Generative protein design

Current State

Generative protein design is rapidly advancing, driven by deep learning and AI. Methods like ProteinMPNN demonstrate superior sequence recovery compared to traditional approaches like Rosetta (Dauparas et al., 2022). Deep generative models, including VAEs, GANs, autoregressive transformers, and diffusion models, are now central to navigating and constructing chemical and proteomic spaces (Das, 2025; Winnifrith et al., 2024). These models are adept at generating novel, yet realistic proteins with desired properties and functions (Winnifrith et al., 2024; Wu et al., 2021).

Strongest Evidence

Deep learning-based methods have shown remarkable performance. ProteinMPNN achieved 52.4% sequence recovery on native protein backbones, significantly outperforming Rosetta's 32.9% (Dauparas et al., 2022). Frameworks like AlphaDesign, combining AlphaFold with autoregressive diffusion models, enable rapid generation and computational validation of proteins with controllable interactions (Jendrusch et al., 2025). Experimentally, de novo designed proteins have successfully created transmembrane β-barrel pores with tailored diameters and geometries (Berhanu et al., 2024) and induced precise RSV-neutralizing antibodies (Sesterhenn et al., 2020). Autoregressive generative models are also proving effective for protein design and variant prediction, especially where robust multiple sequence alignments are lacking (Shin et al., 2021).

Unresolved Uncertainties

Despite progress, challenges remain in consistently achieving precise functional control and high experimental success rates for de novo designed proteins (Sesterhenn et al., 2020). The accuracy of energy functions for computational protein design is still a limiting factor, with issues like parameter overfitting and inability to discriminate incorrect designs (Huang et al., 2019). Furthermore, the increasing sophistication of AI-assisted protein engineering introduces biosecurity concerns, necessitating strengthened nucleic acid biosecurity screening against generative protein design tools (Wittmann et al., 2025).

Why the Topic Matters

Generative protein design holds immense potential to expand the natural protein repertoire, enabling the engineering of new molecules with desirable functions beyond what evolution has produced (Winnifrith et al., 2024). This has broad implications for biocatalysis, therapeutics (e.g., next-generation vaccines), and materials science (Berhanu et al., 2024; Sesterhenn et al., 2020). The ability to design proteins with specific structures and functions is crucial for addressing complex biological and medical challenges, from drug discovery to creating novel enzymes for efficient energy conversion (Das, 2025; Ennist et al., 2022).

Points of Agreement

• Deep learning, especially generative AI models (e.g., VAEs, GANs, autoregressive transformers, diffusion models), has significantly advanced protein sequence and structure design, moving beyond traditional physics-based methods like Rosetta (Dauparas et al., 2022; Das, 2025; Winnifrith et al., 2024; Wu et al., 2021).
• Generative models are highly effective at creating novel, yet realistic proteins with desired properties and functions, achieving experimental success rates nearing 20% (Winnifrith et al., 2024).
• Computational methods like ProteinMPNN demonstrate superior sequence recovery (e.g., 52.4% vs. 32.9% for Rosetta) and are applicable to complex designs involving multiple chains (Dauparas et al., 2022).
• De novo protein design, facilitated by generative AI, enables the engineering of proteins with specific biological functions, such as inducing precise antibody responses for vaccines (Sesterhenn et al., 2020) and sculpting nanopores with tailored sizes and shapes (Berhanu et al., 2024).
• Frameworks like AlphaDesign, which integrate AlphaFold with autoregressive diffusion models, offer generalizable solutions for rapid generation and validation of proteins with controllable interactions and conformations (Jendrusch et al., 2025).

Active Disagreements

• The optimal balance between leveraging evolutionary information and purely generative approaches remains an area of active research. While some methods rely on evolutionary data, others aim to overcome limitations where robust multiple sequence alignments are unavailable (Shin et al., 2021).

