Protein design enables the creation of novel structures and functions beyond those found in nature, with recent progress accelerated by computational modeling and machine learning. However, many automated methods act as black boxes, limiting mechanistic insight. Here we highlight the continuing importance of rational protein design, defined as an approach rooted in physical principles, chemical intuition, and sequence-structure-function relationships. We outline three complementary strategies: backbone-first, sequence-first, and function-first, which provide interpretable design frameworks and enable robust scaffold generation, motif incorporation, and functional engineering. Looking forward, we argue that hybrid workflows combining rational principles with machine learning offer the most promising route to dynamic, explainable, and generalizable protein design.

Engineering protein stability is a powerful strategy across biotechnology and medicine, supporting a broad range of applications such as atomic structure determination, discovery of therapeutic molecules, biomanufacturing, diagnostic reagents, industrial biocatalysis, etc. However, achieving rapid and significant improvements has been historically challenging due to the vast mutational space and the complex interplay of sequence, structure, and function. Indeed, traditional experimental and computational methods often struggle to predict the impact of multiple mutations and effectively integrate diverse data types. To address these limitations, we developed StabLyzeGraph, a novel computational framework powered by Graph Neural Networks (GNNs) for protein mutational analysis and classification of stabilizing mutations. StabLyzeGraph represents proteins as graphs, integrating amino acid physicochemical properties, evolutionary conservation scores, and mapped three-dimensional structural information. The framework consists of a Benchmarking module to evaluate performance, and a Screening module to identify and rank impactful mutations. Benchmarking across 23 diverse datasets demonstrated strong predictive performance, highlighting the GNN's ability to leverage integrated features. Mutational analysis enables the generation and probability scoring of single- and multi-site mutants, demonstrating the model's capacity to classify beneficial combinations of mutants based on learned structural impact rather than mere mutation frequency. StabLyzeGraph also features a user-friendly Graphical User Interface and demonstrates reasonable computational efficiency and scalability for exploring mutational landscapes. This tool provides a robust and versatile approach to accelerate the efficient discovery of stabilizing mutations with tailored properties and represents a step forward in rational protein design, poised to accelerate the creation of novel biologics with enhanced performance. StabLyzeGraph is freely available on GitHub (https://github.com/cosconatilab/StabLyzeGraph) as an open-source tool.

Design, expression, and immunogenicity evaluation of a novel multi-epitope chimeric protein (SOMP) against Salmonella Typhimurium.

Rational design of composite gels demands innovative strategies to enhance structural and functional properties. Dietary fibers (DFs) offer promising potential for reinforcing protein-polysaccharide networks, but their role in soy protein isolate (SPI)-chitosan (CS) binary systems is underexplored. A novel ternary composite gel from SPI, CS and DF was developed to investigate DF-mediated reinforcement mechanisms. The synergistic combination of microbial transglutaminase (MTGase) cross-linking and heat treatment significantly enhanced the water holding capacity (WHC) of SPI-CS-DF, achieving the highest 30.32% increase compared to single treatments. SPI-CS with 1 mg mL-1 citrus dietary fiber (CDF) under the combined treatment exhibited the highest WHC (40.82% increase versus SPI-CS). Furthermore, texture analysis revealed that the addition of 5 mg mL-1 apple dietary fiber (ADF) and 5 mg mL-1 CDF increased gel strength by 37.31% and 36.27%, respectively. However, the addition of oat dietary fiber (ODF) simultaneously reduced the gel strength and hardness. Fourier transform infrared spectroscopy, scanning electron microscopy and confocal laser scanning microscopy demonstrated that ADF and CDF promoted uniform protein networks with porous structures, whereas ODF disrupted matrix continuity. MTGase-treated gels showed higher amide I peak intensity, which exhibited stronger covalent cross-linking. Overall, DF type and concentration are critical to tailoring SPI-CS gel structure and performance. MTGase-heat treatment combined with appropriate DF addition offers an effective strategy to improve WHC and mechanical properties in composite gels. These results provide a theoretical foundation for designing high-performance gel systems with potential applications. © 2026 Society of Chemical Industry.

Rising Star: Rewriting the Code of Life for the Future of Food.

De novo engineering of protein interactions: Retrospective and current advances.

Nanobodies for inflammatory bowel disease: Miniaturizing solutions.

