Protein phase separation characteristic prediction method based on artificial intelligence

Abstract

The invention discloses a protein phase separation characteristic prediction method based on artificial intelligence, particularly relates to the field of protein phase separation characteristic prediction, and is used for solving the problem that an existing method is difficult to fuse protein internal residue interaction structure characteristics, population heritable variation information and experimental phenotype data at the same time in a large-scale missense mutation evaluation scene. And the prediction accuracy and the distinguishing capability are insufficient. The method comprises the following steps: acquiring a wild-type amino acid sequence and population variation data of a target protein, constructing a residue interaction network, identifying a phase separation core interaction cluster, quantifying a network topology disturbance degree caused by missense mutation, combining allele frequency information, constructing a mutation instance graph as input, and calculating the mutation degree of the target protein. And learning and predicting the influence of the mutation on the phase separation characteristic of the protein by using a graph neural network, thereby outputting the phase separation characteristic influence score of each mutation and sorting, and realizing the efficient screening and evaluation of the phase separation related high-risk mutation.

Description

Technical Field

[0001] This invention relates to the field of protein phase separation characteristic prediction technology, and more specifically, to an artificial intelligence-based method for predicting protein phase separation characteristics. Background Technology

[0002] Missense mutations in protein sequences can significantly affect a protein's phase-separation ability by altering the interaction patterns between residues, thereby triggering various diseases such as neurodegenerative diseases, tumors, and inherited metabolic disorders. In precision medicine and functional genomics research, researchers often need to rapidly screen a large number of discovered or newly identified missense mutations to identify high-risk mutations that may significantly affect phase-separation properties, in order to assist in experimental design and disease mechanism research.

[0003] However, existing methods largely rely on low-throughput biochemical experiments or empirical predictions based on single sequence features, making it difficult to simultaneously consider both the overall protein structural interaction characteristics and population genetic background information, resulting in insufficient stability and discriminative power in the prediction results. Furthermore, with the development of population sequencing technology, large-scale population variation databases have accumulated massive amounts of mutation and allele frequency information, but these data have not yet been effectively utilized in phase-separation functional assessment scenarios, making it difficult to form a unified analytical framework with protein internal residue interaction networks and multivalent interaction characteristics.

[0004] Therefore, in applications such as mutation function assessment, disease-related mutation priority ranking, and large-scale mutation screening, there is an urgent need for an intelligent prediction method that can integrate protein residue interaction structural features, population variation information, and known phase separation phenotypic data to achieve systematic assessment and efficient screening of mutations affecting protein phase separation characteristics. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an artificial intelligence-based method for predicting protein phase separation characteristics to address the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An artificial intelligence-based method for predicting protein phase separation characteristics includes the following steps: S1. Obtain the wild-type amino acid sequence of the target protein, and obtain the set of known missense mutations of the target protein and the corresponding allele frequency data based on the population variation database. S2. The wild-type amino acid sequence and homologous sequence are compared. The sequence entropy value of each site is calculated based on the distribution of amino acid types in the comparison results. At the same time, the wild-type residue interaction network is constructed based on the statistical correlation of amino acid types between sites. S3. Identify the amino acid residues corresponding to the nodes with the most connection edges and sequence entropy values ​​below a preset threshold in the wild-type residue interaction network, and define these residues together with their directly interacting neighboring residues as phase-separated core interaction clusters. S4. For each mutant sequence in the missense mutation set, reconstruct the corresponding residue interaction network and calculate the difference between the sum of the shortest path lengths of all node pairs and the wild-type residue interaction network as the network topology perturbation score. S5. Construct a graph neural network prediction model, taking the network topology perturbation score, allele frequency data, and identification information of the phase-separated core interaction clusters corresponding to each mutation as input features, and outputting the phase-separation characteristic influence score of the mutation. S6. Sort all mutations according to the phase separation characteristic impact score output by the graph neural network prediction model, and generate a list of phase separation-related mutations arranged in descending order of potential impact.

