De novo protein design – from new structures to programmable functions

Summary

Methods from artificial intelligence (AI) trained on large datasets of sequences and structures can now “write” proteins with new shapes and molecular functions de novo, without starting from proteins found in nature. In this perspective, I will discuss the state of the field of de novo protein design at the juncture of physics-based modeling approaches and AI. New protein folds and higher-order assemblies can be designed with considerable experimental success rates, and difficult problems requiring tunable control over protein conformations and precise shape complementarity for molecular recognition are coming into reach. Emerging approaches incorporate engineering principles – tunability, controllability, modularity – into the design process from the beginning. Exciting frontiers lie in deconstructing cellular functions with de novo proteins and, conversely, constructing synthetic cellular signaling from the ground up. As methods improve, many more challenges are unsolved.

eTOC/in brief

Advances in artificial intelligence are revolutionizing protein engineering and design. This Perspective discusses the concepts and approaches of de novo protein design, emerging challenges in designing structure and function, and the frontiers that lie ahead in deconstructing cellular processes with de novo proteins.

Introduction

Proteins can accelerate the speed of chemical reactions by many orders of magnitude, convert the energy of light into chemical energy, and regulate the myriads of processes within cells and organisms with the level of accuracy and precision required to sustain life. Because of these powerful functions, natural proteins have long been an attractive target for molecular engineering. The goals of protein engineering range from understanding the mechanisms of molecular and cellular functions to harnessing proteins for practical applications in catalysis, biotechnology, and as precision tools in discovery science and medicine.

The field of protein design is now fundamentally – and practically - rethinking this approach. Rather than reengineering existing proteins, it is becoming possible to build proteins with intricate architectures and functions – as powerful as those in nature, but new and user-programmable – from the ground up. This is the concept of de novo design¹, designing proteins from engineering principles or blueprints, without relying on existing starting points found in nature.

One can of course ask, why would one build everything new, if one can borrow, reuse, and reprogram from nature, or even arrive at functions new to nature despite starting from existing proteins²? Indeed, the approach of evolving or recombining existing protein components for new functions has been incredibly successful^2,3, and de novo design has long lagged behind because of its apparent limitations. Designed proteins, if less active than their natural counterparts, have required extensive screening campaigns to improve activity, and many desired functions seemed out of reach⁴. But if we could design functional proteins completely de novo, from the ground up, without the idiosyncratic features of evolved proteins, there may be several distinct advantages (Fig. 1A). The most obvious one is to enable functions not yet seen in nature (and for which there are no obvious existing starting points for directed evolution). The second advantage is that de novo design could allow us to create proteins that integrate engineering principles – tunability, controllability, modularity – into the design process from the beginning. We could engineer de novo proteins a priori to be (i) tunable, such that it is easy to generate versions with precisely altered biochemical parameters, (ii) controllable, such that protein function is responsive to internal and external stimuli, and (iii) modular, such that we can integrate different functions easily into composite molecular machines and assemblies.

Figure 1. De novo protein design in the age of AI.

(A) Designing proteins de novo (from scratch, without starting from a natural protein) can explore new structures and functions, and design proteins a priori with engineering principles in mind: Proteins could be designed to be tunable in their quantitative properties (rates, affinities, etc.), controllable by arbitrary inputs, and modular such that protein elements can be linked together for diverse input/output behaviors. (B) Advances in AI change the process of de novo protein design. User-defined goals (left) and inputs (middle) are used to generate proteins with new structures and functions (right). Categories 1–4 depict increasingly straightforward prompts leading to increasingly complex design outputs. Blue shading indicates design goals with experimentally validated examples. B-1: AI-based methods to design new protein structures can be unconstrained (generating diverse protein folds; alpha-helices shown in red and beta-strands in yellow) or constrained to diversify a particular fold. B-2: Most current methods to design function specify a “motif” with defined residue positions and orientations in a functional site. In a second step, a protein is generated de novo that surrounds and stabilizes the precise functional site geometry. This process is called “motif scaffolding”. B-3: Advances in AI-based methods are in development that only define the target, and the design method generates a predicted binder. B-4: Starting from a target function (for example converting substrate S to product P), an AI method could generate a protein with the requirements for that function. Currently, protein language models trained on specific protein families or large experimental datasets can generate new sequences with functions similar to those in the training set.

Artificial intelligence (AI) promises a considerable leap in enabling this vision for de novo design. Recent advances in the accuracy of protein structure prediction through deep learning^5–7 have profound influence on the inverse problem, protein design, and are changing how de novo design is conceptualized. Classical approaches to protein design first define a protein backbone structure at the atomic level and then find a sequence that is consistent with that structure⁸. Designing “function” adds a definition of the structure of an active site (typically the relative atomic positioning of key catalytic or binding residues) that is built into a designed protein “scaffold”. Much of the difficulty of designing function lies in the fact that the designed protein needs to adopt the desired functional site structure with extraordinary precision. Even deviations of less than 1 Angstrom in atomic positions can cause the design to fail (if we, for example, think of the precise geometric requirements of hydrogen bonds). Consequently, much of the method development – and the challenge – of de novo design focuses on generating proteins that precisely adopt the desired geometry (specific conformational dynamics and their timescales are other key challenges that I will discuss further below).

In contrast, generative approaches from deep learning offer the possibility, in principle, of designing structure, sequence, and function at the same time. The key conceptual leap seems clear, as structure, sequence, and function are intimately linked. A series of engineering problems of increasing difficulty illustrate the progression of design approaches that are currently being explored (Fig. 1B): (1) If we had a blueprint of the overall architecture of a protein (say, a barrel), could we experimentally realize instances of that architecture that are geometrically diverse (say, barrels of different sizes)? (2) If we had a blueprint of the positions of the most important atoms of a functional site in a protein, could we build a protein around this functional site, without needing to specify the protein fold or architecture that may be optimal for that function? (3) If we just had a function we wanted to design, could we ask a deep learning model to produce both a functional site and a protein sequence and structure model that harbors this site at the same time? (4) Or could we even simply ask the computer to design a protein that functions as desired? The answer to the first two questions is already yes in principle, approaches for the third are in development, and other applications – and more – are coming within reach.

The excitement about these advances in deep learning applied to de novo design does not mean that all problems are solved. Much the opposite, the rapid succession of new methods and their emerging successes in applications shift the focus from simpler design goals to many, often unsolved, larger problems; key and long-standing challenges of accuracy and precision, consideration of protein dynamics and conformational landscapes, and the scale of design problems are increasingly important. I will organize this review around these key challenges, advances, current state, and future opportunities. I will begin with concepts and approaches of de novo protein design, followed by chapters on (i) frontiers in design of new protein structures, (ii) new molecular functions, (iii) de novo proteins interfacing with cellular functions, and (iv) an outlook discussing long-standing and new problems. I will highlight developments in de novo design primarily in the last 5 years; there are many excellent reviews of earlier milestones (see Table 1 for a non-exclusive list of topic-focused reviews).

Table 1.

Recent computational protein design reviews with title, short summary, and reference.

