Methods for protein design
By using AlphaFold to predict protein structural characteristics and Rosetta to optimize amino acid sequences in protein design, combined with molecular dynamics simulation, the problems of insufficient consideration of structural information and low computational efficiency in the existing technology are solved, and efficient protein design and verification are achieved, improving Structural integrity and expression ability.
Patent Information
- Application Number
- CN202380033463.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-04-11
- Filing Date
- 2023-04-11
- Publication Date
- 2024-11-26
AI Technical Summary
Existing protein design methods are difficult to effectively consider structural information, especially in tasks involving tertiary structural constraints, and neural network design methods cannot guarantee the designability of the backbone network, resulting in structural prediction accuracy and computing efficiency problems.
Using a computer-implemented method, the design and verification of proteins are achieved by creating an amino acid sequence pool, using AlphaFold to predict protein structural characteristics, combined with Rosetta and molecular dynamics simulations, the amino acid sequence is optimized to meet the fitness function threshold, and the design and verification of proteins are achieved.
Achieve rapid prediction of new protein monomer structures, able to design proteins that bind to target proteins, and reveal the ability to conformational changes during complex formation, improve structural integrity and expression capabilities, and verify that the designed proteins are under physiological conditions. stability.
Smart Images

Figure CN119032397A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and apparatus for protein design. Background Art
[0002] De novo protein design is a long-standing fundamental goal of synthetic biology, but has been hampered by the difficulty in reliably predicting accurate, high-resolution protein structures from amino acid sequences. Recent advances in the accuracy of protein structure prediction methods, such as AlphaFold (AF), have facilitated proteome-scale structure prediction for monomeric proteins.
[0003] Proteins are the workhorses of most life processes at the cellular scale. Proteins have a wide variety of functions, including catalysis of biochemical reactions, mechanical involvement in cellular motility, and formation of subcellular structures. The central paradigm for understanding protein function is based on the observation that proteins fold into complex but specific three-dimensional structures that vary according to their amino acid sequence, assuming that this is their lowest energy state. Experimental structure determination has therefore been a major pursuit in biology for the past 50 years and has generated more than 10 5 Unambiguously solved structures of proteins have been deposited in the Protein Data Bank (PDB) [1]. These structures reveal diverse folding topologies and geometries of individual proteins classified into 41 architectures, 1390 folding topologies, and 51% of the structures previously extracted from the PDB have formed quaternary structures through oligomerization and complex formation.
[0004] One question is to what extent the amino acid sequence encodes protein structure, and whether protein structure can be predicted from the amino acid sequence. This question has driven a large number of different computational protein structure prediction methods over decades of effort [2]. The pursuit of improved methods is reflected in an independently evaluated biennial all-society competition (CASP) that ranks the accuracy of participants' methods based on determined but unpublished structures, resulting in steady improvements in prediction accuracy [3]. This question has recently culminated in the development of AlphaFold [4] (hereafter referred to as AF). AF is a neural network-based method that achieves atomic accuracy comparable to that of crystal structures. AF has been applied to large-scale proteomes of multiple species, resulting in a new database of predicted structures [5]. Recently, a community-wide evaluation revealed several important applications of AF, ranging from prediction of amino acid variant effects to cryo-EM model construction [6]. Although used for single-chain structure prediction, an unexpected consequence of the AF input protocol allows prediction of non-contiguous chains, enabling the prediction of protein complexes using existing trained networks [7]. This feature has also been applied to large-scale surveys of proteomes to reveal new core eukaryotic complexes [8]. These efforts culminated in the release of AlphaFold Multimer[9], a version of AF specifically trained for complex predictions.
[0005] Despite these advances, the protein folding problem remains much more complex than structure prediction alone. Proteins are not rigid structures, for example under physiological conditions. Proteins exhibit thermodynamic equilibrium between folded and unfolded forms. Many proteins also undergo conformational changes, with a fine balance between their states that they exploit to function
[10] . In addition, there exists a large number of intrinsically disordered proteins (IDPs) that do not exhibit a stable tertiary structure or that fold transiently depending on the environmental context
[10] . A notable biomedical example is the misfolding of amyloid-α helices into β-sheet filaments associated with Alzheimer's disease
[11] . Deformed proteins have also been discovered. These deformed proteins form multiple stable but distinct folds
[12] .
[0006] From a biophysical perspective, protein folding can be understood using the concept of a folding energy funnel within atomic configuration space. Multiple energy minima then correspond to alternating stable states of the protein that transition at characterizable rates. Based on this concept, computational physics methods such as molecular dynamics (MD) simulations provide a way to characterize the dynamics, thermodynamics, and kinetics of conformational transitions
[13] , ab initio folding, disorder transitions, and protein-protein and protein-ligand
[14] binding.
[0007] Despite the complex structures and dynamics of natural proteins, evolution has still only explored a small fraction of the potential protein sequence landscape
[15] . Therefore, there is great potential to unlock the fundamental biophysical principles of protein folding to design and build novel proteins that can exploit this vast space. Recent examples include protein logic gates
[16] , self-assembling systems
[17] , and targeted therapeutics
[18] .
[0008] Until recently, established approaches to protein engineering have focused on tuning naturally occurring proteins through iterative experimental selection processes such as directed evolution
[19] . More recently, computational design methods have enabled de novo protein design, including a full range of functionalities from topological selection rules, protein backbone construction, amino acid sequence optimization, and combinations of these approaches
[20] .
[0009] One state-of-the-art approach in computational protein design is embodied in the Rosetta suite of protein design tools
[21] and Rosetta Remodel
[22] . The Rosetta suite combines tools for all steps from selecting the desired protein topology to designing and validating amino acid folding protein sequences. Amino acid sequence design of proteins in Rosetta Remodel consists of three tasks: topology specification, backbone generation, and fixed backbone design
[22] . After specifying the desired protein topology, the backbone structure of the protein can be generated using matching fragments extracted from existing proteins. Fragments matching the desired secondary structure and set of contacts are selected from the PDB structure. These selected fragments are then sampled using Markov chain Monte Carlo to obtain a pool of candidate backbone geometries. The candidate backbone structures can then be equipped with sequences using a fixed backbone protein design, with the backbone optionally optimized between design steps
[23] . Given a desired backbone geometry, the goal is to find an amino acid sequence that will fold that protein structure. Here, the Rosetta design protocol
[23] first populates the desired backbone with an all-valine sequence and then runs Markov chain Monte Carlo to obtain a low-energy sequence. The general strategy of Rosetta-based protein design has been successfully applied to a variety of design problems. These applications range from the first de novo designed proteins
[23] to synthetic vaccines
[24] , complex assemblies
[25] , and enzymes
[26] . However, Markov Chain Monte Carlo in structure space can be time-consuming and computationally demanding.
[0010] To address this shortcoming of classical protein design, there have been approaches to apply neural networks to various design problems. Some work trains generative models to directly generate protein sequences with desired functions. For example,
[27] trained a language model on the UniProt sequence database to generate sequences with specific functions, and fine-tuned the generative model for specific protein families or functions. In addition to language models, other types of generative models, such as variational autoencoders
[28] and generative adversarial networks
[29] , have also been used to directly generate protein sequences. Although these methods have been successful in generating protein sequences associated with given functions, these methods do not explicitly consider structural information and therefore cannot be applied to protein design tasks involving tertiary structure constraints.
[0011] At the other end of the spectrum of neural network-based protein design, generative models for protein structure generation have been explored
[30] , and generative adversarial networks have been trained to generate realistic backbone distance maps and coordinates.
[31] The same goal was achieved using variational autoencoders. Both approaches have fine-grained control over the designed backbone structure through latent variable manipulation. These methods are viable alternatives to classical backbone design, but they do not guarantee the designability of their predictive backbones.
[0012] Bridging the gap between structure-only and sequence-only approaches,
[32] trained neural networks on the PDB database of protein structures to predict protein sequence given a fixed backbone structure. These methods rely on network components that take into account both protein backbone geometry and protein sequence, including using a local coordinate system for each residue to infer protein geometry.
[33] treat fixed-backbone network design as a constraint satisfaction problem and train their networks as constraint solvers. These methods require a neural network trained for a specific protein design task (fixed-backbone protein design) and are inherently not easily extendable to other design applications without retraining.
[0013] Prior art has also explored repurposing previously trained protein structure or function predictors as part of an in silico screening framework. In these approaches, neural networks are treated as scoring functions to assess the quality of protein designs
[34] . The designs are then refined using gradient-based
[35] , gradient-free
[36] , or neural network-based
[37] optimization.
[0014]
[34] were the first to use an optimization loop involving neural networks for structure prediction of de novo protein design. Since then, their method has been extended to fixed-backbone design
[35] and protein scaffold generation for protein motif stabilization, using Markov Chain Monte Carlo and gradient descent as optimization tools. All of these methods used trRosetta
[38] as a structure predictor.
[36] used a new version of AF as a structure predictor for fixed-backbone protein design using greedy optimization, with sequences initialized from the trained model. Summary of the invention
[0015] In a preferred embodiment, the present invention is a computer-implemented method for protein design.
[0016] A computer-implemented method for designing at least one protein, comprising creating at least one amino acid sequence to be tested, wherein a certain amino acid contained in the amino acid sequence to be tested is selected according to a probability distribution; predicting the structural properties of at least one protein according to the at least one amino acid sequence of the comparison; calculating the fitness function value of the at least one amino acid sequence to be tested based on the structural properties of the at least one protein; and selecting or deselecting the at least one amino acid sequence to be tested according to the value of the fitness function.
[0017] The method may further comprise changing at least one amino acid sequence to be tested by changing at least one amino acid contained in the amino acid sequence to be tested.
[0018] The method may further include repeatedly changing the at least one amino acid sequence until the value of the fitness function exceeds a preselected threshold.
[0019] Predicting the structural properties of at least one protein contained in the amino acid sequence to be tested may include predicting the positions of amino acids relative to each other and / or predicting the full atomic structure of at least one protein.
[0020] The method can optimize amino acid sequences (also called protein sequences) using a gradient-free or gradient-based optimizer. An example of a gradient-free optimizer is an evolutionary algorithm. The optimization produces a backbone (i.e., amino acids or polypeptides connected by peptide bonds, from which side chains branch out) and amino acid sequences that, by construction, are designable and have a high degree of confidence under AF. Amino acid sequences (also called protein sequences) can be automatically searched using an evolutionary algorithm. The alterations can include mutating and / or recombining at least one amino acid sequence to be tested according to the evolutionary algorithm.
[0021] Structural properties can be predicted using AlphaFold. AlphaFold (AF) [4] can be embedded into a design loop for predicting protein structure. For example, by searching for amino acid sequences with high confidence predictions, natural-like protein structures can be predicted, e.g., via AF. A flexible family of fitness functions was designed that encodes a variety of protein design tasks. This family of fitness functions was integrated into the design loop by extensive validation using Rosetta
[21] and / or molecular dynamics simulations.
[0022] The fitness function may be defined based on biological and / or physicochemical properties of at least one protein.
[0023] The fitness function includes one or more fitness function components that represent a biological and / or physicochemical property of at least one protein.
[0024] The method may also include inputting a known structural feature associated with at least one amino acid sequence or target protein and bindable to at least one protein into AlphaFold as a structural template.
[0025] The method is able to rapidly predict new protein monomers starting from random sequences. These new protein monomers can adopt multiple folding modes within the known protein space. The method can be used to design proteins that bind to a pre-specified target protein.
[0026] It has been found that protein monomers show extensive maintenance of structural integrity. The approach also revealed the ability to predict proteins that change conformation upon complex formation, such as the switch from α-helix to β-sheet during amyloid filament formation.
[0027] The method may also include redesigning the at least one amino acid sequence selected to be tested, making the redesigned at least one amino acid sequence more natural, and / or improving the solubility and / or expression ability of the at least one protein.
[0028] The method may also include statistically recovering at least one amino acid sequence to be tested from the predicted structural properties of the at least one protein. Based on the predicted protein structure, the recovered amino acid sequence may be calculated.
[0029] Computationally recovered amino acid sequences can further validate the methods of the present disclosure.
[0030] The method may also include re-predicting the structural properties of at least one protein based on the redesigned at least one amino acid sequence. Based on the restored amino acid sequence, a re-predicted protein sequence may be calculated. The calculation of the re-predicted protein structure can further validate the method of the present disclosure.
[0031] The method may further comprise computationally validating the selected amino acid sequence using molecular dynamics and / or Rosetta ab initio structure predictions.
[0032] The structural integrity of the predicted protein structures was verified and confirmed by standard ab initio folding and structural analysis methods as well as more extensive all-atom molecular dynamics simulations and analysis of the corresponding structural flexibility, intra-monomer and interfacial amino acid contacts.
[0033] The method may further comprise experimentally validating the selected amino acid sequence by determining the solubility and / or expressibility of at least one protein.
