Improved De Novo Peptide Binder Design with Target-Conditioned Inverse Folding

The de novo design of artificial enzymes with customized catalytic functions represents a long-standing challenge in synthetic biology. Recent breakthroughs in deep learning, particularly the rise of Generative Artificial Intelligence (GAI), have transformed enzyme design from structure-centric strategies toward function-oriented paradigms. This review outlines the emerging computational frameworks that now span the entire design pipeline, including active site design, backbone generation, inverse folding, and virtual screening. Detailed description of active site, called a theozyme, is designed to stabilize transition states and can be guided by density functional theory (DFT) calculations that define the geometry of key catalytic components. Guided by the theozyme, GAI approaches such as diffusion and flow-matching models enable the generation of protein backbones pre-configured for catalysis. Inverse folding methods, exemplified by ProteinMPNN and LigandMPNN, further incorporate atomic-level constraints to optimize sequence–function compatibility. To assess and optimize catalytic performance, virtual screening platforms such as PLACER allow evaluation of protein–ligand conformational dynamics under catalytically relevant conditions. Through representative case studies, we illustrate how GAI-driven frameworks facilitate the rational creation of artificial enzymes with architectures distinct from natural homologs, thereby enabling catalytic activities not observed in nature. With the rapid progress and widespread adoption of GAI, we anticipate that de novo enzyme design with customized catalytic functions will soon evolve into a mature and broadly applicable methodology.

NEO-TRA1: A CD25-Targeted De Novo non-Alpha Agonist of the IL-2 Receptor Selectively Expands Regulatory T Cells

Inverse protein folding generates valid amino acid sequences that can fold into a desired protein structure, with recent deep learning advances showing strong potential and competitive performance. However, challenges remain, such as predicting elements with high structural uncertainty, including disordered regions. To tackle such low-confidence residue prediction, we propose a mask-prior-guided denoising diffusion (MapDiff) framework that accurately captures both structural information and residue interactions for inverse protein folding. MapDiff is a discrete diffusion probabilistic model that iteratively generates amino acid sequences with reduced noise, conditioned on a given protein backbone. To incorporate structural information and residue interactions, we have developed a graph-based denoising network with a mask-prior pretraining strategy. Moreover, in the generative process, we combine the denoising diffusion implicit model with Monte-Carlo dropout to reduce uncertainty. Evaluation on four challenging sequence design benchmarks shows that MapDiff substantially outperforms state-of-the-art methods. Furthermore, the in silico sequences generated by MapDiff closely resemble the physico-chemical and structural characteristics of native proteins across different protein families and architectures.

While processing of MHC class II antigens for presentation to helper T-cells is essential for normal immune response, it is also implicated in the pathogenesis of autoimmune disorders and hypersensitivity reactions. Sequence-based computational techniques for predicting HLA-DQ binding peptides have encountered limited success, with few prediction techniques developed using three-dimensional models. We describe a structure-based prediction model for modeling peptide-DQ3.2beta complexes. We have developed a rapid and accurate protocol for docking candidate peptides into the DQ3.2beta receptor and a scoring function to discriminate binders from the background. The scoring function was rigorously trained, tested and validated using experimentally verified DQ3.2beta binding and non-binding peptides obtained from biochemical and functional studies. Our model predicts DQ3.2beta binding peptides with high accuracy [area under the receiver operating characteristic (ROC) curve A(ROC) > 0.90], compared with experimental data. We investigated the binding patterns of DQ3.2beta peptides and illustrate that several registers exist within a candidate binding peptide. Further analysis reveals that peptides with multiple registers occur predominantly for high-affinity binders.

