Amino acid sequence design method targeting a given protein backbone structure
Through deep learning neural network encoder and self-consistent iteration technology, the protein sequence design process is simplified, and the problems of low design success rate and large structural differences in the existing technology are solved, and efficient and stable amino acid sequence design is achieved to ensure that the design of proteins can be Effectively fold into the target structure.
Patent Information
- Application Number
- CN202210091553.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-26
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-01-26
AI Technical Summary
The success rate of automatic design of amino acid sequences under a given protein main chain structure is not high, and it is analyzed that the design of protein structures and templates are quite different, and there is an unreasonable design upstream of the autoregression process when designing long sequences in deep learning models. The effects are clustered and propagated downstream, which is difficult to verify in experiments.
The deep learning neural network encoder is used to encode the local environment of a single central amino acid residue into real-value vectors, and the side chain type is decoded using a pre-trained decoder, and the side chain type is updated through self-consistent iteration, simplifying the sequence design process and avoiding the use of Approximate side chain conformation model, only amino acid type information is considered.
It improves the success rate and structural stability of protein sequence design, reduces the complexity of sequence design, and the designed amino acid sequence can be efficiently folded into the target structure. Experimental verification shows that the difference between the three-dimensional structure and the target structure of the designed protein is within an acceptable range. .
Smart Images

Figure CN114429783B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of protein design, and specifically relates to an amino acid sequence design method with a given protein main chain structure as the target, that is, according to a preset target main chain structure, all or part of the amino acid sequence of the protein is automatically designed. The method uses a pre-trained deep learning neural network encoder to encode the three-dimensional local structural environment of a single central residue into a real-valued vector, and pre-trains an encoder-decoder to decode the vector into the side chain type of the central residue using a decoder. The input of the encoder contains the side chain type information of other residues spatially adjacent to the central residue. In sequence design, we start from an arbitrarily set initial sequence, apply the encoder / decoder to different central residues, and update the side chain type of the central residue according to the decoding output of the local environment of the central residue in the current sequence background; this process is iterated repeatedly to finally generate an amino acid sequence as a design result. Background Art
[0002] Background technology 1: Automatically design amino acid sequences given the main chain structure.
[0003] This technology targets the protein main chain structure given by the user and automatically selects an amino acid sequence so that the molecule with the amino acid sequence can be spontaneously and stably folded into the target three-dimensional protein conformation. The earliest document reporting the successful automatic design of amino acid sequences was Dahiyat et al., Science 278, 82-87 (1997). This work and Rosetta Design (Kuhlman et al., Science 302, 1364-1368 (2003)) developed by Baker and collaborators at the University of Washington, USA, are to achieve sequence design by optimizing the energy function mainly based on the physical model. In addition, the ABACUS method based on statistical energy function (Xiong et al., Nature Communications 5, 5330 (2014) and Xiong et al., Bioinformatics 36, 136-144 (2020)) has also been successfully applied to the automatic design of amino acid sequences for a given main chain. The above method has been applied to protein design and engineering. A large number of experimental results have confirmed that the above method can design amino acid sequences that spontaneously fold into similar target three-dimensional structures. However, the above method still has shortcomings such as low experimental verification success rate and large differences between the designed protein structure and the template.
[0004] Background technology 2: Application of deep learning models to design amino acid sequences.
[0005] There are two main methods reported for designing spontaneously foldable amino acid sequences from scratch using deep learning models, namely, using a. three-dimensional convolutional neural network (Anand et al., Biorxiv, 2021: 2020.01.06.895466) and b. autoregressive model to generate amino acid sequences (Ingraham et al., Advances in Neural Information Processing Systems, 2019, 32: 15820-15831).
[0006] The model using three-dimensional convolutional neural network mainly realizes the redesign of side chain type and conformation by mapping atom types and atomic coordinates to voxel grids. It is mainly limited by the fact that its design process is highly dependent on accurate atomic coordinate information that cannot be obtained during the design process, and the all-atom sequence design process has a high degree of freedom and is difficult to optimize. Although the autoregressive model has achieved significant improvements in indicators such as natural sequence recovery rate, when designing longer sequences, the upstream unreasonable design effects of the autoregressive process will be aggregated and propagated to the downstream process, so it may be difficult to verify in experiments.
[0007] Therefore, there is still a need for a method that can automatically design all or part of the amino acid sequence of a protein based on a preset target main chain structure. Summary of the invention
[0008] In order to solve the technical problems existing in the prior art, the present invention provides the following technical solutions:
[0009] In one aspect, the present invention provides a method for designing an amino acid sequence targeting a given protein main chain structure, characterized in that the method comprises:
[0010] a. Use a deep learning neural network encoder to encode the local environment of a single central amino acid residue into a real-valued vector;
[0011] b. decoding the real-valued vector using a neural network decoder to obtain the side chain type of the central residue;
[0012] c. Pre-train the encoder-decoder using known protein sequence structure data, with the goal of restoring the natural side chain type of the central residue;
[0013] d. Starting from a given protein main chain structure, the pre-trained encoder is applied to different central residues respectively, and the side chain type of the central residue is iteratively updated according to the decoding results to obtain the designed sequence.
[0014] In some embodiments, the input to the encoder includes the local environment of the central residue, which includes the main-chain conformation of the central residue and the following characteristics of residues spatially adjacent to the central residue: position and orientation relative to the central residue, sequence position and side chain type relative to the central residue.
[0015] In some embodiments, the neighboring residues refer to a plurality of residues that are closest to the central residue in a given main chain structure and spatially surround the central residue.