Traditional computational protein design heavily relies on expert-level biological inputs to define structural and functional constraints, posing significant barriers in terms of technical implementation and workflow construction. To address this gap, we capitalize on recent advancements in large language models (LLMs)─which excel at complex reasoning in specialized domains by leveraging knowledge bases to generate expert-grade outputs. In this study, we first propose the protein evolutionary paradigm, a design paradigm that emulates the core logic of natural protein evolution by taking biological function as the ultimate target, achieving progressive optimization of protein sequences under explicit functional and structural constraints through iterative evolutionary refinement. Guided by this paradigm, we present MAESD (Multiagent Evolutionary Framework for Protein Sequence Design), a unified computational framework for function- and structure-constrained evolutionary protein design guided by natural language instructions. This paradigm integrates multiagent collaborative reasoning to bridge the semantic gap between natural language descriptions and biological constraints, while adopting an iterative evolutionary optimization mechanism to ensure the biological plausibility of designed sequences at each iteration. MAESD operates through two core collaborative modules for sequence generation: (1) A semantic-to-biological translation module, which employs LLMs and biological databases to interpret user-provided natural language biological requirements and extract actionable protein design constraints; (2) an evolutionary loop module, which realizes iterative sequence refinement via a ″generation-validation″ cycle─utilizing ProGen2 and ProteinMPNN for sequence generation and integrating structural, energetic, and functional verification to filter and optimize sequences. By fusing natural language understanding with evolutionary computation, MAESD reduces the engineering and implementation burden of protein design workflows by automating pipeline integration and parameter adaptation, while expert biological judgment remains necessary for interpreting results and guiding experimental decisions.

Generative Protein Design for Physiological Intervention: Progress and Horizons.

Epistasis-the context-dependence of mutational effects-is a key driver of protein evolution, influencing adaptive pathways and functional diversity. While specific epistasis arises from direct physical interactions between mutations, non-specific epistasis emerges when a non-linear mapping links a protein's biophysical properties to its function. Enzyme kinetic parameters map directly to free energies, enabling researchers to connect epistasis in these parameters to an enzyme's structural features. Here, we show that this approach is incorrect: enzyme catalytic parameters like kcat and KM inherently exhibit non-specific epistasis due to the multi-state nature of the catalytic cycle. Using enzyme catalytic cycle models, parameterized by free energies of ground and transition states, we simulated 1000 "mutations" or perturbations to the sub-state free energies within the kinetic ensemble. We then combined these mutations, creating one million double mutants with strictly additive free energy effects. Despite the absence of explicit mutational interactions, we observed substantial epistasis in catalytic parameters; its prevalence and complexity increasing with the number of kinetic states in the mechanism. We derived analytical conditions for the emergence of this form of epistasis in a simple kinetic model, demonstrating that non-specific epistasis depends on the relative values of key microscopic rate constants. Finally, we validated our framework by reanalyzing kinetic data for double mutants in Bacillus cereus β-lactamase I and found that reported specific epistasis in catalytic efficiency was substantially stronger than previously inferred, altering mechanistic interpretations. Our results identify an intrinsic, previously unknown source of epistasis that can distort both the magnitude and sign of mutational effects in enzyme kinetics. We provide theoretical and computational tools for recognizing and correcting for this form of non-specific epistasis, enabling accurate mechanistic inference from kinetic data and improving our understanding of the links between epistasis, structure-function relationships, enzyme evolution, and protein design.

T7 RNA polymerase (T7 RNAP) is a foundational enzyme for biotechnology, but its utility for many potential applications is limited by low thermal stability of 43-44°C. While stabilized variants exist, the most stable commercial version has a proprietary sequence. In this work we developed a highly stable T7 RNAP using structure-based computational design. We combined mutations from previous stabilized variants (M5, M8, V7abcd) with new mutations identified by PROSS. These mutations were filtered using data-driven heuristics to preserve function. Our final design, T7T+, contains 30 point mutations from the original T7 RNAP and demonstrates a functional stability (T50) of 54.9°C in a thermal challenge assay, which is 2.4°C higher than the most stable, published open-source variant to date. Circular dichroism spectroscopy showed an apparent melting temperature of 53.8°C. T7T+ retains 59% of wild-type activity at 37°C. 16 of the 18 tested protein designs had higher stability against thermal challenge compared with the genetic background, attesting to the high success rates of existing non deep learning computational methods for the design of stable, functional proteins. A plasmid encoding T7T+ has been deposited in AddGene and is freely available for non-commercial use.

