Protein Engineering, Design and Selection

A challenge in synthetic biology is to transfer the knowledge acquired in vitro to in vivo applications. For instance, binding affinity between peptide binding domains (PBDs) and their binding peptides is usually measured in simple buffers and in the absence of the cell machinery responsible for synthesis and proteostasis. Consequently, there is often a discrepancy between interaction quantified in vitro and in vivo. To identify strongly binding PBD-peptide pairs for in vivo applications, we aim to better understand the determinants of synthetic interaction in living yeast cells. We designed libraries of peptides expected to bind to PBDs based on data collected by phage-display. Using an in vivo protein interaction assay, we assessed the interaction strength of >6,500 PBD-peptide variants in yeast. We found that most substitutions disrupt binding, but that 3.4% increased binding strength to levels higher than expected from the phage-display data. We asked if, by using our mutational data, it was possible to optimize peptide sequences to gain even more binding strength and show additivity effect up to four substitutions. We found that PBD specificity is consistent between in vitro and in vivo assays. We hypothesized that in vivo peptide abundance could contribute to the discrepancy between the PCA and the phage-display results. We performed assays to detect in vivo peptide availability, reporting on peptide abundance, and we found that peptides with a stronger binding also show a higher availability in cells. We show that peptide availability, which seems mainly defined by peptide hydrophobicity, is key for the design of strong synthetic interactions.

Engineering proteins at the scale of modern high-throughput screening technologies remains challenging because most experimental libraries sample only limited regions of sequence space while maintaining relatively low levels of functional diversity. Although recombination-based strategies can introduce substantially larger sequence changes than conventional mutagenesis approaches, it is unclear whether highly recombined enzymes can retain catalytic competence when library sizes approach millions to billions of variants. To investigate this question, we constructed a recombination library by combining ten sequence blocks from ten bacterial PETases, generating a theoretical diversity of 10^8 chimeric enzymes. Screening of yeast-displayed libraries using a fluorophosphonate probe for catalytic-serine reactivity identified substantial populations of catalytically competent variants despite extensive sequence recombination and dozens of amino acid substitutions per chimera. Approximately 2% of sampled variants possessed active-site serine reactivity, corresponding to an estimated 10 catalytically competent enzymes within the full theoretical library. Deep sequencing showed that functional populations remained broadly distributed across sequence space following selection and contained multiple enriched recombination patterns spanning the PETase structure. Secondary lysate screening and enzyme-normalized activity measurements further identified multiple purified PETase chimeras with soluble esterase activity comparable to or greater than the wild type Ideonella sakaiensis PETase on a chromogenic surrogate substrate. Together, these results demonstrate that large-scale recombination can generate millions of diverse and catalytically competent enzymes while remaining compatible with modern high-throughput screening workflows. More broadly, this work establishes recombination-based diversity generation as a scalable strategy for protein engineering and large-scale sequence-function data generation.

Engineering G protein-coupled receptors (GPCRs) for biosensing applications remains challenging due to structural and functional constraints when such proteins are expressed in a heterologous host. OrthoRep offers the continuous accumulation and selection of signal-enhancing mutations in yeast, yet it has not been applied to GPCR evolution. We apply OrthoRep-driven mutagenesis to evolve opioid GPCRs and improve yeast-based biosensors: first, we express the human -opioid receptor (OPRM1) on the OrthoRep p1 plasmid and couple ligand activation to growth. Serial cell passaging of the biosensor generated functionally diverse mutants. However, mutations that decouple receptor activation from selection arose over longer passaging campaigns, halting evolution. We resolve this issue using a site-specific recombinase that re-introduces evolved variants onto the p1 landing pad of an unmutated biosensor, allowing continued selection. This led to an ~18-fold improvement in sensitivity of a mutant OPRM1 over wt-OPRM1. Additional polymerases were added to the platform, broadening the mutational profiles available. Assembling these components forms the platform we call HERO (Heterologous Engineering of Receptors using OrthoRep). Lastly, we deploy HERO using automation to improve the human {delta}-opioid receptor's weak response to an agonist by ~40-fold. Sensitivity also improved for a structurally dissimilar ligand assayed on the same evolved variants.