Protein design concepts and progression of the field
Recent advances in de novo protein design: principles, methods, and applications.	State of protein design before broader adoption of AI-based methods: generation of backbone structures, sequence optimization, design energy functions, and design of molecular functions.	4
De novo Protein Design, a Retrospective	Evolution of the field of de novo protein design, with focus on physicochemical principles, functional helical bundles, membrane proteins, and protein assemblies.	40
A Brief History of de novo Protein Design: Minimal, Rational, and Computational	Progress in protein design illustrated through a timeline of de novo protein structures solved to atomic resolution.	41
Understanding a protein fold: The physics, chemistry, and biology of alpha-helical coiled coils	Progress in understanding and engineering alpha-helical coiled coils including design principles, biological functions, and applications of coiled coils in synthetic biology.	55
Machine / deep learning
Structure-Based Protein Design with Deep Learning	Outline of deep learning approaches to protein design and comparison to prior design methods.	112
Deep Generative Modeling for Protein Design	Comparison of 5 classes of generative models used for protein design.	19
From Sequence to Function through Structure: Deep Learning for Protein Design	Summary and comprehensive tables of recent deep learning methods for (i) fixed backbone sequence design, (ii) structure generation, (iii) sequence generation, and (iv) concomitant design of sequence and structure.	18
Protein-protein interactions
Computational Design of Novel ProteinProtein Interactions - An Overview on Methodological Approaches and Applications	Methods and successful cases of designing protein-protein interactions using (i) template-based approaches (utilizing known protein-protein interactions) and (ii) de novo design.	68
Applications to biological engineering
Computational Protein Design-the next Generation Tool to Expand Synthetic Biology Applications	Summary of computational designs shown to modulate activities in cells, including enzymes, protein specificity engineering, cellular pathway control, and higher-order protein assemblies.	86
Advances in the Computational Design of Small-Molecule-Controlled Protein-Based Circuits for Synthetic Biology	Computational approaches to designing protein-based sensors for small-molecule inputs coupled to functional outputs in cells.	92
Designed protease-based signaling networks	Summary of approaches that have engineered protease-based synthetic circuits for cellular regulation.	113
Design of protein switches
Design Principles of Protein Switches	Applications of switch design inspired by naturally occurring protein switches, and challenges with designing them de novo.	114
Membrane proteins
Principles and Methods in Computational Membrane Protein Design	Overview of innovations in the generation of new membrane protein structures and functions.	115
Computational Design of Transmembrane Proteins	Principles for transmembrane protein design and successful examples.	116
Enzymes
The Road to Fully Programmable Protein Catalysis	Key developments and opportunities in the challenging field of enzyme design.	11

Concepts and approaches of de novo protein design

For several decades, approaches to computational de novo protein design used physics-based approaches and atomistic representations, grounded in structural biology principles and rules derived from naturally occurring protein structures. Now, advances in artificial intelligence are leading to rapid changes in methods. Still, many key concepts of de novo design and important challenges apply to both physics- and AI-based strategies.

Computational protein design as an optimization problem

Computational protein design is most fundamentally formulated as an optimization problem (Fig. 2A): Given a desired structure (and function), design methods seek to predict an optimal sequence that stably adopts that structure (and has that function). De novo design, which I focus on here, does not start from naturally occurring, evolved proteins but aims to expand the space of protein structures, sequences, and functions beyond those seen in nature.

Figure 2. Protein design concepts and approaches.

(A) De novo protein design is formulated as an optimization problem: Given a design objective (a protein with a desired shape and function), find one or more amino acid sequences that have that structure and function. Most design methods divide the process into two steps: First, a structure containing only the polypeptide backbone is generated, and then a sequence is designed for that backbone. For each step, design methods that use atomistic modeling (blue) or AI-based approaches (orange) are indicated. (B) Classical design methods use a “blueprint” defining a protein fold topology (identity and order of secondary structure elements) and then assembles a 3-dimensional backbone from ideal helix, strand, and loop peptide fragments. (C) Backbone generation methods can systematically sample geometries (positions, orientations and sizes of secondary structure elements with varied connecting loops) within a given fold. These methods generate synthetic fold families that, just like evolved protein families, can be optimized for diverse functions. (D) A recent AI-based method, protein diffusion, generates protein backbones through a denoising process from random backbone starting coordinates. This method generates diverse protein folds without having to pre-specify a topology as input.

A key challenge is that the space of potential new sequences and structures is vast, sparsely populated with folded and functional proteins, and poorly mapped. For example, for a small protein of 100 residues there are 20¹⁰⁰ = ~10¹³⁰ sequence possibilities when considering the 20 naturally occurring amino acid types. Since the number of possibilities is larger than the estimated number of atoms in the universe (~10⁸⁰), trying (termed sampling) all these sequences and their possible structures is impossible. Instead, efficient search algorithms are needed to navigate the enormous space of possibilities. At the same time, there are in principle vast numbers of de novo proteins with new sequences, structures, and functions that could be found.

Because functional proteins are rare among all possibilities, we also need rapid methods to distinguish between successful and unsuccessful sequences using computed “scores”. Most design methods have used either empirical or physics-based scoring or energy functions⁹ that aim to estimate protein stability typically by considering atomic packing interactions, hydrogen bonding and electrostatic interactions, and solvation terms. The key challenge is to balance accuracy with speed, and this compromise necessitates approximations. Several sophisticated and well-tested atomistic simulation methods exist that use molecular dynamics with physics-based energy functions or even quantum mechanical calculations. However, each design candidate needs to be evaluated much faster than typically possible with these methods or else the approach is unlikely to find any viable solutions, even computationally. Unfortunately, a step-wise approach, first using approximate scoring functions followed by more accurate refinement, has proven difficult because fast, highly approximate scoring function tend to poorly correlate with the true free energy of proteins. In contrast, statistical approaches that learn from evolutionary sequence patterns¹⁰ and more recent machine learning approaches (discussed below) that take as input even larger amount of data from sequence repositories instead of physics-based scores are revolutionizing the task of finding experimentally viable sequences.

Still the most fundamental and generally unsolved problem is the design of function. As computational design is an optimization problem, we need a quantifiable definition of “function” to optimize towards. Herein lie several challenges. Most fundamentally, such as for an enzyme, we may not have a sufficiently precise description of the requirements for function – such as defined conformational dynamics or electrostatics in an active site – even if we could design these properties accurately (see a recent perspective on challenges in enzyme design¹¹). There are often multiple requirements for function – such as protein stability, the ability to adopt several conformations in a catalytic cycle, their rates of interconversion, specific recognition of desired interaction partners and avoidance of others, and more. Moreover, functional requirements can involve trade-offs, such as activity at the cost of stability, and computational approaches for multi-objective optimization are needed to balance these competing objectives. Finally, our ability to engineer many of these requirements with sufficient accuracy and precision is still limited, a challenge that I will come back to in the chapter on de novo design of molecular functions further below. Dependent on the design goal and the availability of a suitable starting point (a naturally occurring protein) with an activity related to the target function, directed evolution may be the method of choice because the complex optimization criteria are implicitly encoded in an experimental screen for function in the desired context; even novel functions can be reached². On the other hand, the mechanism by which the resulting functional proteins operate may not always be clear, and these proteins could therefore ultimately be less tunable and engineerable if the effects of mutational changes cannot be predicted.

Sequence optimization with atomistic modeling: fixed and flexible backbone design

To make protein design tractable given the challenges of sampling and scoring described above, most design approaches make a key conceptual simplification¹²: They divide the design problem into two steps: the first step generates a protein structure backbone (without a defined sequence), and the second step optimizes a sequence given that backbone (Fig. 2A). The second problem, termed fixed backbone design, was tackled first.

A milestone in fixed backbone design was reached in 1997 with the first complete computational redesign of a backbone structure existing in nature, a 28-residue zinc finger protein⁸. The design used discrete sampling of amino acid side chains with different conformations and residue types, a physics-based scoring function, and a deterministic optimization algorithm that found the global minimum energy sequence. A next milestone was the first computational design of a protein fold not found in nature, Top7. The design process using the modeling program Rosetta to first generate a new protein backbone (I will explain how in a section on structure generation using atomistic modeling further below), followed by iterative cycles of (i) sequence design given a fixed backbone and (ii) backbone minimization given a fixed sequence¹³. The Top7 example illustrates a key concept: Protein backbones are not fixed but they change, albeit often only slightly, when we make sequence changes in design or when proteins perform their functions. Many approaches have been developed to take this backbone flexibility into account in the design process, either by (i) backbone minimization interleaved with fixed backbone design as in the Top7 example¹³, by (ii) sampling small backbone adjustments during design^14,15, or (iii) by pre-generating backbone ensembles onto which sequences are designed and scored^16,17.