[0034] The at least one protein may comprise a monomer, a binding agent that binds to a target protein, a homodimer, a heterodimer, a trimer, a tetramer, a pentamer, an oligomer, or a protein complex. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] For a more complete understanding of the present invention and its advantages, reference is now made to the following description and accompanying drawings, in which:
[0036] Figure 1An overview of the method is shown. (A) The method according to the present disclosure includes defining a fitness function; generating a structure (candidate structure) by generating candidate amino acid sequences using evolutionary search; redesigning the amino acid sequence to make it closer to nature; re-predicting the structure of the redesigned amino acid sequence; filtering the redesigned amino acid sequence by likelihood, solubility and prediction confidence; codon optimizing the remaining amino acid sequence; and sorting the codon-optimized amino acid sequence for downstream experiments. (B) The method includes an optimization cycle. A pool of amino acid sequences (sequence pool) is maintained throughout the design process. AlphaFold predicts the full-atom structure (full-atom geometry) and confidence metrics (pLDDT, pAC) for each sequence. These are combined into a fitness function, which is optimized, such as maximized or minimized, in an optimization cycle. The amino acid sequence (hereinafter also referred to as protein sequence) and the fitness value of the fitness function associated with the amino acid sequence are fed to an optimizer, which updates the sequence pool, for example, through an evolutionary algorithm. If the fitness value associated with the selected amino acid sequence exceeds the required threshold, the amino acid sequence and structure are submitted for verification using Rosetta and molecular dynamics simulation. The fitness function may include one or more fitness function components. The fitness function can be any function of the amino acid sequence, the predicted all-atom structure, and the confidence metric (pLDDT, pAC). Examples of fitness functions include root mean square deviation (RMSD) relative to a reference structure (e.g., native structure, experimentally verified structure, decoy structure, etc.), one or more confidence metrics, or constraints on complex formation and / or conformational changes during complex formation. These and other examples of fitness function components used to construct fitness functions and mathematical expressions of these fitness function components are described in detail below.
[0037] Figure 2 AlphaFold predictions for protein complexes are shown. AlphaFold predictions for dimer (HIV protease 1DMP, de novo designed protein 4PWW), trimer (de novo designed coiled coil 1COI), tetramer (consensus TPR supercoil 2AVP), pentamer (phage lambda tail terminator 3FZB). Predictions (dark grey) aligned and superimposed onto the native structure (light grey). All predictions show low RMSD to the native structure.
[0038] Figure 3De novo monomer designs are shown. (AD) Ab initio computation and molecular dynamics validation of designed monomers ranging from 32 to 256 amino acids in length. The designed monomers (light grey) are overlaid with the lowest energy structure predicted by Rosetta ab initio computation (dark grey) or relaxations of their predicted all-atom structure (structure). Each monomer reports its RMSD relative to the predicted AlphaFold structure, as well as the value of the Rosetta all-atom energy function, also known as the “Rosetta score”, such as REF2015 or energy functions based on electrostatic interactions, van der Waals interactions, and / or statistical physics. The predicted alignment confidence (pAC) (prediction error) is shown for each designed monomer. For validation using Rosetta, the distribution of Rosetta scores and RMSD among AlphaFold structures is shown (Rosetta score distribution): for the first ten relaxed structures (relaxed) starting from ab initio predictions (dark grey) and AlphaFold structures (light grey); for all decoys (i.e. conformations with low free energy) from Rosetta ab initio predictions (ab initio), starting from extended conformations (light grey) and AlphaFold structures (dark grey). For molecular dynamics validation (MD metrics), the Boltzmann-weighted collective variable (CV) distribution of the number of intra-monomer amino acid contacts and backbone RMSF (root mean square fluctuation) of the protein in the 2D landscape relative to the average MD structure is shown (CV distribution). (E) TSNE (t-distributed stochastic neighbor embedding) of designed monomer structures of length 64 amino acids and above. Structures were divided into ten clusters using agglomerative clustering. (F) Representatives of each cluster, showing different structures designed using AlphaFold.
[0039] Figure 4(AC) Rosetta and molecular dynamics validation of designed homodimers of length 32-128 amino acids are shown for de novo dimer design. The designed homodimers are covered by their lowest energy structures, which are caused by Rosetta relaxation of their predicted all-atom structures (complex structures). Each relaxed dimer is reported to AlphaFold structure, Rosetta binding energy and packing statistics with its RMSD. The predicted alignment confidence (pAC) (prediction error) is shown for each designed dimer. For molecular dynamics validation (MD complex metrics), the total amino acid contacts (intra-monomer and interface) and the Boltzmann-weighted CV distribution of RMSF in 2D space of the complex are shown (CV distribution), as well as the distribution of individual monomer contacts and interface contacts. (DF) Rosetta and molecular dynamics validation of designed heterodimers of length 32-128 amino acids. The measurements shown are the same as in (AC). (G) Orthogonality of designed dimers. For two different pairs of designed dimers, the predicted structures of on-target and off-target complexes are shown. For inter-monomer pairs of amino acids, the prediction confidence (prediction error) for on-target dimers was consistently higher than that for off-target dimers.
[0040] Figure 5 De novo oligomer designs are shown. (AG) Rosetta and molecular dynamics validation of the designed trimers, tetramers, and pentamers with monomer lengths of 32-64 amino acids, and a hexamer with a monomer length of 32 amino acids. The designed oligomers are overlaid with the lowest energy structure of the Rosetta relaxation of their predicted all-atom structure (complex structure). Each relaxed oligomer reports its RMSD with respect to the AlphaFold structure, Rosetta binding energy, and packing statistics. The predicted alignment confidence (pAC) (prediction error) is shown for each designed oligomer. For molecular dynamics validation, the Boltzmann-weighted CV distribution of total amino acid contacts (intra-monomer and interface) and RMSF in 2D space of the complex are shown (CV distribution), as well as the distribution of individual monomer contacts and interface contacts. (H) Additional structures.
[0041] Figure 6Binder designs from scratch are shown. (A, B) Rosetta validation of designed binders for proteins previously designed from scratch. Relaxed binders bound to the target protein are covered by their predicted all-atom structures (complex structures). Each relaxed binder reports its RMSD with respect to the AlphaFold structure, Rosetta binding energy, and stacking statistics. The predicted alignment confidence (pAC) (prediction error) of each designed binder is shown. For molecular dynamics validation, the Boltzmann-weighted CV distribution (CV distribution) of the total amino acid contacts (inside the monomer and at the interface) and RMSF of the complex in 2D space are shown, as well as the distribution of single monomer contacts and interface contacts.
[0042] Figure 7 AlphaFold sequence space and conformational changes of de novo designed proteins are shown. (A) AlphaFold predictions of conformational changes as complex structures form. (Left) Predicted structures of monomeric clock protein KaiB (KaiB ground state, KaiBgs) and folding transition stabilized KaiB (KaiBfs), predicted in complex with KaiC (dark grey) and overlaid with native structures of KaiBgs (1R5P) and KaiBfs (5JYT), respectively. Both predictions show low RMSD relative to the corresponding native structures. Overlay of predicted KaiBfs (dark grey) with native KaiBgs (incorrect conformation, light grey) shows high RMSD. (Right) Predicted structures of monomeric amyloid-β and its pentamer (dark grey) with native monomer (1IYT) and oligomer (2MXU). AlphaFold predicts conformational changes of amyloid from α-helical to parallel β-sheet features. (B) Rosetta and molecular dynamics validation of designed oligomers of monomer length 32, showing amyloid-like predicted conformational changes. Designed oligomers (light grey) are overlaid with the lowest energy structure (dark grey) resulting from Rosetta relaxation of their predicted all-atom structure (complex structure). Each relaxed oligomer is reported with its RMSD to AlphaFold structure, Rosetta binding energy, and packing statistics. Monomers are compared to oligomer structures (monomer structures) to illustrate conformational changes upon oligomerization. The predicted alignment confidence (pAC) (prediction error) is shown for each designed oligomer and its constituent monomers. For molecular dynamics validation, the Boltzmann-weighted CV distribution of total amino acid contacts (intra-monomer and interface) and RMSF in 2D space for the complex are shown (CV distribution), as well as the distribution of individual monomer contacts and interface contacts. (C) Additional proteins designed using fitness functions that favor conformational changes upon complex formation.
[0043] Figure 8Validation of designed monomers using Rosetta and molecular dynamics is shown. (AD) Validation of de novo designed monomers of lengths 32, 64, 128, and 256 amino acids using Rosetta and molecular dynamics. (Structure). The lowest Rosetta energy structure is shown overlapped with the predicted AlphaFold structure, with RMSD and Rosetta all-atom score reported. (Prediction error) shows the predicted alignment confidence for each monomer. (Rosetta score distribution) shows the distribution of decoys (i.e., conformations with low free energy) from relaxations and ab initio predictions on Rosetta energies and RMSD to the predicted AlphaFold structure. (Relaxations) shows this distribution for the ten lowest energy ab initio decoys (dark grey) and ten relaxations of the AlphaFold predicted structure (light grey). (Ab initio) shows the distribution of decoys with Rosetta energies <0 REU sampled from the extended conformation (grey) and the AlphaFold structure (red). For molecular dynamics-based validation, (RMSD / F evolution) shows the time evolution of the RMSD (dark grey) and RMSF (light grey) for each monomer during a 100 ns explicit solvent molecular dynamics simulation. (Contact evolution) shows the time evolution of the number of intra-monomer contacts during the 100 ns simulation (grey). (CV distribution) shows the Boltzmann-weighted distribution in the 2D collective variable (CV) landscape of the backbone RMSF and the total contacts (intra-monomer and interface) for all snapshots during the last 50 ns of the simulation. € Additional randomly designed structures for monomers of length 64 and 128 amino acids.
[0044] Fig. 9Validation of dimers designed using Rosetta and molecular dynamics is shown. (A-F) Validation of de novo designed homo- and heterodimers with monomer lengths of 32, 64, and 128 amino acids using Rosetta and molecular dynamics. (Complex structure) shows the superposition of the AlphaFold predicted structure (light grey) with the 10 best relaxations (dark grey) using the Rosetta all-atom scoring function. The RMSD of the relaxed structure is reported for the structure, as well as the Rosetta binding energy dG and packing statistics packstat. (Monomer structure) shows the AlphaFold predicted structure of the monomer (dark grey) superimposed with the structure of the complex (light grey), and the RMSD of the alignment is reported. (Prediction error) shows the predicted alignment confidence for each AlphaFold complex prediction. For molecular dynamics-based validation, (RMSD / F evolution) shows the time evolution of the RMSD (dark grey) and RMSF (light grey) of the entire complex during a 100 ns explicit solvent molecular dynamics simulation. (Contact evolution) shows the time evolution of total intra-monomer contacts and interface contacts during the 100 ns simulation. (CV distribution) shows the Boltzmann-weighted distribution of the backbone RMSF and total contacts (intra-monomer and at the interface) of the whole complex in the 2D collective variable (CV) landscape for all snapshots in the last 50 ns of the simulation. (MD monomer metrics) reports the distribution of the RMSD (monomer RMSD), RMSF (monomer RM SF), monomer contacts and interface contacts for all monomers and interfaces in the complex in the last 50 ns of the simulation.
[0045] Fig.10Validation of designed oligomers using Rosetta and molecular dynamics is shown. (AD) Validation of de novo designed trimers, tetramers, pentamers, and hexamers with monomer lengths of 32 and 64 amino acids using Rosetta and molecular dynamics. (Complex structures) shows the superposition of the AlphaFold predicted structure (light grey) with the 10 best relaxations (dark grey) using the Rosetta all-atom scoring function. The structure reports its RMSD relative to the relaxed structure, as well as Rosetta binding energies and packing statistics packstat for individual monomers with the rest of the complex dG. (Monomer structures) shows the AlphaFold predicted structure of the monomer (dark grey) superimposed with the structure of the complex (light grey), and the RMSD of the alignment is reported. (Prediction error) shows the predicted alignment confidence for each AlphaFold complex prediction. For molecular dynamics-based validation, (RMSD / F evolution) shows the time evolution of the RMSD (dark grey) and RMSF (light grey) of the entire complex during a 100 ns explicit solvent molecular dynamics simulation. (Contact evolution) shows the time evolution of total intra-monomer and interface contacts during the 100 ns simulation. (CV distribution) shows the Boltzmann-weighted distribution of the backbone RMSF and total contacts (intra-monomer and interface) for the entire complex in the 2D collective variable (CV) landscape for all snapshots in the last 50 ns of simulation. (MD monomer metrics) reports the distribution of RMSD (monomer RMSD, blue), RMSF (monomer RMSF, green), monomer contacts (grey), and interface contacts (orange) for all monomers and interfaces in the complex during the last 50 ns of simulation.
[0046] Fig.11Validation of designed binding proteins using Rosetta and molecular dynamics is shown. (A, B) Validation of two previously designed de novo designed binders of length 64 amino acids for target proteins using Rosetta and molecular dynamics. (Target protein) shows the AlphaFold predicted structure of the target protein. (Complex structure) shows the AlphaFold predicted structure of the target protein in complex with the binder (grey) covering 10 best relaxations (color) using the Rosetta all-atom scoring function. The target protein is in blue and all designed binders are in red. The structure reports the RMSD of its relaxed structure, as well as the Rosetta binding energy dG and packing statistics packstat. (Monomer structure) shows the AlphaFold predicted structure of the monomer (dark grey) overlapped with the structure of the complex (light grey), and the RMSD of the alignment is reported. (Prediction error) shows the predicted alignment confidence for each AlphaFold complex prediction. For molecular dynamics-based validation, (RMSD / F evolution) shows the time evolution of the RMSD (dark grey) and RMSF (light grey) of the entire complex during a 100 ns explicit solvent molecular dynamics simulation. (Contact evolution) shows the time evolution of total intra-monomer and interface contacts during the 100 ns simulation. (CV distribution) shows the backbone RMSF of all snapshots in the last 50 ns simulation and the Boltzmann-weighted distribution of total contacts (intra-monomer and interface) for the entire complex in the 2D collective variable (CV) landscape. (MD monomer metrics) reports the distribution of RMSD (monomer RMSD), RMSF (monomer RMSF), monomer contacts, and interface contacts for all monomers and interfaces in the complex during the last 50 ns of the simulation.
