Opportunities and challenges in design and optimization of protein function

Abstract

Protein design methodology has made remarkable progress over the past decade. Historically, the low reliability of purely structure-based design methods limited their application. But recent strategies that combine structure- and sequence-based calculations as well as machine-learning approaches, have dramatically improved success. These approaches have led to design of increasingly complex structures and therapeutically relevant activities. Additionally, protein optimization methods have improved the stability and activity of complex eukaryotic proteins. Thanks to their heightened reliability, design methods have been applied to improve therapeutics, enzymes for green chemistry, and generated vaccines, antivirals and drug-delivery nanostructures. These exciting developments bring protein design closer to becoming a mainstream approach in protein science and engineering. Moreover, they reflect an increased understanding of basic rules that govern the relationship between protein sequence, structure, and function. We note, however, that de novo design is still limited mostly to α-helix bundles, restricting its potential to generate sophisticated enzymes and diverse protein and small-molecule binders. Designing complex protein structures is a challenging next step if the field is to realize its objective of generating new-to-nature activities. Here, we review developments in computational protein optimization and de novo design of function and highlight pressing areas for future research.

Introduction

Life depends on myriads of protein functions. Decades of efforts to understand the determinants of protein function have resulted in a deeper mechanistic understanding of life processes and the generation of new proteins, including life-saving drugs and enzymes that catalyze desirable chemical reactions in an environmentally friendly way^1,2. Protein function is determined by amino acid sequence and structure (Fig. 1A), and understanding these determinants can dramatically improve our ability to control protein activity. Due to the complex relationships between sequence, structure, and function, however, protein-engineering methods have historically relied on experimental screening of natural proteins and iterative mutational processes to discover and optimize biochemical functions^3–6. Although powerful, these approaches exhibit limited reliability, are time-consuming, and can mostly be applied to proteins with functions that are amenable to high-throughput screening^7–9. Additionally, these methods rely on trial and error rather than on direct engineering of the desired properties. The fact that proteins that are critical to research, human health and industry are generated using random processes reveals critical gaps in our understanding of how function is encoded in proteins. One of the most important goals of the protein design field is to bridge these gaps.

Figure 1. Goals of protein design methodology.

A. The fundamental paradigm of protein science is that amino acid sequence determines structure which in turn determines function²⁷. Protein design has been historically dubbed as the inverse folding problem²⁰ of finding the sequence that will fold to a desired structure (step 3. Fold Design). By extension, the problem of finding the sequences and structures that can realize a desired function is the inverse function problem (step 4. Function Design). Computational methods for (1) structure prediction and (2) function annotation have recently made considerable advances. The inverse problems of fold (3) and function (4) design have also made important progress, but significant gaps remain in the design of complex folds and functions relative to those observed in nature. B. A fitness landscape is an abstract model that relates protein variants to their relative activities²⁸. Nearby points on the landscape represent close sequences and the relative heights represent relative fitnesses or activities. Evolution and conventional methods in protein engineering iteratively explore nearby mutants (dashed black lines) starting from a naturally occurring protein (circle close to the viewer). Such methods may find local fitness optima but are restricted in their ability to find distant ones (peak in the background). New computational design methods may be able to reach such solutions which are unlikely to emerge through evolutionary approaches or would emerge only slowly.

The main challenge of computational protein design is to develop reliable models for protein generation and optimization that are based on computable features. Until a decade ago, such methods made only limited contributions to protein engineering due to their low accuracy and reliability^6,10–14. Since then, however, the field has made significant progress, especially in the design and prediction of protein structures¹⁵ (Fig. 1A steps 1 and 3), and we refer the reader to recent reviews on the milestones that enabled these advances^16–18. Given this progress, the next frontier for computational protein design can be dubbed the “inverse function” problem, expanding on the influential paradigm from the early 1990’s that cast design as the “inverse folding” problem (Fig. 1A step 4)^19,20. Whereas the inverse folding problem asks which amino acid sequences fold into a desired three-dimensional structure, the inverse function problem asks how to develop strategies for generating new or improved protein functions. In addition to verifying our understanding of protein design principles, advancing this front will accelerate fundamental and applied discoveries (Fig. 1B) that are crucial to human health, industry, and environmental sustainability.

The advances made over the past decade are mainly due to methods that combine molecular and physical principles with data-based approaches. These have led to higher reliability and broad usefulness. Particularly, stability-design methods have become so reliable that they have been successfully applied to dozens of different protein families, including ones that did not yield to experimental optimization strategies²¹. In addition, de novo protein design, the generation of proteins from scratch, has reached remarkable accuracy and scope, allowing the routine design of small proteins and large self-assembling complexes^22–24. De novo designed proteins have been further designed to generate new binders of proteins or small molecules^23,25,26, advancing toward completely rational generation of new-to-nature activities.

Thus, protein design methodology has come a long way, and some strategies have become integrated in basic and applied protein engineering workflows. Here, we will review the main biophysical challenges that the protein design field has had to address to increase the reliability and scope of these methods. In addition, we will survey some of the opportunities that new methods have already begun to realize while underlining areas where they still struggle. Our perspective is divided into two broad categories of design goals: optimizing existing, typically natural, protein activities, and de novo design of fold and function. We will also review the increasing uses of machine learning in design and highlight areas where protein design methodology may lead to transformative advances in the coming years.

Protein Optimization

Natural evolution adapts proteins to the requirements of their host organisms. The requirements of research and application, however, may differ in important ways from those of the natural host. For instance, the expression levels, stability, specificity, or activity of the natural protein may fall short relative to those needed for human use⁶. Therefore, understanding the rules that govern protein stability and activity has been a long-standing goal of protein science. Protein optimization may be especially challenging because it typically strives to find a protein that excels in several different properties^9,29. Furthermore, large gains in stability and activity often require many mutations³⁰, demanding methods that can accurately predict how changes in sequence and structure impact activity.