Sequence optimization with AI: learning the language of proteins

Increasingly deep learning methods are applied to protein sequence design. AI-based protein structure prediction methods have learned from the vast amount of information in the database of protein structures (PDB) and sequence information for those proteins. Applying similar concepts, protein sequence design methods can learn from the vast amount of information in sequence databases, including those for which there is no structural information.

There are now many different machine learning models that have been developed for protein sequence design and structure generation (for recent reviews see^18,19). Typically, AI methods for sequence design are evaluated by the extent to which the sequences predicted by the model resemble known sequences. A common metric is native sequence recovery, the fraction of predicted amino acid types at each position that are identical to those found in a native (naturally occurring) reference sequence. I will primarily focus the discussion here on AI models that have been experimentally validated. Experimental validation is essential to determine the true success of design methods because only one (or a few) incorrectly predicted amino acids in the core of a designed protein will result in catastrophic experimental failure but only a small decrease in native sequence recovery.

One class of machine learning models that has been successfully applied to protein design are large language models (examples include ProtGPT2²⁰, ESM-2⁷ and ProGen²¹). These models are trained on predicting missing amino acid “letters” in a protein sequence (analogous to language models trained on predicting missing words in a sentence). Once trained, protein language models can generate new protein sequences (just as ChatGPT is trained on text and can generate new text). ESM-2⁷ is a language model trained solely on sequences (not structures) that has been applied to designing new proteins that are stable and monomeric when experimentally tested²². Notably, these proteins are predicted to have diverse structures including ones dissimilar to naturally occurring proteins (albeit there are no experimentally determined structures of these designs to date). These results indicate that the model may have learned an underlying grammar of proteins that generalizes beyond the training examples. ProGen²¹ was similarly trained solely on protein sequences, but in this case from >19,000 protein families including labels of functional properties. For experimental evaluation, ProGen was fine-tuned on enzyme families (or a curated enzyme dataset) to generate designed variants with catalytic parameters similar to the natural proteins, including several with low (down to ~31%) sequence identity to any protein in the training set. Like ESM-2, ProGen does not require a protein structure for design but does require large datasets of sequences for a given protein family. Analogously, a previous machine learning model, UniRep²³, was shown to predict functional properties of proteins to enable variant engineering when fine-tuned on appropriate datasets. A different study showed that language models can be adapted for design of diverse functional sequences without the need for sequence alignments²⁴. This method successfully generated diverse, well-expressing nanobodies for which alignments are difficult because of high diversity in loop lengths and sequences. Language models were also successfully applied to model-guided affinity maturation of antibodies²⁵.

Other models for sequence design take both sequence and three-dimensional structure as input. Given a fixed protein backbone, these models predict amino acid identities using the local structural environment as context²⁶ (sometimes represented as a graph^27,28). ProteinMPNN²⁸ builds on a prior model for graph-based protein design²⁷ and has been extensively validated experimentally on designing proteins with existing and novel folds and large symmetrical protein assemblies. In addition, the model has been fine-tuned to predict the effect of single amino acid point mutations (ThermoMPNN²⁹) using large datasets of stability measurements³⁰. Frame2seq³¹ is a recent model that, in contrast to ProteinMPNN, predicts sequences in a single pass with increased or comparable accuracy but improved speed and a score that reflects prediction accuracy. One important question is to what extent deep learning models generalize, i.e. make predictions outside of the datasets they are trained on. Here, experimental validation suggests that Frame2seq may be able to design stable proteins with undetectable similarity to the starting protein, allowing exploration of novel sequence space. Overall, the high success rates of AI-based sequence design methods in experimental validation (often >10%, in favorable cases >50%) vastly increases the number and types of applications addressable with computational design.

Structure generation using atomistic modeling: design of all major fold classes and symmetrical assemblies

Experimentally validated, state-of-the-art models for de novo protein sequence design, such as ProteinMPNN²⁸ and Frame2seq³¹, require a protein backbone as input. This requirement poses two problems. First, one needs to have a method to generate new protein backbone conformations (Fig. 2). Second, one needs to assess whether these backbones are “designable”, meaning that there exists at least one sequence that stably folds into that structure.

The most obvious way to fulfill the designability criterion is to start with a protein backbone conformation existing in nature and repurposing it for a new function. Indeed, this approach has been successful in many cases. For example, computational design approaches have been developed to redesign enzymes for altered substrate specificity³² and protein-protein interfaces for orthogonal signaling³³. However, seemingly straightforward changes in specificity can be surprisingly difficult to design with computational approaches. A primary reason is the limitation given by the starting backbone conformation. For example, simply replacing a hydrophobic with a polar side chain to interact with a more polar substrate may not place the polar functional group in precisely the correct geometry for optimal hydrogen bonding with the new substrate, and even small deviations in geometry can have detrimental effects on function. For these engineering problems with a close starting point, directed evolution strategies are more suitable.

For generating protein backbone conformations de novo, the problem of designability can be solved in very elegant ways for all-helical structures. Here, breakthroughs were made when applying a set of parametric equations describing the geometry and relative orientation of interacting helices (Crick’s parameterization), which make it straightforward to generate large sets of designable helical coiled-coils. Extensive design and experimental validation studies led to a systematic description of a “periodic table” of coiled-coil architectures³⁴. Crick’s parameterization can be extended to arbitrary helical bundle architectures³⁵ that, when designed and tested in the laboratory, can be extremely thermostable³⁶. Moreover, helical architectures can be spliced together³⁷: The regular geometry of helices allows the alignment of helices in different proteins, leading to a facile method to generate a range of structurally distinct proteins³⁷ and larger helical architectures through fusion of overlapping helical regions. Helical repeat proteins with different curvature³⁸ then allow design of large assemblies with an impressive systematic variation in geometries³⁹. The diversity of designable all-helical structures still underlies many of the successful applications of de novo designs^4,40,41. However, while the problem of designing alpha helical proteins is largely solved due to our understanding of the design rules, more complex functions may require more structurally diverse structures with deviations from canonical helical geometries.

Much progress has also been made with the de novo design of protein folds containing a mixture of alpha helices and beta strands. A typical design process follows a four-step approach: The first step defines a “blueprint” of the desired protein fold topology, defined as the identity and connectivity of alpha-helical and beta-strand secondary structure elements (Fig. 2B). Blueprints allow for the definition of new fold topologies not found in nature¹³. The second step is to assemble a protein backbone from peptide fragments (helices and strands) according to the blueprint and connected by short loops (Fig. 2B). Peptide and loop fragments are typically taken from overrepresented fragments in the PDB, thus ensuring designability at least at the level of local (one-dimensional) sequence-structure compatibility¹³. Designability at the fold level can be assessed by rules found in existing protein topologies, such as organization of secondary structure elements into tertiary motifs⁴². An impressive example was the de novo design of symmetrical TIM-barrel proteins⁴³, a long-standing challenge in design that required specific side chain-backbone hydrogen bonds for defining the strand register between the barrel repeat units to succeed. The third step involves sequence design, often iterated with backbone minimization, as described for Top7¹³ above. This step generates sequences predicted to be optimal for the desired input structure. A final step assesses designed sequences in silico by predicting their structures and comparing the prediction to the intended backbone. Designs passing this test are experimentally validated. These approaches led to the design of diverse alpha-beta protein folds³⁸, and were generalized in methods such as TopoBuilder⁴⁴.