[0047] Fig.12Validation of designed conformationally altered oligomers using Rosetta and molecular dynamics is shown. (AD) Validation of de novo designed trimers, tetramers, and pentamers with a monomer length of 32 amino acids using Rosetta and molecular dynamics. (Complex structures) The AlphaFold predicted structure (light grey) is shown superimposed with the best relaxation (dark) out of 10 relaxations using the Rosetta all-atom score function. The structure reports its RMSD relative to the relaxed structure, as well as Rosetta binding energies and packing statistics packstat for individual monomers with the rest of the complex dG. (Monomer structures) The AlphaFold predicted structure of the monomer (dark grey) is shown superimposed with the structure of the complex (light grey), and the RMSD of the alignment is reported. (Prediction error) shows the predicted alignment confidence of the AlphaFold prediction for each monomer (left) and complex (right). For molecular dynamics-based validation, (RMSD / F evolution) shows the time evolution of the RMSD (dark grey) and RMSF (light grey) for each complex during a 100 ns explicit solvent molecular dynamics simulation. (Contact evolution) shows the time evolution of total intra-monomer and interface contacts during the 100 ns simulation. (CV distribution) shows the backbone RMSF of all snapshots in the last 50 ns simulation and the Boltzmann-weighted distribution of total contacts (intra-monomer and interface) for the entire complex in the 2D collective variable (CV) landscape. (MD monomer metrics) reports the distribution of RMSD (monomer RMSD), RMSF (monomer RMSF), monomer contacts, and interface contacts for all monomers and interfaces in the complex during the last 50 ns of the simulation.
[0048] Fig.13 The forecast run is shown.
[0049] Fig.14(A) Snapshots of a monomer as the fitness function value (pAC) increases during optimization are shown. To increase the fitness function value (pAC) from 0.7 to 1.5, the predicted structure (top) changes. At the same time, the local confidence for each amino acid (pLDDT, top) and the predicted alignment confidence for each pair of amino acids (pAC, bottom) increase. (B) Snapshots of a heterodimer as the fitness function value (pAC) increases during optimization are shown. To increase the fitness function value from 0.8 to 1.5, the structure of the monomer changes, and the monomers are predicted to be closer. At the same time, the alignment confidence predicted per monomer and interface approaches 1.0, indicating that the designed monomers form heterodimers with increasing confidence. (C) Schematic representation of the fitness function components involved in monomer and dimer design. For monomer design, the predicted LDDT is maximized at each sequence position, the pAC is maximized for each pair of amino acids, and the average and maximum protein radii are constrained to ensure compactness. For dimer and oligomer design, pAC is additionally maximized for both the constituent monomers and the entire complex.
[0050] Fig.15 The results of toxin inhibition coupling (TIC) screening of blockers designed in silico according to the present invention are shown. (A) TIC screening of RcaT-Sen2 was performed using an IPTG-inducible high-copy plasmid (donor library) expressing 88 designed blockers. The array donor library (384 density format) was coupled to the E. coli BW25113 receptor carrying the arabinose-inducible RcaT-Sen2 plasmid. These dual-plasmid-carrying transconjugates (i.e., the above-mentioned E. coli receptor had one of the IPTG-inducible plasmids as well as the arabinose-inducible plasmid) were selected, and the colony opacity was extracted using the Iris tool
[58] to assess the suitability of the transconjugates. For each transconjugated strain, the opacity was normalized to the average opacity of the negative control expressing GFP in each plate (which does not affect toxin activity). Mean blocking scores (see x-axis) were calculated by dividing the mean normalized opacity of each strain in the experimental plates (n=2 bioplates) containing arabinose + IPTG (RcaT and blocker induced) by the mean normalized opacity of the control plates (n=2 bioplates) containing only IPTG (blocker induced). Mean blocking scores, neutral scores are shown in grey with SD (cutoff >1.5), and black scores are indicated with p<0.01 after a t-test comparing the mean blocking scores of the experimental and control plates. B Similar results as in (A), but the recipient E. coli BW25113 carried the RcaT-Eco9 plasmid. DETAILED DESCRIPTION
[0051] The method in this paper is a search problem for finding a sequence of a set of amino acids of a protein (amino acid sequence) where the mathematical fitness function exceeds a fixed threshold to fit the design goal of the protein. The mathematical function is a combination of the amino acid sequence that forms the protein, the predicted structure of the protein, and the AF confidence, such as pAE and pLDDT (explained later). The design goal of a protein can be as simple as correct protein folding or some biological function such as binding to a target protein, dimerization, self-assembly, etc. The method uses AF to take into account the properties of the amino acid protein sequence (also called "amino acid sequence") and all atomic protein structures of the amino acid protein sequence, which are integrated into the fitness function to provide a structural prediction of the protein as well as a measure of the prediction confidence.
[0052] Our approach was combined with ab initio structure prediction using Rosetta
[21] and validation with molecular dynamics simulations.
[0053] Figure 1 An outline of the method is shown. An optimization loop is established to iteratively mutate a sequence pool of amino acid sequences, score the sequence pool of amino acid sequences using the output of AF to simulate a fitness function, and subsequently update the sequence pool with the mutated sequences and scores of amino acids, as shown in Figure 1 A. In one aspect, an initial pool of amino acid sequences (or protein sequences of amino acids) is created by uniformly and randomly sampling an amino acid sequence of fixed length from 20 standard amino acids. In another aspect, a certain standard amino acid, such as cysteine, can be excluded from creating the initial pool of amino acid sequences to avoid designing proteins containing disulfide bonds. Alternatively, the initial pool can be initialized or created by sampling from a predefined probability distribution of amino acids or a predefined probability distribution of amino acid sequences, such as given by the statistics of natural amino acids or natural amino acid sequences. In one aspect, during the execution of the method disclosed herein, these predefined probability distributions can be adjusted, for example, based on the properties (such as scores) of the sequence pool of amino acid sequences or based on recent scientific discoveries.
[0054] Compared to the previous method using trRosetta, the AF program is used to provide structure prediction [4]. The optimization objective, i.e., fitness function, is defined as a mathematical function of the protein sequence, the all-atom structure, and the AF confidence. The fitness function can be composed of fitness function components, i.e., mathematical expressions. The fitness function composed of fitness function components can flexibly define the fitness function according to the structural properties and / or biological and / or physicochemical functions of the designed protein. The optimization objective represents various aspects of the protein design sought. The properties and / or biological and / or physicochemical functions may involve the structure of the designed protein, the binding properties of the designed protein, the switching between conformations of the designed protein, the structural changes depending on the environmental conditions of the designed protein, the amino acid distribution of the designed protein (e.g., human-like amino acid distribution for immune evasion), the amino acid sequence distribution of the designed protein (e.g., human-like amino acid sequence distribution), the protein function (e.g., enzymatic activity) including the predicted function predicted by the protein function predictor, the secondary structure of the designed protein, the oligomeric state of the designed protein, and the post-processing modification (PTM) of the designed protein (including the predicted PTM predicted by the PTM predictor).
[0055] The first step in predicting protein structure using AlphaFold is to search for similar protein sequences (similar amino acid sequences) to one or more input protein sequences (input amino acid sequences) in a large database of protein sequences (amino acid sequences). Protein sequences can be of known origin and / or metagenomic origin. Once a similar sequence in the input protein sequence is found, the newly discovered protein sequence is aligned with the input protein sequence. This alignment is achieved by trying to match as many amino acids as possible in the similar protein sequence with the corresponding amino acids in the input protein sequence. Unmatched amino acid sequences in the input protein sequence are assigned a gap "-". This produces a set of multiple sequence alignments. It is understandable that at this stage, the aligned similar protein sequences have no structural information associated with them. In a second optional step, the structures of those aligned protein sequences for which structural information is known can be input into AlphaFold as structural templates.
[0056] The AF program processes a single protein sequence (i.e., a single amino acid sequence) from the initial sequence pool (or the initial pool of amino acid sequences) together with the corresponding one in the multiple sequence alignment (as described above) as the input protein sequence of amino acids (i.e., the input amino acid sequence), and predicts the structure of the single protein sequence by constructing an alignment of the amino acids contained individually in the protein sequence using the multiple sequence alignment [7]. The multiple sequence alignment predicts which amino acids contained in the protein sequence are close to each other (see the discussion of predicted alignment error or pAE below). The method also returns the all-atom structure of the single protein sequence (or single amino acid sequence) and the confidence of the prediction (see Figure 1 A, AlphaFold output). The prediction confidence value is expressed as a combination of the predicted local distance difference test (pLDDT)
[39] and the predicted alignment error (pAE) [4]. The pLDDT test measures the local model quality, while pAE provides confidence for each amino acid pair in the protein sequence. The protein sequence is then optimized to maximize the output fitness function (L) (i.e., the protein sequence, the predicted all-atom structure, and the above confidence - the predicted local distance difference test (pLDDT) and the predicted alignment error (pAE)). This optimization can include minimizing the root mean square deviation (RMSD) of the protein reference structure and maximizing the AF prediction confidence. The RMSD is calculated as the difference between the predicted protein structure based on the known specific backbone structure of the protein. Since this method is applicable to complex structure prediction [6], the fitness function can also constrain the oligomeric state or conformational changes of the protein during binding (see Figure 1 B) In general, any function of protein structure, amino acid sequence, and prediction confidence can be used for optimization.
[0057] For the protein sequence of the amino acids in the sequence pool, the fitness function value of the amino acid protein sequence is calculated. The optimizer further recombines and mutates the protein sequence to explore the sequence space of the protein. "Mutation" refers to replacing one or more amino acids in the protein sequence with different amino acids sampled uniformly at random. "Recombination" means selecting two sequences from the sequence pool and randomly exchanging random fragments of the two sequences. The sequence pool is updated with the recombined and / or mutated sequences.
[0058] An evolutionary algorithm is used to update the sequence pool, for example by mutating and / or recombining protein sequences (amino acid sequences). The following
[40] was used. However, other evolutionary algorithms or other gradient-free or gradient-based optimizers may be substituted as desired.
[0059] The evolutionary algorithm is performed as follows. Using a random sequence pool of N sequences, the fitness function of the amino acid sequence is evaluated to obtain the fitness function value t associated with the amino acid sequence. Next, the sequence with the maximum fitness function value t is determined. maxIn the subsequent step, the fitness function value t≥t in the sequence pool is selected. max – Tolerance t max protein sequences for some suboptimal ranges of tolerance (tolerance t max ) until a certain number "M" of new amino acid sequences (or protein sequences) are identified. The selected amino acid sequences (or protein sequences) are further recombined and mutated, and each recombined / mutated amino acid sequence (or protein sequence) is subjected to a so-called greedy improvement. The term "greedy improvement" means that only those amino acid sequences (or protein sequences) that improve their fitness function value t compared to their unmutated starting amino acid sequence (or protein sequence) are selected. Finally, the top N amino acid sequences (or protein sequences) generated are selected and the algorithm is repeated.
[0060] Throughout the optimization process, the protein changes its structure and both the local confidence based on, for example, pLDDT and the global confidence based on, for example, pAE increase (e.g. Figure 1 C). For a given protein design task, a threshold is chosen above which the amino acid sequence (or amino acid sequences) of the protein is considered to be optimized. Amino acid sequences (or protein sequences) above the threshold are returned from the method, and a subset of the returned amino acid sequences is submitted for validation using, for example, molecular dynamics simulations and Rosetta ab initio structure prediction
[21] . Only a subset is submitted to reduce the time required to process the subset in the Rosetta program.
[0061] Fitness function
[0062] Using gradient-free methods, it is possible to use the protein sequence of amino acids (or amino acid sequences), the all-atom structure, and any function of AF confidence as a fitness function. This flexibility opens up many possibilities for designing amino acid sequences (or protein sequences and protein complexes with specific properties and / or biological and / or physicochemical functions). The fitness function can be specified to represent the goal of the design task.
[0063] A protein complex is a (possibly transient) association of multiple proteins. In the case of this approach, it simplifies to designing multiple protein sequences simultaneously, rather than a single protein sequence.
[0064] The method described herein focuses on the fitness function using the above confidence model (or metric) and protein backbone geometry (or backbone structure). However, this does not limit the present invention, and other fitness functions may also be used.
[0065] The confidence model returns two main confidence measures (or confidence metrics) for protein structure predictions. As mentioned above, these confidence measures are the predicted local distance difference test (pLDDT)
[39] and the predicted alignment error (pAE) [4]. pLDDT provides a local confidence for each amino acid in the protein sequence. pAE provides the prediction error for the position of an amino acid in the local coordinate system of other amino acids. For both confidence measures, the method predicts a binned distribution of values for the amino acids in the protein sequence. The pAE value is converted to the predicted alignment confidence (pAC), i.e., pAC = 1-μ pAE / pA max , where μ pAE is the average pAE, pAE max is the center of the highest pAE interval. Alternatively, the predicted template match score (pTM) can be used
[41] . Optimizing pLDDT and pAE (e.g., high pLDDT and low pAE) provides a prior distribution over the amino acid sequence, based on which a natural-like protein structure can be designed. A natural-like amino acid sequence is one whose amino acid sequence (i.e., or amino acid string) is greater than or equal to the minimum probability under the probability distribution observed in nature. This probability distribution observed in nature can be determined based on the sequences and structures stored in the PDB. Therefore, a natural-like amino acid sequence has
[0066] The fitness function may include a term that maximizes the value in the confidence level or confidence metric, which is given by:
[0067]
[0068] where X = (x, pAE, pLDDT) is a tuple containing the pAE and pLDDT values of the all-atom structure x and the protein sequence, μ pLDDT is the average value of pLDDT, μ pAE is the average value of pAE.