According to the Thermodynamic Hypothesis, the protein native-state energy must be significantly lower than all other states, including misfolded and unfolded ones, for a significant fraction of the protein to fold uniquely into the native state²⁷ (Figure 2A). Thus, general methods for protein design must often comprise elements of both positive and negative design, favoring the desired state and disfavoring competitors, respectively. The fundamental problem that all general protein design strategies face is that only the desired state is defined in atomic detail and is amenable to atomistic calculations, whereas the competing structural states are typically unknown¹⁰. To appreciate the magnitude of the negative-design problem, we may consider that the number of possible undesired states is likely to scale with the exponent of the size of the protein^31,32, and that the median size of proteins is 300 amino acids³³. Given the astronomically large space of combinations of mutations and conformations, ensuring that the desired state exhibits significantly lower energy than any of the myriads of unknown competing states is a daunting task. Indeed, numerous studies have demonstrated that design and engineering efforts often result in proteins that exhibit lower stability than the natural starting point and a tendency to aggregate, misfold, or exhibit conformationally flexible regions^13,34–39.

Figure 2. Computational stability design.

A. Schematic representation of the folding landscape of a marginally stable protein generated using physics-based design methods alone (top) versus physics and phylogeny-driven approaches (bottom). Physics-based methods have access only to the native state and may inadvertently lower the energy of competing (misfolded and unfolded) states, thereby not addressing negative design. Mutations that lower the energy of undesired states are likely to be purged by evolutionary selection; accordingly, a protein designed using a combination of atomistic and phylogenetic constraints may preferentially fold into the native state. E: energy, vdW: van der Waals interactions, elec: Coulomb electrostatic interactions, sol: solvation. B. Stability design of the SARS-CoV-2 Spike protein monomer (PDB entry: 6VYB). A design that encodes 20 mutations (S2D14) in the S2 Spike protein subunit (top left) increased protein yields 11-fold (top right) while improving pseudovirus neutralization titers (pVNT50) of several SARS-CoV-2 variants of concern (bottom)⁶⁷. S-2P is a rationally designed Spike variant that encodes stabilizing mutations Lys986Pro and Val987Pro. Data for panel (B) generously provided by Wayne Harshbarger.

One solution to the problem of encoding negative-design elements is using additional data to guide structure-based calculations. For instance, it is likely that sequence elements that are prone to misfolding and aggregation have been eliminated through natural selection⁴⁰. In an approach called evolution-guided atomistic design, the natural diversity of homologous sequences is analyzed at each position of the target protein to eliminate rare mutations from the design choices before the atomistic design step^21,41. Such filtering implements some aspects of negative design while reducing the design sequence space by many orders of magnitude and focusing it on the space that is more likely to fold stably and accurately^40,42,43. Subsequent atomistic design calculations stabilize the desired state within this reduced sequence space, thus implementing elements of positive design (Fig. 2A).

Another successful approach that improves design success relies on machine-learning inferences applied to experimental data to predict mutations for further experimental screening^44–46. These empirical approaches are powerful^45,47,48 and may become even more so with the advent of deep-learning based Large Language Models (LLMs). However, they demand iterations of mutagenesis and screening for each target protein and apply only to proteins that are amenable to at least medium-throughput screening. The current perspective focuses on structure-based design methods that do not rely on experimentally determined mutation data, and we refer the interested reader to recent reviews on protein engineering by machine-learning methods applied to experimental datasets^44,46.

[H2] Protein stability design

Many natural proteins exhibit only a small energy difference between the native state and the myriads of unfolded or misfolded states and are therefore marginally stable⁴⁰. Marginal stability is often masked in the natural host of the protein due to dedicated cellular machines that promote folding, such as chaperones and proteases⁴⁹. Furthermore, in some cases, marginal stability may be an evolutionarily selected property that ensures fast protein turnover⁵⁰. Yet, marginal stability typically implies low expression levels in heterologous hosts^51–53, limiting usefulness in basic and applied research. The problem of low heterologous expression levels is surprisingly prevalent, as the fraction of cytosolic proteins that are amenable to overexpression is estimated to be <50% of any proteome^54,55, and membrane proteins are typically even less accommodating^48,56,57. Furthermore, it is especially difficult to introduce mutations that improve activities in marginally stable proteins because these mutations may reduce stability below the threshold required for protein folding^8,9,29,58. Therefore, marginal stability is perhaps the most general problem in protein engineering and design.

Due to the very broad impact of protein stability on research and engineering, several structure-based methods have been developed over the past decade to address stability design^42,59–61. The native state of a protein relies on weak and transient forces, and it is the combined effect of thousands of such interactions that favors this state over all others⁴⁰. Thus, stability optimization methods may suggest dozens of mutations relative to the wild-type protein to generate significant improvement in protein stability⁴².

As native-state stability correlates with heterologous expression levels^51–53, stability design approaches often enhance the yields of functional protein. The impact on expression levels has sometimes been remarkable, generating variants with identical or improved functional properties that can be efficiently expressed rapidly and economically^42,62–64. New methods for stability optimization allowed testing hypotheses on the activity of challenging proteins⁶⁵, characterizing the activities of proteins that could not be functionally produced before⁶⁶, reducing the costs of manufacturing therapeutics^62,67, improving the resilience of proteins to environmental conditions^68–71, and engineering improved functions^9,72,73.