Although use of AI-based tools has gained ground in molecular biology research, their reliability remains a matter of debate. Here, we combined AlphaFold 3 modelling with microscale thermophoresis to investigate the interactions between glucan-binding protein C and the biofilm-inhibiting aptamer and its five derivatives. Structural modelling identified a binding motif on the aptamer and suggested a mechanism of action; however, the predictions were only partially supported by the experimental results. The AlphaFold generated confidence metrics did not correlate directly with the measured binding affinities, indicating that confidence scores can help identify plausible interaction interfaces, but they are not reliable predictors of binding strength. These results highlight the importance of integrating structural prediction with biophysical validation to dissect aptamer-protein recognition.

Synthetic biologists aim for standardization and simplicity in order to ensure reproducibility and reliability. Toolkits offer both, making it easier to share DNA sequences between labs while minimizing context dependent effects from sequences surrounding the construct of interest. The Yeast Toolkit (YTK) for the model organism Saccharomyces cerevisiae has found wide adoption within the scientific community. While the Toolkit and its existing extensions offer a large variety of different parts, there is a lack of characterized logic gates and insulator containing backbones for the assembly of genetic circuits with the yeast toolkit syntax. In this work, we present the Yeast Logic Toolkit (YLT). YLT contains 40 pre-assembled backbones with dropouts, 109 parts for NOT/NOR gate assembly, and a subdivided promoter Type 2 part that allows more flexible 5'UTR design. For insulation, the backbones contain spacer sequences and self-cleaving ribozymes which buffer against transcriptional read-through from upstream sequences. Due to our subdivision of Type 2 parts, we are able to manipulate the 5' untranslated region without a cloning scar upstream of the start codon. We also characterized the NOT and NOR gates from the Cello library for yeast in the sequence context of the yeast toolkit cloning scheme. This toolkit extension will make it easier and faster for scientists to assemble genetic circuits in yeast while providing a minimum of context dependent effects such as genetic read-trough due to the standardization of up- and downstream sequences surrounding the transcriptional units.

Human Ca2+/calmodulin (CaM)-dependent protein kinase II alpha (CaMKII) is a highly abundant and conserved protein in the brain. It is a multifunctional serine/threonine kinase that phosphorylates cellular substrates across subcellular compartments, activating signaling pathways essential for long-term memory storage. However, a detailed understanding of protein structure and enzyme-substrate interaction mechanisms require a method to produce active, stable, and pure proteins at scale. Since phosphorylation can alter catalytic activity and substrate affinity, examining the non-phosphorylated state enables an unbiased characterization of basal enzymatic function. Therefore, producing the non-phosphorylated form of the kinase is important for biochemical and structural studies aimed at understanding kinase activation mechanisms. We describe a detailed protocol for the expression and high-yield purification of full-length, non-phosphorylated human CaMKII in bacterial cells. CaMKII was co-expressed with {lambda}-phosphatase to maintain it in a non-phosphorylated state. Protein expression was induced by autoinduction at low temperature to enhance the stability and solubility of CaMKII. Recombinant CaMKII was purified by Ni-NTA affinity chromatography, followed by gel filtration chromatography. Although CaMKII has been considered difficult to purify in its non-phosphorylated form, we obtained it in high yield, and the protein was predominantly non-phosphorylated. Our results show that combining co-expression with autoinduction improves both the solubility and yield of the homogeneous non-phosphorylated form of CaMKII, which could be suitable for subsequent biochemical and structural studies.

Designed binders are routinely filtered on predicted interface confidence, which reports how self-consistently a model places an interface, not where it lies. We designed 200 nanobody loops against the survivin BIR-domain groove, co-folded 196, and checked epitope contact in every predicted complex. Confidence showed no detectable association with epitope engagement ({rho} = +0.100, 95% CI -0.04 to +0.24, n = 196); its highest and lowest quartiles failed alike (odds ratio 0.97, 95% CI 0.41 to 2.26). It tracked instead the run's other confidence estimates ({rho} = +0.61). Twelve-seed replication separated two candidates the criteria had not: one reached the complete epitope in 10 of 12 runs, the other never. Across 155 predictions the top-ranked model was no better than one drawn at random from its five (P = 0.610), though only 3 of 31 failed on every model. Nothing was assayed; nb_cand_057 is an in silico candidate only.