[0069] AF returns the all-atom coordinates of the protein [4], which enables the introduction of geometric (or structural) features into the fitness function in another aspect of the method. Fitness function components (i.e., mathematical expressions) can be introduced into the fitness function, which depends on any geometric features of the protein sequence and its predicted structure. These geometric features are any geometric features of the designed protein that need to be optimized. For example, geometric features include but are not limited to:
[0070] Any weighted sum / product of AlphaFold confidence
[0071] Protein secondary structure determined from predicted 3D structure
[0072] Protein radius of gyration
[0073] Differences between protein structure and some reference protein structures
[0074] Distribution of amino acids in protein sequences
[0075] Predict protein solubility based on protein sequence and structure
[0076] Protein function, such as physicochemical and / or biological function, is predicted by another model based on protein sequence and structure
[0077] For protein complexes: the relative position of the center of mass of each constituent monomer
[0078] • Or any weighted sum / product of these geometric features.
[0079] The main building block that introduces fitness function components into the fitness function is the difference metric between the two sets of coordinates. The following protein structure metrics used in the fitness function are:
[0080] The alignment error [4] is given by:
[0081] AE ij (x, y) = || T xi x i -T yj y i ||
[0082] Where T x,I is to move the vector to the backbone atom x i The coordinate transformation of the local coordinate system at . This is the alignment error used in AF training [4].
[0083] The frame alignment point error (FAPE) is given by [4]:
[0084]
[0085] where c is the cutoff for maximum error, N is the number of amino acids in the protein sequence and clip is a function that cuts off the values at the maximum cutoff. In other words, clip(AE, c) sets all values in the alignment error greater than value c to value c. Frame Alignment Point Error is the main loss function used to train AF. FAPE is a function that computes the difference between two protein structures, invariant to global rotation and translation, but invariant to reflection. The smaller the FAPE value between two protein structures, the more similar these structures are. FAPE is zero if and only if the compared protein structures are equal under global rotation and translation. AF is trained using true structures determined by X-ray crystallography.
[0086] The distance RMSD is given by
[0087] dRMSD(x,y)=RMSD(d(x),d(y))
[0088] The function d(x) calculates the distance between each pair of α-carbons in the amino acid backbone, and is calculated as follows. Given a protein structure, take the 3D coordinates "x" of its backbone α-carbon atoms (i.e., the atoms in the amino acid to which the side chains are attached). The function d(x) calculates the distance between each pair of α-carbons. In fact, this function provides enough information to recover the 3D coordinates of all α-carbon atoms until global translation, rotation, and reflection, which is why the pairwise distances can represent the actual 3D coordinates in the fitness function. If the pairwise distances of two structures are the same, then the actual 3D coordinates are the same before global translation, rotation, and reflection of the protein sequence.
[0089] The template matching score
[41] is a measure of the structural similarity between protein structures:
[0090]
[0091] in and
[0092] L TM (x, y) is an approximation of the template matching score and does not require structural alignment [4]. The template matching score metric is normalized to 1 and provides a smooth fitness function for highly different structures.
[0093] In addition, the compactness fitness function L can be used comp to enforce a general shape constraint on the protein, which penalizes large protein radii:
[0094]
[0095] where μ x Represents the centroid of the coordinates.
[0096] Monomer. To generate globular monomeric proteins, the confidence fitness function (or confidence metric) is combined with a weighted function for compactness fitness as follows:
[0097]
[0098] In the case of elongated structures, this trades compactness for confidence. However, it should be appreciated that this approach can be used for other protein designs and is not limited to globular monomeric proteins. The fitness function for assessing protein globularity is relatively simple. For example, it is not so simple to determine whether a portion of a protein is a transmembrane domain. Given a fitness function that determines this from the protein sequence and / or structure, this approach can be used to design other types of proteins.
[0099] Protein complex. For the design of protein complex design, the same fitness function L mon for the constituent monomers in a protein complex. This ensures that each monomer has a stable, high-confidence structure. Previous work has identified inter-monomer pAE as a predictor of complex formation
[19] . Low inter-monomer pAE (or equivalently, high inter-monomer pAC) corresponds to high confidence in complex predictions. As described above, pAE is the prediction error for the distance and relative orientation of each pair of amino acids in a protein sequence. Assuming that the predictions are calibrated (i.e., the uncertainty in the predictions corresponds well to the true uncertainty), low prediction error pAE in inter-monomer residue pairs corresponds to highly constrained distances and relative orientations between monomers. This is only the case when the proteins in the protein complex interact. For non-interacting proteins, the uncertainty in distances / relative orientations is high because there are no constraints on the distances / orientations between the proteins.
[0100] Therefore, another combination of confidence fitness function and compactness fitness function is imposed on the complex protein, for example:
[0101]
[0102] Where X represents the prediction of the all-atom structure of the complex protein, X m Represent the full atomic structure of each monomer m.
[0103] To guide conformational changes when protein complexes form complexes, a conformational change fitness function L is defined: CC The fitness function L CC can be added to the fitness function L defined above cpx (X,X m )middle:
[0104]
[0105] where monomer(x,m) extracts the coordinates of monomer m from the structure of oligomer x, and maximizing this fitness function minimizes the TM score
[41] between the protein structure as part of a protein complex and its structure as a monomer.
[0106] Protein-protein interactions. The interface pAE (ipAE) is related to the probability of protein-protein interaction. Optimizing the low ipAE between two protein monomers allows the design of protein-protein interactions. In combination with additional confidence and geometric fitness function components that impose constraints (see Table 1 below), de novo monomers, oligomers (L denovo ) and protein binder design (L binder) fitness function. By designing the corresponding fitness function (L change ), one can search for sequences that AF predicts different monomeric and oligomeric states.
[0107] The fitness function component of the fitness function includes multiple confidence metrics (e.g., pAE, interface pAE, pLDDT). The fitness function component of the fitness function includes geometric constraints on the predicted protein structure (radius of gyration R g , residue pair X i and X j The maximum distance d(X i , X j ). The confidence fitness function component provides a prior distribution over the amino acid sequence, ensuring that the generated structural candidates are natural-like (L conf ) and exhibit the desired protein-protein interactions (L ipAE ). Geometric constraints constrain the predicted protein structure to conform to the designed topology and / or geometry. In addition to the components of the fitness function described above, other components are listed and described in the following table.
[0108] Table I
[0109]
[0110]
[0111] Input representation. Amino acid sequences (or protein sequences of amino acids) are represented as one-hot encoded arrays with 20 categories, one for each canonical amino acid. This allows the use of both gradient-free and gradient-based optimizers, since the gradient can be estimated from the one-hot representation using a straight-through estimator or similar methods
[42] . As input to AF, additional multiple sequence alignments are constructed using only a single input sequence, along with a blank template feature. A blank template feature is a blank template feature. Instead of passing one or more templates as input to AlphaFold, no templates are passed to AlphaFold at all. However, AlphaFold still requires an input that contains template information. The use of the term "blank" template feature refers to a template input that does not actually contain a template.
[0112] Sequence template. In order to reduce the size of the search space and achieve precise sequence constraints without introducing additional fitness functions, the optimized amino acid sequence (or protein sequence of amino acids) is represented as a pair (S, T), where S:R V is a unique representation of all V variable amino acids, T:R V →R Nis a template mapping that assembles V variable amino acids into a final evaluated amino acid sequence (or protein sequence) of length N. For example, to design a homodimer with a monomer size of 64, the actual number of variable amino acids is V = 64, and the template T simply converts the input sequence S:R 64 Connect with itself.
[0113] Complex representation. AF was originally designed to accept only a single protein sequence as input. To enable AF to predict the structure of protein complexes, the principles proposed in [7] were followed. In this case, chain breaks for predicting complex protein structures are represented by introducing an increment greater than 32 in the residue index of each additional chain beyond the first. This increment is used to separate residues on different chains by more than 32, which is the maximum relative residue index difference that AF can embed alone [4]. In addition, following [7], multiple sequence alignment features are split so that each monomer has a separate copy of its sequence alignment feature with gaps at the sequence positions of other monomers.
[0114] This can be explained by doing a simulation using the sequence (pseudo fasta format):
[0115] Assume the sequence is as follows
[0116] >1
[0117] MERRYCHRISTMAS
[0118] >2
[0119] HAPPYNEWYEAR
[0120] >3
[0121] HAPPYEASTER
[0122] Let us assume that AlphaFold should predict the structure of its complex and the input to AF will have the following multiple sequence alignment (pseudo fasta format):
[0123] >1:2:3
[0124] MERRYCHRISTMASHAPPYNEWYEARHAPPYEASTER
[0125] >1
[0126] MERRYCHRISTMAS-----------------------
[0127] >2
[0128] ---------------HAPPYNEWYEAR----------
[0129] >3
[0130] --------------------------HAPPYEASTER
[0131] Where "-" represents a gap in the multiple sequence alignment (i.e., unmatched amino acids). That is, for a complex of N proteins, a multiple sequence alignment of N+1 sequences is provided. The first sequence is the sequence connecting all the individual protein sequences. The k+1th sequence is the sequence of the kth input protein, and the number of gaps ("-") on the left is equal to the total length of the 1st to k-1th proteins, and the number of gaps ("-") on the right is equal to the total length of the k+1th to the Nth proteins.
[0132] Design Research
[0133] Optimization. For optimization, an evolutionary strategy optimizer following
[40] was used as described above. The population size was set to 10. During the mutation process, the population size was expanded by two times. Protein sequences with amino acids that were suboptimal at most 10% (in other words, the value of the tolerance range was set to 0.1 as described above) were considered for mutation and recombination. Recombination was performed by crossover with a probability of 10% at each position in the amino acid sequence. Crossover in this case means taking two amino acid sequences and creating a recombinant sequence by initially taking amino acids from the first sequence. At each amino acid, the sequence was switched from the amino acid currently being taken to the other sequence with a probability of 10%.
[0134] For L mon , L cpx L>1.5 and including L cc The optimization is considered complete for sequences with L > 0.9 of the fitness function. Tables II and III contain further parameters of the performed optimization runs.
[0135] AlphaFold configuration. For optimization, AF was configured for single sequence use by disabling ensemble templates, additional MSA features and limiting the number of MSA features to the number of monomers modeled. The number of AF iterations (recycling steps) was used as a parameter for each optimization run (see Tables II and III below). For larger protein complexes, the number of iterations was reduced to two to speed up the computation. All experiments used the parameter set model_1_ptm.
[0136] Table II: De novo protein design runs and parameters.
[0137]
[0138] Table III Binder design runs and parameters.
[0139]
[0140] Rosetta Validation
[0141] The designed protein sequences were validated using the Rosetta protein design and structure prediction tool suite
[21] . Ab initio structure prediction based on fragment assembly
[43] was used as an independent baseline (i.e., independent method) for the designed protein sequences and protein structures.
[0142] Ab initio structure prediction was performed as follows. Protein secondary structure was predicted using PSIPRED
[44] and S4PRED
[45] to provide protein secondary architecture for fragment selection. 3-mer and 9-mer fragments were selected from the Rosetta fragment database based on secondary structure and protein sequence information. Alignment information was not used because the designed protein sequence did not have sufficient homology with the natural protein sequence. Ab initio structure prediction was performed by fragment assembly using the Abinitio Relax protocol. Two starting conformations were evaluated. The first starting conformation started from an extended conformation with all backbone torsion angles set to 180° (in practice, this is the default starting conformation for ab initio structure prediction using the Abinitio Relax protocol), and the second starting conformation started from the AF predicted structure.
[0143] It should be understood that for ab initio structure prediction, there is no a priori information about the protein structure. The prediction process needs to start with some initial conformation, which is a theoretically valid conformation for the amino acid chain. This initial conformation is not necessarily an extended conformation, but an extended conformation is convenient as an initialization because it is easy to set up. In this initial conformation, the protein backbone is also free of conflicts, rather than starting with a random conformation. However, it is important to note that at several steps in the ab initio structure prediction process, the structure will have changed to something that more closely resembles the actual protein structure.
[0144] For both runs for the two starting conformations, 32,000 decoys (conformations with low free energy) were generated. The decoys from both runs were merged and scored using the Rosetta all-atom energy function
[46] . A total of 10 decoys with the lowest energy were selected for relaxation. The decoys were relaxed using the relaxation protocol. In cases where ab initio structure prediction failed to find a minimum energy structure, the AF predicted structure was also relaxed to generate ten additional decoys. The decoys were re-scored using the same Rosetta scoring function to calculate directly comparable energies. It should be understood that ab initio structure prediction and relaxation use different scoring functions in Rosetta. Therefore, in order to be able to compare the structural quality between the two protocols, the output protein structures need to be scored again (re-scored) using the same scoring function. The lowest energy relaxed decoy was selected as the final predicted structure and its Rosetta score and RMSD relative to the AF predicted structure were evaluated.