In the development of vaccine immunogens for clinical use, stability and protein production costs are significant bottlenecks that protein design may alleviate^74–77. For instance, the protein RH5 from the parasite Plasmodium falciparum, the most virulent malaria-causing strain, is the leading candidate for a vaccine that targets the parasite blood stage. But this protein can only be produced in expensive insect cells and denatures at approximately 40°C, whereas vaccines for the developing world require large-scale production and would preferentially not depend on refrigerated transport. A mutant that was designed for higher native-state stability could be robustly expressed in E. coli and exhibited nearly 15°C higher thermal resistance while maintaining the immunogenicity of the wild-type RH5 in animal models. This design lowers costs of production and may enable refrigeration-free transport⁶² and has recently entered Phase 2 clinical trials⁷⁸. Similarly, a designed variant of the SARS-CoV-2 Spike protein exhibited tenfold improved production levels in mammalian cells⁶⁷ (Fig. 2B). Encouragingly, this design elicited a substantial increase in the titer of neutralizing antibodies against both the original SARS-CoV-2 Wuhan strain and subsequent variants of concern, including Omicron. The improvement in neutralization and breadth is likely due to increased conformational stability of the design relative to the natural Spike protein (Fig. 2A), highlighting the important functional implications of specifically stabilizing the protein native state. These results suggest that stability design methods may rapidly, economically, and effectively address both emerging and long-standing global-health challenges.

Stability design has also expanded the feasibility of using enzymes as therapeutics^64,79. A notable example is the bacterial enzyme chondroitinase ABC, which has attracted biomedical research interest for its ability to induce nerve recovery following stroke and spinal-cord injury by degrading the proteoglycan in scar tissue^79–81. The enzyme exhibits limited stability at physiological pH and body temperature, however, and attempts to improve these critical properties have met little success for a decade, perhaps due to its unusual size (>1,000 amino acids). By contrast, by testing merely three designs generated by a webserver, a stabilized variant was produced with a functional half-life of >6 days, compared to under a day for the wild type enzyme, while showing a threefold increase in proteoglycan degradation⁷⁹.

Until recently, atomistic design methods could only be reliably applied to crystallographically determined protein structures due to the high sensitivity of design methods to atomic details⁸². New AI-based structure-prediction methods, such as AlphaFold2⁸³, however, have exhibited remarkable accuracy close to that obtained from crystallographic analysis. This exciting development increases the number of proteins that can be computationally engineered from under 200 thousand experimentally determined entries in the Protein Data Bank (PDB) to the hundreds of millions of protein sequences in genomic databases⁸⁴. In principle, any protein sequence can now be used as a starting point for design to improve stability and expressibility⁶⁶. This new and unexpected capability is especially useful where large numbers of newly discovered proteins must be optimized, such as in antibody⁸⁵ or enzyme discovery workflows⁶⁶ or where speed is crucial, such as in response to a pandemic threat⁷⁵. This ability depends on the accuracy of the prediction tools, and we expect newer generations of structure-prediction methods that model bound ligands and cofactors⁸⁶ to dramatically improve the scope and reliability of design workflows. Furthermore, a future improvement in structure prediction of long loops, especially in antibodies⁸⁷, is likely to enhance design capabilities in this important therapeutic area. Recent design studies have also extended stability design to membrane proteins^88,89 which comprise more than a quarter of every proteome and are typically extremely challenging for heterologous overexpression^56,57 despite their high impact as drug targets.

Several stability design methods were implemented as automated web servers^42,90–92 becoming accessible to scientists with no expertise in computational modeling. These tools have led to dozens of reports with successful case studies that required screening only a handful of designs. To be sure, not all proteins targeted for stability design exhibit improvements⁶³, but given the broad usefulness of current methods, we may be close to the point at which stability design can be applied reliably and universally to any protein class. Beyond the practical implications for accelerating research and technology, this exciting development verifies our understanding of some of the basic rules that determine native-state stability and foldability⁴⁰ even in large eukaryotic proteins that exhibit complex folds and functions.

Enhancing protein activity

Optimizing protein activity is a critical and challenging frontier of design methodology development. Improving affinity, catalytic rate, or substrate selectivity often requires mutations within the active site, but most active-site mutations lead to a significant loss in function. A further complication is that large gains in function often demand several mutations. Active-site positions are proximal, however, and the functional impact of mutations is likely to depend strongly on other mutations in a phenomenon known as epistasis^93,94. Epistasis may severely restrict the emergence of new functions in natural or lab evolution because evolutionary processes rely on the stepwise accumulation of mutations each of which must at least preserve activity²⁸. Desirable mutants that exhibit high activity may be unreachable by such processes if, due to epistasis among the mutations, these processes require passing through low-activity variants^93–97 (Fig 1B). Another severe complication is that functional traits may be mutually antagonistic, such as when mutations that improve activity come at the cost of lower specificity or stability^8,9,29,58. Protein dynamics are a significant source of additional uncertainty in optimizing function. The activity of many enzymes and binding proteins relies on changes in protein conformation between states with similar energy, and typically not all states are characterized in the atomic detail necessary for design⁹⁸. Therefore, although the limitations of protein evolution provide a strong impetus to develop non-evolutionary computational design methods for enhancing activity, the complexities of active-site design entail that such methods are less developed and general than those for stabilizing proteins^43,99,100.

FuncLib is an active-site design method, implemented as an online tool, that addresses some of the complexities of designing active sites⁴³. Instead of accumulating mutations iteratively as evolutionary methods do, it searches for simultaneous combinations of active-site mutations that form stable, low-energy constellations. In practice, the designs exhibit shape and electrostatic diversity in the active site while ensuring that the mutations do not deform the catalytic constellation of residues. FuncLib has typically been applied to generate variants that exhibit diverse activity profiles starting from a natural or engineered enzyme^{43,72,101–103}. In these cases, the method does not model any substrate in the active-site pocket, and yet several studies observed large improvements in activity. Typically, this approach leads to improvements in activities already observed in the parental enzyme, and encouragingly, in a handful of cases new specificities emerged that were not seen in the wild-type protein^102,103.