Uncertain Claims

• The full extent to which generative models can reliably design proteins for any desired biological function, especially complex ones, is still being explored. While successes are notable, the generalizability across all functional landscapes is not yet fully established (Sesterhenn et al., 2020).
• The long-term stability and immunogenicity of de novo designed proteins in in vivo applications, particularly for therapeutics and vaccines, require further extensive validation (Sesterhenn et al., 2020).

Evidence Gaps

• More comprehensive experimental validation is needed across a broader range of protein types and functions to fully characterize the success rates and limitations of various generative design methods (Winnifrith et al., 2024).
• There is a need for standardized benchmarks and evaluation metrics to compare the performance of diverse generative protein design algorithms consistently, especially regarding novel function generation rather than just sequence recovery.
• Further research is required to understand and mitigate the biosecurity risks associated with AI-enabled protein design tools, particularly concerning the potential for creating harmful proteins that evade existing screening methods (Wittmann et al., 2025).

Claim 1: Deep learning models, such as ProteinMPNN, significantly advance protein sequence design, outperforming traditional methods.

• Supporting Evidence: ProteinMPNN, a deep learning-based method, achieved a 52.4% sequence recovery on native protein backbones, demonstrating outstanding performance in both in silico and experimental tests (Dauparas et al., 2022). This method has been instrumental in generating experimentally characterized de novo protein designs, a domain previously dominated by physically-based approaches like Rosetta.
• Limitations: While ProteinMPNN shows strong performance, the complexity of protein folding and function means that even high sequence recovery doesn't guarantee desired functional properties in all cases. Further experimental validation is always required.
• Evidence Strength: Strong.

Claim 2: Generative AI, including autoregressive models, is crucial for designing functional protein sequences and predicting variant effects, especially when evolutionary data is scarce.

• Supporting Evidence: Autoregressive generative models are effective for designing functional sequences and predicting variant effects, particularly in applications where robust multiple sequence alignments are unavailable (Shin et al., 2021). This addresses a significant limitation of traditional methods that rely heavily on evolutionary information.
• Limitations: The effectiveness of these models can still be influenced by the quality and quantity of training data. Designing proteins for entirely novel functions without any evolutionary precedent remains a challenge.
• Evidence Strength: Moderate to Strong.

Claim 3: Generative AI is a disruptive paradigm in molecular science, enabling the algorithmic navigation and construction of chemical and proteomic spaces for drug discovery and protein design.

• Supporting Evidence: Reviews highlight generative AI as a transformative force, using models like VAEs, GANs, and autoregressive models to explore and create novel molecules (Das, 2025; Winnifrith et al., 2024). This paradigm extends protein engineering beyond natural evolution, facilitating the design of proteins with specific desirable properties.
• Limitations: While the theoretical potential is vast, the translational applications are still in early stages for many areas. The complexity of validating all generated designs experimentally remains a bottleneck.
• Evidence Strength: Strong, based on comprehensive reviews and the rapid development in the field.

Claim 4: De novo protein design, facilitated by generative AI, can create novel protein structures like transmembrane β-barrels with tailored properties for specific applications.

• Supporting Evidence: A general approach has been developed to design transmembrane β-barrel pores with different sizes and shapes, demonstrating the ability of de novo design to create novel nanopores beyond naturally occurring channels (Berhanu et al., 2024). This showcases the potential for applications in sensing.
• Limitations: While successful in specific instances, scaling this to a wide range of complex protein structures and functions still requires significant research and development.
• Evidence Strength: Moderate to Strong, based on successful proof-of-concept for specific structures.

The field of generative protein design is experiencing a significant transformation, moving beyond traditional physically-based methods to leverage advanced deep learning and artificial intelligence (AI) approaches. This shift is enabling the creation of novel protein sequences and structures with unprecedented accuracy and functional specificity.

