A protein reverse folding design method and electronic device

Abstract

This invention discloses a protein reverse folding design method, which includes constructing a protein amino acid sequence prediction model. The protein amino acid sequence prediction model is based on a diffusion probability model, specifically utilizing the diffusion and denoising processes of a discrete denoising diffusion probability model to model the amino acid distribution of the protein. The characteristic data of the protein amino acids include amino acid type, physicochemical properties, three-dimensional coordinates, and relative distances along the amino acid sequence. The construction steps of the protein amino acid sequence prediction model include protein data preparation, representing the protein structure as a protein graph, and defining diffusion and denoising generation processes for graph nodes. Defining the diffusion process for graph nodes involves defining a diffusion process for each amino acid node in the protein's amino acid sequence, implemented through a predefined transition probability matrix that describes the transition probability from one amino acid type to another.

Description

Technical Field

[0001] This invention belongs to the field of bioengineering technology, and specifically relates to a protein reverse folding design method based on graph neural networks. Background Technology

[0002] Anti-folding, the prediction of amino acid sequences that will fold into specific 3D protein structures, is crucial for designing new proteins with desired structural and functional properties. However, anti-folding remains challenging due to the vast sequence space to be explored and the complexity of protein folding. Summary of the Invention

[0003] One embodiment of this disclosure is a protein reverse folding design method, which includes a graph-denoised diffusion generative AI model for protein reverse folding.

[0004] The model is based on a diffusion probability model, specifically utilizing the diffusion process and denoising process of a discrete denoising diffusion probability model to model the amino acid distribution of proteins. The characteristic data of protein amino acids include amino acid type, physicochemical properties, three-dimensional coordinates, and relative distances along the amino acid sequence. The model construction steps include protein data preparation, representing the protein structure as a protein graph, and defining a diffusion denoising process for the graph nodes. Attached Figure Description

[0005] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein:

[0006] Figure 1 A schematic diagram of a protein reverse folding graph denoising diffusion generation model according to one embodiment of the present invention. Detailed Implementation

[0007] Recent advances in deep learning have significantly improved the direct mapping from protein structure to amino acid sequences in research on protein backfolding design. However, it still struggles to accurately capture the one-to-many mapping from protein structure to non-unique amino acid sequences. Diffusion probability models, on the other hand, have attracted increasing attention due to their powerful learning capabilities.

[0008] Although diffusion models have been widely used in many fields, their full potential in protein backfolding remains relatively unexplored. Existing sequence design methods are primarily based on language models, including "Masked Language Models" and "autoregressive generative models." However, diffusion probability models employ an iterative prediction approach, generating samples with less noise and better capturing the diversity of real-world data distributions.

[0009] According to one or more embodiments, a method for protein reverse folding design includes modeling the amino acid distribution of a protein using a diffusion process and a denoising process from a discrete denoising diffusion probability model. The steps for constructing the protein amino acid sequence prediction model include protein data preparation, representing the protein structure as a graph, and defining a diffusion denoising generation process for the graph nodes. Amino acid feature data includes amino acid type, physicochemical properties, three-dimensional coordinates, and relative distances on the amino acid sequence.

[0010] The wild-type protein domain dataset, processed through the aforementioned protein data preparation steps, is used as the training dataset for the protein directed evolution model. The protein reverse folding model is trained using a noisy method.

[0011] The model construction method described above defines a diffusion process at the nodes, adding noise to the original amino acid types at the nodes to achieve a uniform distribution across all amino acid types. The isovariant graph neural network denoises the nodes based on the graph's structural information, and new amino acid sequences are generated iteratively at each node through back-diffusion. The generation probability distribution is estimated from the predicted probability using the neural network.

[0012] The model-building method also includes employing a Markov transition matrix, which serves as a guide for the discrete diffusion model, facilitating state transitions. This matrix plays a crucial role in both the diffusion and generation processes.

[0013] The beneficial effects of embodiments of this disclosure include incorporating the secondary structure of the protein as a condition into the sampling process at amino acid nodes. This strategy effectively narrows the exploration space for potential amino acid sequences. Simultaneously, it defines a rapid sampling process for amino acid types, accelerating the sampling speed by 20-100 times.

[0014] Based on one or more embodiments, a graph neural network-based denoising diffusion model, GraDe-IF (Graph Denoising Diffusion for Inverse Protein Folding), is proposed to address the existing protein inverse folding problem. The design of GraDe-IF can be divided into three stages.

[0015] The first stage defines the diffusion process at the nodes.

[0016] A diffusion process is defined for each amino acid node in the protein's amino acid sequence. This diffusion process is a stochastic process that allows the generation of a new amino acid type from the current one. This is achieved by applying a predefined transition probability matrix, which describes the probability of transitioning from one amino acid type to another.