Yeast surface display (YSD) is a versatile platform for protein design and the development of whole-cell biocatalysts. In this study, we engineered the Aga2 anchor subunit, which mediates protein display on the Saccharomyces cerevisiae surface. Using a fungal laccase as a reporter and after only one single round of directed evolution of the Aga2 anchor subunit, we identified an improved Aga2 mutant that enabled a tenfold increase in laccase surface expression (∼26,000 U/g) without compromising enzyme folding or stability. Overall, this work demonstrates a proof of concept for engineering the Aga2 anchor as an effective strategy to enhance laccase surface display in yeast cells, with potential applicability to other enzyme systems.

Protein binders that detect, activate, inhibit, or otherwise modulate their targets are pivotal for biomedical applications. With the increasing accuracy and accessibility of de novo protein design, faster and cheaper experimental screening methods would democratize and accelerate the identification of high-affinity binders. Here we present Cell-Free Two-Hybrid (CF2H), a rapid and sensitive method for detecting high-affinity protein-protein interactions (PPI) that does not require cloning, protein purification nor high-end laboratory equipment. CF2H uses a dimerization-activated DNA binding domain (DBD) fused to prey and bait proteins to trigger transcription upon protein-protein interaction. We demonstrate that CF2H enables the detection of interactions between various types of target and binder proteins such as single-domain antibodies, DARPins, and de novo designed binders. We benchmark CF2H as a screening platform by validating previously reported binders for Mdm2 and discovering high-affinity binders targeting the checkpoint inhibitor PD-L1 in less than 24 hours. Finally, we show that CF2H can be used to characterize small-molecule modulators of PPI and detect protein biomarkers, opening the door for a new class of cell-free biosensors.

Beta-barrel structures are critical components of bacterial outer membranes, where they facilitate transport, cell signaling, antibiotic resistance, and structural integrity. A key feature of beta-barrels is their strand count, which influences pore diameter, binding site locations, and functional properties. However, because of breaks in strands and the presence of strands in periplasmic domains and plug domains, manual counting is inefficient and current algorithms do not accurately determine barrel strand count. To address this, we refined our previous beta-barrel structural assessment tool, PolarBearal, to improve strand number identification in large-scale datasets. To enhance the accuracy of barrel strand number labeling, our updated algorithm integrates three structural criteria, namely inter-residue vector angles, hydrogen-bonding distances, and strand connectivity. Using this algorithm, we labeled strand numbers for 571,760 predicted outer membrane beta-barrel structures obtained from the AlphaFold2 database. Our algorithm has 97% accuracy in strand number assignments, and the resulting dataset facilitates assessment of the homogeneity of strand counts for different types of outer membrane proteins. The strand labeling also provides insights on beta-barrel strand distribution and evolutionary patterns, supporting further research in protein structure prediction and design.

Disordered protein linkers are essential for multidomain protein function and engineering, but quantitative methods for their biophysical characterization remain limited. We combined NMR experiments with molecular dynamics simulations to demonstrate that protein backbone 15N spin relaxation times correlates with backbone rigidities in short disordered linkers. Using a tailored version of the Quality Evaluation Based Simulation Selection (QEBSS) framework, we characterized four model peptides: (GGS)3, (GPS)3, K(AP)5K, and KKEEVKKEEV-(PK)7KEEVKKEEVKK, representing common natural and engineered linker repeats. Glycine-rich sequences showed slight looping tendencies, while proline-containing sequences adopted extended conformations with increased approximate persistence lengths and slower dynamics. Notably, sodium and calcium binding to charged peptides minimally affected rigidity, indicating electrostatics don't dominate linker stiffness. This integrated approach provides quantitative insights into disordered linker properties and MD simulation accuracy, offering biophysical understanding for protein design and machine learning model development.

