A protein fitness prediction method based on deep learning

Abstract

The present invention belongs to the field of life sciences and specifically relates to a method for predicting protein fitness based on deep learning. The method comprises: performing a multiple sequence alignment of a target protein, inferring its evolutionary coupling, and obtaining a local evolutionary representation of homologous protein sequences; utilizing a protein language model to obtain amino acid features of the target protein; performing a positional embedding representation of the amino acid sequence of the target protein to obtain positional features; inputting the local evolutionary representation of the target protein, the amino acid features, and the positional features into a hybrid neural network model to extract features, and performing a fully connected computation on the extracted features to obtain a predicted value for the protein fitness. The present invention proposes a novel hybrid neural network that fully utilizes the local evolutionary information and amino acid feature information of the protein sequence, enabling more efficient extraction and learning of protein sequence data and improving prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of life sciences, and specifically relates to a protein fitness prediction method based on deep learning. Background Art

[0002] Natural proteins often fail to meet the demands of industrial production in terms of stability, tolerance, and selectivity, prompting the search for efficient protein modification methods. Under natural conditions, changes in protein properties or functions typically take a long time. Therefore, directed protein evolution (DRE) has been proposed to accelerate the evolution of natural proteins. DRE primarily involves mimicking and accelerating the protein evolution process in the laboratory through methods such as protein engineering. The target gene is subjected to multiple rounds of repeated mutation and expression, followed by screening based on specific criteria to isolate or enrich for mutants with one or more desired property improvements. Mutant selection is determined by fitness, a measure of the host organism's ability to reproduce in a specific environment in natural evolution. Fitter organisms reproduce faster, and their genes spread throughout the population. When artificial selection is implemented, fitness is defined by the experimenter. Directed evolution of random mutations combined with high-throughput screening has significantly improved enzyme activity, broadening its application in industrial catalysis. While highly effective, DRE strategies still face challenges such as large mutant libraries, a limited number of positive mutations, and difficulty in screening. Appropriate computational tools are also needed to guide the design of mutants and their libraries.

[0003] In recent years, machine learning (ML) has gradually become a new method to assist in the directed evolution of proteins. Machine learning typically uses computer methods to sample combinatorial libraries, thereby achieving greater screening through sequence space searches in each round. In this method, real experimental data from random samples of combinatorial libraries (also called input libraries) are used to train machine learning models, which predict smaller mutant data sets (i.e., prediction libraries). The best-performing mutants are then used as parent sequences for the next round of evolution and mutated at new positions. Deep learning, by training deep neural networks, learns feature hierarchies from low to high, and then performs hierarchical abstract processing on the input data. The original feature data can be mapped into higher-level and more abstract data representations, which can effectively enhance the ability to distinguish and reduce the influence of irrelevant factors. Therefore, deep learning has profoundly changed the field of machine learning. In contrast, traditional learning techniques, such as support vector machines (SVM), Gaussian processes (GP) and artificial neural networks (ANN), strongly rely on manually extracted features. Due to their clear feature encoding principles, these methods may lose sensitive features hidden in the input data; they cannot work by simulating and predicting the fitness of all possible sequences of the target protein after training on a protein sequence database.

[0004] In summary, the existing technical problems are:

[0005] 1. Traditional learning techniques, which manually extract features based on explicit feature encoding principles, will lose sensitive features hidden in the input data;

[0006] 2. Traditional learning techniques cannot work by simulating and predicting the fitness of all possible sequences of the target protein after training on a protein sequence database. Summary of the Invention

[0007] To solve the above technical problems, the present invention proposes a protein fitness prediction method based on deep learning, comprising the following steps:

[0008] S1: Acquire large-scale deep mutation scanning (DMS) datasets and random mutation datasets compiled from published literature;

[0009] S2: Perform multiple sequence alignment of target proteins, infer their evolutionary coupling, and obtain local evolutionary representations of homologous protein sequences;

[0010] S3: Use protein language model to obtain amino acid features of target protein;

[0011] S4: Perform position embedding representation on the amino acid sequence of the target protein to obtain position features;

[0012] S5: The local evolutionary representation, amino acid features, and position features of the target protein are input into the hybrid neural network model to extract features. The extracted features are calculated through full connection to obtain the predicted result value of the protein fitness.