The design of structures with exclusively beta-sheet secondary structures (all-beta proteins) poses distinct challenges. For example, all-beta proteins show a tendency to aggregate. Moreover, attempts to derive parametric design methods, such as for helical bundles, have not been successful. Instead, breakthroughs were made through the realization that beta-barrel structures in nature have defined defects that allow relief of strain that would be present in idealized barrels⁴⁵. This principle allowed the design of a range of beta-barrel geometries and a functional fluorescence-activating beta-barrel⁴⁵. Other design efforts have generated beta-sandwich folds^46,47.

In addition to generating new tertiary structures with different folds, computational design has also been applied to generate quaternary structures. Particularly exciting are the designs of a large variety of symmetrical assemblies with impressive sizes, with important applications as delivery vehicles, reaction compartments, or nanoparticles for vaccines⁴⁸. Designing these assemblies typically involves docking of the component (natural or de novo) monomers in the desired symmetry, and redesign of the resulting interfaces. The design of these architectures is aided by symmetry: any designed interface interaction (if net favorable) will be repeated many times in the assembly, adding up to overall stabilization.

All the structure generation methods discussed above require a desired target structure or blueprint that needs to be prespecified at the start of the design process. The AI-based structure generation methods described in the next section do not have that requirement, opening up new avenues for the formulation of design problems.

Structure generation using AI models: natural and novel folds

The breakthroughs in AI-based methods for protein structure prediction, such as Alphafold2⁵, trRosetta⁴⁹ and RoseTTAFold⁶, have inspired numerous recent advances to invert these models for design: Instead of predicting a structure given a sequence, the task is to generate a structure from scratch and then predict its sequence (methods that generate sequences and structures at the same time are less explored at present). One of the key differences to the parametric or blueprint-based structure generation methods in the previous section is that AI-based methods do not necessarily require definition of the desired protein structure or fold class a priori.

Among the first AI-based approaches that were experimentally validated by de novo design is protein “hallucination”⁵⁰ that inverts the trRosetta structure prediction model for structure generation. Here, sequences are optimized to adopt predicted tertiary structure contact maps that resemble those of natural proteins but are different from those of random sequences. While hallucination generates both backbones and corresponding sequences, many hallucinated designs were not successful when tested experimentally. Considerably higher design success rates were reached when the hallucinated backbones were redesigned with ProteinMPNN in a second step²⁸. The necessity of this second step may reflect the insensitivity of current protein structure prediction methods to amino acid point mutations that can be catastrophic in protein design. Hallucination has been used to generate proteins and symmetrical assemblies with experimentally validated structures^50,51.

More recent AI-based protein design strategies use diffusion models^52–54 borrowed from image generation. Diffusion models start with images that are successively “noised”, followed by training a network on the noised samples to recover the original images. In the case of proteins, diffusion models start with protein structures and add successive noise to the protein coordinates, followed by training to recover the original structures. Using these models for design, one starts from random noise, and the denoising process generates samples of protein structures with properties of those resembling typical proteins (Fig. 2D). One such model, RFdiffusion⁵³, has been used to generate experimentally validated protein monomers, symmetrical assemblies and protein binders, and appears to outperform hallucination-based approaches. Another diffusion model, Chroma⁵⁴, has been used to generate experimentally validated protein monomers. A particularly exciting property of diffusion models is that they can be conditioned in various ways, such as generating particular fold topologies (Fig. 1B – 1) or preserving specified functional sites (Fig. 1B – 2), applications that will be discussed in the chapter on de novo design of molecular functions below.

Frontiers in design of new protein structures

As outlined above, proof-of-principle studies in de novo protein design have built diverse representatives of the major secondary structure architectures of proteins (all-alpha, mixed alpha-beta, and all-beta) as well as impressive higher-order symmetrical assemblies of them. Moreover, new protein structures can now be generated with considerable experimental success rates⁴ (often > 10%), with further increases through the development of recent AI models for both structure generation and sequence design. In this chapter I will focus on frontiers in design of protein structures. I will describe approaches to explore novel fold space, test mechanistic principles through reengineering them, and engineer user-defined shapes tunable for new protein functions. Together, these design strategies begin to build a framework for the de novo design of complex architectures and molecular machines.

Principles through bottom-up construction

While naturally occurring proteins occupy a limited number of protein topologies or folds, the early design success of Top7¹³ demonstrated that a stable new topology not seen in nature could be generated through computational methods. Generalizing this idea, a systematic exploration of alpha-helical coiled-coils led to the design of novel architectures and development of principles to exploit these architectures for diverse functions⁵⁵. Recent advances in AI-based computational protein design now allow in principle to map protein fold space systematically. New backbone generation methods such as GENESIS⁵⁶ are being developed to do so, and could be used to generate novel folds likened to cosmological “dark matter”. Ultimately, more systematic maps of protein fold space could (beyond generating starting materials for engineering) allow for better quantification of designability principles and thereby advance the speed and accuracy of design.

In naturally occurring proteins, functional mechanisms are often coupled in complex ways, reflecting aspects of the history and context in which functions evolved. In contrast, building new functions from the ground up might allow the dissection of principles that are difficult to entangle in evolved systems, such as principles of conformational switching, allosteric control, or mechanical stability. Designing these complex functions de novo is a difficult problem currently but could be reachable in the future.

Finally, a key frontier is the ability to dissect quantitative determinants not only of molecular, but also of cellular, tissue, and organismal functions. Here, de novo designed proteins could be engineered to have precise and systematic variation of molecular properties that in turn tune higher-order biological processes (I will come back to this aspect in a chapter on de novo proteins for cellular functions below).

Precise control over protein geometries: Synthetic fold families for function

Nature does not invent a new protein fold for every new protein function. Instead, existing protein folds are customized and optimized for new functions through changes in fine-grained geometries of functional sites and tuning of relevant protein dynamics. To design biological functions with biologically useful activity and required accuracy, computational design should therefore be able to exert precise control over fold shape as well as functional site geometry and dynamics. Considerable progress has been made with controlling overall course-grained variation of protein folds, as described above. In this section, I will highlight advances with developing methods that allow fine-grained control over the precise geometries of proteins to optimize details of atom-level interactions in functional sites. I define “geometry” as the variation or features including length and orientations of secondary structure elements within a given fold topology (the identity and connectivity of secondary structure elements).

The blueprint structure generation methods (Fig. 2B) described earlier typically generate idealized versions of the targeted fold topology, and although thousands of stable variants can be designed⁵⁷, they often are very similar to each other (1–2 Å root mean squared deviation, RMSD). Several approaches have been developed to instead systematically sample fine-grained geometrical features^58–60 such as pocket shapes⁵⁹. Since a large fraction of evolutionary variation involves diversity in positioning of helical elements, the LUCS sampling method⁵⁸ enables generation of synthetic fold families with tunable geometries through systematic variation of position, orientation, and lengths of helices (Fig. 2C). Several experimentally determined structures showed how the de novo designed proteins with identical fold topology can have large diversity in geometry, in each case in excellent atomistic agreement with the design model. The ability to in principle sample thousands of finely tunable geometries should allow progress with another frontier: the design of defined dynamics and conformational changes (discussed further below).

Complex shapes and blueprints for protein machines

The ability to generate larger protein structures through helical fusions and controllable oligomeric assemblies opens up new avenues to engineer more complex architectures with arbitrary shapes. These shapes could be, for example, the parts of molecular machines and motors (such as rotors and axels), which would need to break symmetry to undergo motion (rotation around the axel). A fascinating example of the design of diverse synthetic protein-based rotor and axel components and their assembly to prototype protein nanomachines⁶¹ was recently described. There are many open challenges such as driving rotation through energy conversion using chemical fuels.