Protein fitness prediction is essential for protein engineering. Recent fitness-landscape benchmarks such as FLIP2 have shown that methods performing well under conventional train-test splits can fail when the test set is shifted out of distribution (OOD), including mutations at unseen positions and variants from unseen protein backbones. In particular, current protein language models (pLMs) often suffer substantial performance degradation under such distribution shifts, limiting their practical use in protein engineering. We therefore develop a framework that augments sequence representations with structural and co-evolutionary information for fitness prediction. Specifically, we pair structure-aware pLMs with features pooled from an AlphaFold-style trunk applied to each sequence and its multiple sequence alignment (MSA), and use a lightweight prediction probe to map these frozen representations to fitness scores. Our framework consistently outperforms the best previously reported FLIP2 scores across several types of distribution shift. We then systematically study the contribution of each component in our framework to generalization. Structural priors yield the largest performance gains on held-out protein backbones, whereas trunk features improve generalization to held-out mutation positions. Rather than proposing a single best method, we characterize which modalities are effective for different protein families and distribution shifts, providing practical guidance for choosing representations according to the generalization regime. These findings support the development of more generalizable fitness predictors that can serve as effective reward functions or oracles for protein design toward desired functions.

Directed evolution enables the discovery of protein mutants with improved fitness through iterative rounds of selection and has been widely applied to adeno-associated virus (AAV) capsid engineering. Machine learning (ML) can augment this process by prioritising mutants for experimental validation, but whether its benefits outweigh the added cost of ML-designed library construction remains unclear. Here, we address this question by considering improvements in manufacturing efficiency of AAV capsids in the context of exosomal encapsulation, which is a technique for improving AAV immunogenicity. We generated a directed evolution dataset comprising 53,974 mutants across three rounds of selection and an independent assessment dataset of 472 ML-designed mutants with detailed profiling. Directed evolution alone yielded a 3-fold improvement, while ML-assisted directed evolution yielded a 41-fold improvement in manufacturing efficiency, demonstrating that ML-assisted directed evolution outperforms directed evolution alone. We show that data quality - specifically the number of selection rounds and the statistical power of mutant counts - has a greater impact on model performance than model architecture. Finally, we provide practical guidance on experimental design and implementation of ML tuning strategies that can augment model performance in protein engineering applications.

Protein-cofactor selectivity prediction is important for enzyme annotation, metabolic engineering, and biocatalyst discovery, but it remains challenging because experimentally supported associations are limited and unevenly distributed, and because biologically relevant differences can be subtler than those captured by general protein-ligand interaction models. Here, we present MINT (Multimodal Interaction Network with Teacher supervision), a multimodal framework for cofactor specificity prediction. MINT combines protein sequence features with cofactor sequence and graph representations and is trained with a gradual adaptation strategy that transfers useful information from a pretrained protein-ligand interaction model to the cofactor prediction task. On a fixed balanced held-out test set, MINT achieved the best overall performance among all compared methods (AUC = 0.932), and this ranking was reproduced in an independent family-stratified evaluation at the natural positive rate (pooled AUC = 0.939). Further analyses suggest that the pretrained model provides transferable structural information, while the gradual training scheme helps adapt that information to cofactor specificity. In a case study of the chemically near-identical NAD/NADP pair, MINT recovered 95.0-98.6% of true binders and resolved a single-phosphate-level distinction with direct relevance to engineering enzyme cofactor specificity. These results show that gradual transfer learning is an effective approach to cofactor specificity prediction when annotation is limited.