A fundamental and challenging task of computational protein studies is to design proteins of desired structures and functions on demand. Data‐driven approaches to protein design have been gaining tremendous momentum, with recent developments concentrated on protein sequence representation and generation by using deep learning language models, structure‐based sequence design or inverse protein folding, and the de novo generation of new protein backbones. Currently, design methods have been assessed mainly by several useful computational metrics. However, these metrics are still highly insufficient for predicting the performance of design methods in wet experiments. Nevertheless, some methods have been verified experimentally, which showed that proteins of novel sequences and structures can be designed with data‐driven models learned from natural proteins. Despite the progress, an important current limitation is the lack of accurate data‐driven approaches to model or design protein dynamics.This article is categorized under: Structure and Mechanism > Computational Biochemistry and Biophysics Data Science > Artificial Intelligence/Machine Learning

The complex of the murine class II histocompatibility molecules I-A(k) with high affinity binding peptides were resistant to denaturation when examined by SDS-polyacrylamide gel electrophoresis at various pH levels. In contrast, complexes made with low affinity binding peptides were highly sensitive to denaturation by SDS. This effect was more pronounced at low pH. Placing a photoactivatable probe at the amino terminus of the peptides resulted in their covalent linkage to soluble I-A(k) molecules. We found an inverse relationship between the capacity of peptides to form SDS-stable complexes with I-A(k) and their extent of covalent association with either the alpha or beta chain. The relationship held true for three different peptides in which the main anchor residues were changed so as to affect their binding affinity for I-A(k) molecules. Thus, high affinity peptides generate a complex in which the motion of their amino termini was restricted, whereas complexes of low affinity peptides are more flexible. In agreement with this observation, complexes of I-A(k) with high affinity peptides were highly resistant to proteolysis, in contrast to those formed with weakly binding peptides, which were more likely to be cleaved. Complexes with low affinity peptides generate a structure with enhanced flexibility as compared with complexes with high affinity peptides.

Artificial intelligencemethods for protein folding and design.

Inverse protein folding concerns the identification of an amino acid sequence that folds to a given structure. Sequence design problems attempt to avoid the apparent difficulty of inverse protein folding by defining an energy that can be minimized to find protein-like sequences. The authors evaluate the practical relevance of two sequence design problems by analyzing their computation complexity. They show that the canonical method of sequence design is intractable, and describe approximation algorithms for this problem. The authors also describe an efficient algorithm that exactly solves the grand canonical method. The analysis shows how sequence design problems can fail to reduce the difficulty of the inverse protein folding problem, and highlights the need to analyze these problems to evaluate their practical relevance.

The inverse protein folding problem is that of designing an amino acid sequence which has a prescribed native protein fold. This problem arises in drug design where a particular structure is necessary to ensure proper protein-protein interactions. Previously, tubular structures for a three-dimensional (3D) hexagonal prism lattice were introduced and their stability was formally proved for simple instances under the hydrophobic-polar (HP) model of Dill. In this article, we generalize the design of tubular structures to allow for much larger variety of designable structures by allowing branching of tubes. Our generalized design could be used to roughly approximate given 3D shapes in the considered lattice. Although the generalized tubular structures are not stable under the HP model, we can prove that a simple instance of generalized tubular structures is structurally stable (all native folds have the designed shape) under a refined version of the HP model, called the HPC model. We conjecture that there is a way to choose which hydrophobic monomers are cysteines in all generalized tubular structures such that the designed proteins are structurally stable under the HPC model.

The inverse protein folding problem on 2D and 3D lattices

Recapitulation and Design of Protein Binding Peptide Structures and Sequences

De Novo Design of Foldable Proteins with Smooth Folding Funnel: Automated Negative Design and Experimental Verification