Further advances in AI-based methods might allow design of complex protein shapes for nanoscale machines and biological patterns by first drawing a component blueprint and assembly plan, followed by custom-optimization of the required protein shapes. In addition, the design could consider the engineering principle of modularity during the design process of these larger assemblies so that they can be built up from plug-and-play pieces.

De novo design of molecular functions

The progress made with the accurate de novo design of new protein folds and diversified shapes and geometries, with success rates approaching >10% or even >50% dependent on the design goal⁴, contrasts with the ongoing challenge of designing new protein functions. Typically, computationally designed proteins provided a starting point with robustly measurable but low activity that would subsequently need to be optimized experimentally to achieve practically useful functions. With the advances of deep learning methods, this paradigm is beginning to change, at least for an initial range of functions. I will first highlight general principles of computational design of function, then outline how AI-based methods are changing the process, and finally describe state-of-the-art applications and frontiers.

Principles for designing function: motifs and scaffolds

Most generally, computational design of function (Fig. 3) involves two steps: The first step defines the requirements for function and the second step optimizes a protein structure and sequence that matches these requirements. With advances in deep learning applied to proteins, how these steps are carried out is changing rapidly, increasingly with notable success rates.

Figure 3. De novo design of molecular functions.

(A) General approach to design molecular functions. (B-C) Design of proteins binding to small molecules, using classical design methods (B) that place target binding sites into pre-generated protein scaffolds, or AI-based approaches (C) that generate new protein backbones around a binding site motif or target. (D-F) Design of proteins binding to target proteins (blue shapes). Regions that are optimized by sequence design are shown as dark red shape. (D) Rotamer interaction field approach⁶⁵. Specific interactions with a target protein surface are identified through docking of disembodied side chains, yielding an interaction field into which pre-existing scaffolds are docked and optimized. (E) Fingerprint approach⁶⁶. Interaction sites on the target are identified by predicting interaction fingerprints using the MaSIF deep-learning method, followed by identification of complementary fingerprints from a library of >400 Million patches. Matching patches are then scaffolded into de novo proteins and optimized. (F) Diffusion approach⁵³. AI-based protein diffusion is used to generate a binding protein with shape complementarity to a pre-specified hotspot on the target. A second step assigns a sequence to the diffused binder backbone.

Most computational approaches to date define the requirements for function as precise and pre-organized active site geometries (Fig. 3A, B). More specifically, these geometries are often defined as the relative position and orientation of functional groups of amino acid residues in a protein active site – for example the positioning of an arginine guanidinium group in suitable hydrogen-bonding geometry and distance to a carboxylate on a protein or small molecule binding partner. The key challenge then is to achieve this precise positioning for multiple interacting groups in a functional site stably designed into a protein scaffold.

Initial successful applications of this concept defined a few functional site geometries, also called motifs, either by rational design of active site interactions or by borrowing motifs from natural proteins, and then transplanted (matched) the motif into a different naturally occurring protein that was used as scaffold⁶². These approaches are principally limited in several ways: First, the precision with which any motif can be accommodated in a scaffold is intrinsically constrained by the available (natural protein) scaffold backbones. To optimize the motif precision, only small adjustments to the backbone were possible in earlier design processes. As a result, the desired geometry was never placed exactly in the desired geometry, often resulting in loss of function. Second, naturally occurring proteins are often only marginally stable. Placing a new functional site into them can therefore lead to unfolding. Third, more complex functional sites with more than 3 to 4 residues can frequently not be matched with reasonable precision to any natural scaffold.

The first and second problems can be addressed by using libraries of de novo designed proteins as scaffolds (Fig. 3B). Approaches where scaffolds can be finely tuned in their geometries, such as helical bundles through parameterization⁶³ or other folds through the structure diversification approaches^58–60 described above, are particularly successful. In addition, de novo designed proteins are often extremely stable, overcoming issues with placing functional sites into them.

The problem of not finding any suitable matches in a given library of pre-generated de novo scaffolds is more complex. To a certain extent, this problem can be overcome by increasing the numbers: generating tens of thousands of potential motifs through computational methods^64,65, and matching these into libraries of hundreds or thousands of scaffolds^65,66.

However, more general approaches that optimize (or even generate) the protein scaffold given a functional site definition, are necessary. Solutions to this “motif scaffolding problem” are in active development using various AI models for proteins. For example, given a motif geometry as input, both protein hallucination⁶⁷ and diffusion⁵³ can in principle generate a suitable scaffold around that motif for a range of scaffolding problems (Fig. 1B – 2; Fig. 3C). The key challenges here are in assessing (i) that the generated protein backbone is indeed designable and (ii) that the precarious details of non-covalent interactions are sufficiently accurate to stabilize the functional site in its desired geometry. Both criteria are currently assessed by predicting the structure of the generated design sequence using an orthogonal deep-learning method that was not used in backbone generation. While a useful computational consistency check, these methods can be insensitive to the effect of small details of interactions. Moreover, most of these methods currently do not explicitly model any non-protein ligands. Nevertheless, the reported success rates with these approaches in functional assays, as detailed for specific applications below, are impressive. Still, few functional designs generated by these methods to date have been validated by high-resolution experimental structures; further data are therefore needed to systematically assess the accuracy of designed functional site geometries.

Finally, AI-based methods should in principle also be applicable to the first step: definition of the requirements for function. An example is the MaSIF (molecular surface interaction fingerprinting) method that captures “chemical fingerprints” of suitable interaction interfaces on a protein target that can be computationally matched with complementary surfaces to generate de novo protein binders⁶⁶ (Fig. 3E). In a different approach, language models trained on protein families appear to encode requirements for function, because these models can be used successfully to generate designed variants with that function²¹.

Molecular recognition: protein-protein interactions

The de novo design of protein binders recognizing target protein partners⁶⁸ has led to exciting applications such as selective cytokine mimics⁶⁹, and protein inhibitors of a histone methyl transferase⁷⁰ and the SARS-Cov2 spike protein⁵³. The first approaches to computational binder design created new interfaces between existing proteins⁷¹ and altered the specificity of existing interfaces⁷². A key development was “hotspot-directed” design⁷³, later generalized using a “rotamer interaction field” approach^45,65: Here, disembodied amino acid side chains are docked against a target surface of interest to identify ideal interactions in a desired surface. In a second step, these docked side chains are incorporated in a scaffold protein to generate a binder (Fig. 3D), first using natural scaffolds and later de novo designed proteins. An impressive larger-scale design study assessed the success rates of this approach⁶⁵. For a panel of 12 targets with different shapes, the computational approach could generate binders for all targets, with success rates for identifying binders in the micromolar range between <0.01 and 1% (using libraries of 15,000 to 100,000 design candidates per target). To achieve nanomolar to picomolar binding affinities, the binders (all hyperstable mini-proteins smaller than 65 amino acid residues) were subsequently optimized using mutational screening.

Recent AI-based methods to protein binder design constitute a step advance (Fig. 3E, F), leading to higher success rates without reliance on large libraries or extensive experimental optimization. For example, RFdiffusion was shown to generate binders in the micromolar range for 5 targets with a 19% estimated success rate, testing fewer than 100 designs per target⁵³. For two targets, low nanomolar binders were identified with no further experimental optimization. Designs generated using the MaSIF method identified binders for 4 targets⁶⁶. For one target, the method yielded a low nanomolar binder without experimental optimization. While the RFdiffusion study above used predefined interaction hotspots on the targets, AI methods such as MaSIF could also be applied to identify good interaction surfaces on targets for which there are no known interaction hotspots. Another promising approach applies iterative design and structure prediction cycles to refine initial designs in a process akin to in silico directed evolution⁷⁴. The Sculptor⁷⁵ method uses deep learning to optimize the backbone conformation of a protein binder for a given target surface. This method addresses a long-standing challenge in computational design: to mimic the ability of antibodies to evolve high shape complementarity to many diverse targets by exploiting the conformational plasticity of loops. In addition, computational design methods such as Sculptor have the advantage over experimental antibody selection methods that the target surface can be specified a priori.