[0008] As a further aspect of the present invention, in step S1, obtaining the wild-type amino acid sequence of the target protein and obtaining the set of known missense mutations of the target protein and the corresponding allele frequency data based on the population variation database specifically includes: Based on the identifier of the target protein, the complete wild-type amino acid sequence is retrieved from a pre-set protein sequence database. All missense mutation records belonging to the protein are obtained from a public database containing information on the frequency of human genome variations. The specific amino acid position where the mutation occurs and the type of the mutated amino acid are extracted and integrated into mutation information. At the same time, the frequency data of this mutation information in the genome of the reference population are obtained. The extracted mutation information and occurrence frequency data are integrated to construct a mutation set and the corresponding allele frequency dataset, and a subset of mutations with known phase separation characteristics is identified in the mutation set.

[0009] As a further aspect of the present invention, in S2, constructing a wild-type residue interaction network specifically includes: Using wild-type amino acid sequences as query sequences, homologous sequence searches are performed in protein sequence databases to obtain a set of homologous protein sequences. Multiple sequence alignment algorithms are then used to align the wild-type sequences with the homologous protein sequences, generating alignment results that characterize the frequency of different amino acid types at each amino acid site. For each amino acid site in the alignment results, the sequence entropy value of the site is calculated based on the frequency distribution of various amino acid types observed at the site. At the same time, the co-occurrence frequency of various amino acid type combinations between any two different amino acid sites in the alignment results is analyzed. The correlation strength between each pair of sites is calculated using mutual information, and a wild-type residue interaction network is constructed with sites as nodes and correlation strength as edges.

[0010] As a further aspect of the present invention, in S3, identifying the amino acid residues corresponding to the nodes with the most connection edges and sequence entropy values ​​below a preset threshold in the wild-type residue interaction network, and defining these residues together with their directly interacting neighboring residues as phase-separated core interaction clusters specifically includes: The number of connection edges corresponding to each node in the wild-type residue interaction network and the sequence entropy value of each amino acid site are obtained. Based on the preset sequence entropy value screening conditions, the nodes that simultaneously satisfy the condition of having the most connection edges and a sequence entropy value lower than the entropy value screening conditions are selected from all nodes. In the wild-type residue interaction network, identify all neighboring nodes that have direct connection edges with the selected nodes. Group the selected nodes and the amino acid residues represented by their corresponding neighboring nodes into a set, which is defined as the phase-separated core interaction cluster.

[0011] As a further aspect of the present invention, in step S4, the corresponding residue interaction network is reconstructed for each mutant sequence in the missense mutation set, and the difference between the sum of the shortest path lengths of all node pairs and the wild-type residue interaction network is calculated as the network topology perturbation score, specifically including: For each mutant amino acid sequence in the missense mutation set, the homologous protein sequence is obtained based on the sequence and multiple sequence alignment is performed. The mutant residue interaction network corresponding to the mutant is constructed. For the wild-type residue interaction network and each mutant residue interaction network, the shortest path length between all possible node pairs in the network is calculated based on the weight of the connecting edges in the network, and the shortest path length of each network is summed to obtain the total shortest path length of each network. The difference between the sum of the shortest path lengths of the mutant residue interaction networks and the sum of the shortest path lengths of the wild-type residue interaction networks is used as the network topology perturbation score corresponding to the mutation.

[0012] As a further aspect of the present invention, in step S5, a graph neural network prediction model is constructed, using the network topology perturbation score, allele frequency data, and identification information of the phase-separated core interaction clusters corresponding to each mutation as input features, and outputting the phase-separation characteristic influence score of the mutation, specifically including: The adjacency matrix of the wild-type residue interaction network is used as the mutation instance graph framework. The allele frequency data corresponding to each node and the Boolean identifier indicating whether the mutation falls into the phase-separated core interaction cluster are used as node features. The network topology perturbation score is used as the global feature. Together, they constitute the graph structure input of the model. A graph neural network model is constructed. The graph structure input is divided into subsets of mutations with phase separation characteristics as inputs. The graph neural network is trained under supervision. The model output represents the score of the influence of phase separation characteristics.

