Origami cranes — with their elegant paper wings and distinct angular beaks — rely on symmetry in their folding pattern to achieve their graceful form. This symmetry mirrors the bilateral balance seen in real birds and enables complexity to emerge in fewer steps. Yet, for the crane’s head to stand apart from its tail, the design must break free from strict symmetry. This subtle shift allows for unique features that perfect symmetry cannot achieve.

Taking inspiration from this concept, two new studies published back-to-back in Nature (Lee et al.; Dowling et al.) from the Baker and King Labs introduce asymmetry into the design of protein nanocages — tiny structures with enormous potential for vaccines and drug delivery. These nanocages are like advanced origami creations, built from many intricately folded protein parts.

Read the authors’ summary of this work, including how the two papers relate, here.

Building larger, more complex structures

“For the last decade, we were limited to building protein nanocages with strict symmetry,” said Neil King, associate professor of biochemistry at UW Medicine. “But in these two papers, we’ve used a principle called pseudosymmetry to design cage-like protein assemblies that are considerably larger than anything we’ve made before. These new types of nanocages could someday serve as highly advanced containers, packaging medicines and delivering them to precise locations in the body, which could make treatments safer and more effective.”

The King Lab has pioneered the design and use of custom protein nanoparticles as vaccine components, leading them to develop the world’s first computationally designed protein medicine with partners at UW Medicine.

Beyond strict symmetry

The scientists drew inspiration from viruses, which are masters of building complex structures with subtle breaks in symmetry. Using pseudosymmetry — where similar but not identical parts combine to form larger assemblies — the team computationally designed and experimentally verified an array of custom nanocages. These included structures with 240, 540, and even 960 protein subunits — by far the largest protein nanocages ever created to date. The largest structure, measuring nearly 100 nanometers across, has a volume 90 times greater than the well-studied AAV viral capsid. 

These new nanocages are not just bigger but also vastly more complex. One assembly contains four distinct protein chains and six unique protein-protein interfaces, all designed entirely from scratch.

“This work revealed secrets of the formation of large virus capsid structures and paved ways to intentionally design similar structures for biomedical applications such as cell targeting and gene delivery,” said Sangmin Lee, a co-lead author and former Baker Lab postdoctoral scholar who is now an assistant professor at Pohang University of Science and Technology (POSTECH).

To achieve this leap, the teams also embraced quasisymmetry — the use of identical protein parts that adopt different shapes depending on their local environment. Designing these shifts in form required extraordinary computational precision, resulting in nanocages with entirely novel architectures.

“Every time I saw one of these new assemblies on the electron microscope display, I had to pause and admire it,” said Ryan Kibler, co-lead author and postdoctoral researcher in the Baker Lab. “Their shapes are so distinct and unusual that it’s obvious they are man-made. This is especially true for the tetrahedral and octahedral types, which, to our knowledge, have never been observed in nature.”

These supersized nanocages represent a major step forward in protein design. Pushing beyond the limits of symmetry opens the door to creating more sophisticated molecular tools that more closely resemble the elegant functions seen in living systems. With their potential as vaccine scaffolds and containers for targeted drug delivery, these intricate assemblies may one day reshape how we treat disease and build nanoscale technologies.

Led by the Institute for Protein Design, this research included collaborators from the the IPD Core R&D Labs, the Stoddard Lab at the Fred Hutchinson Cancer Research Center, the Wysocki Lab at the Ohio State University, and the Veesler Lab at UW Medicine.

Researchers at the Institute for Protein Design have developed a new AI-driven molecular design tool that generates completely novel bioactive peptides. The ability to create such molecules by computation alone may accelerate drug development. 

RFpeptides leverages deep learning to design ring-shaped peptides called macrocycles that bind to disease-associated proteins using only the structure or sequence of a target. This is a departure from traditional peptide discovery methods, which often require extensive screening of vast peptide libraries to identify potential binders. 

Introduced as a preprint, RFpeptides was developed in the laboratories of three IPD faculty members: Gaurav Bhardwaj, Frank DiMaio, and 2024 Nobel laureate David Baker. Trainees Stephen Rettie, David Juergens, and Victor Adebomi led this project.

“RFpeptides extends the AI revolution in biology to the important challenge of peptide design. We hope it will help researchers create peptide-based medicines for a variety of diseases that have no good treatment options today,” said Gaurav Bhardwaj, an assistant professor of medicinal chemistry at the UW School of Pharmacy. 

The University of Washington has licensed RFpeptides to Vilya, an IPD spinout company. Bhardwaj and Baker are co-founders, advisors, and shareholders of the company.

What is RFpeptides?

RFpeptides is a software tool for designing bioactive peptides with precise 3D structures. Peptides are molecules composed of only a few amino acid building blocks. When the first and last of these building blocks link together, the peptide forms a cyclic structure. This often makes the peptide more resistant to degradation and can confer other biochemical benefits, such as a more rigid structure for higher affinity target binding. The precise nature of these binding interactions can activate or deactivate a target protein and thus drive a biological effect. 

How does it work?

RFpeptides builds upon the success of RFdiffusion and introduces key innovations tailored for the specific challenges of macrocyclic peptide design. Pioneered at the IPD, both tools draw inspiration from popular AI image generators that use diffusion models to synthesize new images based on user prompts. With RFdiffusion, a diffusion model sculpts clouds of disconnected amino acids into plausible biochemical structures. To create RFpeptides, the team modified the open-source tool to ensure the first and last amino acids in a designed molecule could form a chemical bond. 

“By expanding the structure modeling capabilities of RoseTTAFold2 and harnessing ProteinMPNN for sequence design, we’ve created a peptide design pipeline that’s both computationally efficient and incredibly accurate,” said associate professor of biochemistry Frank DiMaio. 

Testing RFpeptides

To demonstrate that RFpeptides can produce functional binding peptides for a range of challenging targets, the team selected four proteins implicated in hospital-derived bacterial infection, cancer, and other cellular processes important to human health. They synthesized and tested over a dozen designed macrocycle binders, identifying high-affinity interactions with each target.

“The highlight for me was that we successfully produced a high-affinity binder for a target with no known structure. Starting from just the target’s amino acid sequence, we predicted its structure using AlphaFold 2 and RoseTTAFold 2, designed peptides to bind those predicted structures, and ended up solving the first structure of the pathogenic protein Rhombotarget A,” said co-lead author and graduate student Stephen Rettie. 

“This is quite a neat display of the robustness and generalization capacity these generative models gain during pretraining” said co-lead author David Juergens. “The implications for drug design are really exciting as macrocycles can be customized in many ways that normal peptides and proteins cannot.”

The development of RFpeptides marks another step in leveraging AI to solve complex challenges in biology. While initially tested as a drug design tool, RFpeptides could also be used to create diagnostic reagents and other custom peptides for research challenges beyond medicine.

This project included collaborators from MIT, Forschungszentrum Jülich, Heinrich Heine University Düsseldorf, University College Cork, Tufts University. All funders are listed in the manuscript, which is titled Accurate de novo design of high-affinity protein binding macrocycles using deep learning.

The sense of smell depends on sensitive proteins embedded in the nose. In two studies published recently in Science (1,2), a team led by the Baker Lab has shown that an effectively unlimited number of new protein sensors can now be generated with the help of AI. This is a major step toward more advanced molecular sensing technologies for use in healthcare, pollution monitoring, and more. 

Biochemically speaking, smelling a rose requires two feats. First, floral compounds must bind to olfactory receptors in the nose. This molecular handshake must then produce a reliable signal, leading to the perception of smell.

In the new studies, the team used computers to create custom proteins that bind to specific compounds and transmit stable bioelectronic binding signals across lipid membranes. These sensor proteins are not derived from any found in nature. Instead, they were built from scratch and confirmed to function in the lab.

“When we started with this idea a few years ago, many people thought it was impossible,” said senior author and former Baker Lab postdoc Anastassia Vorobieva, now a group leader at the VIB-VUB Center for Structural Biology in Belgium. “Now we have shown that we can successfully design nanopore proteins with a high success rate and that they can have stable and reproducible conductance.”

This research was led by Baker Lab researchers Samuel Berhanu, Sagardip Majumder, and Linna An. Scientists from the University of Basel, University of Leeds, University of Virginia, Lawrence Berkeley National Laboratory, VIB-VUB, and the Institute for Protein Design Core R&D Labs also contributed to the structural and functional analyses of the new proteins.

The biosensors were created in stages and reported across two papers. Together, these projects yielded entirely synthetic receptors that resemble those in the nose. “This collaboration is a great example of what’s possible with protein design,” said senior author and IPD director David Baker.  “Rather than repurposing biomolecules from nature, we can now create the functions we want from first principles.”

Sculpting new nanopores

In the first study, the researchers designed custom protein nanopores and characterized them in the lab. These tube-like molecules have openings, or pores, less than three nanometers wide — roughly 20,000 times smaller than the diameter of a human hair.

The team showed that these custom nanopores embed into lipid membranes, similar to nasal scent receptors, and function as conduits for electrically charged ions as intended.

“Natural protein nanopores are important tools for DNA sequencing, but only a handful of these delicate molecules are reliable enough to work with. For this project, we created more stable nanopores than anyone’s had access to, and there’s effectively no limit to the number we could make using this design approach,” said Samuel, now a research scientist at IEH Laboratories and Consulting Group.

“This work brings biology and electronics much closer together. We’re now exploring how to incorporate them into bio-electronic devices that can detect tiny traces of chemicals, diagnose diseases, and possibly form critical components of nanoscale filtration devices,” said Sagardip.