In an application of FuncLib to a fungal unspecific peroxygenase (UPO) (Fig 3A), >70% of the experimentally screened designs were functional and several exhibited dramatic differences in enantio- and regioselectivities¹⁰³ (Fig. 3B). The broad versatility of the set of designs was highlighted by two that exhibited opposite enantioselectivity for two different compounds. UPOs are attracting significant interest for their ability to economically oxidize a wide range of aliphatic and aromatic biomolecules¹⁰⁴, and enzymes that exhibit new substrate specificities may find uses in research and the chemical industry. In another demonstration, designs of a bacterial phosphotriesterase (Fig. 3C) exhibited up to 4000-fold improvement in the hydrolysis rate of synthetic venomous nerve agents, reaching the enzymatic efficiency required for therapeutic use⁴³. One of the designs that exhibited large gains in activity comprised four point mutations, and mutational analysis demonstrated that due to epistasis among the designed mutations, no evolutionarily plausible path could accumulate these mutations without going through intermediates that exhibit low activity (Fig. 3D).

Figure 3. Design of enhanced activity.

A. Amino acid residues (orange) in the UPO active site are in close proximity (PDB entry: 6EKZ). Computational design introduces combinations of simultaneous mutations to generate stable and preorganized active sites¹⁰³. Heme shown in cyan; the substrate propranolol in purple. B. Select designs were experimentally tested for C−H oxyfunctionalization of a variety of styrenes. The yellow and blue bars indicate the fraction of enantiomers generated by the designs and starting enzyme (WT). Molecular structures of the products are shown at the top. Designs show striking changes in enantiospecificity (compare, for instance, WT and d28 in the right-most column). C. The active site of a bacterial phosphotriesterase (PDB entry: 1HZY). Wild type residues are in gray sticks, and designed mutations in a quadruple mutant are in cyan. D. Mutations in the active site of the phosphotriesterase show strong epistasis in their impact on the measured activity (2-naphthyl acetate hydrolysis). Each circle represents a phosphotriesterase mutant, and the area of the circle is proportional to the specific activity of the design. The starting enzyme exhibits low specific activity (360 μM s^-1 mg^-1 protein). Each of the point mutants exhibits improved specific activity, but activity declines in the double mutants relative to the His257Trp single mutant. Last, the quadruple mutant (designed enzyme) substantially improves specific activity relative to all single or double mutants. Adapted from ref ⁴³.

These results demonstrate that addressing the challenge posed by epistasis may open the way to dramatically improve existing activities and even generate new ones. Additionally, they show that designing stable and catalytically preorganized active sites is a prerequisite for designing efficient enzymes, in line with theoretical considerations¹⁰⁵. Furthermore, the designs typically do not exhibit stability-activity trade offs, suggesting that computational methods may access high-functioning variants that are unlikely to emerge from applying evolutionary methods. We note, however, that the design methods only increase the likelihood of obtaining highly functional variants that may exhibit substrate specificity differences; they are not sufficient for designing variants with predetermined activity profiles. A future complete computational activity-optimization solution will require advances in ligand docking⁸⁶ and possibly also in transition-state modeling to generate structural models that bias the computational design method in favor of the desired catalytic outcome.

Learning from iterative high-throughput structure-based design

Over the past decade, computational methods have matured to the point that they can generate thousands and even millions of designs through automated pipelines^106–109. These methods and advances in massive DNA synthesis, deep sequencing, experimental screening, and machine learning have enabled the application of statistical learning to draw general lessons about protein folding and activity.

High-throughput structure-based design methods have been used to map sequence and conformation spaces of “miniproteins”¹⁰⁶, proteins typically ≤ 60 amino acids in size. To study the determinants of folding and stability, nearly a million miniproteins were screened for protein stability and foldability¹⁰⁷ using high-throughput proteolytic screening and deep sequencing. This analysis quantified the contribution to stability of molecular properties, such as the compatibility of amino acid choices to the local secondary structure, large hydrophobic cores, charge stabilization, and epistatic mutations. This strategy was also used to design 20 thousand miniproteins to screen for binders of influenza hemagglutinin and botulinum neurotoxin B²⁵. A machine-learning analysis distinguished binders from inert proteins by their stability and contact order at the interface, improving binder recovery success rate by eightfold in a second design round. During the SARS-Cov-2 pandemic, this approach was used to generate novel binders of the Spike receptor binding domain that inhibited ACE2 binding²⁶. This combination of high-throughput design and screening is opening exciting possibilities for complete automation of de novo binder design.

