COVID-19,”
says Teichmann. And in 2023, Nature released
an article collection (see go.nature.com/3vb-
znk7) highlighting progress from HuBMAP and
Science produced a collection detailing the work
of the BICCN (see go.nature.com/3nsf4ys).
Considerable work remains — Teichmann
estimates that it will be at least five years
before the HCA reaches completion. But the
resulting maps will be invaluable when they
arrive. Teichmann, for example, predicts using
atlas data to guide tissue- and cell-specific
drug targeting, while Snyder is eager to learn
how cellular microenvironments inform the
risk and aetiology of complex disorders such
as cancer and irritable bowel syndrome. “Will
we solve that in 2024? I don’t think so — it’s a
multiyear problem,” Snyder says. “But it’s a big
driver for this whole field.”

Nanomaterials printed in 3D
Weird and interesting things can happen
PENG HE

at the nanometre scale. This can make

materials-science predictions difficult, but
A cell atlas of the human lung describes different cell types and how they are regulated. it also means that nanoscale architects can
manufacture lightweight materials with

Nature | Vol 625 | 25 January 2024 | 847

 Work / Technology & tools

Develop your
scientific
career with
our support
MAX SACCONE/GREER LAB

Researchers have crafted microscale metal structures using a hydrogel.

distinctive characteristics such as increased are enthusiastic about the potential to craft
strength, tailored interactions with light or functional nanostructures from rugged,
sound, and enhanced capacity for catalysis high-melting-point metals and alloys.
or energy storage. The final barrier — economics — could be
Several strategies exist for precisely crafting the toughest to break. According to Saha,
such nanomaterials, most of which use lasers the pulsed-laser-based systems used in many
to induce patterned ‘photopolymerization’ of photopolymerization methods cost upwards
light-sensitive materials, and over the past few of US$500,000. But cheaper alternatives are
years, scientists have made considerable head- emerging. For example, physicist Martin
way in overcoming the limitations that have Wegener and his colleagues at the Karlsruhe
impeded broader adoption of these methods. Institute of Technology in Germany have
One is speed. Sourabh Saha, an engineer explored continuous lasers that are cheaper,
at the Georgia Institute of Technology in more compact, and consume less power
Atlanta, says that the assembly of nanostruc- than standard pulsed lasers20. And Greer has
tures using photopolymerization is roughly launched a start-up company to commercial-
three orders of magnitude faster than other ize a process for fabricating nanoarchitected
nanoscale 3D-printing methods. That might be metal sheets that could be suitable for appli-
good enough for lab use, but it’s too slow for cations such as next-generation body armour
large-scale production or industrial processes. or ultra-durable and impact-resistant outer
In 2019, Saha and mechanical engineer Shih- layers for aircraft and other vehicles.
Chi Chen at the Chinese University of Hong
Kong and their colleagues showed that they Michael Eisenstein is a science writer in
could accelerate polymerization by using a Philadelphia, Pennsylvania.
patterned 2D light-sheet rather than a conven-
1. Ferruz, N., Schmidt, S. & Höcker, B. Nature Commun. 13,
tional pulsed laser17. “That increases the rate 4348 (2022).
by a thousand times, and you can still main- 2. Munsamy, G., Lindner, S., Lorenz, P. & Ferruz, N. ZymCTRL:
tain those 100-nanometre features,” says Saha. A Conditional Language Model for the Controllable
Generation of Artificial Enzymes (MLSB, 2022).
Subsequent work from researchers including 3. Watson, J. L. et al. Nature 620, 1089–1100 (2023).
Chen has identified other avenues for faster 4. Ingraham, J. B. et al. Nature 623, 1070–1078 (2023).
nanofabrication18. 5. Krishna, R. et al. Preprint at bioRxiv https://doi.
org/10.1101/2023.10.09.561603 (2023).
Another challenge is that not all materials 6. Rössler, A. et al. Preprint at https://arxiv.org/
can be printed directly through photopoly- abs/1901.08971 (2019).
merization — such as metals. But Julia Greer, 7. Li, Y., Zhang, C., Sun, P., Qi, H. & Lyu, S. Preprint at
https://arxiv.org/abs/2103.02018 (2021).
a materials scientist at the California Institute 8. Yarnall, M. T. N. et al. Nature Biotechnol. 41, 500–512 (2023).
of Technology in Pasadena, has developed a 9. Sun, C. et al. Nature Biotechnol. https://doi.org/10.1038/ Search for your new role
clever workaround. In 2022, she and her col- s41587-023-01769-w (2023).
10. Willett, F. R. et al. Nature 620, 1031–1036 (2023). quickly by discipline,
leagues described a method in which photo­
polymerized hydrogels serve as a microscale
11. Metzger, S. L. et al. Nature 620, 1037–1046 (2023). country, salary and more
12. Sharlene, N. et al. Science 372, 831–836 (2021).
template; these are then infused with metal 13. Weber, M. et al. Nature Biotechnol. 41, 569–576 (2023).
on naturecareers.com
14. Reinhardt, S. C. M. et al. Nature 617, 711–716 (2023).
salts and processed in a way that induces the
15. Shaib, A. H. et al. Preprint at bioRxiv
metal to assume the structure of the template https://doi.org/10.1101/2022.08.03.502284 (2023).
while also shrinking19. Although the technique 16. Sikkema, L. et al. Nature Med. 29, 1563–1577 (2023).
17. Saha, S. K. et al. Science 366, 105–109 (2019).
was initially developed for microscale struc-
18. Ouyang, W. et al. Nature Commun. 14, 1716 (2023).
tures, Greer’s team has also used this strat- 19. Saccone, M. A. et al. Nature 612, 685–690 (2022). 00IFZ | Image credit: gettyimages/skynesher

egy for nanofabrication, and the researchers 20. Hahn, V. et al. Nature Photon. 16, 784–791 (2022).

848 | Nature | Vol 625 | 25 January 2024