A multi-point protein directed evolution design method and device based on a graph neural network

Abstract

The application discloses a multi-point protein directed evolution design method based on a graph neural network, and designs protein directed evolution through a trained protein directed evolution model based on a graph neural network, and comprises the following steps: protein data preparation, representing a protein structure as a protein graph, wherein each node in the graph represents an amino acid, and nodes close to each other are connected by edges; and binding amino acid feature data on the nodes or edges of the graph. The protein directed evolution design method is used for guiding a directed evolution task in protein engineering and designing an effective multi-point protein mutation scheme.

Description

Technical Field

[0001] This invention belongs to the field of bioengineering technology, and specifically relates to a method and apparatus for multi-site protein directed evolution design based on graph neural networks. Background Technology

[0002] Directed evolution of proteins is an important aspect of bioengineering. It improves certain functions of proteins, such as thermal stability, activity, and luminescence, by altering a small number of amino acid types in wild-type proteins. Traditional biological methods, such as high-throughput experiments and rational design, rely on human experience to randomly mutate amino acids. To reduce the enormous search space and experimental costs required by traditional biological methods, some existing deep learning methods use millions of protein sequences or structural data points to train massive models with hundreds of millions of parameters on tens or even hundreds of GPUs. While consuming significant computational resources, these deep learning models cannot guarantee good mutation schemes that meet practical application standards. The reason for this is that existing models do not incorporate as much prior biological or mathematical knowledge as possible, simply handing over the entire learning task to the algorithm.

[0003] Moreover, most of these existing algorithm models follow the autoregressive pattern of language models, which makes it impossible to take into account the synergistic effect or superordinate nature caused by simultaneous mutations at multiple points when designing deep mutations, thus missing some well-performing mutation schemes. Summary of the Invention

[0004] One embodiment of the present invention provides a protein directed evolution design method, which designs protein directed evolution using a trained graph neural network-based protein directed evolution model. This protein directed evolution model is based on a lightweight graph neural network (LGN) architecture and is used to guide directed evolution tasks in protein engineering, designing effective multi-site protein mutation schemes.

[0005] The protein-directed evolution design method includes preparing protein data, representing the protein structure as a protein graph, where each node represents an amino acid, and adjacent nodes are connected by edges; and associating amino acid feature data with the nodes or edges of the graph. The amino acid feature data includes amino acid type, physicochemical properties, three-dimensional coordinates, and relative distances within the amino acid sequence. Attached Figure Description

[0006] 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:

[0007] Figure 1 A schematic diagram of a multi-site protein directed evolution design method according to one embodiment of the present invention. Detailed Implementation

[0008] The embodiments of the present invention address two problems existing in the existing algorithms by designing a lightweight graph neural network representation for protein sequence inference and mutation design tasks.

[0009] According to one or more embodiments, a multi-site protein directed evolution design method based on graph neural networks adopts a lightweight graph neural network architecture, LGN (lightweight graph neural network for short), and its overall training and prediction can be divided into three stages.

[0010] The first stage is data preparation.

[0011] First, the tertiary structure of the protein is modeled into a corresponding graph representation, where each node represents an amino acid, and spatially close nodes (at most k other close nodes are found for each node using the k-nearest neighbor method) are connected by edges.

[0012] To better distinguish different amino acids and describe the relationships between them, prior biological knowledge is used to embed corresponding feature information into nodes and edges. This information includes the type of amino acid, its physicochemical properties, its three-dimensional coordinates in space, and its relative distance within the amino acid sequence. These properties come from different scales, thus characterizing the microenvironment of amino acids across different ranges.

[0013] The reduced-weight CATH dataset can be used as input for subsequent model pre-training. CATH is a free and publicly available protein domain classification dataset that stores a large number of wild-type protein structures and properties measured in the laboratory. After removing protein sequences with a sequence similarity exceeding 40%, the dataset contains more than 30,000 protein domains. Converting each protein domain into a corresponding protein map using the method described above yields the training dataset used for model parameter learning.