MAPK-activated protein kinase 2 (MAPKAPK2, MK2) is a downstream substrate of p38 MAPK that regulates inflammatory signaling and cytokine synthesis, making it an attractive therapeutic target. Although several crystal structures of MAPKAPK2 with ATP-competitive inhibitors have supported drug discovery, structural elucidation of selective non-ATP-competitive inhibitors has remained challenging. MK-25 is a non-ATP-competitive furan-2-carboxamide MAPKAPK2 inhibitor developed at Merck to overcome the selectivity and cellular-efficacy limitations of ATP-competitive chemotypes. Here, we show through isothermal titration calorimetry that MK-25 binds the activated MAPKAPK2, consistent with state-dependent allosteric binding. To determine the binding mode of MK-25, we solved a 2.8 A resolution crystal structure of an activated MAPKAPK2 construct in ternary complex with the ATP-site inhibitor PF-3644022 and a bromo-analogue of MK-25. The ligand occupies a pocket in the C-lobe overlapping the protein-substrate recognition region and formed through rearrangements of D and G helices and the activation segment. These interactions are consistent with the structure-activity relationships reported for the furan-2-carboxamide chemotype and identify a ligandable allosteric pocket that may facilitate future structure-guided design of selective, non-ATP-competitive MAPKAPK2 inhibitors.

A compound's toxicity can be gauged from the toxicities of the compounds it resembles. We showed that in 2003 for rat oral LD50: the compounds in a reference set of 13,645 that most resemble the query form a consortium and vote on its toxicity. This work returns to that method twenty-three years later, on public data, because the reference set used in 2003 is no longer public. Rebuilt on 11,989 curated rat oral LD50 values, the method answers for 67.0% of the held-out compounds carrying an exact value, and 92.1% of those answers land within one log unit; in 2003 the figures were 94% within one log unit on 57.3% of compounds. We then ran it over 28 further endpoints, each recording whether a compound showed a particular toxicity, from cytochrome inhibition to mutagenicity to clinical toxicity: 149,224 measurements on 34,162 compounds. The method answers only when a compound has close enough neighbors, and on those it correctly identifies a median 72.9% of the compounds that showed the toxicity and 96.5% of those that did not. Separately we asked whether a reverse screen carries information about these endpoints beyond molecular size and lipophilicity. We search a compound against a published index of 27,910 ligands solved inside proteins and count the protein families those sites belong to. A compound whose matches reach no protein family inhibits CYP3A4 2.5% of the time; one whose matches span sixteen or more families inhibits 51.2% of the time. Added to heavy atom count, molecular weight and cLogP, the screen raises the mean area under the ROC curve from 0.72 to 0.76 across the 28 endpoints. The OpenADMET consortium has since measured CYP3A4 inhibition on 2,335 compounds, and the same pattern holds. Compounds that resemble known binders of sixteen or more protein families inhibit CYP3A4 24.6% of the time; compounds whose matches reach no protein family inhibit it 7.3% of the time. The more protein families a compound resembles binders of, the more likely it is to inhibit CYP3A4. This association is strong enough to act on before anything is made. Scoring a library on how far its compounds reach across protein families, together with their size and lipophilicity, and setting aside the riskiest fifth removes two thirds of the androgen receptor binders and about half of the CYP3A4 inhibitors. Scoring needs only a two-dimensional structure, with no docking and no conformer generation at query time. Interactive scoring, the data and the pool are at toxpred.ai. Keywords: toxicity prediction; structural similarity; nearest neighbors; reverse screening; target breadth; promiscuity; LD50; cytochrome P450; virtual library triage

Constrained macrocyclic peptides can engage protein surfaces that resist both small molecules and biologics but producing them at library scale still depends on chemical synthesis and macrocyclisation workflows that are slow, specialised and hard to parallelise. We describe a recombinant pipeline for disulphide-stapled cyclic peptides, compatible with 96 well plates, that lowers this barrier. Peptide-encoding sequences are introduced by extension PCR and in vivo assembly (IVA) cloning at about $2 per variant, expressed as small ubiquitin-like modifier (SUMO) fusions secreted to the oxidising periplasm, where the intramolecular disulphide forms spontaneously, and purified by immobilised metal-affinity chromatography (IMAC) with analytical size-exclusion chromatography (SEC). The route from a list of designs to characterised material takes about one week, much of it unattended. Applied to a 60-design library across nine protein targets, the pipeline gave purified products for all 60 designs, at a median total soluble yield of approximately 135 micrograms per 6 mL culture (range 9 to 261) that varied by target. Ellman's assay across 38 designs showed disulphide-formation failure to be sequence-driven and uncorrelated with yield, detecting a failure mode invisible to chromatography. Surface plasmon resonance confirmed target binding and showed that the recombinant SUMO fusion constructs retained affinities comparable to their corresponding peptide-only forms. Complementary to general-purpose recombinant platforms, the pipeline enables higher-throughput disulphide macrocycle screening and is amenable to partially or fully automated workflows.