[0145] Protein-protein interface analysis. Since Rosetta does not allow ab initio structure predictions for protein complexes, except in the case of symmetric homo-oligomers, the interfaces of the designed complexes were evaluated using the Interface Analyzer protocol
[47] . The protein structure was relaxed and the Rosetta binding energies were calculated by removing individual monomers from the complex and then repacking. As a further measure of interface quality, packing statistics were calculated using the PackStat protocol
[48] , which detects unfilled regions within the interface that are inaccessible to solvent.
[0146] Validation based on molecular dynamics simulation
[0147] The designed monomer subsets for protein sequences and higher-order protein complexes were further validated by performing all-atom molecular dynamics (MD) simulations in explicit solvent and analyzing the structural flexibility and the corresponding properties of intra / intra-monomer and interface contacts (for multimeric proteins and protein complexes). Simulations performed in explicit solvent mean that the protein (or protein complex) is "put into a box" where every water molecule and every salt ion is explicitly included. This terminology is in contrast to "implicit solvent" approaches, where water molecules and ions are not simulated explicitly but are modeled by modifying the force field.
[0148] Initial system construction. The standard AMBER force field (ff14sb) was used to describe all protein parameters
[49] . Each protein or protein complex was constructed using atomic TIP3P water and minimal The packing was solvated to form a cubic periodic box and then electroneutrized with an ion concentration of 0.15 M NaCl using standard ion parameters.
[0149] Simulation protocol A standardized minimization, equilibration and simulation protocol consisting of 11 stages was developed for a system of solvated protein structures and solvated protein complex structures. At a specific equilibration stage, a set of constraints (RS) was imposed on each system. These constraints consisted of restraining all heavy atoms (i.e. non-hydrogen atoms) of the protein. Subsequently, each system was minimized in four stages, each with 1500 steps (500 steepest descents + 1000 conjugate gradients), imposing constraints RS with different force constants in each consecutive stage: Stage 1: 10 kcal / molA 2 , second stage: 5kcal / molA 2 , third stage: 1kcal / molA 3 , Stage 4: Unconstrained. MD simulations were performed in all subsequent stages. The SHAKE algorithm was used for all atoms covalently bonded to hydrogen atoms. A time step of 2 fs was used. Long-range Coulomb interactions were handled using a GPU implementation of the particle mesh Ewald summation (PME) method
[50] . In the fifth stage, each system was subjected to RS (k = 10 kcal / molA) within 1 ns. 2 ) was heated from 10K to 300K. Then, a damping constant of γ = 5.0 ps was used. -1 The temperature was maintained at 300 K using a Langevin thermostat, and in stage 6 the system was equilibrated for 1 ns at constant volume, thus in the NVT ensemble (i.e., constant number of particles N, constant volume V, and constant temperature T). Subsequently, the pressure was maintained at 1 atm using a Berendsen barostat with a pressure relaxation time of τp = 1.0 ps, and the system was simulated in the NPT ensemble, with an RS of 100 ps for each subsequent stage: Stage 7: k = 10 kcal / molA 2 , Stage 8: k = 5 kcal / molA 2 , Stage 9: k = 1 kcal / molA 2 In the 10th stage: k = 0.5 kcal / molA 2 Finally, in stage 11, the RS constraints were removed and the systems were simulated in the NPT ensemble for 5 ns. After this, production simulations were performed for 100 ns for each system in the NPT ensemble under the same conditions as in stage 11. Coordinate snapshots of the production simulations were generated every 10 ps, and 10,000 snapshot trajectories were generated for each system.
[0150] Structural flexibility and contact analysis. Structural stability and flexibility were analyzed by calculating the root mean square deviation (RMSD) of the protein backbone atoms relative to the initial predicted protein structure (after aligning the protein backbone) and the root mean square fluctuation (RMSF) relative to the average protein structure in the production MD. In the case of protein complexes, the analysis was performed separately for the entire complex protein and for individual monomeric proteins. The number of side-chain-side-chain intra-monomer contacts was calculated for each snapshot of the production MD using a heavy atom distance threshold of . Similarly, the interfacial contacts were calculated for each interface in the simulated protein complex based on the same criteria. An overall picture of protein contacts was provided by summing the total intra-monomer and interfacial contacts.
[0151] Intra-monomer contacts are contacts within a monomer. These intra-monomer contacts provide an indication of the stability of the monomers. If the number of intra-monomer contacts decreases throughout the MD simulation, it indicates that the protein is unstable, i.e., the protein is tearing apart. Interface contacts represent inter-monomer contacts between each pair of monomers. These interface contacts indicate the stability of the protein-protein interaction. If the number of inter-monomer contacts decreases throughout the MD simulation, it indicates that the monomers are dissociating from each other.
[0152] Finally, for all simulated systems, the global RMSF and the Boltzmann-weighted distribution in the 2D collective variable space of the total contacts (total intra-monomer + total interface) in the last 50 ns of simulation were calculated (G = −k B T ln(ρ)), an approximate potential for the mean force (G) is determined, where ρ is the normalized frequency in the binned 2D landscape.
[0153] Structural Clustering
[0154] Geometricus was used to subdivide the designed protein sequences of 64 amino acids or more into a segment dictionary
[51] . The k-mer method with k = 16 and a cutoff value of The radius method collects fragments. These fragments are shorter chains, or local neighborhoods of amino acids. The fragments collected using the radius method contain all amino acids around a given amino acid, where at least one atom is within the cutoff radius. In practice, this will be the amino acids in the chain surrounding the given amino acid, as well as other amino acids that are close in the 3D structure but not necessarily in the amino acid sequence.
[0155] A set of feature representations was computed for all designed proteins and the dimensionality was reduced using non-negative matrix factorization (NMF) with 50 components
[52] . The computation was as follows. The number of fragments in each protein structure corresponding to each fragment cluster identified by Geometricus was counted. A vector of these counts was returned, which is the feature representation bag. The proteins were grouped into ten clusters using ward-linkage agglomerative clustering on the NMF components.
[0156] Results. We found that AlphaFold can predict the structure of protein complexes. As recently shown [7], AF trained on single-chain protein structures can be used for protein complex structure prediction. To determine whether AF is suitable for designing protein complexes using our approach, we first predicted the structures of small protein dimers, trimers, tetramers, and pentamers ( Figure 2 and Figure 8 ). AF predictions showed low root mean square deviations (RMSDs) from the native structures of the considered protein complexes. While this may be due to the fact that these structures are part of the PDB and therefore part of the AF training dataset [4], coupled with recent work on AF protein complex prediction [6], it provides some indication that AF may be suitable for designing protein complexes using our approach.
[0157] Single sequence optimization design of multiple monomers
[0158] As a first step in AF-based de novo protein design, amino acid sequences of lengths of 32, 64, 128, and 256 amino acids were performed to design monomeric globular proteins ( Figure 3 The designed sequences of the monomeric amino acids showed high confidence in the predicted alignment ( Figure 3 , prediction error), exceeding an average prediction alignment confidence of 0.82 and an average pLDDT of 0.83 (normalized to the [0, 1] range). This suggests that AF assigns high confidence to the predicted structures, as a pLDDT of 0.7 or higher corresponds to confident predictions of backbone structures [4, 5]. To validate the structures predicted by AF, fragment-based ab initio structure predictions were performed using Rosetta
[21] on a randomly selected subset of designs, starting from extended amino acid chains and AF structures (see Figure 8 ). For each monomer’s designed protein sequence, the 10 lowest energy Rosetta structures were extracted. These extracted structures of the monomers as well as the AF structure were extracted using the Rosetta force field
[46] and the lowest energy structure was selected. For most monomer structures, the AF structure (grey, Figure 3The RMSD between (AD), structure) and the lowest energy structure (red) is less than For protein structures whose distance exceeds this threshold ( Figure 3 , C: Result 8), the larger deviation seems to be due to the flexible α-helix, and the structure predicted by AF has a relatively low Rosetta energy.
[0159] To visualize the folding energy landscape of each designed monomer, the distribution of decoys was examined from the Rosetta energies
[46] and the RMSD of the ab initio folded to AF structures ( Figure 3 (AD), Rosetta score distribution). Notably, for monomers of size 32 and 64 amino acids, the relaxations of the predicted structures (blue) have comparable but higher energies than the best 10 decoys predicted ab initio (red). Figure 3 (A, B), Rosetta score distribution, relaxation). Strikingly, for monomers of 128 and 256 amino acids, the relaxed AF structures (blue) are much lower in energy compared to the ab initio decoys (red). Figure 3 (C, D)). This suggests that ab initio structure prediction is unable to find a reasonable minimum energy structure for these monomers. In fact, the decoy distributions for monomers of 32 and 64 amino acids are funnel-shaped, with a minimum at low energy and low RMSD characteristic of protein folding ( Figure 3 (A, B), Rosetta score distribution, ab initio calculation). In contrast, for sizes of 128 and 256 amino acids, the distribution is more scattered, with no clear minimum visible, indicating that ab initio predictions are unable to find a reasonable minimum ( Figure 3 (A, B), Distribution of Rosetta scores, ab initio calculations). However, it was observed that the relaxed AF structure exhibited very low Rosetta energy
[46] , increasing confidence that the designed structure is reasonable.
[0160] As a further validation step, the structural stability of the designed monomers was assessed by performing all-atom molecular dynamics (MD) simulations. Boltzmann-weighted frequency distributions in the 2D ensemble variable (CV) landscape consisting of backbone RMSF and intra-monomer contacts showed that most systems exhibited a well-defined unimodal distribution centered around low RMSF and a large number of contacts. This suggests that the vast majority of consensus samples in the production ensemble sample have a narrow range, consistent with the structural stability ( Figure 3 In addition, in most monomer systems, the time evolution of backbone RMSD relative to the initial AF predicted structure and backbone RMSF relative to the MD average structure are stable at and The following (see Figure 8). Similarly, for simulated monomers, the temporal evolution of intra-monomer contacts remains stable, indicating that the folded structure is maintained. As expected, the number of contacts increases with monomer size, ranging from 50-100, 100-200, 300-400, and 600-800 for monomers of 32, 64, 128, and 256 amino acids in length, respectively (see Figure 8 ).
[0161] To assess the diversity of structures generated by our method, monomers of size 64 amino acids or larger were clustered and representative structures of each cluster were extracted ( Figure 3 (E, F)). Use Geometrics
[51] to embed the structure, perform clustering, and visualize its distribution ( Figure 3 I). By visual inspection, cluster representatives included only α- (clusters 1, 2, 5, 6, 7), only β- (cluster 3), and mixed α-β proteins (clusters 4, 8, 9, 10) ( Figure 3 (F)).
[0162] Search sequence pairs to enable de novo dimer design
[0163] Next, the method was applied to the de novo design of protein homo- and heterodimers. Dimers with monomer sizes of 32, 64, and 128 amino acids were optimized. As with the previous monomer designs, the predicted dimers showed high confidence in the predicted alignment to the AF structure ( Figure 4 (AD), prediction error), average pAC>0.75, average pLDDT>0.83. By visual inspection, the designed interfaces showed a high degree of shape complementarity. Instead of using Rosetta, which is usually more difficult to obtain ab initio structure predictions for protein complexes, Rosetta binding energies between monomers were calculated, as well as interface packing statistics scores
[48] as measures of structural quality. This validation was applied to a subset of the designed dimers (see Fig. 9 The dimer was relaxed using the Rosetta force field
[46] and it was shown that all relaxed structures had lower RMSD than the AF predicted structures ( Figure 4 , (AF), complex structure). In addition, the stacking statistics scores of most dimers are greater than 0.6, which is consistent with the resolution of The packing statistics of the crystal structures are comparable [90,103]. Most structures show Rosetta binding energies better than -40REU, suggesting that the predicted interfaces are stable under Rosetta
[46] .
[0164] MD simulations of dimer and higher-order oligomeric systems allow the analysis of intra-monomer contacts and interfacial contacts between monomers. The global RMSF and the Boltzmann-weighted frequency distribution in CV space of the total contacts (intra-monomer and interfacial) mostly show unimodal distributions with low average RMSF The significant number of total contacts increases with the size of the complex ( Figure 4 , (AF), MD metrics). Many dimeric systems show stable time evolution of global RMSD and RMSF (see Fig. 9 ), although significant deviations were observed in a few dimer systems. Analysis of RMSD, RMSF, intra-monomer and interface contacts showed that monomers in either homodimeric or heterodimeric systems exhibited similar flexibility and number of intra-monomer contacts. A large number of interface contacts were present in each dimer system, although, as expected, these were fewer than the corresponding intra-monomer contacts.
[0165] For synthetic biology applications, orthogonality is a desirable property of dimers
[16] . That is, a pair of monomers designed for dimerization should bind only to its designed partner and not to monomers of other designed dimers. As an initial test of the orthogonality of dimers designed using this method, all monomer combinations for a pair of designed dimers with a monomer size of 32 amino acids were predicted ( Figure 4 (G)). AF predicted the on-target complex with high confidence ( Figure 4 , (G), on-target predictions). In contrast, for off-target complexes, i.e., monomeric complexes that were not designed for dimer formation, the average predicted alignment confidence for inter-monomer amino acid pairs dropped below 0.5. This indicates that AF cannot confidently predict these off-target combinations as dimers, providing preliminary evidence that dimers designed using the approach outlined in this paper can indeed exhibit orthogonality.
[0166] Multiple sequence optimization allows homo-oligomer design
[0167] To demonstrate the feasibility of protein complex design beyond dimers, homo-oligomer designs ranging from trimers to hexamers were performed. The monomer sizes of the trimers, tetramers, and pentamers were 32 and 64 amino acids, respectively, and the hexamer was 32 amino acids. The alignment confidence predicted by AF, Rosetta binding energies, packing statistics, and the behavior of a subset of oligomers under 100 ns molecular dynamics simulations were evaluated as described previously (see Fig.10 The designed structures showed high prediction confidence consistent with most structures at pAC and pLDDT>0.7 ( Figure 5 ).