Lipoprotein lipase (LPL) physically associates with lipoproteins and hydrolyzes triglycerides. To characterize the binding of LPL to lipoproteins, we studied the binding of low density lipoproteins (LDL), apolipoprotein (apo) B17, and various apoB-FLAG (DYKDDDDK octapeptide) chimeras to purified LPL. LDL bound to LPL with high affinity (K(d) values of 10(-12) m) similar to that observed for the binding of LDL to its receptors and 1D1, a monoclonal antibody to LDL, and was greater than its affinity for microsomal triglyceride transfer protein. LDL-LPL binding was sensitive to both salt and detergents, indicating the involvement of both hydrophobic and hydrophilic interactions. In contrast, the N-terminal 17% of apoB interacted with LPL mainly via ionic interactions. Binding of various apoB fusion peptides suggested that LPL bound to apoB at multiple sites within apoB17. Tetrahydrolipstatin, a potent enzyme activity inhibitor, had no effect on apoB-LPL binding, indicating that the enzyme activity was not required for apoB binding. LDL-LPL binding was inhibited by monoclonal antibodies that recognize amino acids 380-410 in the C-terminal region of LPL, a region also shown to interact with heparin and LDL receptor-related protein. The LDL-LPL binding was also inhibited by glycosaminoglycans (GAGs); heparin inhibited the interactions by approximately 50% and removal of trace amounts of heparin from LPL preparations increased LDL binding. Thus, we conclude that the high affinity binding between LPL and lipoproteins involves multiple ionic and hydrophobic interactions, does not require enzyme activity and is modulated by GAGs. It is proposed that LPL contains a surface exposed positively charged amino acid cluster that may be important for various physiological interactions of LPL with different biologically important molecules. Moreover, we postulate that by binding to this cluster, GAGs modulate the association between LDL and LPL and the in vivo metabolism of LPL.

We have measured the binding affinity for five HLA-A alleles: HLA-A1 (A*0101), A2.1 (A*0201), A3 (A*0301), A11 (A*1101), and A24 (A*2401); of a set of all possible nonamer peptides (n = 240) of human papillomavirus type 16 E6 and E7 proteins. High affinity binding peptides were identified for each of the alleles, thus allowing us to select several candidates for CTL-based vaccines. Moreover, this unbiased set of peptides allowed an evaluation of the predictive value of HLA motifs derived either from the analysis of sequencing of pools of naturally processed peptides or from the binding analysis of polyalanine nonameric peptides that differed in the amino acids (aa) present at the anchor positions. Whereas pool sequencing-derived motifs were present in only 27% of high affinity binders, the more expanded motif, based on analysis of different aa substitutions at the anchor positions, was present in 73% of high affinity binders. Furthermore, it was found that the presence of anchor residues in a peptide was in itself not sufficient to determine binding to MHC class I molecules, because the majority of motif-containing peptides failed to bind to the relevant MHC. Finally, specific HLA motifs were used to predict peptide binders of 8, 10, and 11 aa in length. Several high affinity binding peptides were identified for each of the various peptide lengths, indicating a significant size heterogeneity in peptides capable of high affinity binding to HLA-A molecules.

We have screened a variety of L-amino acid peptide libraries against the extracellular domain of the human thrombopoietin (HuTPO) receptor, c-Mpl. A large number of peptide ligands were recovered and categorized into two families. Peptides from each family compete with the binding of HuTPO and with the binding of peptides from the other familiy. Representative peptides were synthesized and found to activate the full-length HuTPO receptor expressed in Ba/F3 cells to promote proliferation. These peptide families show no apparent homology to the primary sequence of TPO. We have focused our optimization efforts on one of the peptides, a linear 14-mer (IEGPTLRQWLAARA) with an IC50 of 2 nM in a competition binding assay and an EC50 of 400 nM in the proliferation assay. In order to enhance the potency of the compound, we constructed dimeric peptides by linking the carboxy-termini of the 14-mers to a lysine branch. These molecules exhibited slightly higher affinity (0.5 nM) and greatly increased potency (0.1 nM). The EC50 of the dimeric peptide was equivalent to that of the 332 aa form of baculovirus-expressed recombinant HuTPO. As previously shown for the erythropoietin-mimetic peptides, the TPO-mimetic peptides probably activate the TPO receptor by binding and inducing receptor dimerization. This supposition is supported by the observation that covalent dimerization of the peptide enhances its potency by 4,000-fold over that of the monomer. The peptide dimer is also active in stimulating in vitro proliferation of progenitors and maturation of megakaryocytes from human bone marrow, and in promoting an increase in platelet count when administered to normal mice.