Natural proteins are typically larger than 200 amino acids³³ and structurally more complex than miniproteins¹¹⁰. The combinatorial assembly and design of enzymes (CADENZ) method generates stable and diverse backbone fragments by selecting dozens of structural fragments from homologous enzymes and optimizing their sequences for maximal stability and compatibility with other fragments¹⁰⁸. It then uses a machine-learning strategy to select designed fragments that would combine to form low-energy full-length proteins that exhibit stable catalytic constellations. CADENZ was used to design a million models of structurally diverse proteins based on glycoside hydrolase structures, and high-throughput activity screening isolated thousands of functional designs that exhibited high diversity (80-160 mutations from any natural enzyme including large insertions and deletions). Comparing features of functional and non-functional designs revealed different physicochemical preferences in active sites versus the rest of the protein. Functional designs were characterized by dense, yet low-energy constellations of active site residues, while the remainder of the protein required compactness, large hydrogen-bond networks and high compatibility between amino acid identity and the local backbone conformation. Implementing these features in a second-generation repertoire yielded more than 10 thousand functional enzymes from a single screen, with thousands of designs that exhibited different substrate specificity from one another¹⁰⁸. Based on similar principles, htFuncLib generates large libraries of multipoint active site mutants that are predicted to form amino acid constellations that do not disrupt the positioning of functional groups in the active site¹⁰⁹. Applied to 27 positions in the chromophore-binding pocket of GFP, the method generated a repertoire encoding approximately 10 million designs from which more than 16 thousand functionally fluorescent designs were retrieved, many of which exhibited more than five active-site mutations. A detailed biophysical analysis revealed dozens of designs with potentially useful changes in spectral properties, including large improvements or changes in brightness, excitation spectra, and thermal stability relative to the starting protein.

Thus, despite the high sensitivity of protein active sites to mutation, high-throughput design studies reveal that the space of active-site functional variants is likely to be large. For instance, despite the long-term interest in engineering GFP, fewer than 100 variants in the chromophore-binding pocket were documented¹¹¹ before the htFuncLib study. It is important to note, however, that while protein active sites may accommodate thousands of variants, these numbers are dwarfed by the vast space of nonfunctional active-site variants (Fig 4). The ability of new energy-based design methods to effectively navigate the complex mutational space of active sites suggests that we have a sound understanding of some of the biophysical underpinnings needed to design improved activity. In particular, the analysis of the newly designed GFP functional variants highlighted the crucial role that protein stability and active-site residue preorganization play in determining whether a protein is functional. On a practical level, design calculations are largely automated, and gene-synthesis costs for such customized libraries are modest¹¹² opening the way for new computational methods to find uses in areas of protein engineering that are amenable to high-throughput screening, such as antibody and binder discovery and optimization. Additionally, we anticipate that the large and diverse datasets that emerge from such screens that systematically relate sequence and structure to stability and activity will drive further improvements in design methodology, including in AI-based design of function.

Figure 4. Design methods must navigate an astronomically large sequence space that is extremely sparse in functional proteins.

(I) The theoretical sequence space of an average-sized protein is huge: for a 350 amino acid protein, 20³⁵⁰ >10⁴⁵⁵ sequences. By designing enzyme domains as in the CADENZ approach, the hypothetical space of possibilities approaches such numbers. Even when designing amino acids only in a functional site (II), as in the case of GFP design (htFuncLib), the hypothetical sequence space is orders of magnitude greater than the number of viral particles in the world¹¹³, and no experimental screening method could search it adequately. In the case of GFP, phylogenetic data and physics-based calculations focus the search and reduce the sequence space by 18 orders of magnitude (III). This still leaves a sequence space that cannot be assayed even by the most high-throughput methods currently available, such as ribosome display. A machine-learning procedure can filter out mutations that do not combine with one another to form stable and foldable proteins, leading to the machine learning (ML) restricted space (IV). This restriction enables constructing an effective library that is enriched with stable and potentially functional designs. From this sequence space of ~10⁷ variants, 10⁴ functional fluorescent proteins were recovered (V)¹⁰⁹ compared to fewer than 100 active-site variants recorded in the fluorescent proteins database¹¹¹.

De novo design of structure and function

All natural proteins are the product of evolution. Due to the limitations of evolutionary processes, new molecular activities only slowly emerge through natural evolution, and many desirable activities may not have been found¹¹⁴. The rationale of de novo design is that these fundamental limitations of evolutionary processes can be overcome by tailoring a protein structure for each desired activity without recourse to natural starting points^17,22. Thus, general methods for de novo function design is the ultimate goal of the “inverse-function” problem (Figure 1A). Although this goal has been attempted for decades, the approaches to achieve it have changed dramatically^16–18. Starting with the manual application of protein design principles to generate new structures^115–117, a period of physics-based approaches followed^10,22 up to the recent application of generative AI^23,118. As the field develops, we are witnessing a surge in the structural complexity and diversity of de novo-designed proteins in step with the adoption of new approaches (Figure 5). While this progress is notable, however, the complexity of de novo-designed folds still pales in comparison to the natural repertoire of structures (Figure 5a). In fact, a very large fraction of designed proteins adopt simple all α helix topologies whereas natural proteins sample a much greater diversity of secondary structures and topologies (Figure 5b). We argue that the simplicity of designed folds constrains the types of functions that they can present. In the following sections, we provide an overview of the achievements in de novo design of fold and function and discuss obstacles that must be overcome to bridge the gap to natural proteins.

Figure 5. Comparison of structural features in natural and de novo designed proteins.

A) De novo proteins are yet to reach the structural complexity of natural ones as measured by size and relative contact order. Structures of a de novo α-helix bundle (PDB: 7CBC) are highlighted versus two natural proteins (PDBs: 3NF4 and 3ZQJ) B) Secondary Structure Element (SSE) content in natural and de novo proteins. De novo proteins are biased towards high α-helix content whereas natural ones are structurally diverse. C) Relative contact order of designed protein structures found in the PDB plotted against their time of publication. Each design is also labeled according to the class of design generation method (PDBs: 1AL1, 1QYS, 3QA9, 3NF4). D) SSE content of natural and de novo binder interfaces (excluding antibodies). De novo binders are biased towards presenting helices at interaction sites compared to natural binders. Antibody binding surfaces, which are not represented in this figure, are dominated by loop regions and almost entirely devoid of helices. E) Number of interfacial residues in natural protein interactions compared to de novo protein binders. F) Hydrogen bonds in natural interfaces compared to de novo interfaces. G) Examples of de novo designed protein-protein interactions: botulinum neurotoxin binder²⁵, influenza binder³⁸, PDL-1 binder and SARS-CoV-2 receptor binding domain (RBD) binder¹⁵⁰. Targets are colored green and binders in red.