Recent Directional Trends and Emerging Methods:

• Deep Learning Dominance: Deep learning models are increasingly outperforming conventional methods in protein sequence design. ProteinMPNN, for instance, demonstrates a 52.4% sequence recovery on native protein backbones, significantly higher than Rosetta's 32.9% (Dauparas et al., 2022). This highlights a strong trend towards AI-driven solutions for sequence generation.
• Generative AI Architectures: The adoption of diverse generative AI architectures is a key trend. This includes variational autoencoders (VAEs), generative adversarial networks (GANs), autoregressive transformers, and score-based denoising diffusion probabilistic models (DDPMs) (Das, 2025; Winnifrith et al., 2024). These models are adept at generating novel, yet realistic, proteins with desired properties.
• Integration with Structure Prediction: Emerging frameworks like AlphaDesign integrate AlphaFold with autoregressive diffusion models to enable rapid generation and computational validation of proteins with controllable interactions, conformations, and oligomeric states (Jendrusch et al., 2025). This synergy between structure prediction and generative design is a powerful development.
• Autoregressive Models for Design and Variant Prediction: Autoregressive generative models are gaining traction for designing functional sequences and predicting the effects of variations, particularly in applications where robust multiple sequence alignments are challenging, such as for disordered proteins or highly variable regions like antibody complementarity determining regions (Shin et al., 2021).

Applications:

• Therapeutics and Vaccinology: De novo protein design is proving crucial for inducing precise antibody responses. For example, the TopoBuilder algorithm has been used to engineer epitope-focused immunogens that induced robust RSV-neutralizing antibodies in animal models (Sesterhenn et al., 2020). Generative AI is also a frontier in drug discovery (Das, 2025).
• Biocatalysis and Materials Science: The ability to design proteins with specific functions extends to creating new enzymes for efficient solar-to-fuel energy conversion (Ennist et al., 2022) and sculpting conducting nanopore size and shape for sensing applications (Berhanu et al., 2024).

Signals Needing More Validation:

• Experimental Validation of AI-Designed Proteins: While computational performance is high, the experimental characterization of de novo protein designs generated by deep learning methods, though increasing, still requires more extensive validation to match the breadth of physically-based approaches (Dauparas et al., 2022).
• Biosecurity Implications: The power of AI-assisted protein engineering introduces new biosecurity challenges. Efforts are underway to strengthen nucleic acid biosecurity screening against generative protein design tools that could create variants of proteins of concern (Wittmann et al., 2025). This area requires ongoing vigilance and validation of screening methods.

Synthesis

Generative protein design has shifted from physics-based methods like Rosetta to deep learning-based approaches, significantly improving performance and success rates. Recent advances leverage generative AI architectures, including autoregressive models, variational autoencoders (VAEs), and diffusion models, to design novel protein sequences and structures with unprecedented accuracy. Models like ProteinMPNN have demonstrated superior sequence recovery compared to traditional tools, while frameworks like AlphaDesign integrate structure prediction (AlphaFold) with generative processes for rapid, controllable design. This paradigm shift enables the creation of functional proteins for complex applications, achieving experimental success rates nearing 20% and expanding the accessible repertoire of protein structures and functions beyond what is found in nature.

Methods & Materials

Key methods in generative protein design are increasingly dominated by deep learning. ProteinMPNN is a prominent deep learning method for sequence design on fixed backbones, outperforming older methods. Autoregressive models, adapted from natural language processing, generate sequences amino acid by amino acid, proving useful for designing variable regions like those in antibodies. Diffusion models, such as those used in AlphaDesign, work by iteratively 'denoising' a random state into a coherent protein structure and sequence. These computational tools are often coupled with experimental validation techniques, including X-ray crystallography and nuclear magnetic resonance (NMR), to confirm that the designed proteins fold into their intended structures and exhibit the desired functions.

Applications

Generative methods are being applied across biotechnology and medicine. In therapeutics, they are used to design epitope-focused immunogens for vaccines, as demonstrated with respiratory syncytial virus (RSV), inducing precise and potent neutralizing antibody responses. In materials science, researchers are sculpting transmembrane β-barrels to create nanopores with custom geometries for sensing applications. Other applications include the de novo design of photochemical reaction centers for solar-to-fuel energy conversion and the engineering of novel enzymes for biocatalysis. The ability to design proteins with controllable interactions and conformations opens new avenues for drug discovery and synthetic biology.