[0014] The second stage is model pre-training. The transformed CATH dataset is used as input to train the corresponding graph convolutional neural network, which learns the mapping rules from amino acid nodes to protein space.

[0015] Considering the scarcity of mutation experiment data, a self-supervised learning framework was first used to pre-train the model, allowing it to learn the common language of wild-type proteins. Given that protein mutation in nature is equivalent to random evolution, the learning task was set as recovering noisy amino acid sequences. Specifically, the amino acid types at nodes were first perturbed, and features strongly correlated with amino acid types (such as solvent accessible surface area, or SASA) were masked. Then, this noisy dataset was fed into equivariant graph convolution (EGC) layers to encode the microstructure of each amino acid.

[0016] Here, EGC can be stacked in multiple layers. For a graph with n nodes,

[0017] 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 ,······].

[0018] The goal of the l-th layer EGC is to transform the three inputs above to obtain the information of the (l+1)-th layer, i.e., H. l+1 ,

[0019] Two new variables H l+1 , The calculation method is as follows:

[0020]

[0021]

[0022]

[0023] 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.

[0024] 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 ijThe 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.

[0025] After several layers of EGC, each amino acid can be represented as a vector. This vector, when decoded by a fully connected layer, outputs the corresponding predicted label, such as the amino acid type without noise. To improve the expressive power of each vector encoding, a multi-task learning strategy is used. This strategy not only recovers the amino acid type but also predicts the SASA and B-factor (temperature factor, an indicator that describes protein flexibility) for each amino acid.

[0026] The loss function is defined as L total =L aa +λ1L SASA +λ2L B-factor ,

[0027] Where λ1 and λ2 are hyperparameters used to balance the importance of different terms in the loss function. aa The difference between the predicted amino acid type and the actual amino acid type is calculated using the cross-entropy loss function; L SASA and L B-factor This represents the loss value corresponding to the difference between the prediction and the actual value for SASA and B-factor, calculated using the mean squared error.

[0028] Considering that the ultimate goal of the model is to study directed evolution, that is, to modify the original amino acid into a different and better amino acid type, rather than simply restoring the original amino acid type, L is calculated. aa In this case, the BLOSUM62 matrix (blockssubstitution matrix, a commonly used amino acid substitution scoring matrix used to describe the similarity between different amino acid types) was used for label smoothing to enhance the robustness of the model and the diversity of predicted amino acid types.

[0029] Modified L aa Defined as:

[0030] Among them, u(y aa |X i This refers to the amino acid type distribution defined by BLOSUM62.

[0031] The third stage is the mutation scheme evaluation. Depending on whether the protein to be improved has already had a mutation score obtained from a small number of wet experiments, this can be divided into two schemes.

[0032] In the absence of any wet experimental data, for novel proteins and mutant properties, the wild-type (or improved prototype) tertiary structure is first converted into a corresponding protein graph representation and input into a pre-trained graph neural network (LGN). This yields a 20-dimensional vector output (corresponding to 20 amino acid types), representing the predicted probability of the amino acid type at each node. The mutation score is calculated by comparing this probability with the logarithm of the probability of the presence of the wild-type protein.

[0033] Right now

[0034] This scoring function is called the log odds ratio.

[0035] The inference results based on the pre-trained model can provide single-site and multi-site mutation schemes for designing new proteins and functions without any wet experimental data. The quality of the mutation schemes obtained based on this approach requires only about 1% of the computational resources of existing models to achieve similar predictive results. Here, "multi-site" means that multiple amino acids can be changed at once when mutating them. Therefore, this embodiment has no limitation on the number of amino acids changed; both single-site and multi-site mutations can be performed.