[H2] De novo fold design

The de novo design problem is typically formulated in two steps: backbone generation and sequence search^24,119 (Fig. 1A). To generate a plausible backbone, most design methods use regular structural motifs of α helices and β sheets connected by short loops to create idealized topologies dominated by densely packed secondary structures^22–24,119. In such simple topologies, the sequence-structure rules are well understood, simplifying the design process^24,120–122. Moreover, to increase the likelihood of achieving the desired structure, fold design can use features that exclude alternative conformations, for instance, by using short segments to connect secondary structure elements, by ensuring high compatibility between amino acid type and secondary structure, and by introducing mutations that relieve backbone and residue strain in the desired state^{106,107,120,122}.

A guiding principle for backbone design is the notion of “backbone designability”^123,124. A designable backbone is one for which it is possible to find a sequence of natural amino acids that yields a unique structure. Designable backbones usually contain common structural motifs found in nature, both at the level of the regular secondary elements and the 3D arrangement of those elements, increasing the likelihood of producing a stable protein^125,126 without explicitly encoding negative-design principles²⁴. The resulting protein structures can be thought of as “idealized” versions of natural protein folds. In these versions, the secondary structures are canonical, well-packed, and unstrained to promote stability and foldability. The ability to automatically design thousands of such proteins^106,107 is a remarkable demonstration of the success of protein design methods in capturing some of the features that are critical for protein folding. Nevertheless, idealized designs lack the structural irregularities that are hallmarks of functional motifs in natural proteins, such as surface cavities, kinked secondary structure elements, and desolvated polar groups¹¹⁰. Because of their regularity, many designed proteins exhibit high stability, but methods for designing sophisticated activities in de novo proteins are limited. Part of the functional design challenge is that natural proteins are much larger and topologically more complex than those that have been the subject of design studies¹¹⁰. Moreover, function in natural proteins often depends on large and structured loop regions^110,127, and such regions continue to pose a severe challenge to structure prediction and design methods⁸⁷.

Following backbone design, a compatible sequence is generated. Physics-based methods have been widely applied in this field. They compute an energetically favorable amino acid for each position of a computed backbone structure. Given the high dimensionality of the sequence search problem, it is impossible to enumerate the full space of sequences¹²⁸, and heuristic^129,130 sampling approaches^129,130 are used to find low-energy sequences. Physics-based methods have enabled several landmark designs, including proteins with folds not previously found in nature²⁴, idealized folds based on natural ones¹²⁰, a large number of miniproteins and peptides¹³¹, and the exploration of new geometries of secondary structure elements in protein folds¹³².

In recent years, the toolbox for protein design has been expanded with deep learning methods. Typically, these methods learn structural and amino acid patterns from natural protein structures and use the learned distributions to generate new backbones and matching amino-acid sequences. In contrast to physics-based methods, deep learning approaches need huge amounts of high quality data for training, and the basis for their design decisions is often opaque, making it challenging to interpret the results in biophysical terms based on molecular features. Nevertheless, machine learning methods are more efficient that physics-based ones and scale well with the number of amino acids. Similar to physics-based methods, most deep learning methods rely on discrete steps of structure and sequence design, though recent developments design structure and sequence in a single step^133,134. Once sequences are generated their quality is typically assessed by verifying that AlphaFold2 predicts a model structure that closely matches the designed one^{23,118,135–141}.

The deep learning-based de novo design approaches have demonstrated remarkable accuracy in generating protein folds, such as β-barrels^133,142,143, immunoglobulin-like folds¹⁴³, α-β folds^23,133,143, water-soluble analogs of membrane proteins¹⁴³, and new-to-nature structures^{23,118,136,144}, including proteins with more than 300 amino acids²³. This is a clear breakthrough relative to the achievements of the physics-based approaches for which such challenging folds and large proteins were either extremely difficult or beyond the scope¹⁴⁵. While it remains unclear what are the precise methodological aspects that have led to these achievements, one possibility is that deep learning strategies optimize the probability of a sequence to fold into the desired conformation relative to alternatives, thereby implicitly capturing some negative design principles¹⁴⁶.

Two popular generative models are currently diffusion-based models and protein language models. Diffusion-based models can rapidly sample many backone conformations by gradually denoising a randomly generated constellation of backbone atoms^23,147,148. By contrast, protein language models capture evolutionary patterns in sequences that can be used to generate novel sequences. Even though these are trained on natural sequences, they may generalize beyond their training set to generate de novo proteins¹³³. Notably, methods development is progressing quickly, and many new learning paradigms are being introduced¹⁴⁹.

The automation of de novo design workflows enables statistical comparison of computed and natural proteins to understand the current limitations of design methodology (Figure 5 and Supplementary Box 1). Clearly, de novo-designed proteins tend to be smaller than natural ones, and their relative contact order, which quantifies fold complexity, is also smaller (Fig. 5a). Particularly, de novo designs are dominated by secondary structures and exhibit a much higher fraction of α helices compared to natural proteins (Figure 5b), as also observed in previous analyses^17,18. The dominance of α helices can be explained by the high stability and foldability of these structural elements due to their highly predictable pattern of stabilizing hydrogen bonds and reduced degrees of freedom; these features facilitate accurate design in the absence of an explicit way to encode negative design principles. β-strands are often more difficult to generate due to the requirement to design non-canonical structural features that guard against the tendency of β strands to aggregate^116,120. Additionally, designed folds exhibit lower loop content than natural proteins, despite the fact that loop regions often form essential parts of active sites¹¹⁰. Despite these critical gaps between de novo designed and natural proteins, we observe a clear trend towards greater fold complexity in designs over the years, in particular with the advent of deep learning-based design strategies around 2020 (Fig. 5c). This is an encouraging sign that perhaps some of the gaps in fold complexity between designed and natural folds (which are still dramatic) will be overcome with more sophisticated design methods.