Limitations & Future Directions

Despite rapid progress, significant challenges remain. A primary limitation is ensuring that computationally designed proteins are functionally active in vitro and in vivo, as high structural accuracy does not guarantee biological function. The vastness of the protein sequence-structure space makes comprehensive exploration computationally expensive. Furthermore, the power of these tools raises biosecurity concerns, as they could potentially be used to design harmful proteins that evade current screening protocols. Future work will focus on improving functional prediction, co-designing sequence and structure simultaneously, and developing robust biosecurity measures to mitigate misuse.

Update Triggers

This review should be updated upon the emergence of:

• Novel generative architectures that significantly outperform current models like diffusion or autoregressive transformers.
• Reports of high-throughput experimental validation of thousands of de novo designs, confirming high success rates for complex functions.
• Breakthroughs in designing multi-protein complexes or molecular machines with dynamic functions.
• Development of standardized benchmarks and robust biosecurity screening tools specifically for AI-generated proteins.

Near-Term Research Directions

The immediate focus should be on refining and integrating diverse generative models. While methods like ProteinMPNN (Dauparas et al., 2022) and autoregressive models (Shin et al., 2021) have shown success, the next step is to hybridize these with diffusion models and frameworks like AlphaDesign (Jendrusch et al., 2025) to better control for specific structural and functional constraints. A key area is improving the design of complex functionalities beyond simple structure, such as enzymatic activity, allosteric regulation, and precise epitope binding for vaccines (Sesterhenn et al., 2020). Another priority is developing models that require less evolutionary data, making them applicable to truly novel folds and disordered proteins.

Validation Needs

High-throughput experimental validation is the primary bottleneck. Current computational design outpaces our ability to synthesize and test proteins. We need scalable pipelines for gene synthesis, protein expression, and functional screening. Biophysical characterization (e.g., NMR, crystallography) remains crucial for confirming that designed structures match computational models, as demonstrated in the design of transmembrane pores (Berhanu et al., 2024). Developing rapid, low-cost functional assays is essential to close the design-build-test-learn loop.

Datasets and Benchmarks

The field requires standardized, comprehensive benchmark datasets that go beyond sequence recovery on native backbones. A critical need is a public repository of both successful and failed de novo designs, complete with experimental characterization data. This would enable more robust model training and evaluation. Benchmarks should assess not just structural accuracy but also functional success, stability, and expressibility. The development of improved energy functions like EvoEF2 (Huang et al., 2019) provides a foundation, but these need to be integrated with deep learning models and validated against large-scale experimental data.

Translation Barriers

Key barriers to clinical and industrial translation include manufacturability (expression and purification yields), immunogenicity of de novo proteins, and long-term stability. For therapeutics, ensuring designed proteins have the required specificity and avoid off-target effects is paramount. Furthermore, as generative tools become more powerful, addressing biosecurity concerns is critical. Strengthening nucleic acid synthesis screening protocols is necessary to prevent the misuse of these technologies for creating proteins of concern (Wittmann et al., 2025).

High-Value Experiments

1. Design of Multi-State Functional Proteins: Create proteins that switch between distinct conformational states in response to specific triggers (e.g., ligand binding, light), which is a hallmark of complex biological machinery.
2. De Novo Enzyme Design for Non-Natural Reactions: Engineer enzymes that catalyze industrially relevant chemical reactions not found in nature, moving beyond modifying existing scaffolds.
3. Programmable Protein Assemblies: Design protein components that self-assemble into complex, pre-defined nanostructures or materials with novel properties, such as photochemical reaction centers (Ennist et al., 2022).
4. Head-to-Head Benchmarking: A large-scale, community-wide experiment comparing leading generative models (e.g., ProteinMPNN, RFdiffusion, AlphaDesign) on a standardized set of challenging design tasks, with all designs synthesized and tested under identical conditions.