[0036] Furthermore, with a small amount of wet experimental test results, the model designed in this invention can be further fine-tuned and optimized, greatly improving the performance of the initial model. The neural network structure LGN+ of the fine-tuned model is set up the same as the pre-trained model in the part that extracts amino acid vector representations. For a new protein, it first extracts the predicted probabilities of its 20-dimensional amino acid types, and then calculates the score of the newly generated sequence after modifying certain amino acids. Compared to the pre-trained model, the fine-tuned model LGN+ is mainly improved in the following two aspects:

[0037] First, the 20-dimensional amino acid type prediction probability y aa A learnable fully connected layer is then added to learn a scoring function applicable to a specific protein (the input protein, such as green fluorescent protein, or GFP) and its function, replacing the original log-probability scoring. The scoring function is fitted by the fully connected layer, and its output scalar can be directly used as a score for the mutation scheme.

[0038] Secondly, when extracting latent vector representations for nodes, the parameters learned by the pre-trained model can be used as initial values, and the learnable parameters within the neural network layers can be adjusted using new protein mutant data. Here, extraction refers to the vector representation of amino acids obtained from the pre-trained model, which can be further transformed and mapped to obtain a 20-dimensional output, or other outputs (such as SASA).

[0039] The pre-trained version of LGN and the fine-tuned version of LGN+ proposed in this invention were tested on single-site and multi-site mutation designs. All data used were publicly available protein mutation scanning datasets (deep mutagenesis scanning, DMS) from previous studies. To demonstrate the superiority of this invention over existing methods, it was compared with some representative deep learning algorithms.

[0040] Experiments show that the finely tuned LGN+ can achieve performance improvements of over 100% compared to existing methods while consuming only one-tenth or even one-hundredth of the computational resources and training data. The table below shows the test results on three single-site (RASH_HUMAN, HG_FLU, MK01_HUMAN) and three multi-site (RRM, GFP, CAPSD) protein mutation test sets. Considering that test scores from different laboratories, different proteins, and different protein functions are not directly comparable, this disclosure uses Spearman's correlation to evaluate the ranking correlation between the predicted mutation sample list and the true list. This evaluation score is a decimal between -1 and 1, with a result closer to 1 indicating better prediction performance. The comparative test results are shown in the table below.

[0041]

[0042] The advantages of this invention are that, compared to existing deep learning models, it significantly reduces the required computational resources and training data, and can simultaneously infer the common probability distribution of all sites on an entire amino acid sequence. Furthermore, leveraging the initial model's fast inference speed and low training data requirements, this invention further designs a fine-tunable advanced version, LGN+.

[0043] The proposed solution has been validated on real deep mutational scanning (DMS) data, and the predicted mutation schemes show a near-perfect positive correlation with the wet-lab-based test results. Compared with other existing deep learning solutions, the performance improvement of this invention can reach over 100%.

[0044] If the units involved in this invention are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0045] 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 method of protein directed evolution design, comprising, designing protein directed evolution by a trained protein directed evolution model based on graph neural network, the constructing step of the protein directed evolution model comprises: protein data preparation, representing protein structure as a protein graph, each node in the graph representing an amino acid, and similar nodes being connected by edges; binding amino acid feature data on the nodes or edges of the graph, the amino acid feature data including amino acid type, physicochemical property, three-dimensional coordinates, and relative distance in the amino acid sequence; using a wild-type protein domain data set processed by the protein data preparation step as a training data set for the protein directed evolution model; pre-training the protein directed evolution model using self-supervised learning, the pre-training comprising perturbing the amino acid type on the nodes, and masking features strongly correlated with the amino acid type, and delivering the obtained data set to a rotation and translation equivariant graph convolutional layer EGC to encode the microstructure of each amino acid.

2. The method of claim 1, wherein, The protein directed evolution design method performs mutation scoring on the output of the protein directed evolution model, i.e., calculates a log probability score, wherein, is the predicted probability of the amino acid type at the node.

3. The method of claim 1, wherein, the protein directed evolution model comprises a fully connected layer for mutation scoring, which outputs the mutation score of the protein directed evolution model.

4. A protein directed evolution design apparatus, characterized by, The apparatus comprises a memory; and a processor coupled to the memory, configured to execute instructions stored in the memory to implement the method of any one of claims 1 to 3.

5. A storage medium having stored thereon a computer program, characterized in that The program is executed by the processor to implement the method of any one of claims 1 to 3.

Images