[0017] However, this process is not blind. To incorporate prior biological information, a BLOSUM (Blocks Substitution Matrix) is used. The BLOSUM matrix is ​​a widely used amino acid substitution probability matrix, where each element represents the frequency with which two amino acids are substituted during protein evolution. Specifically, high-frequency substitutions correspond to high scores, while low-frequency substitutions correspond to low scores. By integrating the BLOSUM matrix into the transition probability matrix, it ensures that the resulting new amino acid types are biologically feasible.

[0018] After defining the diffusion process at the nodes, this process can be used to diffuse and add noise to amino acid sequences. This means that a new amino acid sequence can be generated in the space of amino acid types, which has a certain degree of randomness but also follows prior biological knowledge.

[0019] The second stage is the training of the denoising model.

[0020] After obtaining the noisy protein map through the diffusion process, a tool is needed to denoise it. Here, an Equivariant Graph Neural Network (EGNN) is chosen. EGNN is a powerful deep learning model that demonstrates excellent performance when processing data with graph structures. The EGNN here is obtained by stacking multiple layers of EGC. For a graph with n nodes,

[0021] In the l-th layer, the input comes from the hidden node in the previous layer. Node coordinate embedding And edge information E = [····,e ij,······].

[0022] The goal of the l-th layer EGC is to transform the above three inputs to obtain the information of the (l+1)-th layer, i.e., H. l+1 M l+1 =EGC[H l ,X pos M l ], where H l+1 M l+1 H l ,X pos M l These are the node features of layer l+1, the edge features of layer l+1, the node features of layer l, the 3D coordinates of the node, and the edge features of layer l, respectively.

[0023] Two new variables H l+1 M l Each component The calculation method is as follows:

[0024]

[0025]

[0026]

[0027] Where, φ e and φ h φ represents the message passing rules for edges and nodes based on a multilayer perceptron (MLP), respectively. x Used to embed vector m ij The projection is a scalar value.

[0028] Specifically, EGC first uses edge attributes and Euclidean distance between nodes to aggregate the representation of node pairs, and then utilizes the Euclidean distance of first-order neighbors and the scalar φ. x (m ij The 3D position of the next-level node is updated. The hidden embedding of node i is updated by aggregating the hidden representation of node i in the previous level and the information of its first-order neighbors. Through this information aggregation rule, EGC maintains the variance invariance of rotation and translation in 3D space and the permutation invariance of nodes.

[0029] After going through several layers of EGC, each amino acid can be represented as a vector. After decoding it in a fully connected layer, the corresponding predicted label can be output, such as the amino acid type without noise.

[0030] However, training EGNN requires a large amount of training data. Here, the CATH dataset is chosen as the training data. CATH is a free protein domain classification database containing a large number of wild-type protein structures and properties measured in the laboratory. After removing protein sequences with a sequence similarity of more than 40%, the CATH dataset contains more than 30,000 protein domains.

[0031] To transform these protein domains into a form that the model can process, the diffusion process described above is used to convert each protein domain into a corresponding protein map. This results in a large training dataset of protein maps. This training dataset is then used to pre-train the EGNN model. During training, EGNN learns how to extract useful information from noisy protein maps while removing the noise.

[0032] The third stage involves sampling to generate amino acid sequences.

[0033] Once the diffusion process at the nodes is defined, and the isotropic graph neural network is used for network denoising, these two components are combined to generate new amino acid sequences. In this process, an amino acid sequence is first initialized. This initial sequence can be randomly generated or use some common amino acid sequences. Then, the trained isotropic graph neural network is used to predict a new amino acid sequence. This predicted sequence represents the model's best guess for a given backbone.

[0034] Then, the predicted amino acid sequence is noise-added using a defined transition matrix. During this process, a new amino acid is selected at each position according to the probability distribution of the transition matrix, and this new amino acid replaces the original amino acid. Repeating this process allows sampling an amino acid sequence from the backbone. After multiple iterations, an amino acid sequence that has undergone repeated denoising and noise addition is obtained. This sequence generation process combines prior biological knowledge, the denoising capabilities of isovariant graph neural networks, and random sampling through the transition matrix. Thus, a new amino acid sequence corresponding to a given backbone is obtained. This generation process allows for the generation of possible new amino acid sequences from known protein structures, providing new possibilities for understanding protein structure and function. Furthermore, this generation process provides new tools for protein design and engineering, enabling the generation of proteins with specific properties as needed. Figure 1As shown in the figure, Diffusion Process, Denoising Process, prior, condition, vector, summation, network propagator, aggregator, and Equivariant GNN are all part of the equivariant graph neural network.