Recent advances in artificial intelligence have enabled models to capture protein sequence and structural features. We present PeTriPPI2, a hybrid framework for protein--protein interaction (PPI) prediction that combines sequence embeddings from ESM-2 with structural representations from PeTriBERT, an encoder originally designed for inverse folding. We evaluate this sequence--structure approach on the Pinder dataset. On a test set of 2,342 protein pairs, the selected checkpoint achieves 0.898 accuracy, 0.917 precision, 0.876 recall, and 0.896 F1 score for the interacting class (AUROC 0.963). Compared with SpatialPPIv2 on the same test data, PeTriPPI2 has higher precision (0.917 versus 0.886) but lower recall (0.876 versus 0.942) and F1 score (0.896 versus 0.913). Component ablations indicate that both the sequence and structural pathways contribute to this configuration.

In this work, we focus on predicting fold-switching phenomena observed in certain protein structures. Proteins that undergo fold-switching are called metamorphic proteins. We model fold-switch prediction in proteins as a position/residue-wise classification task. We use various features extracted from a protein's sequence as well as structure to train the classifiers. We propose a novel class of handcrafted parametrised features that encode the structural context of a residue. We also use embeddings from existing protein language models (PLMs) to train the classifiers. We train binary and 3-class classifiers, with the 3-class classifiers providing granular information on the type of fold-switch. In our computations, higher classification performance is achieved for binary prediction as compared to 3-class granular prediction. The handcrafted features perform comparably with the PLM embeddings on our standard test set. A rigorous error analysis by us during the review period revealed a critical data-leakage vulnerability in the metamorphic protein test data. This analysis also revealed that the PLM-based classifier suffers sequence-memorisation bias and fails to generalise to novel fold-switching sequences. On the other hand, one of our domain-informed handcrafted feature based classifier detects fold-switching without relying on sequence memory.

Lenalidomide (Revlimid), bortezomib (Velcade), and dexamethasone, are used alone or in combination (RVd therapy) as first-line therapies for hematologic malignancies including the plasma cell dyscrasia, Multiple Myeloma (MM). The effects of RVd treatment on tumor cells have been documented, but little is known about their impacts on "healthy" immune cells. New therapeutics, including chimeric antigen receptor T cell (CAR-T) and bispecific T-cell engagers (BiTEs) depend on robust T cell function but these are often harvested for cellular therapies after or in tandem with RVd therapy. Understanding the molecular effects of these drugs on healthy T cells is therefore of particular importance. Since the molecular effects of each drug includes regulation of key transcription factors, we used a tri-modal, single-cell assay (TEA-seq), simultaneously profiling mRNA transcripts, cell surface proteins, and chromatin accessibility in primary human peripheral blood T cells treated for 4, 24, or 72 hours in vitro. Synergies and conflicts potentially arising from combinatorial therapy were identified and validated by flow cytometry. Doublet and triplet drug combinations were used to assess effects on cell-surface marker expression and on T cell activation. Our results suggest that T cell function can be optimized by the administration of these drugs individually and that sequential administration may offer an opportunity to enhance T cell function by affecting localization and altering T cell states. This work provides those interested in immunology, oncology, and cell therapies an opportunity to explore molecular pathways that could be manipulated by these commonly used drugs to modify T cell function.

Protein language models (PLMs) have transformed our ability to learn from evolutionary sequence space, but protein engineering ultimately asks a different question: not what evolution selected, but what we should build next. Zero-shot likelihoods therefore provide useful, but not universal, measures of fitness and can misalign with engineering objectives. Experimental supervision redirects these priors towards properties of interest, enabling target-specific prediction and closed-loop optimization. PLMs thereby complement structure-based design: structural methods provide geometric control, while PLMs integrate experimental feedback to optimize functional and developability properties. Realizing this potential requires evaluation beyond retrospective predictive accuracy, towards extrapolation, multi-objective optimization, and prospective experimental success. We argue for recurring competitions combining scalable predictive benchmarks with prospective experimental challenges. Ultimately, progress should be measured not by predicting existing experiments, but by enabling successful new ones.