Artificial intelligence (AI) is revolutionizing protein science and transforming the fields of analytical and bioanalytical chemistry by harnessing advanced machine learning and deep learning techniques to address longstanding challenges. AI can now predict protein structures from amino acid sequences with near-experimental accuracy, as exemplified by breakthroughs such as AlphaFold2, significantly enhancing our understanding of protein function, dynamics, and interactions. In analytical chemistry, AI enables high-throughput protein characterization, structural analysis, and real-time data interpretation. In bioanalytical chemistry, it supports precise biomarker identification, protein quantification, and modeling of complex protein–protein interactions. Beyond structure prediction, AI accelerates the design of novel proteins and enzymes, facilitates proteomic data analysis for biomarker discovery, and aids drug development. While challenges remain in modeling dynamic systems and intrinsically disordered regions, the integration of AI promises to revolutionize analytical and bioanalytical methodologies, improve precision, and drive innovations in drug discovery, synthetic biology, and personalized medicine, positioning AI as a cornerstone of modern protein research.

Type 3 (T3) Cu proteins play essential roles in binding and activating molecular oxygen (O2) and are prevalent across all domains of life. Despite sharing the same coordination motif, T3 Cu proteins display divergent functions: hemocyanin transports O2, while tyrosinase catalyzes the hydroxylation of monophenols and the subsequent oxidation of diphenols and catechol oxidase oxidizes only diphenols. Here, we report the design and characterization of a di-Cu protein (Cu-HC4) inspired by the active sites of natural T3 Cu proteins to investigate the structural features that facilitate catalytic oxidase activity. Cu-HC4 is roughly 1/5th the size of the commercially available mushroom tyrosinase and shares only around 20% sequence identity with the T3 Cu protein templates. Notably, Cu-HC4 displays high thermostability and exhibits diphenol oxidation activity at ambient and elevated temperatures (≥60 °C). Cu-HC4 also initiates the formation of melanin polymers, mimicking melanin biosynthesis of natural tyrosinases. Mechanistic investigations demonstrate that Cu-HC4 utilizes both Cu centers cooperatively for diphenol oxidation and requires O2 for catalysis like natural Cu oxidases but follows a distinct catalytic pathway compared to those enzymes. Cryo-EM characterization of a tetrameric form of HC4 reveals slight deviations in the relative positions of the active site His residues that may account for differences in reactivity between Cu-HC4 and natural T3 Cu enzymes.

Multivalent interactions mediated by multidomain proteins are pivotal in numerous biological processes. However, the thermodynamic intricacies of the domain interactions within such complexes remain elusive. In this study, we employed surface plasmon resonance to explore the temperature-dependent kinetics of multidomain protein interactions across various valences from monomers to tetramers. Rigorous screening of protein-peptide binding pairs fused with discrete multivalent protein scaffolds facilitated the selection of candidates with minimal nonspecific interactions and suitable monomer binding kinetics that could be extended to higher valences. We developed a theoretical model to extract the thermodynamic quantities for both inter- and intramolecular interactions. By employing initial rate analysis, we could extract thermodynamic quantities describing complicated interactions between multivalent proteins. Our analysis provides novel insights into the thermodynamics of intramolecular interactions in multivalent protein complexes with implications for protein design and engineering.

Geometric deep learning assists protein engineering. Opportunities and Challenges.

Light-sensitive proteins allow organisms to perceive and respond to their environment, and have diversified over billions of years. Among these, Light-Oxygen-Voltage (LOV) domains are widespread photosensors that control diverse physiological processes and are increasingly used in optogenetics. Yet, the evolutionary constraints that shaped their protein dynamics and thereby their functional diversity remain poorly resolved. Here we systematically characterize the dynamics of 21 natural LOV core domains, significantly extending the spectroscopically resolved catalog through the addition of 18 previously unstudied variants. Using time-resolved spectroscopy, we uncover an exceptional kinetic diversity spanning from picoseconds to days and identify distinct functional clusters within the LOV family. These clusters reflect evolutionary branching, including a divergence of ≈1.0 billion years between investigatedLOV variants from plants and ≈0.4 billion years of separation within one of these functional clusters. Individual variants with extreme photocycles emerge as promising anchor points for optogenetic applications, ranging from highly efficient adduct formation to ultrafast recovery. Beyond natural diversity, we introduce a LOV domain generated by artificial intelligence-guided protein design. Despite being sequentially remote from its maternal template, this variant retains core photocycle function while exhibiting unique biophysical properties, thereby occupying a new region on the biophysical landscape. Our work emphasizes how billions of years of evolution defined LOV protein dynamics, and how protein design can expand this repertoire, engineering next-generation optogenetic tools.