Binding to chemicals

The second study introduced a novel approach for creating proteins that bind to specific chemicals. 

The team chose four challenging small molecules as binding targets. These included cholic acid, an important marker for liver disease; methotrexate, an anti-folate cancer treatment that requires regular blood monitoring; and thyroxine, a hormone used to diagnose thyroid conditions. 

“It hasn’t been easy to detect chemicals like these, but we succeeded in generating proteins that recognize each of our four targets,” said Linna. “The same approach could be used to create sensors for virtually any small molecule.”

Using a variety of deep-learning tools, the team generated “pseudocycle” proteins. These consist of repeating structural units that surround central binding pockets of varying shapes. They then docked target chemicals into these pockets and optimized the interaction surfaces using LigandMPNN, a deep-learning tool expanded from ProteinMPNN. The highest affinity binding proteins were identified via laboratory screening.

Creating synthetic sensors

By combining the small molecule binders within the new nanopores, the team created proteins that change conductance in response to target molecules — essentially creating synthetic tools that act like olfactory receptors.

Linna believes that synthetic nanopore sensors are just one technology that will emerge from this research. “When I talk to other scientists, they’re excited by how these binding proteins may be used in many different detection systems,” she commented. In addition to the nanopore sensors, the team describes chemically induced dimerization (CID) systems that could be used to control cancer cell therapies and more.

This research opens the door to developing highly specific biosensors for detecting disease biomarkers, monitoring environmental pollutants, and advancing personal healthcare devices. The collaboration underscores the interdisciplinary nature of modern science and the potential for computational protein design to revolutionize how we create solutions to important challenges in medicine and beyond.

This research was supported by several federal, private, and philanthropic organizations. All funding sources are listed in the manuscripts.

Berhanu S, Majumder S, Müntener T, et al. Sculpting conducting nanopore size and shape through de novo protein design. Science, 18 July 2024

Ah L, et al. Binding and sensing diverse small molecules using shape-complementary pseudocycles. Science, 18 July 2024

This advance could allow scientists to create cheaper alternatives to antibodies for disease detection and treatment.

This week we report in Nature an AI-enabled advance in biotechnology with implications for drug development, disease detection, and environmental monitoring. Using a combination of traditional and deep learning based molecular design approaches, we’ve created proteins that bind with exceptionally high affinity and specificity to a variety of challenging biomarkers, including human hormones. Notably, we achieved what we believe to be the highest binding affinity ever reported between a computer-generated biomolecule and its target.

AI-enabled protein design software in action. Beginning with a desired binding target (pink) and a cloud of disconnected amino acids, RFdiffusion iteratively sculpts a new protein structure that cradles the target peptide. At the end, ProteinMPNN assigns amino acid side chains to the new protein structure, yielding a complete protein molecule. Laboratory tests reveal this protein binds its target with the highest affinity ever reported to date for a computer-generated protein without any experimental optimization.

This project was led by Baker Lab members Susana Vazquez-Torres, Preetham Venkatesh, and Phil Leung, PhD. It included collaborators from UW Medicine and the University of Copenhagen, as well as from our Core R&D Labs.

Targeting helical peptides

The team set out to create proteins that could bind to glucagon, neuropeptide Y, parathyroid hormone, and other helical peptide targets. Such molecules, crucial in biological systems, are especially challenging for drugs and diagnostic tools to recognize as they often lack stable molecular structures. Antibodies can be used to detect some of these targets but are often costly to produce and have limited shelf lives.

“There are many diseases that are difficult to treat today simply because it is so challenging to detect certain molecules in the body. As tools for diagnosis, designed proteins may offer a more cost-effective alternative to antibodies,” explains Venkatesh.

The role of generative AI

The study introduces a novel way of using RFdiffusion, a generative model for creating new protein shapes, in conjunction with the sequence-design tool ProteinMPNN. Developed in the Baker Lab, these programs allow scientists to create functional proteins more efficiently than ever before. By combining these tools in new ways, the team was able to create binding proteins by using limited target information, such as a peptide’s amino acid sequence alone.

“We’re witnessing an exciting era in protein design, where advanced artificial intelligence tools, like the ones featured in our study, are accelerating the improvement of protein activity. This breakthrough is set to redefine the landscape of biotechnology,” notes Vazquez-Torres.

Measuring binding in the lab

In collaboration with the Joseph Rogers Lab at the University of Copenhagen and Andrew Hoofnagle Lab at UW Medicine, we conducted laboratory tests to validate the new biodesign methods.

Mass spectrometry was used to detect designed proteins that bind to low-concentration peptides in human serum, thereby demonstrating the potential for sensitive and accurate disease diagnostics. Additionally, the proteins were found to retain their target binding abilities despite harsh conditions including high heat, a crucial attribute for real-world application.

From binders to biosensors

With improved methods for creating binding proteins in place, the team turned to the challenge of designing new biosensors. To make sensors that could detect parathyroid hormone (PTH), we grafted a high-affinity PTH binder into our previously reported lucCage biosensor system. The best-performing biosensor lit up when mixed with PTH, displaying a 21-fold increase in bioluminescence.

“The ability to design proteins with such high affinity and specificity opens up a world of possibilities, from new disease treatments to advanced diagnostics,” concludes senior author and Institute director David Baker.

Funding

This work was supported by the  National Institutes of Health (T1D U01 DK121289, U19 AG065156, K99EB031913, P30 GM124165), National Science Foundation (EF-2021552), Department of Energy (BER-ERCAP0022018; DE-AC02-06CH11357), European Molecular Biology Organization (ALTF 292-2022), Washington State General Operating Fund, Amgen, Audacious Project, AWS, Bill and Melinda Gates Foundation (INV-010680), Donald and Jo Anne Petersen, Howard Hughes Medical Institute, Microsoft, Novo Nordisk Foundation (NNF19OC0054441), Open Philanthropy Project, and Partnership for Clean Competition.

Torres SV, Leung PJY, Venkatesh P, et al. De novo design of high-affinity binders of bioactive helical peptides. Nature. 2023 [Open Access, Accelerated]

Today we published new research that may one day be applied to help remove large amounts of excess carbon from the environment. 

In a paper appearing in Nature Communications, we show that custom proteins can drive the formation of carbon-rich minerals in laboratory settings. This offers a potential pathway for enhanced carbon storage via engineered organisms. This research was performed in collaboration with the Center for the Science of Synthesis Across Scales (CSSAS), which is co-led by the UW and Pacific Northwest National Laboratory (PNNL).

“In nature, proteins are behind the growth of bones, shells, and other durable materials like limestone. We’re now exploring how custom proteins can be used to create a wide range of new, robust materials. This study is an important first step, showing we can guide mineral formation in novel ways,” explains senior author David Baker.

The project was led by Baker Lab members Fatima Davila-Hernandez and Harley Pyles, and PNNL postdoctoral research associate Biao Jin of the De Yoreo research group. Jim De Yoreo is a Battelle Fellow  at PNNL, an affiliate professor of materials science and engineering and of chemistry at the UW, and the Deputy Director of CSSAS.

The team used advanced molecular design software to create proteins with custom lengths and surface chemistries. They also showed that these designed features can influence the formation and growth of calcium carbonate crystals, a key component of seashells and limestone. The process by which biomolecules such as proteins promote mineral growth is called biomineralization.

Biology’s Role in Earth’s Carbon Cycle

Our research paves the way for future studies on calcium carbonate formation under biological conditions, which is a critical part of the global carbon cycle. This line of research could also help in the fight against climate change by offering new routes to large-scale and long-term carbon sequestration.

“Our designed proteins are an important step toward mimicking nature’s efficiency in carbon storage. While we’re not yet at nature’s level, we are gaining valuable insights that may inform the development of practical, scalable solutions for long-term carbon storage,” said Pyles.

Whether in human teeth or aquatic microbes, nature’s mineral-forming proteins contain irregular features that make them difficult for scientists to study, even with advanced instruments or the latest protein modeling tools like RoseTTAFold All-Atom. By designing new mineral-forming proteins from scratch, we hope to enable controlled research that will reveal the interplay between the living and nonliving worlds.

Addressing Climate Change

Combatting human-caused climate change will require many innovative solutions, including finding effective ways to remove carbon pollution from the environment. Organisms like kelp, algae, and trees naturally store carbon atoms in their tissues through photosynthesis, but their capacity for long-term storage is limited. In contrast, some aquatic microorganisms convert carbon in ocean water into calcium carbonate crystals using biomineralization, leading to sedimentation and limestone formation. However, this natural process is slow.

Integrating designed proteins into living organisms may one day accelerate limestone formation in the ocean, potentially transforming billions of tons of carbon pollution into enduring mineral deposits. Rigorous evaluations of the feasibility and environmental impacts of this research will be needed before any real-world application.

Keep an eye out for more updates from us and our colleagues as we continue to explore the potential of protein design to create solutions across medicine, technology, and sustainability.

Join Us

Are you intrigued by the prospect of using protein design to solve modern challenges? We invite you to join us or support our work.

Funding

This work was supported by numerous funding sources, including the United States Department of Energy Office of Science and The Audacious Project at the Institute for Protein Design. A full list of funders is provided in the paper.

Davila-Hernandez FA, Jin B, Pyles H, et al. Directing polymorph specific calcium carbonate formation with de novo protein templates. Nature Communications. 14 Dec 2023 [Open Access]

In a recent preprint, we describe deep-learning breakthroughs that significantly expand the types of proteins and protein assemblies that can be modeled and designed using computers.