[0016] In some embodiments, the neighboring residues are the k residues that are closest to the Cα atom of the central residue in a given main chain structure, except for the central residue, and the Cα atom of the central residue is ranked by distance, wherein k is an integer, such as 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 20, 22, 23, 24, 25, 26, 27, 28, 29, 30.
[0017] In some embodiments, the decoder is trained using the Adam optimizer implemented in PyTorch with parameters β1=0.9, β2=0.99 and a learning rate ranging from 10 -4 Gradually reduce to 10 -6 .
[0018] In some embodiments, the PDB structures used for training and testing are 25,234 X-ray structures selected using the PISCES server with a resolution cutoff of The sequence identity cutoff was 50%.
[0019] In some embodiments, a Transformer architecture deep learning encoder is used to encode the local environment information around a single central amino acid residue and a decoder is used to decode it to complete the pre-training task, and the central residue vector before decoding is a real-valued vector representing the local environment.
[0020] In some embodiments, each data block of the Transformer is composed of a multi-head self-attention module, a layer normalization transformation module, a feedforward module and a second layer normalization transformation module in sequence.
[0021] In some embodiments, the number of data blocks of the Transformer is 12.
[0022] In some embodiments, the decoding targets of the decoder include properties of the central residue, and the properties of the central residue include side chain type, secondary structure state, solvent accessible area, side chain torsion angle and B factor of main chain atoms.
[0023] In some embodiments, the real-valued vector is decoded through a single-layer fully connected network.
[0024] In some embodiments, the output of the decoder is a vector of "logits" values, which are converted to "probabilities" through a Softmax function.
[0025] In some embodiments, the probability of a side chain type is Wherein a and b both represent the side chain type.
[0026] In some embodiments, the decoding loss of the decoder is measured using a cross entropy loss function, i.e., for each training item
[0027]
[0028] Where a is the side chain type index, y = (y0, y1, ..., y 20 ), is the one-hot vector generated by the natural side chain type, p=(p0,p1,……,p 20 ) is the vector of predicted probabilities.
[0029] In some embodiments, the total loss of the decoder is
[0030]
[0031] Where λ1, λ2, λ3 and λ4 are trade-off parameters for weighing different losses, L AA refers to the amino acid type loss function, L SS3 / L SS8 Refers to the protein secondary structure type loss function, L SASA refers to the solvent accessible area loss function, L Bfactor Refers to the B-factor loss function, L χ1 / L χ2 Refers to the amino acid side chain conformation loss function; pre-training of the encoder-decoder network is achieved by minimizing the loss function.
[0032] In some embodiments, two different schemes are used to update the side chain type of the central residue during iterations. The first scheme selects the side chain type with the highest probability predicted by the decoder, and the second scheme samples the side chain types according to the probability distribution output by the encoder.
[0033] In some embodiments, the decoder gives the energy of 20 amino acids for the central residue, and the energy is normalized to calculate the probability of each type of amino acid.
[0034] In some embodiments, to control the diversity of sampled side chain types, the following formula maps the original probability distribution predicted by the decoder into a new probability distribution,
[0035]
[0036] Where P'a represents the new probability of mapping, P a and P b They represent the original probabilities of residue types a and b output by the decoder, respectively, and α=1.5.
[0037] In some embodiments, the number of iterations is greater than 350.
[0038] In some embodiments, the number of iterations is 1000.
[0039] In some embodiments, in each iteration, calculations for all selected residues are performed in parallel.
[0040] In some embodiments, the evolution of the negative logarithm of the designed side chain type probability (-logP value) is monitored during the iterative run. If the -logP value of the residue decreases and converges to a plateau value, the sequence whose total -logP recorded during the run decreases and converges to the plateau value is used as the final designed sequence; if not all positions can fully converge in the sequence design run, the sequence with the lowest total -logP recorded during the run is used as the final design result.
[0041] definition
[0042] ABACUS-R: the name of the model of the present invention, which is an improved version of the ABACUS model, and R represents local environmental information.
[0043] Deep learning: A method of machine learning that attempts to achieve a high level of abstraction of data using multiple processing layers that contain complex structures or consist of multiple nonlinear transformations.
[0044] Self-consistent iteration: Iteratively updates the amino acid type of the sequence until each residue selects the best amino acid type based on the surrounding environment.
[0045] Central residue: The most central residue in the local environment.
[0046] Local environment: The local environment of the protein composed of the k nearest amino acids centered on the central residue.
[0047] Real-valued vector: A vector consisting of multiple real-valued components.
[0048] Maximum iterative self-consistency: As many residues as possible in the protein sequence will no longer change due to the possible changes of surrounding residues in the previous round of iterations during the iterative process of selecting the amino acid type with the highest probability based on its local environment.
[0049] Solvent accessible area: is the area of the solute molecule surface that can be contacted by solvent molecules.
[0050] B factor: It describes the dispersion of electron density in X-ray crystal diffraction analysis results. In theory, the B factor can represent the static or dynamic flexibility of atoms.
[0051] Non-redundant PDB structure: protein PDB structure after removing redundant sequences.
[0052] Transformer: A deep learning architecture used in this paper as an encoder for encoding local context.
[0053] Model eval : Model parameters obtained by training with the training set and used to compare and evaluate model performance.
[0054] Model final : Model parameters obtained by training with the training set and actually used to design protein sequence performance.
[0055] Unbiased computational evaluation: that is, using an independent test set for test comparison.
[0056] Side chain type: amino acid type.
[0057] Secondary structure state: refers to the spatial direction and regular cyclic arrangement of protein polypeptide chains along the main chain backbone.