[0168] (AG) prediction error), low RMSD of Rosetta relaxed structures Good binding energy <-40REU, stacking statistics >0.59, comparable to native protein complexes
[53] ( Figure 5 (AG, complex structure).
[0169] MD simulations of homo-oligomers show similar properties to dimers, with time evolution showing generally modest RMSD and RMSF and a large number of intra-monomer and interfacial contacts (see Fig.10 ). The Boltzmann-weighted distribution of RMSF in contact space again shows a unimodal distribution, with the mean RMSF remaining constant at The total contact scales with the size of the complex. Analysis of the contacts within individual monomers and at interfaces shows that almost all systems have significant contacts between each monomer and interface ( Figure 5 (AG)MD indicator).
[0170] Interestingly, inspection of the designed protein complexes revealed a portion of extended oligomers with exposed interfaces on both sides ( Figure 5 (GH)). The designed amino acid sequences are likely to oligomerize into larger assemblies. Correspondingly, MD simulations of these complexes revealed that, as expected, all monomers maintained intra-monomer contacts and all but one interface maintained sequential interfacial contacts. This suggests that, in addition to simple oligomers, large-scale assemblies in AF can be designed and validated.
[0171] Design using fixed monomers to find binders for target proteins
[0172] As a natural extension of using AF for dimer and oligomer design, AF has potential applications in biologics design and synthetic biology, where proteins that bind to a fixed target protein (e.g., a fixed monomer, also called a template) are designed. Figure 6 and 12 , two proteins of 64 amino acids in length were selected. These two proteins were previously designed as target proteins and exhibited significantly different folds ( Figure 6 (A, B) Target protein). While optimizing the amino acid sequence of the designed binding protein, the amino acid sequence of the target protein was fixed during the design process. The predicted structure of the designed binding protein showed high alignment confidence > 0.81, good Rosetta binding energy < -58REU, and packing statistics > 0.67, which are comparable to the natural protein interface
[48] ( Figure 6 (A, B)). This indicates that the designed binders form dimers with the target protein. By visual inspection, it was found that the binders designed for each target protein appeared to bind at the same interface, indicating a preference for binding sites during the design process.
[0173] MD simulations of two binders designed for each of the two target proteins showed a unimodal distribution in the 2D CV space of the global RMSF total contacts and an average RMSF The average number of total contacts is between 300 and 400, consisting of a large number of individual intra-monomer (100-200) and interfacial (40-120) contacts.
[0174] For the design of binders, AF provides a template amino acid sequence of the target protein, whose structure should remain fixed throughout the design process. The provision of the template allows the design of binders without the need for the MSA input of the target protein. In order to minimize computation, the amino acid sequence of the target protein is trimmed around the desired binding site. The trimmed amino acid sequence is added to AF as a template. These templates can fix the computational cost of designing one or more binders that bind to the target protein, regardless of the size of the target protein.
[0175] Sequence design reveals features of AlphaFold's conformational changes
[0176] To determine whether AF sequence space contains information about conformational changes of proteins upon interaction with other proteins, the structures of two proteins known to exist in different monomeric and oligomeric states were predicted: KaiB, a circadian clock protein
[54] , and amyloid β, a protein involved in Alzheimer's disease
[55] .
[0177] As a monomer, KaiB adopts its ground state structure KaiBgs
[54] . After complexing with KaiC, KaiB transforms into a folding transition stable state KaiBfs
[54] . The AF prediction of KaiB as a monomer is in good agreement with the native KaiBgs structure (RMSD ), while the predicted structure of the KaiB and KaiC complex is in good agreement with the native structure of KaiBfs (RM SD ). However, the structural alignment of the predicted KaiB and KaiC complexes showed that the structure of KaiBgs has a higher RMSD ( Figure 7 (A) ), indicating that AF prediction captures part of the conformational changes between KaiBgs and KaiBfs.
[0178] In the monomeric state, native amyloid-β forms an α-helical structure, which changes into a parallel β-sheet structure after aggregation ( Figure 7 (a) gray)
[105] . Although in RMSD Although not as accurate in terms of the α-helical monomer state and the parallel β-sheet oligomer state, AF captures the transition between the α-helical monomer state and the parallel β-sheet oligomer state very well (Fig. (A) red). This suggests that AF may have learned to predict conformational changes upon protein binding in a limited way.
[0179] In addition, a set of oligomer designs that exhibit conformational changes upon complex formation were identified ( Figure 7 (B) and 12 designed structures show the α-helix from the monomeric state ( Figure 7 (B), monomeric structure) to a stack of parallel β-sheets characteristic of amyloid
[55] . Under AF, both monomeric and oligomeric states show high confidence in the predicted alignment ( Figure 7 (B) Prediction error), the binding energy of the oligomeric state is <-88 REU, and the packing statistics of all oligomers are >0.65. This indicates that the complex is stable under the Rosetta force field
[86] .
[0180] Similarly, MD simulations show that the RMSF contact space is dominated by low RMSF and a large number of contacts (100-500) centered on a clear minimum. However, as mentioned before, these conformationally switched open oligomers do differ significantly compared to the other oligomers, including the open oligomers that maintain the conformation. While openness is captured by a large number of contacts in all but one of the continuous interfaces, the number of interfacial contacts (40-80) is similar to or greater than the number of contacts within the monomer (10-60). This confirms the elongated conformational structure of the monomer in the oligomeric state.
[0181] Heterodimers and homo-oligomers were designed that displayed conformational changes upon complex formation. Proteins that exhibited the desired conformational changes were found by maximizing the TM score between the monomeric and oligomeric states during the optimization process ( Figure 7 (C)). Interestingly, the resulting proteins were found to undergo conformational changes beyond the amyloid transition found in previously designed experiments, suggesting that a wider variety of conformationally changing proteins exist in AF sequence space.
[0182] Structure to sequence recovery
[0183] To generate improved sequences for structural candidates recovered from AF, a conditional autoregressive diffusion model (ADM) was trained on protein sequences and structures in the PDB. The conditional ADM can learn to recover the amino acid sequence of a protein given its predicted structure by optimizing the following loss:
[0184]
[0185] Among them, p θ represents the output distribution of protein sequences in the model, pdata represents data distribution, s represents protein sequence, x represents protein structure, and p mask (s′|s) represents the distribution of masked amino acid sequences based on the baseline true value sequence s. mask The distribution of is chosen to be uniform over the mask sequence. That is, the number of mask positions is the same as the number of mask positions, which are sampled uniformly at random. θ (s′|s, x) is parameterized as
[0186] Π i p θ (s i |s′,x)
[0187] The product of independent assortative distributions for each amino acid is a convention for modeling masked sequences. Our ADM architecture is based on the message passing block of ProteinMPNN (Protein Message Passing Neural Network). Our model consists of 6 MPNN (Message Passing Neural Network) blocks with 30 nearest neighbors, 128 node features, and 128 edge features. Before passing the protein backbone to MPNN, Gaussian noise is added to all atomic positions with mean μ = 0.0 and standard deviation In contrast to ProteinMPNN, node features are initialized by embedding the masked amino acid sequence, and the model is trained to directly predict the independent probability distribution of each masked amino acid in the masked amino acid sequence. This makes it easier to implement complex dependencies between amino acids in homo-oligomers or repeated proteins during sampling. The training set consists of structures from the PDB with a resolution higher than Deposited until January 2021. The validation and test sets come from structures deposited in the PDB before January 2022 with less than 10% sequence identity to the training set. The training set is clustered at 40% sequence identity. Structures are sampled from the training set with probability inversely proportional to the cluster size, without replacement. Structures are accumulated in batches until adding additional chains would make the batch size exceed 10k amino acids. Batched models are padded with 10k amino acids. To simplify implementation, the losses are averaged over all amino acids in a batch rather than per sequence. The model was trained for 250k steps using Adam (Kingma and Ba, 2014) with hyperparameters β1=0.9 and β2=0.98, with linear warmup to a final learning rate of 1e-3 in the first 10k steps.
[0188] In order to recover (or redesign) the sequence of the predicted structure, for example, sampled from AF and subject to constraints, the masked sequence for ADM inference is initialized. At all positions where the initial sequence is not specified to be kept (or fixed), the sequence is masked. For example, in the case of protein binder design, the sequence of the binder is completely masked, while the sequence of the target protein remains intact (or fixed). A random masked position of the binder sequence is sampled. For this random position, a probability distribution is calculated, where the probability of forbidden amino acids is set to zero. The probability distribution is rescaled by the inverse temperature of 10, and the amino acids at the random position are sampled. If the sampled random position is associated with one or more sequence constraints, amino acids are sampled or set at those positions affected by one or more sequence constraints. The process is repeated until there are no more masked amino acid residues. In this way, 100 sequences are sampled for each predicted protein structure. The sampled sequences are ranked according to the likelihood of each sampled sequence under the model. For all subsequent steps in the method, the sampled sequence with the highest likelihood can be used.
[0189] To ensure that the redesigned (or restored) amino acid sequence still folds as expected, the redesigned (or restored) sequence was re-predicted using AlphaFold for 12 cycles of iterations. For those re-predicted protein structures and discarded amino acid sequences, the average pLDDT and pTM were calculated if the product pLDDT·pTM<0.5. The RMSD between the initial prediction and the re-predicted protein structure and the CamSol solubility score of the protein structure were calculated (Sormanni et al., 2015). All remaining re-predicted protein structures were ranked according to pLDDT·pTM, RMSD, and CamSol solubility score. The top-ranked structures among the re-predicted protein structures were selected for experimental verification.
[0190] RcaT inhibitor design
[0191] The structure of RcaT-Sen2 was predicted using ColabFold, a method with 96 recycling steps and Early stopping. A crop of 100 amino acids was extracted around a putative active site as a template for fixed target design. AlphaDesign was run to generate predicted (or candidate) structures of 50 and 100 amino acids with fitness L binder (X|RcaT-Sen2; active site), 2 recycling steps, suboptimal 0.1, population size 10. Based on the predicted (candidate) protein structure, 100 amino acid sequences were generated using the autoregressive diffusion model while keeping the target amino acid sequence unchanged. The amino acid sequence with the highest probability was retained, while the predicted (candidate) protein structure was discarded. Protein structures were re-predicted using AF in 12 recycling steps, and re-predicted (candidate) protein structures with pLDDT·pTM<0.5 were discarded. For the re-predicted protein structures that were not discarded, the CamSol solubility score was calculated. From these re-predicted protein structures, protein structures with CamSol solubility score <1 were discarded. The remaining re-predicted (candidate) protein structures were codon-optimized for expression in Escherichia coli.
[0192] In the phenotypic experimental validation of the designed binders (see Fig.15 ), a total of 88 proteins (50 or 100 amino acids in length) were designed in silico to bind to the active site and block the activity of the growth inhibitory toxin RcaT-Sen2, which is part of the Retron-Sen2 toxin antitoxin system (Retron TA)
[56] . Of these, 16 were also designed to additionally bind to the active site of RcaT-Eco1, a different RcaT protein from Escherichia coli (21.5% identical to RcaT-Gen2 at the protein level). To experimentally validate the functionality, solubility and / or expression capacity of the designed blockers, the ability of the designed blockers to inhibit the growth inhibitory activity of RcaT-Sen2 was tested using a high-throughput reverse genetics screening method called toxin inhibition binding (TIC)
[56] . In TIC, a mobilizable gene donor plasmid is bound on agar plates to a large-scale receptor of E. coli transformed with the toxin plasmid. Typically, the strain expressing the toxin will not grow, but growth will be restored if the mobilizable plasmid encodes the toxin blocker.
[0193] The gene donor library was constructed by cloning each of the 88 blockers designed according to the present disclosure into an IPTG-inducible high-copy vector by Golden Gate assembly
[57] . The gene donor strain was arrayed as 384 colonies on agar plates and coupled to the E. coli BW25113 receptor transformed with an arabinose-inducible vector carrying RcaT-Gen2, and the transconjugates were then tested for their ability to grow when co-expressing the RcaT-Sen2 toxin with the designed blockers.
[0194] To assess the promiscuity of the designed blockers, the gene donor library was tested against RcaT-Eco9, a toxin that is 49% identical to RcaT-Gen2 from Escherichia coli
[56] . One quarter of the designed binders (22) restored bacterial growth in the presence of RcaT-Sen2 to varying degrees ( Figure 1A). Two of the popular RcaT binders, “cpx-50-nr2-run_5_0” and “50aa_1_55”, completely abolished the toxicity of RcaT-Sen2 and fully restored the growth of the transconjugate. Although primarily designed to target RcaT-Sen2, 15 constructs were also able to inhibit the toxicity of RcaT-Eco9 ( Figure 1 B). Although the two toxins are not identical, their active sites are structurally similar, so functional blockers may bind to both active sites (no blockers were found to inhibit only RcaT-Eco9). Notably, the top 5 hits for RcaT-Eco9 validated blockers were designed to bind to both RcaT-Sen2 and RcaT-Eco 1 active sites, which may explain their apparent promiscuity. In summary, proteins designed according to the present disclosure can inhibit the activity of RcaT toxins in vivo, supporting the use of this tool for in silico design of protein binders.