This work updates two of our most powerful software tools, RoseTTAFold and RFdiffusion, by training them to operate on additional types of molecules beyond just proteins. Through these efforts, we aim to both increase our understanding of life and accelerate the design of new proteins with advanced functions.

This project was led by postdoctoral scholar Jue Wang and graduate students Rohith Krishna and Woody Ahern, who are all members of the Baker Lab at the Institute for Protein Design at the University of Washington School of Medicine.

RoseTTAFold All-Atom models more of life’s molecules

The team began by making changes to RoseTTAFold, which, like AlphaFold, is a highly accurate deep-learning program for modeling protein structures. These programs were developed to model only biomolecules made entirely of amino acids. The new upgrades now allow RoseTTAFold All-Atom to model full biological assemblies that contain many different types of molecules, including proteins, DNA, RNA, small molecules, metals, and other bonded atoms, including covalent modifications of proteins.

This is important for understanding biology because proteins rarely function on their own and instead must interact with other non-protein compounds. RoseTTAFold All-Atom may also benefit drug discovery research by allowing scientists to model how proteins and small-molecule drugs interact.

“We’ve expanded our modeling capabilities beyond amino acids, which should bring clarity to new aspects of molecular biology. It’s a bit like switching from black and white to a color TV,” said Krishna.

RFdiffusion All-Atom generates proteins with advanced functions

This upgrade to RoseTTAFold allowed the team to create a new version of RFdiffusion, which is our powerful deep-learning program that can generate new protein structures in a manner similar to how DALL-E or Midjourney generate art.

Through extensive laboratory testing, we confirmed that the updated design tool, dubbed RFdiffusion All-Atom, yields proteins with advanced functions, including the ability to bind to specific small molecules like heme or the heart disease drug digoxigenin.

“Our improved protein design software will allow us and others to create an even wider range of versatile and functional molecules,” said Ahern. “I can’t wait to see all the ways scientists will use these tools.”

The ability to create proteins that are programmed to bind to specific target molecules sets the stage for the development of new diagnostics, gene-editing technologies, enzymes, and more. The ability to model full biomolecular systems should considerably extend the power of deep learning tools for protein structure prediction, leading to a richer understanding of biology. 

This research was posted on bioRvix on Oct. 9 and has not yet been peer-reviewed. The initial development and applications of RoseTTAFold were reported in Science in 2021. The initial development and applications of RFdiffusion were reported in Nature in 2022.

Funding

This work was supported by The Audacious Project, Howard Hughes Medical Institute, Bill and Melinda Gates Foundation (OPP1156262), Open Philanthropy (INV-010680), Schmidt Futures, Helmsley Charitable Trust (2019PG-T1D026), Juvenile Diabetes Research Foundation (2-SRA-2018-605-Q-R), Microsoft, Amgen, Seoul National University, University of Sheffield, Washington State General Operating Fund, Defense Threat Reduction Agency (HDTRA1-19-1-0003), Department of Health and Human Services (75N93022C00036), National Energy Research Scientific Computing Center (BER-ERCAP0022018), National Institutes of Health, European Research Council, Royal Society University Research Fellowship (URF\R1\191548), and the Human Frontiers Science Program (RGP0061/2019).

Today we report in Nature the design of proteins that recognize and bind to the so-called “intrinsically disordered regions” of proteins and peptides. The body produces such disordered molecules naturally, but many have been linked to health disorders, including myeloma and other cancers.

“Disordered proteins play important roles in biology. By designing new proteins that latch onto them, we may finally be able to diagnose and treat many burdensome diseases,” said senior author David Baker, director of the Institute for Protein Design

Traditionally, scientists who seek to develop new drugs focus on targets that have a defined structure. Antibody therapies for COVID-19, for example, work by recognizing the tell-tale shape of the Spike protein that protrudes from the surface of the coronavirus.

But how can a drug recognize a shape-shifting molecule in the body?

“To create proteins that bind to intrinsically disordered targets, we chose to go after something that’s common to all peptides and proteins: their chemical backbone.”

— Kejia Wu, co-lead author and Baker Lab graduate student

Proteins, and their shorter cousins peptides, are chains of chemicals called amino acids. The body uses an alphabet of 20 amino acids to construct its roughly 20,000 unique proteins. But every amino acid inside every protein and peptide chain shares certain chemical features. Whether those chains are structured or unstructured, this common chemical backbone remains.

Wu, together with co-lead author and recent Baker Lab postdoctoral scholar Hua Bai, devised a new protein design strategy that focuses on the inherent chemical features found in the backbone of every intrinsically disordered protein and peptide. Using computational protein design, they were able to create new proteins that bind to disordered peptides and proteins.

When tested in the lab, the best of their new proteins latched on to a disordered target for over 2,000 seconds. In one experiment, the team also showed that it was possible to design proteins that recognize and bind to a specific human protein that is known to contain an intrinsically disordered region. That target protein, called ​​ZFC3H1, may be a biomarker for cancer.

The team also show that their new method can be used to recognize specific linear peptide sequences, which could have great utility in medical and proteomic research, including for peptide sequencing.

Read the manuscript:
Wu K., Bai H., et al. De novo design of modular peptide-binding proteins by superhelical matching. Nature, 2023

The de novo design of small proteins with beta-barrel topologies has been a challenge for computational design due to the complexity inherent in these folds. In a new study appearing in PNAS, a team led by Baker Lab research scientist David E. Kim describes the successful design and characterization of four different classes of small beta barrels using Rosetta energy-based methods and deep learning approaches.

The designed beta barrels exhibited high thermal stability and were observed to fold into their desired structures. This study provides insights into the determinants of folding and design of this important class of proteins and offers new routes to the design of high-affinity binding proteins by increasing the protein shape-space available for designing binders to protein targets of interest.

In the paper, the authors also discuss the relative strengths and weaknesses of traditional protein design using Rosetta and more recent deep-learning approaches, including hallucination. Their best results came from combining both methods: using hallucination for backbones and Rosetta for sequence design and structure relaxation. This combination led to well-folded proteins in 90 percent of cases compared with just 30 percent of cases when only Rosetta was used. However, recent work from the Baker Lab suggests that machine learning methods may soon take over the sequence design step as well.

Read the full report:
De novo design of small beta barrel proteins (Open Access)

Many useful proteins — including those that make up the King Lab’s nanoparticle vaccine platform — must be secreted from cells, but this can prove difficult. In a new paper appearing in PNAS, a team led by King Lab postdoctoral scholars John Wang and Alena Khmelinskaia developed a new computer program called Degreaser which can identify and remove hidden transmembrane sequences that hinder secretion.

The team showed that Degreaser can improve the secretion of existing nanoparticles without reducing their stability. They also used Degreaser to create new nanoparticles that secrete as well as natural ones. This new program and the improved nanoparticles could be useful in medicine and biotechnology. 

Read the full report:

Improving the secretion of designed protein assemblies through negative design of cryptic transmembrane domains (Open Access)

Under ideal conditions, cryo-electron microscopy can be used to determine protein structures at near-atomic resolution. But conditions are not always perfect. To help researchers make use of medium-resolution cryo-electron density maps, scientists in the DiMaio Lab have developed EMERALD, which is a new software tool that can accurately and automatically produce deposition-ready small molecule models into cryoEM maps. 

Benchmarked on over 1,000 ligand-bound proteins, EMERALD identified a confident solution in 62% of EMDB entries. In some cases, it identified alternate models that were supported by external data. Led by graduate student Andrew Muenks, this research was published in Nature Communications and included collaborators from the Veesler Lab at UW.

All methods described in the paper are available as part of Rosetta, using weekly releases after February 5, 2023 (version 2023.06 or later). The Rosetta XML files and flags for running all the refinements discussed in this manuscript are included in Methods. A demo for running EMERALD is included in Supplementary Data 2.

Read the full report:

Automatic and accurate ligand structure determination guided by cryo-electron microscopy maps (Open Access)

Today we report in Nature [PDF] the computational design of highly efficient enzymes unlike any found in nature. Laboratory testing confirms that the new light-emitting enzymes, called luciferases, can recognize specific chemical substrates and catalyze the emission of photons very efficiently. This is an important step in the field of protein design as enzymes have many uses across biotechnology, medicine, environmental remediation, and manufacturing.

This project — which was powered by deep-learning software developed at the Institute for Protein Design — was led by two postdoctoral scholars in the Baker Lab, Andy Hsien-Wei Yeh and Christoffer Norn, and included collaborators in the Houk Lab at UCLA.

“Living organisms are remarkable chemists. Rather than relying on toxic compounds or extreme heat, they use enzymes to break down or build up whatever they need under gentle conditions. New enzymes could put renewable chemicals and biofuels within reach.” 

— David Baker, Director of the Institute for Protein Design

To create new luciferases, the team first selected chemicals called luciferins that they wanted the proteins to act upon. They then used software to generate thousands of possible protein structures that might react with those chemicals. The best-performing enzymes emit enough light to be seen by the naked eye.

Hallucinating protein folds

Rather than modifying existing proteins, the team devised a new deep-learning-based protein design strategy dubbed ‘family-wide hallucination.’ By integrating unconstrained de novo design and fixed backbone sequence-design approaches, this approach can generate an essentially unlimited number of never-before-seen proteins that have a desired fold.

Family-wide hallucination uses the de novo sequence and structure discovery capability of unconstrained protein hallucination for loop and variable regions, and structure-guided sequence optimization for core regions.

“Enzyme design has frustrated scientists for decades as it is one of the most challenging tasks in all of biochemistry. But with our new tools for generating fit-for-function scaffolds, I’m excited to see how much progress we can make in designing enzymes that have practical applications in medicine and biotechnology.”

