Overview

Main Conclusion

The authors' view of the field is clear: three long-standing challenges are now close to being solved, or have largely been solved. These are the design of new protein structures, the design of new protein assemblies and the design of protein binders. This does not mean every design succeeds or that experimental validation is no longer needed. It means the core methodologies have become stable enough, and success rates practical enough, for application-focused teams to begin using them.

Figure 1. Workflow of protein design.

Basic structure and folding design is now relatively mature: a new sequence can be designed to fold into a specified three-dimensional structure. Protein assemblies can be designed so that multiple subunits self-assemble into cages, two-dimensional arrays, fibers, crystals and other nanomaterials. Protein binders can be designed as small and stable proteins that recognize viral proteins, toxins, receptors or disease-associated surfaces. Enzymes, switches and molecular machines are advancing rapidly, but remain more difficult because they must not only look correct but also move, transmit energy and catalyze high-barrier reactions.

In short, the key question is shifting from whether proteins can be designed to what should be designed and for which real-world problem. That is a sign of a maturing field.

How Deep Learning Changed the Field

Early computational protein design relied heavily on physical and chemical models. Rosetta was a representative example: it sampled possible protein backbones, searched for amino-acid sequences compatible with those backbones and scored designs using energy functions that considered van der Waals forces, hydrogen bonds, electrostatics, torsion angles and solvation. This strategy was grounded in the Anfinsen hypothesis: in principle, the amino-acid sequence encodes the native folded structure.

This route was powerful. Top7, reported in 2003, was a landmark because it represented a protein fold absent from nature whose crystal structure closely matched the computational model. It showed that de novo protein structure design was not science fiction.

Figure 2. Progressive denoising in the diffusion model enables the generation of target-binding proteins.

In recent years, however, the field has changed more deeply. AlphaFold, RoseTTAFold and related models demonstrated that deep learning can infer how sequence maps to structure from large structural datasets. Protein design then began to shift from computing low-energy structures to generating high-probability sequence-structure combinations. RFdiffusion and ProteinMPNN are two key tools in this workflow: RFdiffusion imagines a plausible backbone under functional constraints, while ProteinMPNN assigns sequences likely to fold into that backbone. AlphaFold or RoseTTAFold can then be used to check whether the design is structurally plausible before synthesis and experimental validation.

The important point is that AI has not removed experiments. Instead, progress depends on a design-build-test loop. Models propose candidates, experiments reveal which ones are real, and failures expose model blind spots. What has changed is that experiments increasingly test directed computational candidates rather than blindly screening astronomical random libraries.

Mature Task 1: New Folds and New Structures

The most fundamental protein design problem is inverse folding: given a desired three-dimensional shape, can a sequence be found that reliably folds into it? The review argues that this problem is now largely mature. Over more than two decades, researchers have designed TIM barrels, helical repeat proteins, beta barrels and many folds not found in nature. Deep learning has expanded the sampling space and made many formerly expert-driven steps more automated.

The design space now extends beyond ordinary soluble proteins. Membrane proteins can also be designed de novo, including alpha-helical transmembrane structures, beta-barrel pores and channels. This matters because membrane proteins are central to drug targets, ion channels, molecular sensors and nanopore sequencing. If pore size, interior chemistry and selectivity can be designed, new technologies for protein sequencing and general molecular sensing may become possible.

Protein design has also moved toward non-natural amino acids, D-amino acids, cyclic peptides and oligomers closer to small-molecule drug space. The boundary between protein design and medicinal chemistry is becoming less rigid.

Mature Task 2: Protein Nanomaterials

Designing a single folded protein is only the first step. A more advanced goal is to make multiple proteins self-assemble into defined nanocages, fibers, two-dimensional lattices or three-dimensional crystals. The intuitive analogy is building blocks: proteins become programmable modules whose interfaces drive them to assemble into target structures.

Figure 3. Overview of protein design challenges and design of protein folds and assemblies.