[0058] Side chain torsion angle: refers to the angle between the atoms in the amino acid side chain.
[0059] Logits value: the value directly output by the model before normalization.
[0060] Softmax function: A function used to process the relationship between logits and P.
[0061] CATH category: Different categories of status obtained based on the three-dimensional structure of proteins. All proteins are divided into four types in CATH, namely, all-α type, all-β type, α+β type, and low secondary structure type.
[0062] The relationship between -logits and -logP values:
[0063] Cα-Cα distance: the distance between two Cα atoms in amino acids.
[0064] Rigid body rotation: Two rigid bodies rotate relative to each other.
[0065] One-hot vector: A vector in which only one position has a value of 1 and all other positions are 0.
[0066] "MASKED" type: a placeholder is used to represent the unknown amino acid type.
[0067] CαRMSD: Determine the distances between N pairs of Cα atoms.
[0068] The method of the present invention only uses the side chain type of the environmental residue as the input of the encoder. In the sequence design process, only the side chain type needs to be updated, and there is no need to reconstruct and optimize the side chain conformation, which significantly reduces the complexity of the sequence design problem. Therefore, the whole sequence can be simply optimized by an iterative method. The obtained sequence has a high success rate of folding into a target structure and high structural stability.
[0069] The present invention models and characterizes the local environment of proteins and redesigns the amino acid sequence of the main chain of the target protein.
[0070] In the modeling and characterization stage of the local environment of the protein, the inventors applied a deep neural network model (Transformer) to perform multi-task learning on the three-dimensional local structural environment of a single central residue using natural protein structure data with sequence redundancy removed, in order to obtain a real-valued vector that can characterize the local three-dimensional environment of the amino acid, and applied simple linear changes to restore central residue information including amino acid type information, side chain conformation, solvent accessible area, and B factor.
[0071] In the full sequence design stage, the pre-trained model obtained in the previous step is used to iteratively design the amino acid type for each residue site of the main chain structure of the target protein. In the process of iterative optimization of each amino acid site, since each amino acid site can be repeatedly updated, the most convergent amino acid sequence can be obtained in the end. In addition, users can design more diverse amino acid sequences by controlling the entropy of the distribution of 20 amino acid types at each residue position.
[0072] In the technical content of the present invention, the main distinguishing features from the background technology three-dimensional convolutional neural network are as follows: the background technology needs to use an approximate amino acid side chain conformation model to design the sequence, that is, it needs to obtain accurate atomic coordinate position information, which is not available when the protein sequence is designed from scratch. The main distinguishing features from the background technology autoregressive model are: the present invention uses self-consistent iteration in the amino acid sequence design process to avoid the accumulation of upstream erroneous design effects in the autoregressive process and affect the downstream amino acid design, thereby designing a more stable amino acid sequence.
[0073] Beneficial Effects
[0074] Compared to the three-dimensional convolutional neural network, which is the only background technology outside the present invention that has been publicly experimentally verified to be able to design main chains from scratch, the present invention (ABACUS-R) encodes the three-dimensional local structural environment of a single central residue as a real-valued vector, and pre-trains a decoder to decode the vector into the side chain type of the central residue, and designs the protein sequence through self-consistent iteration. The present invention avoids the use of approximate protein side chain models in the process of designing protein sequences, and implicitly characterizes the protein side chain conformation information, that is, the model does not use the specific position or dihedral information of the amino acid side chain atoms, but only considers the amino acid type information, which simplifies the optimization process of sequence design, so that diverse and stable amino acid sequences can be designed quickly. In addition, the real-valued vector encoding the three-dimensional local structural environment of proteins in the present invention can also provide assistance for other downstream protein engineering tasks.
[0075] By adopting the method of the present invention, under the same comparison conditions, the full sequence recovery rate index of ABACUS-R in natural templates is about 44%, which is higher than all currently reported models.
[0076] In experimental verification, ABACUS-R has been shown to be able to design foldable artificial proteins from scratch for multiple natural protein main chain structures. Among the 57 designed proteins tested, 49 have been experimentally proven to form soluble folded monomeric three-dimensional structures, and the high-resolution crystal structures of five of them were solved experimentally, with the structural differences from the designed models being within 1.3 fold and 0.6 fold, respectively. This is much lower than the error of protein structures and models designed and solved by ABACUS (Xiong et al., Nature Communications 5, 5330 (2014) and Xiong et al., Bioinformatics 36, 136-144 (2020)) or Rosetta (Kuhlman et al., Science 302, 1364-1368 (2003)). The above results all show that ABACUS-R is a better protein sequence design method. BRIEF DESCRIPTION OF THE DRAWINGS
[0077] Figure 1 A self-consistent iterative process for designing a full sequence of a given target main chain structure according to an embodiment of the present invention is shown.
[0078] Figure 2 It is shown that ABACUS-R is used to design corresponding sequences for a given target backbone according to an embodiment of the present invention.
[0079] Figure 3 The gel filtration chromatogram of 1r26-A3 designed by ABACUS-R according to the present invention is shown. 1H-NMR diagram and X-ray structure analysis diagram.
[0080] Figure 4 The gel filtration chromatogram of 1r26-A6 designed by ABACUS-R according to the present invention is shown. 1 H-NMR diagram and X-ray structure analysis diagram.
[0081] Figure 5 The gel filtration chromatogram of 1r26-A7 designed by ABACUS-R according to the present invention is shown. 1 H-NMR diagram and X-ray structure analysis diagram.