[0013] Preferably, the S2 specifically includes:

[0014] HHblits in HH-Suite was used to search for protein sequences that were homologous to the target protein in the random mutation datasets compiled from large-scale deep mutation scanning datasets and published literature, and the homologous protein sequences were formatted into the A3M multiple sequence alignment format. CCMPred used Markov random fields to learn the A3M multiple sequence alignment of homologous protein sequences to identify evolutionary coupling and obtain the coupling matrix e. ij and the site preference vector e i , according to the site preference vector e i and the coupling matrix e ij Construct a data representation of protein co-evolution information and obtain a local evolutionary representation of homologous protein sequences.

[0015] Furthermore, CCMPred uses Markov random fields to learn multiple sequence alignments of homologous sequences to identify evolutionary coupling, which is expressed as:

[0016]

[0017] Where L(e) represents the evolutionary coupling probability, e i represents the site preference vector, e ij represents the coupling matrix, represents the i-th amino acid in the n-th sequence, represents the jth amino acid in the nth sequence, Z represents the normalization constant, N is the number of homologous sequences, L is the number of columns in the multiple sequence alignment, and exp() represents the exponential function with the natural constant e as the base.

[0018] Furthermore, according to the site preference vector e i and the coupling matrix e ij Construct a data representation of protein co-evolution information and obtain the local evolution representation of homologous protein sequences, which is expressed as:

[0019] V i =[e i (x i ),e i1 (x i ,x1),ei2 (x i ,x2),e i3 (x i ,x3),...,e iL (x i ,x L )]

[0020] Among them, V i The data representation of protein co-evolution information is the local evolution representation, e i represents the site preference vector, e iL represents the L-th coupling matrix, x i represents the i-th amino acid, x L Indicates the Lth amino acid.

[0021] Preferably, the protein language model is used to obtain the amino acid features of the target protein, specifically including:

[0022] The protein language model ESM-1b generates a 1280-dimensional vector representation for each amino acid of the target protein based on the input target protein sequence, and obtains amino acid features containing information about biological properties.

[0023] Preferably, the S4 specifically includes:

[0024] The amino acid sequence of the target protein is passed through a 20-dimensional embedding layer to obtain amino acid embedding features, and position embedding is performed on the embedded features to obtain position features.

[0025] Preferably, the hybrid neural network model includes: a bidirectional gated recurrent unit, a TextCNN layer, a BN layer, two fully connected layers, and a dropout layer.

[0026] Preferably, the S5 specifically includes:

[0027] The local evolutionary representation and amino acid features of the target protein are concatenated after linear dimensionality reduction, the position features and the reduced features are connected in sequence order, the connected features are input into a bidirectional gated recurrent unit to generate a hidden state vector for each amino acid in the sequence, the hidden state vector is input into the TextCNN layer, the sequence feature representation is extracted through convolution and pooling operations, the extracted sequence features are batch normalized through a BN layer, the normalized sequence feature representation is passed through two fully connected layers with a tanh activation function, and a dropout layer is used in the two fully connected layers to prevent overfitting to obtain the predicted fitness value.

[0028] Beneficial effects of the present invention:

[0029] 1. Make full use of the local evolutionary information and amino acid feature information of protein sequences;

[0030] 2. This paper proposes a new hybrid neural network that fuses TextCNN and bidirectional GRU. The fused network can more effectively extract and learn protein sequence data, improve prediction accuracy, and has good generalization performance for different data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of the present invention;

[0032] Figure 2 This is a structural diagram of the neural network model of the present invention. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0034] A protein fitness prediction method based on deep learning, such as Figure 1 As shown, the following steps are included:

[0035] S1: Acquire large-scale deep mutation scanning (DMS) datasets and random mutation datasets compiled from published literature;

[0036] S2: Perform multiple sequence alignment of target proteins, infer their evolutionary coupling, and obtain local evolutionary representations of homologous protein sequences;

[0037] S3: Use protein language model to obtain amino acid features of target protein;

[0038] S4: Perform position embedding representation on the amino acid sequence of the target protein to obtain position features;

[0039] S5: The local evolutionary representation, amino acid features, and position features of the target protein are input into the hybrid neural network model to extract features. The extracted features are calculated through full connection to obtain the predicted result value of the protein fitness.