[0035] The following is a test example of the protein backfolding design method disclosed herein.

[0036] The protein refolding task described in this disclosure was tested on multiple GraDe-IF datasets. This disclosure uses a CATH v4.3.0-based classification system, dividing proteins according to CATH topology to obtain 18,024 proteins for training, 608 proteins for validation, and 1,120 proteins for testing. The model was tested on three different categories: short-sequence proteins (sequence length less than 100), single-stranded proteins, and all proteins. Embodiments of this disclosure employ two metrics to quantify the quality of the recovered protein sequence: perplexity and recovery rate. Perplexity measures how well the model's predicted amino acid (AA) probabilities match the actual amino acid at each position in the sequence. Lower perplexity indicates a better fit to the data. Recovery rate assesses the model's ability to recover the correct amino acid sequence based on the protein's 3D structure. Recovery rate is typically calculated as the proportion of amino acids in the predicted sequence that match the original sequence. A higher recovery rate indicates a stronger ability of the model to predict the original sequence from the structure.

[0037] In the backfolded portion, this disclosure generates high-confidence sequences by comparing the performance of protein recovery on CATH. Recovery rates are improved by 4.2% and 5.4% for single-stranded proteins and short sequences, respectively. This disclosure is also evaluated on the TS50 and T500 datasets.

[0038] In the classification of recovery performance based on hidden and surface amino acids, this disclosure found that more conserved core residues exhibited higher native sequence recovery rates. Conversely, active surface amino acids exhibited lower sequence recovery rates. In all three protein sequence classes, the recovery rate of internal residues significantly exceeded that of external residues, with the recovery rate increasing with the number of contacts. This disclosure also demonstrates recovery rates for different secondary structures, achieving high recovery rates on most secondary structures.

[0039]

[0040] This disclosure, for the first time, attempts to provide a diffusion model for backfolding, modeling the backfolding problem as a denoising problem, where randomly assigned amino acid types in the protein (backbone) map are restored to wild-type. In practical backfolding tasks, this model achieves state-of-the-art results in terms of recovery rates for both single-stranded proteins and short sequences, especially in conserved regions within proteins. Furthermore, the predicted structure of the generated sequence is identical to that of the local sequence. To achieve the desired functionality, this disclosure innovatively allows the model to depend on both secondary and tertiary structures simultaneously, conditionally expressed in the form of residue maps and corresponding node features.

[0041] Compared to existing protein refolding algorithms, this disclosure significantly reduces the required computational resources and training data, and can effectively infer the common probability distribution of the entire amino acid sequence. Furthermore, this invention leverages the fast inference speed and low training data requirements of its initial model to design a fine-tunable advanced version to handle more complex protein refolding scenarios.

[0042] The design method and model proposed in this disclosure have been validated on real-world protein data, and the derived protein reverse folding pathways show a high degree of consistency with experimental results.

[0043] The embodiments disclosed herein can be implemented as software functional units and sold or used as independent products. Such software products can be stored in a computer-readable storage medium. This medium includes, but is not limited to, USB flash drives, external hard drives, read-only memory (ROM), random access memory (RAM), disks, or optical discs. The computer software product contains several instructions that enable a computer device (such as a personal computer, server, or network device) to execute all or part of the steps described in the embodiments of this invention.

[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A protein backfolding design method, characterized in that, This method includes constructing a protein amino acid sequence prediction model. The protein amino acid sequence prediction model is based on a diffusion probability model, which uses the diffusion and denoising processes of a discrete denoised diffusion probability model to model the amino acid distribution of the protein. The steps for constructing the protein amino acid sequence prediction model include protein data preparation, representing the protein structure as a protein graph, and defining a diffusion denoising generation process for the graph nodes. The characteristic data of protein graph nodes include amino acid type, physicochemical properties, three-dimensional coordinates, and relative distances on the amino acid sequence. A diffusion process is defined for each node in the graph, that is, a diffusion process is defined for each amino acid node in the protein's amino acid sequence. This is achieved through a predefined transition probability matrix, which describes the transition probability from one amino acid type to another. After defining the diffusion process at the graph nodes, this process is used to diffuse and add noise to the amino acid sequence. The BLOSUM matrix is ​​then integrated into the transition probability matrix to obtain a protein graph with noise. A denoising model was used to denoise the noisy protein map obtained after the diffusion process, resulting in a new amino acid sequence. The denoising model is based on an equivariant graph neural network and is obtained through training data.

2. The method according to claim 1, characterized in that, The protein domain dataset processed by the protein data preparation steps is used as the training dataset for the protein amino acid sequence prediction model.

3. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method as described in any one of claims 1-2.

Images