[0082] Figure 6 The gel filtration chromatogram of 1r26-B4 designed by ABACUS-R according to the present invention is shown. 1 H-NMR diagram and X-ray structure analysis diagram.
[0083] Figure 7 The gel filtration chromatogram of 1cy5-A7 designed by ABACUS-R according to the present invention is shown. 1 H-NMR diagram and X-ray structure analysis diagram.
[0084] Figure 8 The gel filtration chromatogram of 1ubq-A4 designed by ABACUS-R according to the present invention is shown. 1 H-NMR diagram and X-ray structure analysis diagram. DETAILED DESCRIPTION
[0085] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.
[0086] Example 1 Construction of encoder-decoder network
[0087] The encoder part of the encoder-decoder network of ABACUS-R is a Transformer, whose input includes the side chain types and 3D main chain structure information of all other residues that are structurally close to the central residue. It should be emphasized that the side chain type of the central residue and the side chain conformation of the neighboring residues are not used as input. The output of the encoder constitutes the required vector representation, which is decoded into various properties of the central residue. This encoder-decoder network was trained on the PDB structures of the selected set. Figure 1A self-consistent iteration for designing a full sequence of a given target main chain structure is shown. The method can start with a randomly selected initial sequence, and then for one or more randomly selected central residues, a pre-trained encoder-decoder network is applied to its local environment (depending on the side chain types of the surrounding residues in the current sequence) to update the type of the central residue. In successive iterations, this operation is repeated for different central residues until the side chain types of as many residues as possible no longer change (i.e., convergence has been achieved).
[0088] Example 2 Training of encoder-decoder network
[0089] A set of non-redundant PDB structures has been used to train the encoder-decoder network. We split the selected PDB structures into training and test sets in two different ways, thus learning two sets of network parameters. The first set of network parameters (Model eval ) was obtained by using about 95% of the protein structures for training and the remaining about 5% of the proteins for testing, the structures used for testing belong to the single-domain topology type (according to the CATH4.2 classification of protein structures in the dataset of the selected structures). With this selection of test proteins, no test structure belongs to the same CATH topology as the training structure. Therefore, Model eval Can be used for unbiased computational evaluation. The second set of network parameters Model final It is learned by randomly splitting protein structures into approximately 95% for training and 5% for testing, ignoring their CATH structural classification. final Sequences have been used to design the experimental assays of the present invention.
[0090] The present invention adopts a multi-task learning strategy to train the encoder, in which multiple attributes of the central residue are regarded as decoding targets, including side chain type, secondary structure state (three-state and eight-state schemes for classifying secondary structure types are considered simultaneously), solvent accessible area (SASA), side chain torsion angles (i.e., χ1 and χ2) and main chain atom B factors.
[0091] Model eval and Model final The decoder can recover the side chain type of the central residue with an accuracy of about 50%. In addition, for a large portion of decoding results that did not recover the actual natural side chain type, the decoder produced side chain types with biochemical properties similar to the natural side chain types (e.g., lysine vs. arginine, phenylalanine vs. tyrosine). Notably, the decoding accuracy of side chain type was significantly improved by adding the decoding accuracy of properties other than side chain type to the overall loss of network training.
[0092] For the categorical side chain type attributes, the actual output of the decoder is a vector of "logits" values, which are converted to (normalized) "probabilities" via a Softmax function. In order to check whether negative values of the logits values can be interpreted as the effective "energy" of different residue types, we selected a protein mutant dataset with experimentally measured protein stability changes and checked the Pearson correlation coefficient between the stability changes of different mutants and the change in -ΔΔlogits (measured by ΔΔG). The value of 0.51 obtained is comparable to that reported by deep learning networks specifically trained to reproduce the effects of mutations on protein stability. This observation supports the use of "-logits" values as an alternative value for the effective "energy" of side chain types in a specific local three-dimensional environment.
[0093] Example 3 Convergence of Sequence Design Iterations
[0094] Application Model of the present invention eval And self-consistent iteration to get from Model eval All sequences were designed for the 100 target structures obtained from the test set. These target structures cover the three main CATH categories. For each target structure, 10 sequences were designed using 10 different runs, each starting from a different random initial sequence. Since the iterative method is actually a greedy algorithm that maximizes the side chain type (prediction) probability, we monitored the evolution of the negative logarithm of the designed side chain type probability (-logP value) during the iterative runs. The average per-residue -logP value decreased and converged to a plateau value. At the same time, the side chain types of most residues converged to the corresponding types in the final sequence. For all target structures, the iterative runs can produce self-consistent sequences that either fully converged or converged to positions with only a very small number of fluctuations. If not all positions can be fully converged in the sequence design run, the sequence with the lowest total -logP recorded during the run is taken as the final design result. This choice is for convenience and is not critical, because the number of non-converged positions and the fluctuations in -logP between the sequences recorded in the stable phase are very small, indicating that all these sequences are equally acceptable. Nevertheless, we checked the non-converged residues in several target structures (results not shown) and found that they were usually groups of two to three residues that interacted with each other but had little interaction outside the corresponding group. In the final iteration, the decoded side chain type combinations for such groups oscillated between several possibilities instead of converging to a single combination, indicating that the optimal side chain type cannot be selected for all residues in such a group simultaneously.
[0095] For the same target structure, runs starting from different random starting sequences often result in sequences that are not identical but still highly similar. For the 100 targets considered here, the sequence identities between the (almost fully converged) self-consistent sequences from different runs range from 0.76 to 0.89, and the -logits values for each residue vary only within a narrow range (which is smaller than the difference between the -logits values of the corresponding designed and native sequences). Therefore, sequences designed by different runs can be considered equally credible.