– Christoffer Norn, co-lead author

To add active sites into the hallucinated protein folds, the team chose to precisely place a positively charged guanidinium group of an arginine residue to stabilize negative charges present on the reaction’s transition state. Additional active site residues were also designed.

Let there be light

When manufactured and tested in the laboratory, the researchers identified three active enzymes among their initial designs. They named the best-performing one LuxSit, a play on UW’s Latin motto lux sit, which roughly translates to ‘let light exist’.

LuxSit has many properties that make it an attractive tool for biotechnological research. At just 117 residues, it is smaller than any known luciferase. Incubation of the enzyme with its synthetic luciferin substrate diphenylterazine (DTZ) resulted in blue luminescence at 480 nanometers, which is consistent with the substrate’s chemiluminescence spectrum. And the protein was found to remain partially folded under near-boiling conditions.

“We were able to design very efficient enzymes from scratch on the computer, as opposed to relying on enzymes found in nature. This breakthrough means that custom enzymes for almost any chemical reaction could, in principle, be designed.”

– Andy Hsien-Wei Yeh, co-lead author

Refinement of LuxSit led to dramatic improvements in performance. An optimized enzyme, dubbed LuxSit-i, generated enough light to be visible to the naked eye. It was found to be brighter than the natural luciferase enzyme found in the glowing sea pansy Renilla reniformis.

The power of deep learning

The team went on to design additional luciferases that recognize another synthetic luciferin substrate, 2-deoxycoelenterazine or h-CTZ. 

Because the molecular shape of h-CTZ differs from DTZ, they again used family-wide hallucination to generate custom protein scaffolds. To create active sites, precise arrangements of histidine and arginine side chains were installed. These active site features were modeled after those observed to be most successful in the first round of luciferase design. 

Rather than RosettaDesign, the team turned to the recently developed ProteinMPNN tool to come up with the remaining amino acid sequences of the new enzymes. ProteinMPNN is a sequence design tool powered by deep learning that runs in about one second, which is more than 200 times faster than the previous best software. Its results are superior to prior tools, and it requires no expert customization to run.

Out of the 46 designed h-CTZ catalysts tested in the lab, two were found to have measurable luciferase activity. This marks a hundred-fold increase in success rates — from 0.04% (3/7,648 for DTZ) to 4.35% (2/46 for h-CTZ) — for this second round of de novo enzyme design. This improvement is likely due to the knowledge gained during the first design round coupled with the increased performance of ProteinMPNN.

Until now, computational enzyme design had been limited by the number of available protein scaffolds and the extreme difficulty of placing enzyme active sites in them. The use of deep learning to produce large numbers of custom protein scaffolds, together with next-generation tools for protein sequence design, has set the stage for a new era in enzyme design.

Funding

This work was supported by the Howard Hughes Medical Institute, National Institutes of Health (K99EB031913), the United World Antiviral Research Network, National Institute of Allergy and Infectious Disease (1 U01 AI151698-01), Audacious Project at the Institute for Protein Design, Open Philanthropy Project Improving Protein Design Fund, Novo Nordisk Foundation (NNF18OC0030446), National Science Foundation (CHE-1764328, OCI-1053575), and Eric and Wendy Schmidt by recommendation of the Schmidt Futures program. The National Natural Science Foundation of China (22103060) provided partial computational resources.

Update (July 2023): Our manuscript on the development of RFdiffusion has been published in Nature.

A team led by scientists from the Baker Lab has created a powerful new way of designing proteins that combines structure prediction networks and generative diffusion models.

The team demonstrated extremely high computational success using the new method and experimentally tested hundreds of A.I.-generated proteins, finding that many may be useful as medications, vaccines, or even new nanomaterials.

Originally appearing as a preprint, this research is now available in Nature. Additional applications of RFdiffusion are also described in a companion preprint.

Drawing inspiration from DALL-E

The software tool DALL-E produces high-quality images that have never existed before using a machine-learning tool called a diffusion model, which is an algorithm that specializes in adding and removing noise.

Diffusion models for image generation begin with grainy bits of static and gradually remove noise until a clear picture is formed. Additional pieces of software guide this denoising process so that the new images end up matching what was asked for.

Drawing inspiration from this work, we have developed a guided diffusion model for generating new protein molecules called RFdiffusion. With prior protein design software, tens of thousands of designed molecules may have to be tested before finding a single one that performs as intended. Using the new method, the team had to test as little as one per design challenge.

RFdiffusion was developed by a team of computational biologists from UW Medicine, Columbia University, and MIT. The project was led by Joseph Watson, David Juergens, Nathaniel Bennett, Brian Trippe, Jason Yim, Helen Eisenach, Woody Ahern.

A new state-of-the-art

RFdiffusion outperforms existing protein design methods across a broad range of problems. These include topology-constrained protein monomer design, protein binder design, symmetric oligomer design, enzyme active site scaffolding, and symmetric motif scaffolding for therapeutic and metal-binding protein design.

So far, we have used RFdiffusion to generate ultra-high affinity binders and a series of novel symmetric assemblies experimentally confirmed by electron microscopy. 

“These works reveal just how powerful diffusion models can be for protein design,” says Watson. “It’s extremely exciting,” added Juergens, “and it’s really just the beginning.

Watson JL, Juergens D, Bennett NR, Trippe BL, Yim J, Eisenach HE, Ahern W, et al. De novo design of protein structure and function with RFdiffusion. Nature. 2023

Vázquez Torres S, Leung PY, Lutz ID, Venkatesh P, et al. De novo design of high-affinity protein binders to bioactive helical peptides. bioRxiv. 2022.

Researchers at the Institute for Protein Design have discovered how to create peptides that slip through membranes and enter cells. This drug design breakthrough may lead to new medications for a wide variety of health disorders, including cancer, infection, and inflammation. This research appears in the journal Cell [PDF].

“This new ability to design membrane-permeable peptides with high structural accuracy opens the door to a new class of medicines that combine the advantages of traditional small-molecule drugs and larger protein therapeutics,” said senior author David Baker, director of the Institute.

Peptide drugs are made from the same building blocks as proteins. Unlike many traditional small molecule medicines, peptides can bind protein targets in the body with great precision, promising fewer side effects. 

Gaurav Bhardwaj, Adam Moyer, Naozumi Hiranuma, Patrick Salveson, and David Baker at the UW Medicine Institute for Protein Design have co-founded a new company, Vilya, Inc., with ARCH Venture Partners that is licensing the platform and molecules described in the paper.

“We know that peptides can be excellent medicines, but a big problem is that they don’t get into cells. There are a lot of great drug targets inside our cells, and if we can get in there, that space opens up.”

 Discovering how to send peptides through membranes was a chemical problem. “Membranes are made of lipids, and most peptides have chemical features that cause them to hold onto water molecules. Dragging these water molecules through the lipids is difficult,” explains Bhardwaj. The scientists tried several solutions. They first crafted peptides with chemical features that reduce interactions with water. In another approach, they designed peptides that could change shapes while crossing membranes.

To evaluate their new peptide design strategies, the team synthesized over 180 custom molecules. Laboratory tests on artificial membranes revealed that most of the peptides could pass through lipids as desired. Further tests involving gut epithelial cells led the scientists to believe that some of the new peptides could make the jump from the stomach into the bloodstream. Animal studies confirmed that when swallowed, some of the peptides could efficiently move out of the gut, cross several membranes, and enter living cells.

The Core Research Labs at the Institute for Protein Design determined high-resolution structures for 36 of the new compounds, confirming that the peptides adopted the precise shapes that the designers intended.

“These molecules are promising starting points for future drugs. My lab is now working to turn them into antibiotics, antivirals, and cancer treatments,” said Bhardwaj.

Bhardwaj believes it may now be possible to create peptide drugs that treat COVID-19. “One of the most obvious drug targets is located inside infected cells. If we could shut down that enzyme, that would prevent the virus from creating more copies of itself.”

Antibiotics are another area of focus. “Antibiotic resistance is becoming a problem for most drugs, and the discovery of new antibiotics from nature has been slow,” explains Bhardwaj. “We are using rational design to try to create new peptide-based antibiotics. Here, the fact that these molecules can be swallowed rather than injected is going to be very important for their effective use and broader adoption ”

Scientists at the University of Washington developed the design methods, created the new peptides, and performed some laboratory tests. Permeability using gut epithelial cells was measured in the David Craik lab at the University of Queensland. Structural characterization of the new peptides was performed by the Gaetano Montelione lab at Rensselaer Polytechnic Institute.

The paper, “Accurate de novo design of membrane-traversing macrocycles,” appears online in Cell on August 29. The research was supported by The Audacious Project; Gates Ventures; Eric and Wendy Schmidt by recommendation of the Schmidt Futures; the Nordstrom Barrier Institute for Protein Design Directors Fund; Wu Tsai Translational Fund; Bill and Melinda Gates Foundation (INV-010680; Open Philanthropy Project; Takeda Pharmaceuticals; Howard Hughes Medical Institute; Washington State Supplement Funding; Department of Defense; Simons Foundation; Defense Threat Reduction Agency (HDTRA1-19-1-0003); National Institutes of Health (R01AG063845, R01GM120574, R35-GM141818, F32GM120791-02); and Washington Research Foundation.

Vilya Launches to Design a New Transformational Class of Medicines That Precisely Target Disease Biology (businesswire.com)

Today we report in Science [PDF] the development of artificial intelligence software that can create proteins that may be useful as vaccines, cancer treatments, or even tools for pulling carbon pollution out of the air.