Protein inverse folding, which infers compatible amino acid sequences from a given backbone structure, is central to rational protein design. Current methods rely largely on sequence recovery rate as the evaluation metric, emphasizing structural compatibility but seldom addressing function. Yet functional activity depends on the proper network of non-covalent interactions-protein-ligand contacts being a prime example. We argue that function-oriented protein design is governed by three complementary constraints: structural, interaction, and evolutionary, none of which are handled in a unified manner by existing pretrained models. To bridge this gap, we introduce zPolaron, a graph neural network built on protein-ligand complexes for functional protein design. zPolaron represents residues and ligand atoms as a heterogeneous graph and jointly encodes structural, interaction, and evolutionary information by fusing physical-interaction-enhanced graph representations with ESM-2 embeddings. Across multiple fitness prediction benchmarks, zPolaron outperforms all baselines, with gains evident across structural, interaction, and evolutionary dimensions. Wet-lab experiments further confirm its utility in functional enzyme screening and design: the optimized CalA6-D205A-S160G mutant achieves a 17.84-fold improvement in activity, while the CalB-I121M mutant yields a 1.91-fold increase. Together, these results show that integrating interaction and evolutionary information fills the missing puzzle of inverse folding, enhancing protein fitness prediction beyond structural compatibility. zPolaron is available at https://github.com/zelixirSH/zPolaron.

We describe the concept for a new modality to create multi-specific therapeutics that utilizes modified nucleic acids as a scaffold to concatenate multiple warheads (hence called 'catamers'). Catamers are assembled via Watson-Crick interactions into multi-specific molecules from individual components such as ligand-binding modules, protease-cleavage sites, and toxins. Each warhead is covalently bound to an oligonucleotide, which in turn is connected to the catamer scaffold. Post-synthesis of the individual catamer components, they are annealed together such that Watson-Crick base pairs on the oligonucleotides dictate the organization of the warheads on the complete catamer. Notably, catamers are not aptamers, though they can contain aptamers as warheads, along with small molecules and peptides. By design, catamers are intended to address three key problems that have hamstrung the field of multi-specifics: (i) optimization rate; (ii) manufacturing cost; and (iii) unpredictable immunogenicity (Amash et al., 2024 PMID: 39189686). Catamers are purely synthetic and do not require cultured cells to produce. These advantages are based on the underlying design of the catamer and will require experimental validation in future studies. As a proof-of-concept catamer, we focus on a prostate-specific membrane antigen (PSMA)-targeted toxin conjugate. In molecular simulations, this catamer displays structural characteristics to support its intended pharmacology. Importantly, the design addresses one of the key shortcomings of antibody drug-conjugates (ADCs): off-target toxicity caused by release of the toxin after pinocytosis by healthy cells in contact with the blood.

Structure-guided mutation design requires relating a proposed substitution to its molecular environment to anticipate its effects. Existing evaluations of large language models (LLMs) in protein engineering emphasize sequence-based prediction or tool-integrated workflows, leaving unclear whether LLMs themselves can infer mutation effects from structural evidence without dedicated predictors. We evaluate physical intuition-qualitative reasoning that connects structural observations with molecular interaction principles-using a benchmark that tests the same substitution in the same protein across different binding partners, with predictions based on local atomic coordinates and no dedicated mutation-effect predictors. In the tested comparisons, the best-performing LLM achieved group accuracy comparable to the evaluated energy predictors and Biomni, an LLM-driven tool workflow, while the methods differed in their prediction patterns. Most evaluated LLMs achieved higher group accuracy for larger affinity improvements. Structural interventions support selective dependence on mutation-site and partner information, while assessable structural claims largely agree with the supplied coordinates. These findings support direct LLM structural reasoning as a basis for developing rational mutation-design tools that connect molecular observations with partner-specific effects.