[0096] Example 4 Comparison of designed sequences and natural sequences
[0097] For 100 target structures, the average recovery rate between the designed sequences and the corresponding natural sequences was 43.1% ± 5.4% (the recovery rate ranged from 32% to 61% for different targets). The recovery rate of natural side chain types did not show a strong dependence on the CATH type of the overall structure, but the recovery rate tended to decrease with increasing solvent accessibility. The amino acid composition of the designed sequences and the natural sequences was highly similar: the correlation coefficient between the frequencies of the 20 amino acids in the designed sequences and the frequencies of the amino acids in the natural sequences was 0.93. However, some side chain types, including glutamic acid, alanine and lysine, were more frequently used in the designed sequences than in the natural sequences, while the frequencies of amino acids such as glutamine, histidine and methionine were significantly less in the designed sequences than in the natural sequences.
[0098] Furthermore, we further used Rosetta to compare the Rosetta single residue energies between the designed sequence and the native sequence calculated after backbone relaxation. For the same target backbone structure, the Rosetta single residue energies of the designed sequence and the native sequence were similar. Therefore, using Rosetta energy as an independent indicator, the designed sequence is suitable as the native sequence for the corresponding target structure.
[0099] Despite the high similarity between the designed and natural sequences in Rosetta energy, the single-residue -logits values of the designed sequences are much lower than those of the natural sequences. This is because the self-consistent iteration effectively minimizes the -logits values of the designed sequences. The -logits metric has a component that is orthogonal to the Rosetta energy. The partial orthogonality between the two metrics can also be illustrated by the correlation between each residue -logits value and the Rosetta energy of the designed and natural sequences (Pearson correlation coefficients are 0.24 and 0.17, respectively).
[0100] Example 5 Experimental analysis of designed sequences
[0101] Sequences were designed for three natural backbones using ABACUS-R (PDB IDs of the targets are 1r26, 1cy5, and 1ubq). These backbones were chosen because they have been used to evaluate the ABACUS model. To evaluate this approach in an unbiased manner, the automatically generated sequences were used directly for experimentation without any post-design selection or adjustment. Two batches of experiments were performed using two different protocols. The first protocol used the self-consistent iteration described above to maximize convergence on the side chain types of all residues. Although this protocol effectively minimized the -logits value, the designed sequences were highly similar to each other (the identity between different sequences designed for the same backbone structure was approximately 80%). In the second protocol, the side chain type predicted by the decoder with the highest probability was not selected. Instead, the design of the side chain type was performed by probabilistic sampling within 20 amino acids. The second protocol designed sequences that differed greatly from each other but had relatively high -logits.
[0102] Of the first 27 sequences designed, 26 have been successfully expressed in E. coli. All expressed proteins can be purified in soluble form. We have conducted a series of experimental analyses on the purified proteins, including gel filtration chromatography, one-dimensional NMR spectroscopy, HSQC spectroscopy, DSC scanning, protein crystallization and X-ray structure analysis.
[0103] All purified proteins from the first batch were assayed by gel filtration chromatography and 1 H NMR spectra showed that all these proteins were monomeric and could be folded into a good three-dimensional structure. The thermal stability of the five proteins was determined by DSC experiments (97 to 117°C), indicating that all of these proteins had high stability. We used X-ray crystallography to resolve the high-resolution structures of four of the proteins, including three structures designed with 1r26 as the main chain and one structure designed with 1cy5 as the main chain. All resolved structures were highly consistent with the corresponding design targets (RMSD of Cα atoms was 2.34 Å). For proteins designed with 1ubq as the main chain, we only obtained one crystallized protein (1ubq-A4), but the crystal was a twin crystal and its X-ray structure could not be resolved. 1 The results of H NMR, HSQC spectrum and DSC scanning demonstrated that the protein could fold into a well-defined three-dimensional structure.
[0104] In the second batch of experiments, we designed proteins using the second scheme and obtained sequences with greater sequence differences. The identity between different sequences designed for the same target is about 58%. Its average -logits value is slightly higher than the -logits value of the first batch of sequences. Of the 30 proteins experimentally verified, 25 proteins were successfully expressed in E. coli and 23 were purified in soluble form. Gel filtration chromatography, 1 NMR data and HSQC spectra indicate that all these proteins are monomers and can fold into a well-defined three-dimensional structure. DSC curves of five proteins covering three target backbones indicate that these proteins have high thermal stability (85 to 118°C). We solved a high-resolution structure of the second batch of proteins (1r26-B4), which has a Cα RMSD of 2.37% with the target backbone of 1r26. Therefore, although the modified protocol resulted in a decrease in the success rate, the structures of these successfully designed proteins were consistent with the design target structure, although their thermal denaturation temperatures Tm seemed to be slightly lower than those of the proteins designed using the most convergent approach for the same target.
[0105] Example 6 Protein amino acid sequence design method:
[0106] 1. 3D local environment modeling
[0107] The set of residues that form the three-dimensional local environment of the central residue is defined as k neighboring residues based on the Cα-Cα distance. As input to the encoder, we consider three types of features of each neighboring or surrounding residue: position and orientation relative to the central residue (X SPA ), relative to the sequence position of the central residue (X RSP ) and side chain type (X AA ). The original input of different types of information is defined as follows.