This project was led by Jue Wang, Doug Tischer, and Joseph L. Watson, who are postdoctoral scholars at UW Medicine, as well as Sidney Lisanza and David Juergens, who are graduate students at UW Medicine. Senior authors include Sergey Ovchinnikov, a John Harvard Distinguished Science Fellow at Harvard University, and David Baker, professor of biochemistry, HHMI Investigator, and director of the Institute for Protein Design at UW Medicine.

“The proteins we find in nature are amazing molecules, but designed proteins can do so much more,” said Baker. “In this work, we show that machine learning can be used to design proteins with a wide variety of functions.”

Training new neural networks

Inspired by how machine learning algorithms can generate stories or even images from prompts, the team set out to build similar software for designing new proteins. “The idea is the same: neural networks can be trained to see patterns in data. Once trained, you can give it a prompt and see if it can generate an elegant solution. Often the results are compelling — or even beautiful,” said lead author Joseph Watson.

The team trained multiple neural networks using information from the Protein Data Bank, which is a public repository of hundreds of thousands of protein structures from across all kingdoms of life. The neural networks that resulted have surprised even the scientists who created them.

The team developed two approaches for designing proteins with new functions. The first, dubbed “hallucination” is akin to DALL-E or other generative A.I. tools that produce new output based on simple prompts. The second, dubbed “inpainting,” is analogous to the autocomplete feature found in modern search bars and email clients.

“Most people can come up with new images of cats or write a paragraph from a prompt if asked, but with protein design, the human brain cannot do what computers now can,” said lead author Jue Wang. “Humans just cannot imagine what the solution might look like, but we have set up machines that do.”

Starting with gibberish

To explain how the neural networks ‘hallucinate’ a new protein, the team compares it to how it might write a book: “You start with a random assortment of words — total gibberish. Then you impose a requirement such as that in the opening paragraph, it needs to be a dark and stormy night. Then the computer will change the words one at a time and ask itself ‘Does this make my story make more sense?’ If it does, it keeps the changes until a complete story is written,” explains Wang.

Both books and proteins can be understood as long sequences of letters. In the case of proteins, each letter corresponds to a chemical building block called an amino acid. Beginning with a random chain of amino acids, the software mutates the sequence over and over until a final sequence that encodes the desired function is generated. These final amino acid sequences encode proteins that can then be manufactured and studied in the laboratory.

Autocomplete for proteins

The team also showed that neural networks can fill in missing pieces of a protein structure in only a few seconds. Such software could aid in the development of new medicines.

“With autocomplete, or “protein Inpainting”, we start with the key features we want to see in a new protein, then let the software come up with the rest. Those features can be known binding motifs or even enzyme active sites,” explains Watson. Laboratory testing revealed that many proteins generated through hallucination and inpainting functioned as intended. This included novel proteins that can bind metals as well as those that bind the anti-cancer receptor PD-1.

Creating new vaccines

The new neural networks can generate several different kinds of proteins in as little as one second. Some include potential vaccines for the deadly respiratory virus RSV.

All vaccines work by presenting a piece of a pathogen to the immune system. Scientists often know which piece would work best, but creating a vaccine that achieves a desired molecular shape can be challenging. Using the new neural networks, the team prompted a computer to create new proteins that included the necessary pathogen fragment as part of their final structure. The software was free to create any supporting structures around the key fragment, yielding several potential vaccines with diverse molecular shapes.

When tested in the lab, the team found that known antibodies against RSV stuck to three of their hallucinated proteins. This confirms that the new proteins adopted their intended shapes and suggests they may be viable vaccine candidates that could prompt the body to generate its own highly specific antibodies. Additional testing, including in animals, is still needed.

“I started working on the vaccine stuff just as a way to test our new methods, but in the middle of working on the project, my two-year-old son got infected by RSV and spent an evening in the ER to have his lungs cleared. It made me realize that even the ‘test’ problems we were working on were actually quite meaningful,” said Wang.

“These are very powerful new approaches, but there is still much room for improvement,” said Baker, who was a recipient of the 2021 Breakthrough Prize in Life Sciences. “Designing high activity enzymes, for example, is still very challenging. But every month our methods just keep getting better! Deep learning transformed protein structure prediction in the past two years, we are now in the midst of a similar transformation of protein design.”

Funding 

Compute resources for this work were donated by Microsoft and Amazon Web Services. Funding was provided by the Audacious Project at the Institute for Protein Design; Microsoft; Eric and Wendy Schmidt by recommendation of the Schmidt Futures; the DARPA Synergistic Discovery and Design project (HR001117S0003 contract FA8750-17-C-0219); the DARPA Harnessing Enzymatic Activity for Lifesaving Remedies project (HR001120S0052 contract HR0011-21-2-0012); the Washington Research Foundation; the Open Philanthropy Project Improving Protein Design Fund; Amgen; the Human Frontier Science Program Cross Disciplinary Fellowship (LT000395/2020-C) and EMBO Non-Stipendiary Fellowship (ALTF 1047-2019); the EMBO Fellowship (ALTF 191-2021); the European Molecular Biology Organization (ALTF 139-2018); the “la Caixa” Foundation; the National Institute of Allergy and Infectious Diseases (HHSN272201700059C), the National Institutes for Health (DP5OD026389); the National Science Foundation (MCB 2032259); the Howard Hughes Medical Institute, the National Institute on Aging (5U19AG065156); the National Cancer Institute (R01CA240339); the Swiss National Science Foundation; the Swiss National Center of Competence for Molecular Systems Engineering; the Swiss National Center of Competence in Chemical Biology; and the European Research Council (716058).

Today we report in Science the design of rotary devices made from custom proteins. These microscopic “axles” and “rotors” come together to form spinning assemblies, rather than being locked in just one orientation. Such mechanical coupling is a key feature of any machine.

The new axle-rotor devices — which are each about a billion times smaller than a poppy seed — were designed on computers, produced inside living cells, and studied in the lab.

This research paves the way for a new generation of nanoscale machines in which the motion of the components is powered by solar energy or chemical fuel.

The project was led by biochemist Alexis Courbet, a postdoctoral scholar in Baker lab, and by Jesse Hansen, a recent graduate student in the laboratory of Justin Kohlman, an associate professor of biochemistry at UW Medicine.

“One of our goals is to create nanomachines that might one day circulate through the blood and autonomously remove unwanted plaques or even cancer cells,” said Courbet. “We know that very complex machines can be assembled from simple parts.”

Scientists from the Kohlman and Veesler labs at UW Medicine used electron microscopes to visualize the rotation of the new protein devices.

This work was supported in part by The Audacious Project at the Institute for Protein Design, Open Philanthropy, National Science Foundation, and a Washington Research Foundation Senior Fellowship. A full list of funders can be found in the manuscript.

Today we report in Nature a new method for generating protein drugs. Using Rosetta-based design, an international team designed molecules that can target important proteins in the body, such as the insulin receptor, as well as proteins on the surface of viruses. This solves a long-standing challenge in drug development and may lead to new treatments for cancer, diabetes, infection, inflammation, and beyond.

This work was led by two Baker lab postdoctoral scholars – Longxing Cao, PhD, and Brian Coventry, PhD – who combined recent advances in computational protein design to arrive at a strategy for creating new proteins that bind molecular targets in a manner similar to antibodies. They developed software that can scan a target molecule, identify potential binding sites, generate proteins targeting those sites, and then screen from millions of candidate binding proteins to identify those most likely to function.

The team generated high-affinity binding proteins against 12 distinct molecular targets. These targets include important cellular receptors such as TrkA, EGFR, Tie2, and the insulin receptor, as well proteins on the surface of the influenza virus and SARS-CoV-2 (the virus that causes COVID-19).

“When it comes to creating new drugs, there are easy targets and there are hard targets,” said Cao, who is now an assistant professor at Westlake University. “In this paper, we show that even very hard targets are amenable to this approach. We were able to make binding proteins to some targets that had no known binding partners or antibodies.”

In total, the team produced over half a million candidate binding proteins for the 12 selected molecular targets. Data collected on this large pool of candidate binding proteins was used to improve the overall method.

“We look forward to seeing how these molecules might be used in a clinical context, and more importantly how this new method of designing protein drugs might lead to even more promising compounds in the future,” said Coventry.

The research team included scientists from the University of Washington School of Medicine, Yale University School of Medicine, Stanford University School of Medicine, Ghent University, The Scripps Research Institute, and the National Cancer Institute, among other institutions.

This work was supported in part by The Audacious Project at the Institute for Protein Design, Open Philanthropy Project, National Institutes of Health (HHSN272201700059C, R01AI140245, R01AI150855, R01AG063845), Defense Advanced Research Project Agency (HR0011835403 contract FA8750-17-C-0219), Defense Threat Reduction Agency (HDTRA1-16-C-0029), Schmidt Futures, Gates Ventures, Donald and Jo Anne Petersen Endowment, and an Azure computing gift for COVID-19 research provided by Microsoft.

A new approach for creating custom protein complexes yields asymmetric assemblies with interchangeable parts.

Today we report in Science the design of new protein assemblies made from modular parts. These complexes — which adopt linear, branching, or closed-loop architectures — contain up to six unique proteins, each of which remains folded and soluble in the absence of any binding partners. Baker lab postdoctoral scholars Danny Sahtoe and Florian Praetorius led this work.

Proteins in living cells often come together to form complexes that carry out vital functions. A key feature of these sophisticated assemblies is their ability to exchange parts as needed. To replicate DNA, for example, dozens of unique proteins in a cell spontaneously form into clamps, clamp-loading assemblies, and multi-chain enzyme complexes.