Despite significant advances in binder design, not all challenges are addressed. Key difficulties include the design of binders for target surfaces that are highly flexible or very polar. Nevertheless, progress is being made with explicitly considering flexibility in molecular recognition⁷⁶ and biased design for polar contacts in the interface⁶⁶. It will be interesting to analyze the growing number of successfully designed de novo binders for privileged interfaces or interaction modes. As yet, helical interaction surfaces on the designed binder are overrepresented (although not exclusively). Helices are more designable owing to their regular geometries, well-known design rules, and the intrinsic property that backbone donor and acceptor groups are internally satisfied; hence, the detrimental effect of unsatisfied buried hydrogen bonding donors and acceptors in helical interface is minimized.

Molecular recognition: protein-small molecule interactions

Small-molecule recognition is key to numerous protein functions including catalysis and signaling. Design of proteins binding to small molecules has remained a difficult problem, with few examples of engineering small molecule binding sites de novo into existing proteins^77,78, as well as de novo designed helical bundles⁶³ and a beta-barrel⁴⁵. In particular, highly polar or flexible small molecules are more challenging targets due to the difficulty of optimizing the precise geometries of polar contacts or the entropic penalties incurred when binding ligands with many rotatable bonds. Overall, the achieved affinities are typically in the micromolar or high nanomolar range before experimental optimization. Nevertheless, these approaches offer exciting opportunities for design of small-molecule-induced assemblies to control extra- and intracellular signaling processes.

Several deep-learning approaches have been proposed to scaffold motifs for interactions with small molecules. To date, many studies report in silico benchmarks. Experimental success (although no experimentally determined structures) has been reported for scaffolding metal binding sites⁵³. Very recently, an all-atom version of RFdiffusion, RFdiffusionAA,⁷⁹ has been applied to design proteins binding to the therapeutic digoxigenin, the enzymatic cofactor heme, and other targets. For digoxigenin, ~4,400 designs were experimentally screened to identify three designs that showed enrichment in a yeast display assay, with one design binding in the nanomolar range. While these are currently modest success rates, an exciting aspect of the method is that it could achieve high shape complementarity to small molecules by simply defining the target without having to pre-generate a binding motif. It will be interesting to compare the agreement between the AI-generated design models and experimentally determined structures for these emerging design methods.

Multi-objective optimization: Conformational changes and switches

The functions of evolved proteins are typically complex and composite, such as coupling binding to conformational changes, or posttranslational modifications to changes in activity. To ultimately match and surpass the advanced functions of natural proteins, de novo design approaches must be able to optimize over these different objectives. Such approaches are at early stages, with some notable advances.

One frontier area is to design tunable conformational switches by optimizing single sequences over multiple conformational states. Pioneering examples led to the design of a protein that switches between two different secondary structures⁸⁰ and proteins that have different designed conformations distinguished by alternative states of a tryptophan side chain moving on the millisecond time scale⁸¹. Most recently, switches have been designed that upon peptide binding interconvert between two different structured states related by an overall hinge-motion of two helical subdomains⁸². This latter application was enabled by the ability of the AI-based sequence design model ProteinMPNN²⁸ to optimize sequences while simultaneously considering two different structures. In the case of hinge-proteins, the problem is simplified since most intramolecular interactions stay the same except certain intra-domain interactions altered by the hinge.

For some naturally occurring protein switches, AlphaFold2 can recapitulate alternative states among the different generated model predictions⁸³. It is an open question to what extent AI-based structure prediction methods can predict and design multiple states de novo, without having been trained on natural examples of a given conformational switch.

Ideally, computational methods should be able to accurately predict the underlying distributions of conformations, and efforts to develop such methods are underway^84,85. It will be exciting to see applications of these concepts to the de novo design of conformational switches and other advanced functions that require explicit consideration of conformational changes or allosteric effects⁷⁶. The area of multi-objective designs of conformational switches is likely to see further advances in the design of more complex, composite protein functions de novo.

De novo proteins for cellular functions

Synthetic signaling systems that can control biological processes (chimeric antigen receptors are a prominent example) have many significant applications in fundamental biology, bioengineering, and medicine. The vast majority of such signaling systems built to date have used naturally occurring components (genetic elements and proteins) and recombined or reprogrammed them for new functions^3,86. The increasing success of de novo protein design now allows, in principle, to build protein signaling systems entirely from the ground up. Unlike natural proteins that are evolved to function in specific contexts, de novo proteins could be engineered a priori with context-independent function that allow tunability and modular behavior (Fig. 1). In addition, new functions not yet seen in nature may become accessible. De novo proteins could be engineered to sense new signals, integrate signals and perform logic, and precisely regulate downstream biological behaviors (Fig. 4). For each of these functions, computational methods could generate elementary components with tunable properties (such as binding on- and off-kinetics, diverse assembly geometries, etc.), and these components could be linked together in a modular fashion to generate diverse signaling behaviors. In this section I will describe progress with computational engineering of proteins for cellular functions, from reprogramming existing proteins to designing components de novo. I will also highlight how engineering principles of modularity and tunability are being incorporated into the design process, and how designs are interfaced with cellular processes to dissect principles of regulation.

Figure 4. De novo design to control cellular functions.

(A) Computational design of small-molecule sensors that couple auxin ligand binding to conformational change and fluorescence energy transfer (FRET)⁸⁷ (left) or metabolite-induced protein-protein dimerization to split reporter complementation⁷⁸ (right). (B) Different quantitative behaviors for CID systems. Top: “ratchet” mechanism, where ligand binding leads to a conformational change in one protein that creates a composite binding interface for the second protein. Bottom: “molecular glue” mechanism where the small molecule can bind either partner. This mechanism can lead to “bandpass filter” behavior where complex formation is low at high ligand concentrations because each of the two protein partners are bound by a different ligand molecule. (C) Mechanism of the de novo designed LOCKR system, where an output element is buried but can be displaced by a competing key element, leading to an output. (D) Application of the Co-LOCKR system to perform logic operations based on the composition of receptors present on the cell surface⁹⁷.

Design of sensors and actuators with diverse inputs and tunable outputs

The ability to sense and respond to molecular signals is a fundamental ability of all living systems, and engineering it de novo could advance many areas of science, technology, and medicine. Examples include metabolic engineering, by monitoring intermediates in production of industrially valuable chemicals; cell signaling, by creating tools to dissect normal and disease processes with improved precision; and cancer treatment, by achieving tight regulation of advanced therapies such as CAR-T-cells. An exciting example of a computationally designed sensor that functions at the organism level to track the distribution of the plant signaling molecule auxin in plant roots in real time was recently described⁸⁷ (Fig. 4A).

A key challenge in designing new sensor/actuator systems is to develop generalizable ways to couple detection (sensing) of a signal to a cellular output response (actuation). Unless the signal is intracellular or readily traverses a cell membrane, engineered sensor/actuator systems must transmit the signal from the outside of the cell to the inside. No entirely de novo engineered transmembrane signaling system exists yet. Nevertheless, progress has been made with reengineering existing transmembrane signaling systems to modulate allosteric signal transduction⁸⁸ and quaternary structure changes⁸⁹ in GPCRs.