Earlier approaches used symmetry equations, rigid fusion and symmetric docking. With deep learning, RFdiffusion, hallucination and reinforcement-learning strategies can generate backbones under target symmetry constraints more flexibly. ProteinMPNN then designs sequences, and structure-prediction tools evaluate whether the assembly is likely to form.

One symbolic example is the SKYCovione COVID-19 vaccine, described in the review as the first clinically approved human medicine based on de novo design. It uses an engineered icosahedral nanoparticle to display antigen in a highly ordered manner. This shows that de novo designed proteins are no longer just elegant structures; they have entered real medical products.

Protein nanomaterials may support vaccines, drug delivery, structural biology scaffolds, pH-responsive particles, receptor-clustering arrays and reconfigurable two-dimensional systems that convert into cages in response to paired binders. Because proteins can be genetically encoded and expressed in cells, protein nanotechnology has advantages that complement DNA nanotechnology.

Mature Task 3: Protein Binders

Protein binders are perhaps the part of this review closest to an industrial inflection point. A binding protein is a small and stable protein designed to attach to a specific site on a target protein. It may block viral entry, neutralize toxins, activate or inhibit receptors, or recruit disease-associated proteins for degradation.

Antibodies can perform similar roles, but de novo binders can be smaller, more stable, more modular, easier to multimerize and sometimes better suited for local delivery. The review notes that combinations of Rosetta, RFdiffusion and ProteinMPNN have produced more than 200 experimentally validated protein-target binders, and methods such as BindCraft are rapidly expanding this capability.

The most important conclusion is that the central problem for binders is moving from how to design them to which target and function to pursue. The next bottlenecks may be disease biology, target selection, delivery route, clinical development and regulatory strategy. Highly polar target surfaces remain challenging, affinity and success rates can still improve, and translation from animal models to human patients remains a much larger hurdle.

Small-Molecule Binders and Sensors

Beyond protein targets, researchers also want proteins that recognize small molecules such as drugs, metabolites, toxins or in vivo reporter molecules. The review discusses binders for digoxigenin, methotrexate, apixaban, cholic acid, bilin and cortisol, as well as a plant sensor for fentanyl.

The value of this design class is not only binding a small molecule. The binding event can be converted into a readable signal or a controllable action. A protein that dimerizes only in the presence of a small molecule can become a chemically induced dimerization system. A binding module inserted into a nanopore can become a gated sensor. An approved drug can become a trigger for orthogonal control of cellular signaling.

This shows that de novo design is not just solving isolated target problems. It is building reusable molecular logic components that can be connected to sensors, cell signaling pathways, nanopores or synthetic-biology circuits.

Enzyme Design: Powerful but Still Difficult

If a binder is designed to attach to a target, an enzyme must make chemistry happen. This is much harder. A high-performance enzyme must place the substrate correctly, stabilize the transition state, position catalytic residues precisely, control water and proton transfer and often coordinate metal ions, heme or other cofactors.

Early computational enzyme design often began with an idealized active site: define the catalytic residues needed around a transition state, then search existing scaffolds for geometry-compatible positions. This produced designed retro-aldolases and Kemp elimination enzymes, but activities were often low and extensive directed evolution was required to approach natural enzyme performance.

New generative models change the workflow. Instead of placing an active site into an existing scaffold, models can generate a protein backbone around the idealized catalytic site. RFdiffusion2 can generate supporting structures under catalytic-residue, metal-ion or atom-level geometric constraints, while models such as PLACER help evaluate whether the active site is preorganized across reaction steps. Serine hydrolases, metallo-hydrolases, heme peroxidases and de novo luciferases have all advanced, but the review remains cautious: enzyme design has not yet reached the maturity of folding, assembly or binder design.

Next Frontiers: Switches, Logic and Molecular Machines

The most exciting future is not just static structure design, but protein systems that switch states, transmit signals, respond to environments and perform actions. Natural protein machines use chemical fuel to move, synthesize molecules and maintain cellular quality control. Artificial systems must combine binding, conformational change, catalysis and assembly to approach that complexity.