“We see many areas in synthetic biology that would benefit from the ability to pair up proteins in structurally defined ways, but there is currently no easy way to do this,” said Praetorius. “For this project, we wanted to create stable proteins that could spontaneously and rapidly assemble into predictable configurations upon mixing. These could then be fused to other molecules to drive them together.”

Implicit negative design

To begin, the team first sought to create new pairs of proteins that would bind reliably to their cognate partners but not themselves. Such stable heterodimers could then serve as junction points for larger multi-protein architectures.

“To create stable heterodimers, we turned to negative design,” explains Sahtoe. “We decided to introduce three properties into our proteins that make self-association unlikely. First, we made our proteins stable by including real hydrophobic cores. Second, we designed interfaces that require two beta-sheets to pair up, creating unique and continuous beta-sheets across the heterodimer interfaces. And finally, we used Rosetta to model the possible unwanted homooligomeric states and then used that information to prevent those states.”

Using these design principles, the team generated several new heterodimeric proteins that assemble rapidly, even in crowded cell lysates. High-resolution crystal structures of three such dimers confirmed the intended binding modes.

These well-behaved proteins were then recombined and fused to create new “connector proteins” that could bind partners at either end. The team went on to create branched, star-shaped, and angled connectors as well. These modular parts can in principle be mixed and matched to create roughly 500 unique multi-chain assemblies.

The new self-assembling protein parts function as designed in living cells, mediating the assembly of liquid-liquid condensates or more solid assemblies depending on the interaction affinities. The team also showed that assemblies that had already formed could be reconfigured by mixing in new components.

This work was supported by EMBO long-term fellowships, Human Frontiers Science Program long-term fellowships, a Washington Research Foundation Innovation fellowship, and by NIH, DARPA, HHMI, Open Philanthropy Project, The Audacious Project, and Eric and Wendy Schmidt by recommendation of the Schmidt Futures. All data are available in the main text or supplementary materials. Design scripts, protein sequences, design models, and models of assemblies are also available through Zenodo.

Just as convincing images of cats can be created using artificial intelligence, new proteins can now be made using similar tools. In a new report in Nature, we describe the development of a neural network that “hallucinates” proteins with new, stable structures.

“For this project, we made up completely random protein sequences and introduced mutations into them until our neural network predicted that they would fold into stable structures,” said co-lead author Ivan Anishchenko, PhD, an acting instructor in the Baker lab at the Institute for Protein Design. “At no point did we guide the software toward a particular outcome — these new proteins are just what a computer dreams up.”

In the future, it should be possible to steer the artificial intelligence so that it generates new proteins with useful features. “We’d like to use deep learning to design proteins with function, including protein-based drugs, enzymes, you name it,” said co-lead author Sam Pellock, a postdoctoral scholar in the Baker lab.

The research team, which included scientists from UW Medicine, Harvard University, and Rensselaer Polytechnic Institute (RPI), generated two thousand new protein sequences that were predicted to fold. Over 100 of these were produced in the laboratory and studied. Detailed analysis on three such proteins confirmed that the shapes predicted by the computer were indeed realized in the lab.

“Our NMR studies, along with X-ray crystal structures determined by the University of Washington team, demonstrate the remarkable accuracy of protein designs created by the hallucination approach”, said co-author Theresa Ramelot, a senior research scientist at RPI in Troy, New York.

Gaetano Montelione, a co-author and professor of chemistry and chemical biology at RPI, notes “The hallucination approach builds on observations we made together with the Baker lab revealing that protein structure prediction with deep learning can be quite accurate even for a single protein sequence with no natural relatives. The potential to hallucinate brand new proteins that bind particular biomolecules or form desired enzymatic active sites is very exciting”.

“This approach greatly simplifies protein design,” said senior author David Baker. “Before, to create a new protein with a particular shape, people first carefully studied related structures in nature to come up with a set of rules that were then applied in the design process. New sets of rules were needed for each new type of fold. Here, by using a deep-learning network that already captures general principles of protein structure, we eliminate the need for fold-specific rules and open up the possibility of focusing on just the functional parts of a protein directly.”

“Exploring how to best use this strategy for specific applications is now an active area of research, and this is where I expect the next breakthroughs,” said Baker.

Funding was provided by the National Science Foundation (1937533, MCB2032259), National Institutes of Health (DP5OD026389, GM120574, P30GM124165, S10OD021527), Department of Energy (DE-AC02-06CH11357) Open Philanthropy, Eric and Wendy Schmidt by recommendation of the Schmidt Futures program, Audacious Project, Washington Research Foundation, Novo Nordisk Foundation, and Howard Hughes Medical Institute. The authors also acknowledge computing resources from the University of Washington and Rosetta@Home volunteers.

A team led by scientsts in the Baker lab has combined recent advances in evolutionary analysis and deep learning to build three-dimensional models of how most proteins in eukaryotes interact. This breakthrough has significant implications for understanding the biochemical processes that are common to all animals, plants, and fungi. This open-access work appears in Science.

“To really understand the cellular conditions that give rise to health and disease, it’s essential to know how different proteins in a cell work together. In this paper, we provide detailed information on protein interactions for nearly every core process in eukaryotic cells. This includes over a hundred interactions that have never been seen before.”

Proteins are the workhorses of all cells, but they rarely act alone. Different proteins often must fit together to form precise “complexes” that carry out specific tasks, including reading genes, digesting nutrients, and responding to signals from neighboring cells and the outside world. When protein complexes malfunction, disease can result.

“This work shows that deep learning can now generate real insights into decades-old questions in biology —not just what a particular protein looks like, but also which proteins come together to interact,” said senior author Qian Cong, an assistant professor in the department of biophysics at the University of Texas Southwestern Medical Center.

To exhaustively map the interactions that give rise to protein complexes, a team of structural biologists from UW Medicine, University of Texas Southwestern Medical Center, Harvard University, and other several institutions examined all known gene sequences in yeast. Using advanced statistical analyses, they identified pairs of genes that naturally acquire mutations in a linked fashion. They reasoned that such shared mutations are a sign that the proteins encoded by the genes must physically interact.

The researchers also used new deep-learning software to model the three-dimensional shapes of these interacting proteins. RoseTTAFold, invented at UW Medicine, and AlphaFold, invented by the Alphabet subsidiary DeepMind, were both used to generate hundreds of detailed pictures of protein complexes.

“As computer methods become more powerful, it is easier than ever to generate large amounts of scientific data, but to make sense of it still requires scientific experts,” said senior author David Baker, director of the Institute for Protein Design. “So we recruited a village of expert biologists to interpret our 3D protein models. This is community science at its best.”

The hundreds of newly identified protein complexes provide rich insights into how cells function. For example, one complex contains the protein RAD51, which is known to play a key role in DNA repair and cancer progression in humans. Another includes the poorly understood enzyme glycosylphosphatidylinositol transamidase, which has been implicated in neurodevelopmental disorders and cancer in humans. Understanding how these and other proteins interact may open the door to the development of new medications for a wide range of health disorders.

The protein structures generated in this work are available to download from the ModelArchive. The authors thank and remember John Westbrook at the Protein Data Bank for his support in establishing formats and software code to allow efficient deposition of the models into the archive; John sadly passed during the preparation of this manuscript.

The project was led by Ian Humphreys, Aditya Krishnakumar, and Minkyung Baek at UW Medicine as well as Jimin Pei at the University of Texas Southwestern Medical Center. Collaborating institutions include UW Medicine, UT Southwestern, Harvard University, Wayne State University, Cornell University, MRC Laboratory of Molecular Biology, Memorial Sloan Kettering Cancer Center, Gerstner Sloan Kettering Graduate School of Biomedical Sciences, Fred Hutchinson Cancer Research Center, Columbia University, University of Würzburg, St Jude Children’s Research Hospital, FIRC Institute of Molecular Oncology, and Istituto di Genetica Molecolare, Consiglio Nazionale delle Ricerche.

This work was supported by Microsoft, Amgen, Southwestern Medical Foundation, The Washington Research Foundation, Howard Hughes Medical Institute, National Science Foundation (DBI 1937533), National Institutes of Health (R35GM118026, R01CA221858, R35GM136258, R21AI156595), UK Medical Research Council (MRC_UP_1201/10), HHMI Gilliam Fellowship, the Deutsche Forschungsgemeinschaft (KI-562/11-1, KI-562/7-1), AIRC investigator and the European Research Council Consolidator (IG23710 and 682190), Defense Threat Reduction Agency (HDTRA1-21-1-0007), Cancer Prevention and Research Institute of Texas (RP210041), and National Energy Research Scientific Computing Center.

Today we report the development and initial applications of RoseTTAFold, a software tool that uses deep learning to quickly and accurately predict protein structures based on limited information. Without the aid of such software, it can take years of laboratory work to determine the structure of just one protein. With RoseTTAFold, a protein structure can be computed in as little as ten minutes on a single gaming computer. This work was led by Baker lab postdoctoral scholar Minkyung Baek, Ph.D.

RoseTTAFold is a “three-track” neural network, meaning it simultaneously considers patterns in protein sequences, how a protein’s amino acids interact with one another, and a protein’s possible three-dimensional structure. In this architecture, one-, two-, and three-dimensional information flows back and forth, allowing the network to collectively reason about the relationship between a protein’s chemical parts and its folded structure.

As reported in Science, our team used RoseTTAFold to compute hundreds of new protein structures, including many poorly understood proteins from the human genome. We also generated structures directly relevant to human health, including for proteins associated with problematic lipid metabolism, inflammation disorders, and cancer cell growth. And we show that RoseTTAFold can be used to build models of complex biological assemblies in a fraction of the time previously required.