De novo design of function

The ultimate goal of protein design is developing general methods that can be reliably applied to generate desired functions without recourse to natural starting points, thus fully addressing the “inverse function” problem (Fig. 1A). While this grand goal is far from being achieved, one important step toward it is embedding functional motifs from natural proteins into computationally designed ones. Furthermore, the de novo-designed parts of the protein can be sculpted to accommodate the functional site in an optimal way. This task has been achieved by both physics-based and deep learning-based approaches, resulting in proteins that present viral epitopes for vaccines, metal-binding proteins, enzymes, and protein-binding sites^141,151,152.

Beyond incorporating a known functional motif into a designed protein, de novo design also aims to design functional proteins completely from scratch. The main focus in recent years has been on protein or small-molecule binders. In the case of the latter, success came from physics-based methods where ligands are modeled simultaneously with interacting side chains and then used to search for protein structures that can implement such arrangements, resulting in high-affinity binders¹⁵³. The challenge for protein binders, on the other hand, is that protein-protein interactions must strike a delicate balance between opposing biophysical contributions, such as favorable electrostatic and hydrogen-bond interactions and the unfavorable desolvation penalty of shielding polar groups from solvent. Furthermore, the protein interaction surfaces must be carefully optimized to exhibit high shape and electrostatic complementarity and minimal strain¹⁵⁴. Nonetheless, physics-based methods have been used to design protein binders for different targets, such as cell-surface receptors and viral proteins^26,38,155. Recently, deep learning-based methods have also made important contributions to protein binder design. The MaSIF framework trained a geometric deep learning framework on protein surfaces for prediction and design tasks¹⁵⁶. This approach enabled accurate prediction of the most likely protein interaction sites on a protein surface and the design of protein binders¹⁵⁰. Additionally, diffusion-based generative approaches showed remarkable success in the design of de novo binders, including several cases of high-affinity binders²³.

Although the number of successfully designed binders is small, in the following we compare several structural features between designs and natural binders. Similar to our analysis of protein folds, here too, we find that designed binding surfaces rely heavily on regular secondary structures, particularly α helices (Figure 5d). Furthermore, these surfaces tend to be smaller (Figure 5e) and rely on fewer hydrogen bonds than natural ones, highlighting the difficulties in designing precise hydrogen-bond interactions across large and complementary protein surfaces (Figure 5f). We note that the distribution of natural proteins we analyzed excluded antibodies. The dominance of loop regions in antibody binding sites would greatly accentuate the differences we observe here.

We conclude that despite the significant progress made in de novo binder design in recent years, there is a dramatic gap in the types of molecular interactions that can be generated compared to those observed in nature that has not been closed over the past decade¹¹. Closing this gap is likely to demand a dedicated focus on de novo design of nonideal structure elements, such as those seen in antibodies and enzyme active sites. Designing such elements, however, may come at the price of lowering the stability and foldability observed in de novo proteins and may require explicitly encoding negative design principles to ensure accuracy. We expect that some of the advantages of high stability observed in de novo designs relative to natural proteins may have to be sacrificed to design more complex structures that encode sophisticated functions.

Emerging applications for de novo designed proteins

Recent years have seen an explosion of novel applications of de novo designed proteins. In this section we present a few that highlight innovations in the design approaches and these are summerized in Figure 6. One of the areas where de novo designs may have clear advantages over natural proteins is in protein-based therapeutics, particularly antivirals. In this area, the small size, high stability, and relatively economical and scalable production of designs compared to antibodies are all appealing and have inspired several successful design campaigns against influenza^23,38 and SARS-CoV-2^155,157. Designs were generated to act through several inhibitory mechanisms including conformational inhibition³⁸ of target viral proteins and directly blocking the interactions between the virus and the human cell surface receptor^155,157,158, demonstrating the versatility of design approaches.

Figure 6. Several practical applications of de novo protein design.

De novo designed proteins present a range of possibilities for the development of new molecules with biotechnological applications. Some popular areas of research are antivirals (PDB: 3R2X), cancer therapies (PDB: 7JH5), protein based therapies (PDB: 2B5I), protein switches (PDB: 6IWB), drug delivery vehicles (PDB: 6VFI), vaccines (PDB: 3LHP) and biosensors (PDB: 7AYE). Natural proteins (targets) are colored green and the designed proteins are colored red.

Another notable example of computationally designed protein-based drugs are immunomodulatory interleukins¹⁵⁹. Designs for this task were generated by grafting a natural interleukin binding site followed by de novo design of the supporting scaffold. The resulting design exhibited enhanced affinity and selectivity and increased biological activity compared to natural interleukins in vivo. Immunomodulatory proteins, specifically cytokines, are an attractive target for design because the natural pleiotropy of interleukins across different receptors suggests many possibilities for fine-tuning their potency and specificity¹⁶⁰.