[0108] X SPA The definition of each component includes: (1) the Cα-Cα distance from the environment residue to the central residue; (2) the Cartesian coordinates of the environment residue Cα in a local coordinate system defined by the backbone N, Cα, and C atoms of the central residue (in this coordinate system, Cα is located at the origin, the Cα-N direction is along the x-axis, and the N-Cα-C plane corresponds to the xy plane); (3) a rigid body rotation to align the directions of the two local coordinate systems of the central residue and the environment residue. The Cα-Cα distance is mapped to to A vector of 16 components around 16 distance values in the range. Cartesian coordinates are used directly. Rigid body rotations are mapped to 3D vectors whose direction corresponds to the rotation axis and whose length corresponds to the rotation angle. After this mapping, the X SPA It is a 22-dimensional real-valued vector.
[0109] X RSP The composition of is defined based on the positional differences between the neighboring residues and the central residue, i.e., Δi = ii central .X RSP is a one-hot vector encoding 129 states, each integer value of Δi from -64 to +63 corresponds to one of the 128 states, and the remaining state encodes whether the corresponding neighboring residue is on the same peptide chain as the central residue.
[0110] X AA is a one-hot vector that encodes 20 side chain types, plus a special "MASKED" type (for central residues).
[0111] The features of the central residue are mapped to a vector of the same dimension as the neighboring residues, except that its X SPA All components of a vector are zero, and its X AA The values are mapped to their "MASKED" type. In addition, the central residue (X BB ) as the main chain conformation. BB The components are 15 main chain torsion angles φ i-2 ,ψ i-2 ,ω i-2 ...φ i+2 ,ψ i+2 ,ω i+2 , where i is the index of the central residue position. The angle value is divided by π to obtain a value in the range [-1, +1). In order to keep the input dimensions of the vectors encoding the central residue and the adjacent residues the same, X BB The vector is also formally considered for all neighboring residues, but all its components have the value zero.
[0112] To generate the input for the Transformer encoder from the above raw input feature vector, the one-hot encoded feature vector X AA and X RSP A linear transformation is performed, and the eigenvector X with real-valued components SPA and X BB Use linear layer to transform. The intermediate vectors obtained are denoted as E AA 、E RSP 、E SPA and E BB , which are simply concatenated together to form the overall input vector E, i.e.
[0113] E AA =W AA X AA ,
[0114] E RSP =WRSP X RSP ,
[0115] E SPA =W SPA X SPA +b SPA ,
[0116] E BB =W BB X BB +b BB ,
[0117] E=[E AA E RSP E SPA E BB ],
[0118] Where W AA , W RSP , W SPA , W BB and the bias vector b SPA and b BB Consists of trainable elements, shared among all residues.
[0119] 2. Structure of Transformer-based Encoder
[0120] We note that the set of E vectors for all residues is {En; n = 0, 1, 2, ..., k}, where E0 is the vector of the center residue. These vector sets are fed into a Transformer consisting of 12 data blocks. Each data block of the Transformer consists of a multi-head self-attention module, a layer normalization transformation module, a feed-forward module, and a second layer normalization transformation module in sequence.
[0121]
[0122]
[0123]
[0124]
[0125] (where the residue index n ranges from 0 to k, the module index i ranges from 1 to 12, and the MHA(·,·), LN(·), and FFN(·) operations refer to the multi-head self-attention module, the feed-forward module, and the layer normalization transformation module, respectively. From the output of the above Transformer, It can be used as a general representation of the local environment of a central residue.
[0126] 3. Center Residue Decoder and Training Loss
[0127] By Different attributes of the central residue are decoded by different decoders, and each decoder network has only a single hidden layer.
[0128] For the decoder of the side chain type, its output layer contains 20 perception nodes, and the output value (logits) is transformed by using Softmax to produce the predicted probabilities of 20 side chain types, that is, for side chain type a
[0129]
[0130] The cross entropy loss function is used to measure the decoding loss, that is, for each training item
[0131]
[0132] Where a is the side chain type index, y = (y0, y1, ..., y20) is the one-hot vector generated by the natural side chain type, and p = (p0, p1, ..., p20) is the vector of predicted probabilities.
[0133] The decoder and loss for the secondary structure (SS) state are defined in the same way as for the side chain type. Both three-state and eight-state classification methods are used. The cross entropy loss is L SS3 and L SS8 For numerical attributes including solvent accessible area (SASA) and X-ray B-factor, we first normalize these attributes using the mean and standard deviation calculated from the training data. and is defined as the L-1 error of regression. The total loss of multi-task learning is given by
[0134]
[0135] Where λ1, λ2, λ3 and λ4 are trade-off parameters for weighing different losses, L AA refers to the amino acid type loss function, L SS3 / L SS8 Refers to the protein secondary structure type loss function, L SASA refers to the solvent accessible area loss function, L Bfactor Refers to the B-factor loss function, L χ1 , L χ2 Refers to the amino acid side chain conformation loss function.
[0136] 4. Self-consistent iteration of overall sequence design
[0137] For a given target backbone, we start from an initial sequence and randomly select side chain types for all residues. In each iteration, we randomly select some residues, treat each residue as a central residue, apply the pre-trained encoder-decoder network to its current three-dimensional local environment, and update the side chain type of the central residue according to the output of the decoder (side chain type). In one iteration, the calculations for all selected residues are performed in parallel to take advantage of parallel computing to speed up convergence. The initial number of residues selected for parallel side chain type updates is set to 80% of the sequence length, and this number is gradually reduced in subsequent iterations until the residues are considered one by one in the final iteration.
[0138] During the iterations, we monitored the total -logP value during the evolution, which is defined as
[0139]
[0140] Where L is the length of the sequence, i is the residue index, a i is the index of the residue type at position i in the current sequence, is the probability of the corresponding side chain type predicted by the encoder-decoder network according to the current local environment of residue i (depending on the current side chain types of i's neighboring residues).