[0195] Computer protein structure prediction and experimental design
[0196] Complex formation and conformational change predictions were performed using Colabfold. Predictions were performed using five AF pre-trained parameter sets (model_1_ptm, model_2_ptm, model_3_ptm, model_4_ptm, and model_5_ptm), and models were selected based on the highest predicted template match score. Complex queries were predicted without paired MSA. Structural alignment of the resulting top models was performed using PyMol for RMSD reporting. Fig.13 The parameters of the prediction runs on proteins with PDB identifiers IR5P, 5JYT, 1IYT, 2MXU are summarized. For optimization, the evolutionary optimizer was used. The population size was set to 10. During the mutation process, the population size was expanded by a factor of 2. No more than 10% of suboptimal sequences were considered for mutation and recombination. Recombination was performed with a probability of 10% for crossover at each sequence position. For L mon , L cpx L>1.5 sequence and including L ccThe optimization was considered complete for sequences with L>0.9 of the fitness function included. Tables II and III (see above) contain further parameters of the performed optimization runs. AF was configured for single sequence use by disabling collections, templates, additional MSA features and limiting the number of MSA features to the number of modeled monomers. The number of AF iterations (recycling steps) was given as a parameter for each optimization run (see Tables II and III above). For larger protein complexes, the number of iterations was reduced to 2 to speed up the calculations. The AF pre-trained parameter set model_1_ptm (pre-trained using as many sequences from the MSA and structural templates as possible, called model_1, and fine-tuned to predict template matching scores) was used for all experiments. Finally, designed proteins of length 64 and above were subdivided into a dictionary of fragments using Geometricus. The k-mer method with k=16 and a cutoff of The fragments were collected using the radius method. A set of feature representations was calculated for all designed proteins and the dimensionality was reduced using non-negative matrix factorization (NMF) with 50 components. The proteins were separated into 10 clusters using a forward linkage agglomerative hierarchical clustering algorithm on the NMF components.
[0197] in conclusion
[0198] This paper discloses a de novo protein design method based on the optimization of amino acid protein sequences using an evolutionary algorithm. AlphaFold (AF) [4] has been embedded into the design loop as a prediction predictor. The output prediction confidences (or confidence metrics) of AF contribute to the optimization, namely the predicted local distance difference test (pLDDT)
[39] and the predicted alignment error (pAE)
[16] . A set of flexible fitness functions (or fitness function components, see Table I and the paragraphs preceding Table I) are designed to encode a variety of design tasks, as well as an extensible platform for the development of further fitness functions to solve customized design problems. The method includes steps for post-processing the designed protein sequences to make them more like natural proteins and to improve solubility and expression when such proteins are experimentally synthesized. This enables a range of applications, including the de novo design of protein monomers, dimers, oligomers, context-dependent conformational transitions, and binders targeting proteins.
[0199] The predicted protein structures were extensively validated using the Rosetta suite of protein design and structure prediction tools
[21] . Ab initio structure prediction based on fragment assembly
[43] was used as an independent baseline for the designed protein structures. In addition, a more rigorous validation protocol was developed using all-atom molecular dynamics (MD) simulations that go beyond the computational techniques traditionally used for structure prediction evaluation. MD simulations enable extensive exploration of the assumed native state; therefore, instability increases with increased overall structural flexibility, loss of internal contacts within protein monomers, and / or loss of interfacial contacts within complexes.
[0200] A subset of the designed protein binders were also experimentally validated using an established phenotypic high-throughput reverse genetics screening method called toxin inhibition binding (TIC)
[56] . For a set of 88 proteins designed to bind to the toxin protein RcaT-Sen2, 25% restored bacterial growth to some extent, while 2 completely abolished the activity of RcaT-Sen2. Some constructs even showed more extensive abrogation of the toxicity of a related but not identical toxin (RcaT-Eco9). This demonstrated first that the binders designed using this approach were sufficiently soluble and expressible in vivo to be used in such experiments, and further that the computational design approach produced significant binders that were able to completely inhibit the corresponding function of the target protein.
[0201] The approach presented here has been applied to design de novo monomeric proteins based on a fitness function starting from completely random amino acid sequences of length 32–256 amino acids, in which pLDDT and pAE were combined. This approach yielded a series of structurally stable de novo designed monomeric proteins with different folds. Using the function of AF to predict protein complexes, complexity predictions were made for a number of systems that showed good structural consistency. By further specifying functions based on globular compactness and complexity predictions, a range of stable de novo protein complexes could be designed, including homodimers, heterodimers, and homo-oligomers ranging from trimers to hexamers. Orthogonality between pairs of dimers could be demonstrated, so the approach outlined in this paper may be applicable to design mutually exclusive combinations, such as in protein logic gates
[16] . In addition, a number of open predicted complexes have been observed, providing a potential avenue for the design of self-assembling systems
[17] .
[0202] A particularly interesting subset of predicted protein structures exhibits striking conformational changes between monomeric and oligomeric states. Structural predictions of existing protein systems known to change conformation and / or fold between monomeric and oligomeric forms, including the amyloid-α-β switch, suggest that AFs inherently contain features of conformational change. Conformational transitions between monomeric and oligomeric forms have been observed in several de novo designed open oligomeric systems. By defining an objective function to maximize the structural differences between monomeric and oligomeric forms, this approach can redesign oligomers that exhibit such conformational switching properties. Context-dependent conformational transitions are a desirable feature in synthetic biology applications. For example, proteins can be designed to self-assemble inside rather than outside cells by designing membrane-permeable α-helices that spontaneously switch to membrane-impermeable β-sheet filaments when the accumulated intracellular concentration drives the equilibrium toward the oligomeric state.
[0203] This approach enables the selection of unmodified proteins in the design cycle, so by combining this selection with de novo designed proteins starting from random sequences of amino acids, it is possible to design monomeric binders for target proteins, which may be pre-specified. Applying this approach to a selected set of target proteins allows the design of stable binders that exhibit significant interfacial contacts at the same interface of a given target protein. This approach can be further optimized for therapeutic applications in potent biodesign.
[0204] Acknowledgements
[0205] The inventors would like to thank the German Federal Ministry of Education and Research (BMBF) (de.NBI project: 031A537B) and the Volkswagen Foundation (contract 95826) and the Volkswagen Foundation "Experiment! Funding Program" (grant number 93874-1) for their support. The inventors would like to thank Alessio Ling Jie Yang, Jacob Bobonis, Carlos Geert Pieter Voogdt and Athanasios Typas for conducting the experimental validation studies.
[0206] References
[0207] [1] Christine Zardecki, Chenghua Shao, Maria Voigt, and Stephen K. Burley. Protein data bank: 50 years of macromolecular structures enabling research and education. The FASEB Journal, 35, 2021.
[0208] [2] Brian Kuhlman and Philip Bradley. Advances in protein structure prediction and design. Nature Reviews Molecular Cell Biology, 20:681-697, 2019.
[0209] [3]Andriy Kryshtafovych, Torsten Schwede, Maya Topf, Krzysztof Fidelis, and John Moult. Critical assessment of methods of protein structure prediction (CA-P)-Round XIV. Proteins, 2021.
[0210] [4] John M Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvu-nakool, Russ Bates, Augustin Zídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A Kohl, Andy Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David A. Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W. Senior, Koray Kavukcuoglu, Pushmeet Kohli, and Demis Hassabis. Highly accurate protein structure prediction with alphafold. Nature, 596:583–589, 2021.
[0211] [5]Kathryn Tunyasuvunakool, Jonas Adler, Zachary Wu, Tim Green, Michal Zielinski, Augustin Alex Bridgland, Andrew Cowie, Clemens Meyer, Agata Laydon, et al. Highly accurate protein structure prediction for the human proteome. Nature, 596(7873):590-596, 2021.
[0212] [6]Mehmet Akdel, Douglas Eduardo Valente Pires, Eduard Porta Pardo, Jürgen Arthur O. Zalevsky, Bálint Mészáros, Patrick Bryant, Lydia L. Good, Roman A. Laskowski, Gabriele Pozzati, Aditi Shenoy, Wensi Zhu, Petras J. Kundrotas, Victoria Ruiz Serra, Carlos HM Rodrigues, Alistair SDunham, David Burke, Neera Borkakoti, Sameer Velankar, Adam Frost, Kresten Lindorff-Larsen, Alfonso Valencia, Sergey Ovchinnikov, Janani Durairaj, David B. Ascher, Janet M Thornton, Norman E. Davey, Amelie Stein, Arne Elofsson, Tristan I. Croll, and Pedro Astructural biology community assessment of alphafold 2 applications. bioRxiv, 2021.
[0213] [7] Milot Mirdita, Sergey Ovchinnikov, and Martin Steinegger. Colabfold-making protein folding accessible to all. bioRxiv, 2021.
[0214] [8] Ian R. Humphreys, Jimin Pei, Minkyung Baek, Aditya Krishnakumar, Ivan Anishchenko, Sergey Ovchinnikov, Jing Zhang, Travis J. Ness, Sudeep Banjade, Saket Bagde, Viktoriya G. Stancheva, Xiao-Han Li, Kaixian Liu, Zhi Zheng, Daniel J. Barrero, Upasana Roy, Israel S. Fernández, Barnabas Szakal, Dana Branzei, Eric C. Greene, Sue Biggins, Scott Keeney, Elizabeth A. Miller, J. Christopher Fromme, Tamara L. Hendrickson, Qian Cong, and David Baker. Structures of core eukaryotic protein complexes. bioRxiv, 2021.
[0215] [9] Richard Evans, Michael O'Neill, Alexander Pritzel, Natasha Antropova, Andrew Senior, Tim Green, Augustin Zídek, Russ Bates, Sam Blackwell, Jason Yim, Olaf Ronneberger, Sebastian Bodenstein, Michal Zielinski, Alex Bridgland, Anna Potapenko, Andrew Cowie, Kathryn Tunyasuvunakool, Rishub Jain, Ellen Clancy, Pushmeet Kohli, John M Jumper, and Demis Hassabis. Protein complex prediction with alphafold-multimer. bioRxiv, 2021.
[0216]
[10] David D Boehr, Ruth Nussinov, and Peter E Wright. The role of dynamic conformational ensembles in biomolecular recognition. Nature chemical biology, 5(11):789–796, 2009.
[0217]
[11] Subramanian Vivekanandan, Jeffrey R Brender, Shirley Y Lee, and Ayyalusamy Ramamoorthy. A partially folded structure of amyloid-beta(1–40) in an aqueous environment. Biochemical and biophysical research communications, 411(2):312–316, 2011.
[0218]
[12] Kulkarni Madhurima, Bodhisatwa Nandi, and Ashok Sekhar. Metamorphic proteins: the janus proteins of structural biology. Open Biology, 11(4): 210012, 2021.
[0219]
[13] S Kashif Sadiq, Frank Noé, and Gianni De Fabritiis. Kinetic characterization of the critical step in HIV-1 protease maturation. Proceedings of the National Academy of Sciences, 109(50):20449–20454, 2012.
[0220]
[14] Neil J Bruce, Gaurav K Ganotra, Daria B Kokh, S Kashif Sadiq, and Rebecca C Wade. New approaches for computing ligand–receptor binding kinetics. Current opinion in structural biology, 49:1–10, 2018.
[0221]
[15] Po-Ssu Huang, Scott E Boyken, and David Baker. The coming of age of denovo protein design. Nature, 537(7620):320–327, 2016.
[0222]
[16] Zibo Chen, Ryan D Kibler, Andrew Hunt, Florian Busch, Jocelynn Pearl, Mengxuan Jia, Zachary L VanAernum, Basile IM Wicky, Galen Dods, Hanna Liao, et al. Denovo design of protein logic gates. Science, 368(6486):78–84, 2020.
[0223]
[17] Zibo Chen, Matthew C Johnson, Jiajun Chen, Matthew J Bick, Scott E Boyken, Baihan Lin, James J De Yoreo, Justin M Kollman, David Baker, and Frank DiMaio. Self-assembling 2D arrays with de novo protein building blocks. Journal of the American Chemical Society, 141(22):8891–8895, 2019.
[0224]
[18] Aaron Chevalier, Daniel-Adriano Silva, Gabriel J Rocklin, Derrick R Hicks, Renan Vergara, Patience Murapa, Steffen M Bernard, Lu Zhang, Kwok-Ho Lam, Guorui Yao, et al. Massively parallel de novo protein design for targeted therapeutics. Nature, 550(7674):74–79, 2017.
[0225]
[19] Michael J Dougherty and Frances H Arnold. Directed evolution: new parts and optimized function. Current opinion in biotechnology, 20(4): 486–491, 2009.
[0226]
[20] Xingjie Pan and Tanja Kortemme. Recent advances in de novo protein design: Principles, methods, and applications. Journal of Biological Chemistry, 100558, 2021.
[0227]
[21] Andrew Leaver-Fay, Michael D. Tyka, Steven M. Lewis, Oliver F. Lange, James M Thompson, Ron Jacak, Kristian Kaufman, Paul D. Renfrew, Colin A. Smith, William Sheffler, Ian W. Davis, Seth Cooper, Adrien Treuille, Daniel J. Mandell, Florian Richter, Yih-En Andrew Ban, Sarel Jacob Fleishman, Jacob E. Corn, David E. Kim, Sergey Lyskov, Monica Berrondo, Stuart G. Mentzer, Zoran Popovic, James J Havranek, John Karanicolas, Rhiju Das, Jens Meiler, Tanja Kortemme, Jeffrey J. Gray, Brian Kuhlman, David Baker, and Philip Bradley. Rosetta3: an object-oriented software suite for the simulation and design of macromolecules. Methods in enzymology, 487:545–74, 2011.