Ideally, an engineered system should be specific to the signal but modular in its output response, such that a given input signal can be linked to a variety of output responses that could be changed without having to reengineer the entire system. One architecture that fulfils these criteria is chemically-induced heterodimerization (CID). Here, two components of a sensor preferentially heterodimerize in the presence of a small molecule, which can be linked in a modular fashion to complementation of a functional output reporter. Many suitable split reporters exist that activate, for example, fluorescence, enzyme activity, or most generally expression of any gene or gene combination. CID systems can be entirely intracellular, but can also provide a coupling mechanism across the membrane when sensing triggers the preferential assembly of transmembrane proteins with domains on either side of the membrane. Several CID systems have been rationally engineered based on selecting binders to drug-bound proteins as starting points^90–92. To date, one modular sense/response system has been built by de novo computational design of a small molecule recognition site⁷⁸ (Fig. 4A), albeit by engineering it into an existing protein-protein interface to create a CID system. The synthetic system showed dose-response behavior in cells to detect a metabolic intermediate produced via an engineered pathway. The output response could be exchanged in a modular fashion, and a crystal structure of the assembly showed good agreement with the computational design model.

Advances in computational design now pave the way to design CID systems with tunable binding behaviors entirely de novo. Moreover, the specific architecture of CID systems can determine different input/output behaviors (Fig. 4B). For example, CID systems can exhibit a “bandpass filter” response⁹³, where the signal is high only at intermediate signal concentrations but low otherwise. Other CID systems can show “molecular ratchet” responses that shift the response amplitude and sensitivity dependent on the concentrations of the CID components⁹⁴. Modeling the quantitative response of different CID architectures creates exciting opportunities to realize different input / output behaviors with engineered systems. Looking into the future beyond CIDs, one could imagine creating de novo sensors and actuators for diverse inputs such as peptides, pH, light, ionic strength, temperature, and mechanical force.

Regulation and logic

Another key property of all living systems is the ability to integrate signals and make decisions. Cellular decision making takes place in complex signaling networks, where not all interactions and their functions are known. Synthetic signaling systems offer the advantage of simplifying feedback and regulatory mechanisms such that they can be finely tuned and robustly controlled.

A pioneering study engineered de novo helical bundle proteins such that they could be embedded into positive and negative feedback system controlling both natural signaling (the yeast mating pathway) and synthetic gene circuits⁹⁵. The regulation mechanisms were based on a protein domain replacement strategy in the de novo designed LOCKR (latching orthogonal cage-key proteins) system⁹⁶ (Fig. 4C). Here, an output element located on a helix is buried inside a de novo helical bundle but can be displaced by a “key”, a helical input element, that competes with the locked helix. Feedback mechanisms could be engineered by designing a degronLOCKR, where the output element is a protein motif important in regulation of degradation (degron). The degron is exposed in the presence of the input signal (the key) and targets a fused cargo protein to the proteasome. The system was shown to be tunable, even in a combinatorial fashion, by modulating the key’s production (via an inducible promoter) or the key’s binding affinity to the degronLOCKR (by changing the length of the key).

A different study implemented colocalization-dependent regulation (Co-LOCKR) that perform AND, OR, and NOT Boolean logic operations in response to combinations of molecules present at the cell surface⁹⁷ (Fig. 4D). Other de novo proteins that have been used to implement Boolean logic in cells include sets of helical bundle heterodimers with engineered specificities mediated by hydrogen bonding networks linked to a split luciferase reporter or transcriptional regulators⁹⁸ and designed coiled-coil dimerization domains linked to split proteases⁹⁹.

Self-assembly and localization in cells

There has been long-standing interest in the signaling properties of cellular assemblies, ranging from higher-order oligomers to membrane-less compartments. Engineering such systems de novo could both contribute to deconstructing the function of natural systems as well as exploit specific characteristics such as signal amplification. Efforts to engineer de novo proteins that self-assemble in cells are beginning to emerge. For example, de novo helical proteins were designed to assemble into membraneless organelles whose assembly dynamics can be controlled. One assembly was shown to co-compartmentalize an enzyme pair to improve product formation¹⁰⁰. In a second example, pairs of de novo designed symmetric protein homo-oligomers each comprising 2–120 individual protein components were shown to assemble in mammalian cells into protein networks whose mechanical properties could be tuned intracellularly¹⁰¹. A third study designed de novo single-pass alpha-helical transmembrane domains that assemble into defined dimers, trimers, and tetramers¹⁰². These and similar designs could be used to probe how defined changes in valences and geometries of protein signaling assemblies affect biological responses.

Other approaches are beginning to engineer cellular delivery by designing de novo binders to transmembrane receptors triggering endocytosis¹⁰³. Another study developed a de novo designed system with dual function for both delivery and subcellular localization¹⁰⁴. Here, an arginine-rich peptide is designed to penetrate the cell and subsequently bind a complementary acidic partner that can be fused to various other proteins to control subcellular localization.

Interfacing with and deconstructing biological processes

Ultimately, to deconstruct and regulate complex biological processes, de novo engineered systems must have robust interfaces with complex biological machinery. One way to do so is to use de novo designed proteins as assembly parts for downstream biological processes. Here the de novo components could provide controllable inputs (such as the CID systems discussed above^78,92), tunable assembly kinetics, or defined geometries. Design efforts with these goals are beginning to emerge and provide new tools to probe necessary and sufficient parts of natural signaling. For example, extracellular 2-dimensional arrays of de novo designed proteins have been used as assembly parts¹⁰⁵ linked to intracellular proteins of interest via transmembrane helices. Inducible extracellular assembly promoted intracellular clustering, which was then used to trigger polarity of protein targets in mammalian cells and dissect regulatory events sufficient for cytoskeleton polarization. In another example, de novo designed proteins were used to change valences and geometries of synthetic cell surface receptor ligands. Here, de novo designed cyclic homo-oligomers with up to eight subunits were linked to a de novo designed FGF receptor binding protein and applied to probe and manipulate FGF signaling¹⁰⁶. Notably, defined oligomers are uniquely engineerable with designed proteins, in contrast to standard antibody reagents or natural binders.

Engineering principles

Increasingly, de novo design studies adopt strategies to engineer protein functions that can be readily expanded beyond a single example into families of de novo proteins that could be used as elementary components in engineering larger, compositive synthetic signaling systems. Consider for example, instead of building one sensor for a specific signal, building a family of sensors for that signal that have different input/output characteristics. Another example would be to engineer a set of signaling assemblies with the same architecture but controllable by different signals. A third example would be a set of de novo protein-protein interaction elements with different on- and off-kinetics or oligomeric assembly properties, which can be linked together in combinatorial and modular fashion. Ideally, all of these could be combined to construct signaling systems with desired “off-the-shelf” characteristics and not requiring extensive re-optimization in each specific context. Table 2 summarizes examples of emerging approaches to designing extendable systems to be tunable, controllable, and composable.

Table 2.

Protein systems engineered to be tunable, controllable, and composable, with publication title, short summary, and reference.