Structure-based vaccine immunogen design is a very promising area, including for de novo design approaches^151,161. Vaccine immunogen stabilization⁶², grafting of neutralization epitopes on stable de novo designed scaffolds^141,151,162, and even targeting specific antibody germline responses through design have been recently explored¹⁶³. Many of the designed immunogens were tested in non-human primates showing the induction of neutralizing antibodies, and germline targeting immunogens have since entered human clinical trials showing engagement of the desired antibodies¹⁶⁴. One of the most impressive recent advances has been the generation of vaccine presentation and delivery vehicles in the form of self-assembled protein nanoparticles. These designed protein-based nanoparticles have shown a breadth of applications related to different viruses (RSV, HIV, SARS-CoV-2, and influenza), consistently improving the efficacy of the immune responses and enabling more economical manufacturing options^165–168. Remarkably, a designed nanoparticle targeting SARS-CoV-2 has recently been approved for human use^168,169.

Biosensing is another area for which protein design is particularly well suited. In such designs the molecular recognition module needs to be designed in a way that binding of the analyte impacts the conformation of the sensor in measurable ways. Currently, such approaches are more readily applicable to detecting the presence of proteins than small molecules given that larger conformational changes are more readily triggered by protein-protein binding events. Nevertheless some progress has also been made in small molecule sensing where ligand receptors can be embedded in multidomain architectures and used for quantifying drug dosage¹⁷⁰, or an array of small protein sensors produced to mimic the multifactorial sensing of olfactory system¹⁷¹.

One of the most innovative areas in modern biomedicine is cell-based therapeutics as a strategy to treat non-solid tumors. Cell-based therapies are a class of living drugs that can leverage protein design and synthetic biology. The engineering of chimeric antigen receptor T-cells (CAR-Ts) has been a dynamic field where many concepts that could benefit from sophisticated protein design have been attempted. Specific examples arise from the need to control the activity of CAR-T cells to avoid toxicity¹⁷². For instance, small molecule switches have been used for cancer therapies through design of receptor activation in vivo, thereby adding a critical safety switch for clinical usage¹⁷³. Additionally, de novo designed transmembrane segments that were incorporated into receptors in CAR-T cells were shown to generate more specific signaling than the standard segments used in such constructs¹⁷⁴. This led to lower release of inflammatory cytokines and may lead to safer CAR-T cell treatments. One of the biggest problems of CAR-T cells is the lack of specific antigens on the tumor cells. Using computational design strategies, multiple antigen binders were combined into “logic gates” that increased the tissue specificity of the engineered cells¹⁷⁵.

To conclude, de novo protein design is increasingly finding uses in areas at the forefront of therapeutic engineering. The inherent features of de novo designs provide improved stability and manufacturability compared to conventional modalities.

Outlook

Protein design has made huge progress over the past decade. From concentrating efforts on understanding the fundamental principles that govern sequence-structure relationships¹²⁰, the field has matured to the point that thousands of unique protein structures can be designed using automated workflows^25,106,107. In addition, despite concerns that design methods were not accurate enough for optimizing natural protein activities^10,12,30, new methods that combine physics- and data-based calculations are making significant inroads into applied protein engineering. Recent algorithms can generate variants of natural proteins that encode dozens of simultaneous mutations to significantly increase stability and desirable activities^{42,43,108,109}. Such methods have dramatically accelerated the development of proteins for research and therapeutic or industrially relevant applications^62,67,79,102 and have been applied by scientists with no expertise in computational modeling. Encouragingly, the broad spectrum of proteins and functions that have been optimized using protein design suggests that one of the critical goals of protein engineering — i.e., a general strategy for protein optimization that can be universally applied — may be within reach.

Methods for designing protein function from scratch (the “inverse function problem”) are, however, limited in scope. Despite advances in de novo binder design, the resulting proteins still exhibit limited structural diversity, and targeting complex protein surfaces is still beyond reach. Furthermore, only rudimentary enzymatic activities can be designed de novo¹⁵². We argue that addressing these important challenges requires advancing beyond the idealized proteins that are typically generated today and approach the complexity of natural binders and enzymes, such as antibodies, TIM-barrels, and β propellers. Designing complex structures without compromising stability and foldability is, in our view, the most pressing challenge for de novo design. To overcome this challenge, we may need methods that encode negative design principles either through new data-driven approaches or, more realistically, by combining AI-based methods and physics-based ones. A solution to this outstanding challenge will open numerous opportunities for addressing problems in research, drug development, and sustainable chemical manufacturing, among many other urgent basic and applied topics.

Glossary.

Fold design

Designing a backbone that shares no significant sequence homology with natural proteins. Sometimes denoted de novo design.

Function design

Implementing a new function into a protein scaffold.

Structure-based design

Design based on computed or experimentally determined molecular structures using physical principles.

Protein backbone

The protein mainchain of amino acids connected through covalent amide linakages. Also denoted scaffold.

Protein optimization

Design with the goal of optimizing desired protein functional aspect(s), such as thermodynamic and kinetic stability, production yields, catalytic efficiency, binding affinity, and specificity.

Stability design

Design with the goal of improving protein thermodynamic and kinetic stability.

Sequence space

The theoretical space of possible combinations of protein sequence changes. This space is often too large for experimental or computational enumeration, and design methods must find ways to restrict and sample it efficiently.

Positive design

Designing elements that improve the stability of a desired structural state.

Negative design

Designing elements that destabilize undesired (e.g., nonfunctional or aggregation-prone) states.

Epistasis

Non-additive effects of combinatorial mutations; for instance, when mutations are toleated in combination, but not individually, or vice versa.

Backbone generation

Generating a spatial arrangement of the protein backbone excluding the amino acid sidechains.

Idealized topology

Simplified geometric representation of protein structure, mostly comprising secondary structure elements connected by short linkers.

Backbone designability

The ability of amino acid sequences to fold into the desired backbone. A backbone that has many solutions is highly designable.

Relative Contact Order (RCO)

Represents the relative complexity of a protein fold. Computed as the extent to which amino acids that are far in the primary sequence are physically close in the 3D structure.