“In just the last month, over 4,500 proteins have been submitted to our new web server, and we have made the RoseTTAFold code available through the GitHub website. We hope this new tool will continue to benefit the entire research community”.

This work was supported in part by Microsoft, Open Philanthropy Project, Schmidt Futures, Washington Research Foundation, National Science Foundation, Wellcome Trust, and the National Institute of Health. A full list of supporters is available in the manuscript.

IPD researchers together with collaborators at the National Institutes of Health have developed experimental flu shots that protect animals from a wide variety of seasonal and pandemic influenza strains. The lead vaccine candidate has entered human clinical testing at the NIH. If it proves safe and effective, these next-generation influenza vaccines may replace current seasonal options by providing protection against many more strains that current vaccines do not adequately cover.

A study detailing how the new flu vaccines were designed and how they protect mice, ferrets, and nonhuman primates appears in the March 24 edition of Nature. This work was led by researchers at the Institute for Protein Design and the Vaccine Research Center, which is part of the National Institute of Allergy and Infectious Diseases at the NIH.

Influenza virus causes an estimated 290,000–650,000 deaths per year. Available flu vaccines, which need to be taken seasonally, often fail to protect against many circulating flu strains that cause illness, and the threat of another influenza pandemic looms.

“Most flu shots available today are quadrivalent, meaning they are made from four different flu strains. Each year, the World Health Organization makes a bet on which four strains will be most prevalent, but those predictions can be more or less accurate. This is why we often end up with ‘mismatched’ flu shots that are still helpful but only partially effective,” said lead author Daniel Ellis, a research scientist in the laboratory of Neil King at the IPD.

To create improved influenza vaccines, the team attached hemagglutinin proteins from four different influenza viruses to custom-made protein nanoparticles. This approach enabled an unprecedented level of control over the molecular configuration of the resulting vaccine and yielded an improved immune response compared to conventional flu shots. The new nanoparticle vaccines, which contain the same four hemagglutinin proteins of commercially available quadrivalent influenza vaccines, elicited neutralizing antibody responses to vaccine-matched strains that were equivalent or superior to the commercial vaccines in mice, ferrets, and nonhuman primates. The nanoparticle vaccines—but not the commercial vaccines —also induced protective antibody responses to viruses not contained in the vaccine formulation. These include avian influenza viruses H5N1 and H7N9, which are considered pandemic threats.

“The responses that our vaccine gives against strain-matched viruses are really strong, and the additional coverage we saw against mismatched strains could lower the risk of a bad flu season,” said Ellis.

This study was supported by the National Institutes of Health (R01GM120553, HHSN272201700059C as well as intramural funding to the Vaccine Research Center), Open Philanthropy Project, Audacious Project, Burroughs Wellcome Fund, University of Washington Arnold and Mabel Beckman cryo-EM center, and a Pew Biomedical Scholars Award.

Scientists at the IPD and Ovchinnikov lab at Harvard have applied deep learning to the challenge of protein design, yielding a new way to quickly create protein sequences that fold up as desired. This breakthrough has broad implications for the development of protein-based medicines and vaccines.

Computational protein design has primarily focused on finding amino acid sequences that encode very low energy target structures. However, what is most relevant during folding is not the absolute energy of the folded state, but the energy difference between the folded state and the lowest-lying alternative states. In a new publication in PNAS, a team led by IPD postdoctoral scholars Christoffer Norn and Basile Wicky describe a deep learning approach that captures the entire folding landscape. They also show that it can enhance current protein design methods.

Read the paper: Norn C, Wicky BIM, et al. Protein sequence design by explicit energy landscape optimization. PNAS. March 2021  [PDF] 

This work builds on a recently described convolutional neural network called trRosetta that predicts residue-residue distances and orientations from input sets of aligned sequences. Combining these predictions with Rosetta energy minimization yielded excellent predictions of structures in benchmark cases. While co-evolution between pairs of positions in the input multiple sequence alignments is critical for accurate prediction of structures of naturally occurring proteins, it is not necessary for more “ideal” designed proteins: accurate predictions of the latter can be obtained from single sequences.

Because distance and orientation predictions are probabilistic, the team rationalized that they might inherently contain information about alternative conformations, and thus provide more information about design success than classical energy calculations. Moreover, because these predictions can be obtained rapidly for an input sequence on a single GPU, they reasoned that it should be possible to use the network to directly design sequences that fold into a desired structure by maximizing the probability of the observed residue-residue distances and orientations versus all others.

In stark contrast to energy-based sequence design approaches which have characterized the field to date, sequence design using trRosetta has the remarkable ability to capture properties of the entire energy landscape and consider alternative states that can reduce the occupancy of the desired target structure. Such implicit considerations of the full landscape are almost impossible to achieve with atomistic models without employing extremely CPU-intensive calculations, including large-scale Rosetta ab initio structure prediction and molecular dynamics simulations on very long time-scales. On the other hand, because of the lower resolution of the trRosetta method, it is less accurate in the immediate vicinity of the folded structure. Integration of trRosetta design into Rosetta all-atom calculations combines the strong features of both approaches. More generally, this work demonstrates how deep-learning methods can complement detailed physically based models by capturing higher-level properties normally only accessible through large-scale simulations.

This project was supported by the National Institutes of Health, Howard Hughes Medical Institute, Audacious Project, and by Eric and Wendy Schmidt, by recommendation of the Schmidt Futures program. Christoffer Norn is supported by a grant from the Novo Nordisk Foundation. Basile Wicky is an EMBO long-term fellow. The source code for the study is available at https://github.com/gjoni/trDesign

Biochemists have created barrel-shaped proteins that embed into lipid membranes, expanding the bioengineering toolkit.

In a milestone for biomolecular design, a team of scientists has succeeded in creating new proteins that adopt one of the most complex folds known to molecular biology. These designer proteins were shown in the lab to spontaneously fold into their intended structures and embed into lipid membranes. Appearing in the journal Science [PDF], this research opens the door to the construction of custom nanoscale tools for advanced filtration and DNA sequencing.

“Right now scientists all over the world are using protein nanopores as part of their effort to sequence genetic material from the pandemic coronavirus and discover mutant strains. For this project, we wanted to design new nanopore proteins completely from scratch that could serve as starting points for a wide range of future applications, including improved DNA sequencing” said lead author Anastassia Vorobieva, a recent postdoctoral scholar in the laboratory of David Baker, director of the Institute for Protein Design at the University of Washington School of Medicine.

Bacteria are encased in a very specialized membrane, called the outer membrane, which protects them from the outside world. Proteins that embed into these membranes facilitate the movement of specific chemicals into and out of the cell. Such natural protein pores share a similar nanoscale structure: a flat sheet of protein that curls in on itself to form a barrel, through which other molecules — including nutrients, vitamins, and even strands of DNA — can pass. This is known as a transmembrane beta-barrel.

To create new transmembrane beta-barrels, Vorobieva and colleagues used molecular design software to draft possible structures. Though they drew inspiration from proteins found throughout the living world, they arrived at sequences that differ from any known before. Their most successful designer proteins contain eight ribbon-like strands that fold into a compact barrel structure that stands just three nanometers tall.

“We began with a relatively simple notion about what would make the proteins fold. But when we tested these initial hypotheses, nothing worked at all. That was very frustrating. We didn’t assume we would get it right the first time, but we did think we could get some information back that would tell us how to move forward. Instead, I had to go back and look carefully at how nature solves this problem. The key was to try to detect patterns in those proteins. It was a really difficult thing to do.”

Researchers in the laboratory of Sheena Radford, Astbury professor of biophysics at the Astbury Centre for Structural Molecular Biology at the University of Leeds, tested whether improved versions of the designer proteins could embed into artificial lipid membranes. They found that they could do so efficiently without the help of any accessory proteins. This is in marked contrast to how natural transmembrane beta barrels fold.

“These designed proteins are interesting from a basic science perspective because they have no evolutionary history,” said Radford, a specialist in protein folding. “By studying them, we can discover some of the essential features that enable transmembrane beta-barrel proteins to fold into a membrane.”

Binyong Liang, an assistant professor working within the laboratory of Lukas Tamm at the University of Virginia School of Medicine, used nuclear magnetic resonance to confirm that the new barrels folded as intended.

This work is the latest achievement in the rapidly progressing field of protein design. In recent years, scientists at the Institute for Protein Design have created innovative vaccines, experimental cancer treatments, and sensors capable of detecting antibodies against COVID-19. The ability to design new proteins from scratch with new functions has implications for diagnosing and treating a wide range of diseases, as well as for materials science.

“With this type of research, it helps to understand a bit about how evolution works at the molecular level, but we are also trying to see beyond that. That’s really the challenge of protein design,” said lead author Vorobieva.

The paper, “De novo design of transmembrane beta-barrels,” appears in the February 19 edition of Science. The research team included scientists from UW Medicine, University of Virginia School of Medicine, University of Leeds, Johns Hopkins University, and The Ohio State University. This work was supported by the National Institutes of Health, Howard Hughes Medical Institute, Fulbright Belgium and Luxembourg, Wellcome Trust, Biotechnology and Biological Sciences Research Council, Medical Research Council, Open Philanthropy Project, Air Force Office of Scientific Research, Nordstrom Barrier Fund, and Eric and Wendy Schmidt by recommendation of the Schmidt Futures program​.

Today we report the design of a new class of protein material that interacts with living cells without being absorbed by them. These large, flat arrays built from multiple protein parts can influence cell signaling by clustering and anchoring cell surface receptors. This breakthrough could have far-reaching implications for stem cell research and enable the development of new materials designed to modulate the behavior of living systems.