[0141] When generating the sequences for examination, two different schemes have been applied to update the side chain types of the central residues during iterations. In the first "normal" scheme, the side chain type with the highest probability (as predicted by the corresponding decoder) is always chosen. This scheme was used to generate the most self-consistent full sequences examined in the first batch of experiments. In the second scheme, the side chain types are sampled according to the probability distribution derived from the encoder output. In order to control the diversity of the sampled side chain types, we map the original distribution predicted by the decoder into a new distribution using the following formula,
[0142]
[0143] The value of the parameter α controls the diversity of sampled side chain types: when α takes very large positive values, the first and second schemes are consistent; lowering the value of α increases the diversity of designed sequences, but also leads to an increase in the -logits metric. To design sequences for examination in the second batch of experiments (for sequences designed for the same target, whose inter-sequence identity is about 60%), we used a value of 1.5 for α, which still resulted in -logits values lower than those of natural sequences.
[0144] 5. Other implementation and calculation details
[0145] The ABACUS-R model has been implemented in Python. The encoder-decoder network has been implemented using the PyTorch package. The hyperparameters that have been explored during model optimization include the number of nearest neighbor residues to consider (i.e., the value of k), the number of data blocks for the Transformer encoder, the dimensionality of all internal layers or modules, and the weights of different losses. Some of the final values used are: k is 20; the number of data blocks for the Transformer is 12; λ1, λ2, λ3, and λ4 are 0.2, 0.2, 0.2, and 0.5. The network has been trained by using the Adam optimizer implemented in PyTorch with parameters β1=0.9, β2=0.99 and a learning rate ranging from 10 -4 Gradually reduce to 10 -6 The PDB structures used for training and testing are 25234 X-ray structures selected using the PISCES server with a resolution cutoff of The sequence identity cutoff was 50%.
[0146] The sequence design runs for the target proteins considered in this article typically require about 350 to 400 iterations to reach convergence. For simplicity, the present invention uses 1000 iterations for all targets. For a target of about 100 residues, 1000 iterations take about 30 seconds on a computer equipped with a GeForce RTX 3090 GPU.
[0147] Rosetta energies were calculated using the RosettaDesign program version 3.12. Since ABACUS-R does not explicitly reconstruct side chain conformations, all side chains of all evaluated sequences (i.e., native and designed sequences) were restacked using the FixBB option of Rosetta, and the complete structures were then relaxed using Rosetta RelaxBB to determine the minimum energy.
[0148] 6. Amino Acid Sequence Verification Method
[0149] The DNA sequence of the designed protein was constructed into the NdeI and Xho sites of pET-22b(+) by General Biotech and Qingke Biotech. The plasmid encoding the designed sequence was transformed into Escherichia coli BL21(DE3) cells. 600 =0.7, 1 mM IPTG was used to induce protein expression at 16°C for 20 hours. 1 H- 15N HSQC NMR studies were performed by culturing the bacteria in an inorganic medium (24 g / L KH2PO4, 5 g / L NaOH, 0.5 g / L NH4Cl, 2.2 mM MgSO4, 0.1 mM CaCl2, and 2.5 g / L glucose) using 15 NH4Cl was used as the sole nitrogen source to prepare 15 The cells were harvested and sonicated in a buffer containing 20 mM Tris and 500 mM NaCl, pH 7.8. The soluble supernatant was purified by Ni 2+ The protein was purified by affinity chromatography and concentrated in a buffer containing 20 mM Tris, 300 mM NaCl, and 1 mM EDTA, pH 7.8, and then used Gel filtration chromatography was performed using a Purifier Systems (GE Healthcare) Superdex 75 column.
[0150] The thermal stability of the designed proteins was evaluated by micro differential scanning calorimetry (nanoDSC) (TA Instruments, DE). Protein samples were prepared in a buffer containing 300 mM NaCl, 20 mM Tris, 1 mM EDTA, pH 8.0, with a protein concentration of 3-6 mg / mL. The protein melting temperature was monitored from 25 to 125 °C at a heating rate of 1 °C / min in a 0.33 mL cell. The reversibility of protein denaturation was measured by reheating and scanning 3 times. The nanoDSC scans were background corrected and analyzed using Launch NanoAnalyze software.
[0151] All NMR data were collected on a room temperature Bruker DMRX600 spectrometer. 2D HSQC-NMR samples typically contained 0.35–0.5 mM 15 N-labeled proteins, 25 mM NaH2PO4, 150 mM NaCl, 2 mM EDTA (pH 6.9) and 10% (v / v) D2 O. Data were processed using the programs NMRDraw / NMRPipe and SPARKY.
[0152] Using the screening kit, the purified and concentrated protein (15-20 mg / mL) was screened for crystallization by hanging drop vapor diffusion at 289 K. Crystals for data collection were obtained in 24-48 hours in PEG1500 for 1r26-A3, 2.5 M ammonium sulfate and 0.1 M BIS-TRIS propane, pH 7.0 for 1r26-A6, 2% v / v 1.4-dioxane, 0.1 M BICINE pH 9.0, 10% w / v PEG 20000 for 1r26-A7, and 27% w / v PEG 2000MME and 0.1 M sodium cacodylate, pH 6.5 for 1r26-B4. 1cy5-A7 crystals appeared in a buffer containing 12% w / v polyethylene glycol 3,350 and 4% v / v Tacsimate, pH 7.0 for 7 days. All crystals were briefly soaked in a storage solution supplemented with 30% glycerol (v / v) and then quickly frozen in liquid nitrogen. Diffraction data were collected at the BL19U1 and BL02U1 beamlines of the Shanghai Light Source. For 1r26-A3, 1r26-A6, 1r26-A7, the data sets were processed using the HKL2000 program, and for 1cy5-A7 and 1r26-B4, the data sets were processed using the XDS program. Molecular replacement was performed using MOLREP using the designed target structure as a search template. The final structure was optimized by PHENIX.