[0228]
[22] Po-Ssu Huang, Yih-En Andrew Ban, Florian Richter, Ingemar André, Robert M. Vernon, William R. Schief, and David Baker. Rosetta remodel: A generalized framework for flexible backbone protein design. PLoS ONE, 6, 2011.
[0229]
[23] Brian Kuhlman, Gautam Dantas, Gregory C. Ireton, Gabriele Varani, Barry L. Stoddard, and David Baker. Design of a novel globular protein fold with atomic-level accuracy. Science, 302:1364–1368, 2003.
[0230]
[24] Bruno E. Correia, John T. Bates, Rebecca J. Loomis, Gretchen Baneyx, Chris Carrico, Joseph G. Jardine, Peter Rupert, Colin E. Correnti, Oleksandr Kalyuzhniy, Vinayak Vittal, Mary J. Connell, Eric Stevens, Alexandria Schroeter, Man Chen, Skye MacPherson, Andreia M. Serra, Yumiko Adachi, Margaret A. Holmes, Yuxing Li, Rachel E. Klevit, Barney S. Graham, Richard T. Wyatt, David Baker, Roland K. Strong, James E. Crowe, Philip R. Johnson, and William R. Schief. Proof of principle for epitope-focused vaccine design. Nature, 507:201–206, 2014.
[0231]
[25] Yang Hsia, Jacob B. Bale, Shane Gonen, Dan Shi, William Sheffler, Kimberly K. Fong, Una Nattermann, Chunfu Xu, Po-Ssu Huang, Rashmi Ravichandran, Sue Yi, Trisha N. Davis, Tamir Gonen, Neil P. King, and David Baker. Design of a hyperstable 60-subunit protein icosahedron. Nature, 535:136–139, 2016.
[0232]
[26] Florian Richter, Andrew Leaver-Fay, Sagar D. Khare, Sinisa Bjelic, and David Baker. De novo enzyme design using rosetta3. PLoS ONE, 6, 2011.
[0233]
[27] Ali Madani, Bryan McCann, Nikhil Naik, Nitish Shirish Keskar, Namrata Anand, Raphael R. Eguchi, Po-Ssu Huang, and Richard Socher. Progen: Language modeling for protein generation. bioRxiv, 2020.
[0234]
[28] Alex Hawkins-Hooker, Florence Depardieu, Sebastien Baur, Guillaume Couairon, Arthur Chen, and David Bikard. Generating functional protein variants with variational autoencoders. PLoS Computational Biology, 17, 2021.
[0235]
[29] Donatas Repecka, Vykintas Jauniskis, Laurynas Karpus, Elzbieta Rembeza, Jan Zrimec, Simona Poviloniene˙, Irmantas Rokaitis, Audrius Laurynenas, Wissam Abuajwa, Otto Savolainen, Rolandas Meskys, Martin KM Engqvist, and Aleksej Zelezniak. Expanding functional protein sequence space using generative adversarial networks. bioRxiv, 2019.
[0236]
[30] Namrata An and Possu Huang. Generative modeling for protein structures. NeurIPS, 2018.
[0237]
[31] Hao Huang, Boulbaba Ben Amor, Xichan Lin, Fan Zhu, and Yi Fang. G-vae, a geometric convolutional vae for protein structure generation. ArXiv, abs / 2106, 11920, 2021.
[0238]
[32] Jingxue Wang, Huali Cao, John Zeng Hui Zhang, and Yifei Qi. Computational protein design using deep learning neural networks. Scientific Reports, 8, 2018.
[0239]
[33] Alexey Strokach, David Becerra, Carles Corbi-Verge, Albert Perez-Riba, and Philip M. Kim. Computational protein design with deep learning neural networks. Cell systems, 2020.
[0240]
[34] Ivan Anishchenko, Tamuka Martin Chidyausiku, Sergey Ovchinnikov, Samuel J Pellock, and David Baker. De novo protein design by deep network hallucination. bioRxiv, 2020.
[35] Christoffer H Norn, Basile IM Wicky, David Juergens, Sirui Liu, David E. Kim, Doug K Tischer, Brian Koepnick, Ivan V. Anishchenko, David Baker, and Sergey Ovchinnikov. Protein sequence design by conformational landscape optimization. Proceedings of the National Academy of Sciences of the United States of America, 118, 2021.
[36] Lewis Moffat, Joe G Greener, and David T Jones. Using alphafold for rapid and accurate fixed backbone protein design. bioRxiv, 2021.
[0241]
[37] Ethan C. Alley, Grigory Khimulya, Surojit Biswas, Mohammed AlQuraishi, and George M. Church. Unified rational protein engineering with sequence-based deep representation learning. Nature Methods, 1–8, 2019.
[0242]
[38] Jianyi Yang, Ivan V. Anishchenko, Hahnbeom Park, Zhenling Peng, Sergey Ovchinnikov, and David Baker. Improved protein structure prediction using predicted inter residue orientations. Proceedings of the National Academy of Sciences, 117:1496–1503, 2020.
[0243]
[39] Valerio Mariani, Marco Biasini, Alessandro Barbato, and Torsten Schwede. ldt: a local superposition-free score for comparing protein structures and models using distance difference tests. Bioinformatics, 29:2722–2728, 2013.
[40] S. Sinai, Richard Wang, Alexander Whatley, Stewart Slocum, Elina Locane, and Eric D. Kelsic. Adalead: A simple and robust adaptive greedy search algorithm for sequence design. ArXiv, abs / 2010,02141, 2020.
[0244]
[41] Yang Zhang and Jeffrey Skolnick. Scoring function for automated assessment of protein structure template quality. Proteins: Structure, 57, 2004.
[0245]
[42] Eric Jang, Shixiang Shane Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. ArXiv, abs / 1611,01144, 2017.
[0246]
[43] Kim T. Simons, Richard Bonneau, Ingo Ruczinski, and David Baker. Abinitio protein structure prediction of casp iii targets using rosetta. Proteins: Structure, 37, 1999.
[0247]
[44] DT Jones. Protein secondary structure prediction based on position-specific scoring matrices. Journal of molecular biology, 292 2:195–202, 1999.
[0248]
[45] Lewis Moffat and David T. Jones. A deep semi-supervised framework for accurate modelling of orphan sequences. bioRxiv, 2020.
[0249]
[46] Rebecca F. Alford, Andrew Leaver-Fay, Jeliazko R. Jeliazkov, Matthew J. O’Meara, Frank Dimaio, Hahnbeom Park, Maxim V. Shapovalov, Paul D. Renfrew, Vikram Khipple Mulligan, Kalli Kappel, Jason W. Labonte, Michael S. Pacella, Richard Bonneau, Philip Bradley, Roland L. Dunbrack, Rhiju Das, David Baker, Brian Kuhlman, Tanja Kortemme, and Jeffrey J. Gray. The rosetta all-atom energy function for macromolecular modeling and design. Journal of chemical theory and computation, 13 6:3031–3048, 2017.
[0250]
[47] Steven M. Lewis and Brian Kuhlman. Anchored design of protein-protein interfaces. PLoS ONE, 6, 2011.
[0251]
[48] William Sheffler and David Baker. Rosettaholes: Rapid assessment of protein core packing for structure prediction, refinement, design, and validation. Protein Science, 18, 2009.
[0252]
[49] James Maier, Carmenza Martinez, Koushik Kasavajhala, Lauren Wickstrom, Kevin Hauser, and Carlos Simmerling. ff14sb: Improving the accuracy of protein side chain and backbone parameters from ff99sb. Journal of chemical theory and computation, 11 8:3696–713, 2015.
[0253]
[50] Ulrich Essmann, Lalith E. Perera, Max L. Berkowitz, Thomas A. Darden, Hsing-Chou Lee, and Lee G. Pedersen. A smooth particle mesh Ewald method. Journal of Chemical Physics, 103:8577–8593, 1995.
[0254]
[51] Janani Durairaj, Mehmet Akdel, Dick de Ridder, and Aalt DJ van Dijk. Geometricus represents protein structures as shape-mers derived from moment invariants. Bioinformatics, vol. 36 Suppl_2:i718–i725, 2020.
[0255]
[52] Daniel D. Lee and H. Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401:788–791, 1999.
[0256]
[53] Nicholas F. Polizzi, Yibing Wu, Thomas Lemmin, Alison M. Maxwell, Shao-Qing Zhang, Jeff Rawson, David N. Beratan, Michael J. Therien, and William F. DeGrado. De novo design of a hyperstable non-natural protein-ligand complex with sub- De novo design of ultrastable non-native protein-ligand complexes with sub-precise accuracy. Nature chemistry, 9 12:1157–1164, 2017.
[0257]
[54] Roger Tseng, Nicolette F Goularte, Archana G. Chavan, Jansen Luu, Susan E Cohen, Yong-Gang Chang, Joel Heisler, Sheng Li, Alicia K. Michael, Sarvind Tripathi, Susan S. Golden, Andy Li Wang, and Carrie L. Partch. Structural basis of the day-night transition in a bacterial circadian clock. Science, 355:1174–1180, 2017.
[0258]
[55] Rodrigo Gallardo, Neil A. Ranson, and Sheena E. Radford. Amyloid structures: much more than just a cross-βfold. Current opinion in structural biology, 60:7–16, 2019.
[0259]
[56] Bobonis, J., Mitosch, K., Mateus, A., Karcher, N., Kritikos, G., Selkrig, J., Zietek, M., Monzon, V., Pfalz, B., Garcia-Santamarina, S., Galardini, M., Sueki, A., Kobayashi, C., Stein, F., Bateman, A., Zeller, G., Savitski, M.M., Elfenbein, J.R., Andrews-Polymenis, H.L., Typas, A., 2022. Bacterial retrons encode phage-defending tripartite toxin–antitoxin systems. Nature, 1–7, https: / / doi.org / 10.1038 / s41586-022-05091-4.
[0260]
[57] Engler, C., Kandzia, R., Marillonnet, S., 2008. A One Pot, One Step, Precision Cloning Method with High Throughput Capability. PLoS ONE 3, e3647, https: / / doi.org / 10.1371 / journal.pone.0003647.
[0261]
[58] Kritikos, G., Banzhaf, M., Herrera-Dominguez, L., Koumoutsi, A., Wartel, M., Zietek, M., Typas, A., 2017. A tool named Iris for versatile high-throughput phenotyping in microorganisms. Nature Microbiology 2, 17014, https: / / doi.org / 10.1038 / nmicrobiol.2017.14.
Claims
1. A computer-implemented method for designing at least one protein, comprising: - creating at least one amino acid sequence to be tested, wherein the amino acids contained in the amino acid sequence to be tested are selected according to a probability distribution; - predicting structural properties of said at least one protein based on said at least one amino acid sequence; - calculating the value of the fitness function of said at least one amino acid sequence to be tested based on said structural properties of said at least one protein; - selecting or deselecting said at least one amino acid sequence to be tested according to the value of said fitness function.
2. The method of claim 1, further comprising modifying the at least one amino acid sequence to be tested by changing at least one amino acid contained in the amino acid sequence to be tested.
3. The method of claim 2, further comprising repeating said altering of said at least one amino acid sequence until the value of said fitness function exceeds a preselected threshold.
4. The method according to any one of claims 1 to 3, wherein: The prediction of the structural properties of the at least one protein contained in the amino acid sequence to be tested comprises predicting the positions of amino acids relative to each other and / or predicting the full atomic structure of the at least one protein.
5. The method according to any one of claims 1 to 4, wherein the alteration comprises mutation and / or recombination of the at least one amino acid sequence to be tested.
6. A method according to any one of claims 1 to 5, wherein the structural properties are predicted using AlphaFold.
7. The method according to any one of claims 1 to 6, wherein the fitness function is defined according to biological and / or physicochemical properties of the at least one protein.
8. The method according to any one of claims 1 to 7, wherein the fitness function comprises one or more fitness function components representing biological and / or physicochemical properties of the at least one protein.
9. The method according to any one of claims 1 to 8 also includes inputting known structural properties associated with the at least one amino acid sequence or with a target protein to which the at least one protein can bind into AlphaFold as a structural template.
10. The method of any one of claims 1 to 9, further comprising redesigning the at least one amino acid sequence selected to be tested so that the redesigned at least one amino acid sequence is more natural and / or improves the solubility and / or expression ability of the at least one protein.
11. The method of claim 10, further comprising re-predicting structural properties of the at least one protein based on the redesigned at least one amino acid sequence.
12. The method according to any one of claims 1 to 11, further comprising statistically restoring the at least one amino acid sequence to be tested according to the predicted structural properties of the at least one protein.
13. The method of any one of claims 1 to 12, further comprising computationally validating the selected amino acid sequence using molecular dynamics and / or Rosetta ab initio structure prediction.
14. The method of any one of claims 1 to 13, further comprising experimentally validating the selected amino acid sequence by determining the solubility and / or expressibility of the at least one protein.
15. The method of any one of claims 1 to 14, wherein the at least one protein comprises a monomer, a homodimer, a heterodimer, a trimer, a tetramer, a pentamer, an oligomer, a protein complex, a component of a protein complex, a binding agent that binds to a target protein, a protein that exhibits multiple conformations.
Citation Information
Cited By
Antibacterial peptide de novo design method and system based on spectrum structure-activity relationship
CN121054095A