[0040] Deep-mutational scanning (DMS) involves high-throughput synthesis of various genetic mutations to build a mutation library and determine the effects of mutations on protein function. The resulting DMS datasets, along with random mutation datasets compiled from published literature, include target protein sequences, protein mutant sequences, and their corresponding fitness (label values). The target protein's fitness is 1. If the mutant's fitness is higher than the target protein's, it indicates that the mutant has improved the target protein's function better than the target protein.

[0041] The S2 specifically includes:

[0042] HHblits in HH-Suite was used to search for protein sequences that were homologous to the target protein in the random mutation datasets compiled from large-scale deep mutation scanning datasets and published literature, and the homologous protein sequences were formatted into the A3M multiple sequence alignment (MSA) format. CCMPred used Markov random fields to learn the A3M multiple sequence alignment of homologous protein sequences to identify evolutionary coupling and obtain the coupling matrix e. ij and the site preference vector e i , according to the site preference vector e i and the coupling matrix e ij Construct a data representation of protein co-evolution information and obtain a local evolutionary representation of homologous protein sequences.

[0043] CCMPred uses Markov random fields to learn the MSA of homologous sequences to identify evolutionary coupling, expressed as:

[0044]

[0045] Where L(e) represents the evolutionary coupling probability, e i represents the site preference vector, e ij represents the coupling matrix, represents the i-th amino acid in the n-th sequence, represents the jth amino acid in the nth sequence, Z represents the normalization constant, N is the number of homologous sequences, L is the number of columns in the multiple sequence alignment, and exp() represents the exponential function with the natural constant e as the base.

[0046] According to the site preference vector e i and the coupling matrix e ij Construct a data representation of protein co-evolution information and obtain the local evolution representation of homologous protein sequences, which is expressed as:

[0047] V i =[e i (x i),e i1 (x i ,x1),e i2 (x i ,x2),e i3 (x i ,x3),...,e iL (x i ,x L )]

[0048] Among them, V i The data representation of protein co-evolution information is the local evolution representation, e i represents the site preference vector, e iL represents the L-th coupling matrix, x i represents the i-th amino acid, x L Indicates the Lth amino acid.

[0049] ESM-1b is a protein language model trained using unsupervised learning on 86 billion amino acids from 250 million protein sequences spanning evolutionary diversity. The resulting model incorporates information about biological properties in its representations, which are learned solely from sequence data. For an input sequence, ESM-1b generates a 1280-dimensional vector representation for each amino acid. The reprojected ESM-1b representation is called the amino acid feature representation.

[0050] The protein language model is used to obtain the amino acid features of the target protein, including:

[0051] The protein language model ESM-1b generates a 1280-dimensional vector representation for each amino acid of the target protein based on the input target protein sequence, and obtains amino acid features containing information about biological properties.

[0052] The S4 specifically includes:

[0053] The amino acids in the input sequence are represented using a word embedding similar to word embedding in text processing. The dimension of this word embedding is 20. The word embedding only records the information of each amino acid, but not the position of these amino acids in the sequence. Therefore, position embedding is added after the word embedding layer to supplement the position information. The combined vector is called a position feature.

[0054] The amino acids in the input sequence are represented using a word embedding similar to word embedding in text processing. The dimension of this word embedding is 20. The word embedding only records the information of each amino acid, but not the position of these amino acids in the sequence. Therefore, position embedding is added after the word embedding layer to supplement the position information. The combined vector is called a position feature.

[0055] The hybrid neural network model, such as Figure 2 As shown, it includes: bidirectional gated recurrent unit, TextCNN layer, BN layer, two fully connected layers, and dropout layer.

[0056] The model was trained using the Adam optimizer with default parameters, and the mean square error (MSE) was used as the loss function of the model.

[0057] The loss function of the model is:

[0058]

[0059] Among them, MSE represents the loss function of the model; y is the actual label value; f(x) is the predicted value of the current model training; and n is the number of samples.