[0153] Example 7 Designed amino acid sequence and its characteristics
[0154] The present invention screened out the following design sequence, and its characteristics are as follows:
[0155] 1r26-A3 (SEQ ID NO: 1): Cα RMSD with 1r26 is The soluble protein was expressed in Escherichia coli with a thermal denaturation temperature of 98°C. Crystallization was performed, and crystals were obtained and the structure was analyzed. The protein structure was analyzed and the protein database PDB ID was 7VQL.
[0156] 1r26-A6 (SEQ ID NO: 2): Cα RMSD with 1r26 is The soluble protein was expressed in Escherichia coli with a thermal denaturation temperature of 108°C. Crystallization was performed, and crystals were obtained and the structure was analyzed. The protein structure was analyzed and the protein database PDB ID was 7VQV.
[0157] 1r26-A7 (SEQ ID NO: 3): Cα RMSD with 1r26 is The soluble protein was expressed in Escherichia coli with a thermal denaturation temperature of 107°C. Crystallization was performed, and crystals were obtained and the structure was analyzed. The protein structure was analyzed and the protein database PDB ID was 7VQW.
[0158] 1r26-B4 (SEQ ID NO: 4): Cα RMSD with 1r26 is The soluble form of the protein was expressed in E. coli, crystallized, and the crystals were obtained and the structure was analyzed. The protein structure was analyzed and the protein database PDB ID is 7VU4
[0159] 1cy5-A7 (SEQ ID NO: 5): Cα RMSD with 1cy5 is , a soluble form of the protein was expressed in Escherichia coli with a thermal denaturation temperature of 97°C. Crystallization was performed, crystals were obtained and the structure was elucidated. The elucidated protein structure has a protein database PDB ID of 7VTY.
[0160] 1ubq-A4 (SEQ ID NO: 6): A soluble protein was expressed in E. coli with a thermal denaturation temperature of 117°C. Crystallization was performed to obtain crystals.
[0161] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for designing an amino acid sequence targeting a given protein main chain structure, characterized in that: The method comprises: a. using a deep learning neural network encoder to encode the local environment of a single central amino acid residue into a real-valued vector, wherein the input of the encoder includes the local environment of the central residue, wherein the local environment includes the main-chain conformation of the central residue and the following characteristics of the residues spatially adjacent to the central residue: position and orientation relative to the central residue, sequence position and side chain type relative to the central residue; b. decoding the real-valued vector using a neural network decoder to obtain the side chain type of the central residue; c. Pre-train the encoder-decoder using known protein sequence structure data, with the goal of restoring the natural side chain type of the central residue; d. Starting from a given protein main chain structure, the pre-trained encoder is applied to different central residues respectively, and the side chain type of the central residue is iteratively updated according to the decoding results to obtain the designed sequence.
2. The amino acid sequence design method targeting a given protein main chain structure according to claim 1, characterized in that: in, The neighboring residues refer to a plurality of residues that are closest to the central residue in a given main chain structure and spatially surround the central residue.
3. The amino acid sequence design method targeting a given protein main chain structure according to claim 1, characterized in that: The local environment information of a single central residue is encoded using an encoder and decoded using a decoder to complete the pre-training task. The central residue vector before decoding is a real-valued vector representing the local environment.
4. The amino acid sequence design method targeting a given protein main chain structure according to any one of claims 1 to 3, characterized in that: The decoding target of the decoder is the side chain type of the central residue; during pre-training, other information besides the side chain type of the central residue is added as the decoding target, including the secondary structure state, the solvent accessible area, the side chain torsion angle and the B factor of the main chain atom, for learning to improve the encoding-decoding accuracy.
5. The amino acid sequence design method targeting a given protein main chain structure according to any one of claims 1 to 3, characterized in that: The input of the encoder does not use the side chain conformation information or side chain atom position information of the adjacent residues, so there is no need to construct a spatial structure model of the amino acid residue side chains when the encoder-decoder is used for amino acid sequence design.
6. The amino acid sequence design method targeting a given protein main chain structure according to any one of claims 1 to 3, characterized in that: The output of the decoder is a vector of "logits" values, which are converted to "probabilities" through a Softmax function.
7. The amino acid sequence design method targeting a given protein main chain structure according to any one of claims 1 to 3, characterized in that: Two different schemes were used to update the side chain type of the central residue during iterations. The first scheme selected the side chain type with the highest probability predicted by the decoder, and the second scheme sampled the side chain types according to the probability distribution output by the encoder.
8. The amino acid sequence design method targeting a given protein main chain structure according to any one of claims 1 to 3, characterized in that: In each iteration, calculations for all selected residues are performed in parallel.
9. The amino acid sequence design method targeting a given protein main chain structure according to any one of claims 1 to 3, characterized in that: During the iterative run, the evolution of the negative logarithm of the probability of the designed side chain type (-logP value) is monitored. If the -logP value of the residue decreases and converges to a plateau value, the sequence whose total -logP value recorded during the run decreases and converges to the plateau value is taken as the final designed sequence; If not all positions converged completely during a sequence design run, the sequence with the lowest total -logP recorded during the run was taken as the final design result.