The research, which appears in the January 6 edition of Nature [PDF], was led by the Baker lab at UW and the Derivery lab at the Medical Research Council Laboratory of Molecular Biology in Cambridge, UK. 

Cells commonly terminate signaling by absorbing both an activated receptor and the molecule that stimulated it, targeting both for destruction inside the cell. “This tendency of cells to internalize receptors likely lowers the efficiency of immunotherapies,” said Emmanuel Derivery, assistant professor at the MRC Laboratory of Molecular Biology. “Indeed, when antibody drugs bind their target receptors and then become internalized and degraded, more antibody must always be injected.”

To create a way around this, Baker lab postdoctoral scholar Ariel Ben-Sasson designed new proteins that can assemble into large, flat patches upon an external cue. This molecular scaffolding was then further engineered to display signaling molecules. Graduate student Joseph Watson of the Derivery lab showed that such protein materials could latch onto cells, activate surface receptors, and resist being absorbed by the cell for hours or even days. By swapping out which cell surface receptors were targeted by the patch, the researchers showed that different cell types could be activated.

“We now have a tool that can interact with any type of cells in a very specific way. This is what is exciting about protein engineering — it opens fields that people may not expect.”

“This work paves the way towards a synthetic cell biology, where a new generation of multi-protein materials can be designed to control the complex behavior of cells,” said David Baker, professor of biochemistry and director of IPD.

According to co-author Hannele Ruohola-Baker, professor of biochemistry and associate director of the UW Institute for Stem Cell and Regenerative Medicine, versions of these new materials could eventually help physicians alleviate the dangers of sepsis by controlling the inflammatory response to infection and even enable entirely new ways to treat COVID-19, heart disease, and diabetes, and even mitigate the downstream effects of strokes, including Alzheimer’s disease.  “This breakthrough helps pave the way for the use of synthetic cell biology in regenerative medicine,” said Ruohola-Baker.

This research was supported by the UK Medical Research Council, UK Engineering and Physical Sciences Research Council, Wellcome Trust, Human Frontier Science Program, Howard Hughes Medical Institute, US National Institutes of Health, US Department of Energy Office of Basic Energy Sciences Biomolecular Materials Program at Pacific Northwest National Laboratory, Medimmune, and Infinitus.

In a new paper [PDF] appearing in Science, a team of IPD researchers together with colleagues at UW Medicine and Fred Hutchinson Cancer Research Center demonstrate a new way to precisely target cells — including those that look almost exactly like their neighbors. They designed nanoscale devices made of synthetic proteins that target a therapeutic agent only to cells with specific, predetermined combinations of cell surface markers. 

These ‘molecular computers,’ which are based on the LOCKR system, operate all on their own, searching out the cells that they were programmed to find.

“We were trying to solve a key problem in medicine, which is how to target specific cells in a complex environment. Unfortunately, most cells lack a single surface marker that is unique to just them. So to improve cell targeting, we created a way to direct almost any biological function to any cell by going after combinations of cell surface markers.”

The tool they created is called Co-LOCKR, or Colocalization-dependant Latching Orthogonal Cage/Key pRoteins. It consists of multiple synthetic proteins that, when separated, do nothing. But when the pieces come together on the surface of a targeted cell, they change shape, activating a sort of molecular beacon.

The presence of these beacons on a cell surface can guide a predetermined biological activity — like cell killing — to a specific, targeted cell. 

The team showed that Co-LOCKR can focus the cell-killing activity of CAR T cells. In the lab, they mixed Co-LOCKR proteins, CAR T cells, and a soup of potential target cells — some had just one marker, others had two or three. Only the cells with the predetermined marker combination were killed by the T cells. If a cell also had a predetermined “healthy marker”, that cell would be spared.

“T cells are extremely efficient killers, so the fact that we can limit their activity on cells with the wrong combination of antigens yet still rapidly eliminate cells with the correct combination is game-changing.”

This cell-targeting strategy relies entirely on proteins, which sets it apart from most other methods that rely on engineered cells and operate on slower timescales.

This research was conducted at the University of Washington School of Medicine Institute for Protein Design, the Immunotherapy Integrated Research Center at Fred Hutchinson Cancer Research Center, and the University of Washington Department of Bioengineering.

The co-lead authors of this work are Marc J. Lajoie (supported by a Washington Research Foundation Innovation Postdoctoral Fellowship and a Cancer Research Institute Irvington Fellowship from the Cancer Research Institute), Scott E. Boyken (supported by the Burroughs Wellcome Fund Career Award at the Scientific Interface), and Alexander I. Salter (supported by the Hearst Foundation and Fred Hutchinson Cancer Research Center Interdisciplinary Training Grant in Cancer Research). This work was also supported by the National Institutes of Health (R01 CA114536, NIGMS T32GM008268, 1R21CA232430-01, T32CA080416), the National Science Foundation (CHE-1629214), the Defense Threat Reduction Agency (HDTRA1-18-1-0001), the Nordstrom Barrier IPD Directors Fund, the Hearst Foundation, the Washington Research Foundation and Translational Research Fund, the Howard Hughes Medical Institute, the Open Philanthropy Project, and The Audacious Project organized by TED.

The same basic tools that allow computers to function are now being used to control life at the molecular level, with implications for future medicines and synthetic biology. 

Together with collaborators, we have created artificial proteins that function as molecular logic gates. These tools, like their electronic counterparts, can be used to program the behavior of more complex systems. The team showed that the new designer proteins can regulate gene expression inside human T-cells, a development that may improve the safety and durability of future cell-based therapies.

“Bioengineers have made logic gates out of DNA, RNA and modified natural proteins before, but these are far from ideal. Our logic gates built from de novo designed proteins are more modular and versatile, and can be used in a wide range of biomedical applications” said David Baker.

Whether electronic or biological, logic gates sense and respond to signals in predetermined ways. One of the simplest is the AND gate; it only produces output when one input AND another are present. For example, when typing on a keyboard, pressing the Shift key AND the A key produces an uppercase letter A. Logic gates made from biological parts aim to bring this level of control into bioengineered systems. With the right gates operating inside living cells, inputs such as the presence of two different molecules — or one and not the other — can cause a cell to produce a specific output, such as activating or suppressing a gene.

“The whole Apollo 11 Guidance Computer was built from electronic NOR gates,” said lead author Zibo Chen, a recent UW graduate student. “We succeeded in making protein-based NOR gates. They are not as complicated as NASA’s guidance computers, but nevertheless are a key step toward programming complex biological circuits from scratch.”

Recruiting a patient’s own immune cells in the fight against cancer has worked for certain forms of the disease. But targeting solid tumors with this so-called CAR-T cell therapy approach has proven challenging. Scientists think part of the reason why has to do with T cell exhaustion. Genetically altered T cells can only fight for so long before they stop working, but with protein logic gates that respond to exhaustion signals, the team from UW Medicine hopes to prolong the activity of CAR T cells.

“Longer-lived T cells that are better programmed for each patient would mean more effective personalized medicine,” said Chen.

This work was led by researchers at the UW Medicine Institute for Protein Design. It also included biochemists from Northwestern University, The Ohio State University, Altius Institute for Biomedical Sciences and UC San Francisco.

Read the full report in Science   PDF

Today we report the design of protein sequences that adopt more than one well-folded structure, reminiscent of viral fusion proteins. This research moves us closer to creating artificial protein systems with reliable moving parts.

In nature, many proteins change shape in response to their environment. This plasticity is often linked to biological function. While computational protein design has been used to create molecules that fold to a single stable state and to re-engineer natural proteins to alter their dynamics or fold, the design from scratch of closely related sequences that adopt well-defined but divergent structures has remained an outstanding challenge.

To create shape-shifting proteins, a team led by recent Baker lab postdoc Kathy Wei, Ph.D., began by identifying sets of amino acid sequences predicted to fold into very different structures — in this case, pairs of cylindrical helical bundles with different lengths.

“We knew from the beginning that we wanted a sequence to transform between a short state with helical “arms” pointed “down” and a long state with helical “arms” pointed “up”. The plan was to use established protocols to first design different proteins that are in each of the two states and then mutate the sequences of these two starting points toward each other until we found a sequence that could fold into both states,” said Wei.

After rounds of design on the computer and testing in the lab, the team succeeded in creating a single molecule that could be seen in both states.

“One of the main challenges for this project was finding a way to tell if the proteins took on the shape they were designed to be in. High-throughput screening methods tend to rely on an enzymatic property of a protein. Since these designed proteins only differed in their shapes, we had to use crystallography and NMR to check their folding, which is a slow process and not guaranteed to yield results.”

“While we found a really promising protein sequence that we can measure in both of the designed states, it’s surprisingly much less dynamic than we would’ve expected. Next, we want to understand how to make the conformational changes more dynamic and how to trigger them in a controlled manner.“

The team included scientists from the University of Washington, UC Berkeley, UC Santa Cruz, and Stanford. Their work was supported by the NIH, DOE, HHMI and the Chan Zuckerberg Biohub.

Computational design of closely related proteins that adopt two well-defined but structurally divergent folds. Kathy Y. Wei, Danai Moschidi, Matthew J. Bick, Santrupti Nerli, Andrew C. McShan, Lauren P. Carter, Po-Ssu Huang, Daniel A. Fletcher, Nikolaos G. Sgourakis, Scott E. Boyken, and David Baker. PNAS.

Read the full report: https://www.pnas.org/content/early/2020/03/17/1914808117   PDF