The review highlights protein-nucleic acid interactions, multi-state protein switches and nanomachines. New methods may produce more compact and modular DNA-binding proteins, protein-RNA structures and functional complexes. LOCKR-type systems can convert input binding events into luminescence, fluorescence, degradation, relocalization or Boolean logic. Hinge-like switches, allosterically regulated assemblies and conditionally released cytokine mimics point toward programmable therapeutic systems.

The review also mentions small axial and rotor-like protein assemblies with internal degrees of freedom. These are still far from ATP synthase or myosin, but they indicate a possible future in which proteins are designed as nanoscale mechanical devices driven by light or chemical fuel.

Beyond Medicine: Materials, Mineralization and Artificial Photosynthesis

The review does not limit protein design to therapeutics. It also emphasizes sustainability and materials science. Biomineralization is one example: bones, teeth and shells show that living systems can control inorganic nucleation and growth with proteins. Designed protein scaffolds may one day recognize inorganic surfaces or crystal lattices, guide calcium carbonate formation for carbon dioxide fixation or template semiconductor materials such as zinc oxide.

Artificial photosynthesis is another long-term direction. Natural photosynthesis is complex and efficient but constrained by evolutionary history. By designing simpler and more robust modules, researchers may expand plant use of near-infrared light, reconstruct chlorophyll special pairs or assemble multinuclear metal clusters for water oxidation and other multi-electron reactions.

The broader message is that proteins are not only therapeutic molecules. They can be programmable materials platforms that self-assemble, recognize chemical environments, can be produced by cells and bridge organic and inorganic worlds.

Risks and Bottlenecks

The review also emphasizes boundaries and risks. Immunogenicity remains a central question for any de novo protein entering the human body. Early data suggest that many designed proteins may have relatively low immune responses because they are stable and soluble, but this cannot replace rigorous assessment. T-cell epitope minimization, long-term safety and population diversity all need systematic study.

Manufacturing is another bottleneck. Medical products require high-purity production, and non-medical materials or industrial catalysts may require bulk protein production. Many designed proteins express well, remain soluble and are thermostable, but not every design is automatically manufacturable.

Experimental evaluation also becomes more difficult as the field moves from binding to dynamic switching, complex catalysis and multi-state machines. New high-throughput assays, functional screens and in vivo validation platforms will be as important as the models themselves.

Finally, misuse and intellectual property cannot be ignored. Tools that design therapeutic proteins could in principle be misused, even if the authors argue that the benefits for protection and treatment are likely to outweigh hypothetical risks. Easier sequence redesign will also complicate patent boundaries and ownership questions.

Overview of What CD ComputaBio Can Provide

De novo protein design programs require coordination among structure modeling, foldability assessment, interface design, molecular dynamics, protein therapeutic design, delivery strategy and immunogenicity risk evaluation. CD ComputaBio supports research-stage projects that connect computational design with practical validation and application decisions. The following related capabilities were selected from the attached website outline because they directly match the technical themes discussed on this page.

Contact Us

Contact us to discuss how CD ComputaBio can support your de novo protein design project, from structure modeling and binder or assembly design to stability analysis, delivery-system planning and early developability assessment. Our team can help translate a biological goal into a computational workflow, prioritize candidates for wet-lab validation and connect design outputs with practical experimental decisions. For more information, please visit CD ComputaBio or submit an inquiry through the website contact channel.

References:

1. Yang W, Wang S, Lee G R, et al. The past, present and future of de novo protein design. Nature, 2026, 652(8112): 1139-1152. https://doi.org/10.1038/s41586-026-10328-7.

• De Novo Design of Miniature Metallo Ketoreductases
• Nature Highlights Baker Lab's New Breakthrough: Single-Protein Artificial Virus-Like Particles Reach 220 nm in Diameter
• Nature Portfolio Journal Highlights Folddisco, a 20× Faster and 4× More Storage-Efficient Tool for Large-Scale Protein Structural Motif Search
• Next-Generation Peptidomimetic Therapeutics: Backbone Modification, Molecular Engineering Strategies, and Biotechnological Advances