Families of components with tunable properties
Expanding the space of protein geometries by computational design of de novo fold families	De novo protein fold families with finely tunable shapes through systematic variation of helical elements.	58
An enumerative algorithm for de novo design of proteins with diverse pocket structures	Families of de novo NTF2 fold proteins with pockets tunable for ligand binding.	59
De novo design of protein homo-oligomers with modular hydrogen-bond networkmediated specificity	Alpha-helical homo-oligomers with diverse interaction specificity determined by central hydrogen-bond networks.	117
Programmable design of orthogonal protein heterodimers	Orthogonal 4-helix protein heterodimers of two helical hairpins, with interaction specificity determined by hydrogen-bond networks.	118
De novo design of bioactive protein switches	Orthogonal LOCKR systems that function in vitro, in yeast and in mammalian cells.	96
Reconfigurable asymmetric protein assemblies through implicit negative design	Families of beta sheet-mediated heterodimers with diverse on- and off-rates.	39
Controllability
Reprogramming an ATP-driven protein machine into a light-gated nanocage	Generalizable strategy to control reversible shape changes of a protein nanocage through light-triggered conformational switching of a covalently attached azobenzene linker.	119
Computational design of a modular protein sense-response system	Control of protein-protein assembly through de novo design of a small molecule binding site into a protein-protein interface.	78
A rational blueprint for the design of chemically-controlled protein switches	Computational protein design strategy to repurpose drug-inhibited protein-protein interactions as OFF- and ON-switches.	91
Rational Design of Chemically Controlled Antibodies and Protein Therapeutics	Design and application of small-molecule-controlled switchable protein therapeutics using an engineered OFF-switch system91 (above).	120
Designed protein logic to target cells with precise combinations of surface antigens	Application of the LOCKR systems96 (above) as colocalization- dependent protein switches (Co-LOCKR) that can perform Boolean logic operations on the cell surface.	97
Modularity
Computational design of a modular protein sense-response system	A de novo designed chemically-induced heterodimerization system78 (above) can be linked to diverse modular split response systems.	78
Reconfigurable asymmetric protein assemblies through implicit negative design	Tunable beta-sheet heterodimers39 (above) can be assembled into complexes with up to 6 different components.	39
Modular and Tunable Biological Feedback Control Using a de Novo Protein Switch	The LOCKR system96 (above) can be modularly recombined and rationally tuned to implement feedback control of endogenous signaling pathways and synthetic gene circuits.	95
De novo design of modular and tunable protein biosensors	The LOCKR system96 (above) can be adapted into modular biosensors for diverse proteins.	121

Challenges and next opportunities

Advances in AI are revolutionizing protein design, and new methods are emerging rapidly. Currently, successful experimental applications address relatively simple problems, such as design of idealized folds (still with an overrepresentation of all-helical proteins), symmetrical assemblies, and protein-protein interfaces – albeit most recently with examples of remarkable shape complementarity. The increasing success rates of these applications are bringing important, long-standing challenges, such as design of precise geometries of polar functional sites and dynamical proteins, into reach. Latest developments such as protein diffusion models that model not just the protein backbone but all atoms including side chains and ligands⁷⁹ can be used to generate proteins de novo around small molecule ligands, albeit still requiring screening of relatively large numbers of designs. Further-reaching design goals such as molecular machines are coming into reach, and more complex composite functions can be deconstructed into designable components and implemented⁶¹.

Deep learning and data

The step-change with AI-based protein structure prediction required vast datasets of protein structures and sequences. In principle, function is also encoded in these structures and sequences, and this encoding has been used by machine learning models to generate functional proteins^21,23–25. However, precise requirements for specific target activities and dynamics are more difficult to extract for desired properties where we lack informative datasets. Herein lie both significant challenges and opportunities for advances reachable by deep learning. Integration with approaches for quantitative measurements of functional parameters at scale seems to be one promising avenue. There are exciting opportunities for new capabilities to generate robust and accurate large-scale datasets that validate designs and probe their stability^30,57, as well as recent high-throughput methods for rapid determination of rate constants and affinities¹⁰⁷.

Multiple objectives and energy landscapes

Advanced protein functions will most likely involve integration of properties, such as the cycles of molecular recognition, resulting conformational changes, and exposure of new recognition sites exhibited by naturally occurring protein switches (such as regulatory GTPases). More generally, diverse inputs modulate protein functions through shifts in their free energy landscapes. Ideally, new methods should be capable of shaping specific properties of these landscapes – such as multiple defined minima and the barriers between them – during the design process. There are numerous challenges with such an approach that would explicitly consider these aspects of function, including methods and informative data at sufficient scales to train models as well as characterize functional designs. Progress will also require approaches that can simultaneously quantify and optimize these multiple objectives. Such multi-objective optimization could be contrasted to – or integrated with - designs that deconstruct coupled functions to make them modular and individually tunable, such as the sense-response systems discussed above that combine separate modules for sensing and responding.

Extracting principles

As designing functional proteins beyond simpler model systems becomes possible, extracting principles becomes important. In particular, principles are needed such that de novo protein systems are actually tunable, modular and controllable without extensive trial-and-error or individual optimization for new contexts (such as cell type). Ideally, designs would be the result of directed and interpretable optimization (not a black box) that can systematically vary desired properties.

Since its first applications, the field of protein design has promised fundamental insights into sequence – structure – function – dynamics relationships, “learning by building”. The growing power of engineering protein components de novo provides different opportunities to also probe the functional principles of proteins embedded in complex interconnected biology. At the same time, these directions will also accelerate the engineering of advanced cellular functions with de novo components, with ultimate applications to cell therapies.

Emergent properties and advanced cellular functions with de novo components

The interactions and modular combinations of naturally occurring proteins lead to emergent cellular behavior that is not displayed by the individual components alone. For example, systems of proteins undergoing reversible covalent modification (e.g., phosphorylation) with opposing regulators (e.g., kinases and phosphatases) can show ultrasensitive switching, meaning that a small change in the concentration or activity of a regulator can cause a sharp change in output (modified protein)¹⁰⁸. In nature, such switches are assembled into cascades for signal amplification¹⁰⁸. As a second example, interlinked positive and negative regulation can control cell “states”, meaning long-term, stable patterns of gene expression¹⁰⁹, that can be responsive to environmental signals. Already, the modularity of existing proteins can be used to reprogram advanced cellular functions¹¹⁰, and machine learning can guide modular engineering¹¹¹. The concept of composing protein systems from de novo designed elements should allow bottom-up design to make these advanced biological functions engineerable. This approach should allow both deconstruction and construction.

Conclusion

The field of computational de novo design is making a step change into a new beginning. Advances in AI applied to protein design now make many, albeit relatively simple, design goals easier and more successful. Versatile protein folds and even large protein assemblies – which already have exciting clinical applications as vaccines – can be engineered with high structural accuracy. It is increasingly possible to engineer de novo proteins that bind tightly to user-specified surfaces on target proteins. Applications of these de novo binding proteins range from probes for fundamental cell biology to therapeutic candidates. Long-standing goals of de novo design, such as proteins sensing new small molecule signals, design of advanced functions involving conformational changes and allostery, and engineering emergent behaviors such as ultrasensitive switching, still pose significant challenges but are coming within reach. Progress is also being made with interfacing designed systems with biology, for example to control the geometry, localization, and timing of cellular assembly processes.

Numerous exciting challenges lie ahead. Current frontiers include prediction of protein behavior beyond structure: quantitative parameters such as binding affinities, conformational dynamics, and ultimately cellular functions. Advances in deep learning will require informative data at sufficient scales to enable accurate design of these behaviors. Advanced protein functions are often composite, coupling input signals to diverse functional outputs; predictive design should hence be capable of integrating multiple objectives. Extracting principles from data is important to make desired protein properties indeed engineerable. New opportunities lie in building complex functions from the ground up. Here, de novo proteins could be designed a priori with engineering principles of tunability, controllability, and modularity. Families of such de novo components with tunable and controllable properties could be recombined to generate diverse behaviors. Interfacing these de novo systems with biological processes could enable both deconstructing cellular functions and controlling them. The rapidly evolving field of de novo protein design provides an exciting environment for the creativity of scientists and engineers to address the many more unsolved than “solved” challenges at the interfaces of biological and new-to-nature functions.