[0060] The S5 specifically includes:

[0061] The local evolutionary representation and amino acid features of the target protein are concatenated after linear dimensionality reduction, the position features and the reduced features are connected in sequence order, the connected features are input into a bidirectional gated recurrent unit to generate a hidden state vector for each amino acid in the sequence, the hidden state vector is input into the TextCNN layer, the sequence feature representation is extracted through convolution and pooling operations, the extracted sequence features are batch normalized through a BN layer, the normalized sequence feature representation is passed through two fully connected layers with a tanh activation function, and a dropout layer is used in the two fully connected layers to prevent overfitting to obtain the predicted fitness value.

[0062] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A protein fitness prediction method based on deep learning, characterized in that: include: S1: Obtain deep mutation scanning datasets and random mutation datasets compiled from published literature; S2: Perform multiple sequence alignments of target proteins based on proteins in deep mutation scanning datasets and random mutation datasets curated from published literature, infer their evolutionary coupling, and obtain local evolutionary representations of homologous protein sequences; S3: Use protein language model to obtain amino acid features of target protein; S4: Perform position embedding representation on the amino acid sequence of the target protein to obtain position features; S5: Input the local evolutionary representation, amino acid features, and position features of the target protein into the hybrid neural network model to extract features. The extracted features are then subjected to full-connection calculations to obtain the predicted value of protein fitness. The local evolutionary representation and amino acid features of the target protein are concatenated after linear dimensionality reduction, the position features and the reduced dimensionality features are connected in sequence order, the connected features are input into a bidirectional gated recurrent unit to generate a hidden state vector for each amino acid in the sequence, the hidden state vector is input into the TextCNN layer, the sequence features are extracted through convolution and pooling operations, the extracted sequence features are batch normalized through a BN layer, the normalized sequence features are passed through two fully connected layers with a tanh activation function, and a dropout layer is used in the two fully connected layers to prevent overfitting to obtain the predicted fitness value.

2. A protein fitness prediction method based on deep learning according to claim 1, characterized in that: The S2 specifically includes: HHblits in HH-Suite was used to search for homologous protein sequences of the target protein in the random mutation dataset compiled from the deep mutation scanning dataset and the published literature, and the searched homologous protein sequences were formatted into the A3M multiple sequence alignment format. CCMPred used Markov random fields to learn the multiple sequence alignment of homologous protein sequences to identify evolutionary coupling and obtain the coupling matrix e. ij and the site preference vector e i , according to the site preference vector e i and the coupling matrix e ij Construct a data representation of protein co-evolution information and obtain a local evolutionary representation of homologous protein sequences.

3. A protein fitness prediction method based on deep learning according to claim 2, characterized in that: CCMPred uses Markov random fields to learn multiple sequence alignments of homologous sequences to identify evolutionary coupling, expressed as: Where L(e) represents the evolutionary coupling probability, e i represents the site preference vector, e ij represents the coupling matrix, represents the i-th amino acid in the n-th sequence, represents the jth amino acid in the nth sequence, Z represents the normalization constant, N is the number of homologous sequences, L is the number of columns in the multiple sequence alignment, and exp() represents the exponential function with the natural constant e as the base.

4. A protein fitness prediction method based on deep learning according to claim 2, characterized in that: According to the site preference vector e i and the first coupling matrix e ij Construct the data representation of protein co-evolution information, expressed as: V i =[and i (x i ),And i1 (x i ,x1),e i2 (x i ,x2),and i3 (x i ,x3),...,and iL (x i ,x L )] Among them, V i The data representation of protein co-evolution information is the local evolution representation, e i represents the site preference vector, e iL represents the L-th coupling matrix, x i represents the i-th amino acid, x L Indicates the Lth amino acid.

5. A protein fitness prediction method based on deep learning according to claim 1, characterized in that: The protein language model is used to obtain the amino acid features of the target protein, including: The protein language model ESM-1b generates a 1280-dimensional vector representation for each amino acid of the target protein based on the input target protein sequence, and obtains amino acid features containing information about biological properties.

6. A protein fitness prediction method based on deep learning according to claim 1, characterized in that: The S4 specifically includes: The amino acid sequence of the target protein is passed through a 20-dimensional embedding layer to obtain amino acid embedding features, and position embedding is performed on the embedded features to obtain position features.

7. A protein fitness prediction method based on deep learning according to claim 1, characterized in that: The hybrid neural network model includes: a bidirectional gated recurrent unit, a TextCNN layer, a BN layer, two fully connected layers, and a dropout layer.

Images