[0013] As a further aspect of the present invention, the phase separation characteristic influence score is obtained by numerically labeling the phase separation phenotype corresponding to the mutation subset in the graph neural network model training with quantitative experimental data, and then normalizing the data to map it into a fixed numerical range.

[0014] As a further aspect of the present invention, in step S6, sorting all mutations according to the phase separation characteristic impact score output by the graph neural network prediction model to generate a list of phase separation-related mutations arranged in descending order of potential impact specifically includes: The mutation set other than the mutation subset with known phase separation characteristics is input into the graph neural network prediction model to obtain the corresponding phase separation characteristic influence score set. The phase separation characteristic influence scores of the mutation subset with known phase separation characteristics are merged to form a complete score set. All missense mutation instances in all mutation sets are arranged in descending order of the values ​​in the complete score set to generate a structured list.

[0015] The technical effects and advantages of the protein phase separation characteristic prediction method based on artificial intelligence of the present invention are as follows: This invention constructs an intelligent prediction process that integrates protein residue interaction networks, population genetic variation information, and experimental phenotypic annotations, enabling a systematic assessment of the impact of missense mutations on protein phase separation characteristics. It can provide stable and discriminative prediction results in large-scale mutation screening scenarios.

[0016] Compared to traditional methods based solely on amino acid sequences or single physicochemical characteristics, this invention introduces residue interaction networks and network topology perturbation scores to quantitatively characterize the overall interaction structure changes caused by mutations, thus more realistically reflecting the impact mechanism of mutations on multivalent interaction systems. Simultaneously, incorporating population genetic constraints such as allele frequencies into the prediction model effectively reduces the risk of misjudging highly tolerant mutations in the population and improves the accuracy of identifying potentially pathogenic or functionally critical mutations. By modeling mutation instance graphs using graph neural networks, this invention can learn the complex relationships between mutation nodes and their structural environment within a fixed topological framework, avoiding the subjectivity and limitations of manually set rules. This method is not only applicable to risk assessment of known phase-separation-related mutations but can also rank and prioritize new mutations without experimental validation, significantly improving the efficiency of mutation function research, disease mechanism analysis, and experimental resource allocation, demonstrating good versatility, scalability, and practical application value. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a protein phase separation characteristic prediction method based on artificial intelligence according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 Figure 1 This invention presents a method for predicting protein phase separation characteristics based on artificial intelligence, which includes the following steps: S1. Obtain the wild-type amino acid sequence of the target protein, and obtain the set of known missense mutations of the target protein and the corresponding allele frequency data based on the population variation database. S2. The wild-type amino acid sequence and homologous sequence are compared. The sequence entropy value of each site is calculated based on the distribution of amino acid types in the comparison results. At the same time, the wild-type residue interaction network is constructed based on the statistical correlation of amino acid types between sites. S3. Identify the amino acid residues corresponding to the nodes with the most connection edges and sequence entropy values ​​below a preset threshold in the wild-type residue interaction network, and define these residues together with their directly interacting neighboring residues as phase-separated core interaction clusters. S4. For each mutant sequence in the missense mutation set, reconstruct the corresponding residue interaction network and calculate the difference between the sum of the shortest path lengths of all node pairs and the wild-type residue interaction network as the network topology perturbation score. S5. Construct a graph neural network prediction model, taking the network topology perturbation score, allele frequency data, and identification information of the phase-separated core interaction clusters corresponding to each mutation as input features, and outputting the phase-separation characteristic influence score of the mutation. S6. Sort all mutations according to the phase separation characteristic impact score output by the graph neural network prediction model, and generate a list of phase separation-related mutations arranged in descending order of potential impact.

[0020] In step S1, the wild-type amino acid sequence of the target protein is obtained, and the set of known missense mutations of the target protein and the corresponding allele frequency data are obtained based on the population variation database.

[0021] A unique identifier for the target protein is defined, using an internationally recognized protein entry numbering format for precise matching and retrieval in a pre-built protein sequence database. This database pre-stores a large number of manually verified protein entries, each corresponding to a unique protein identifier and its standardized wild-type amino acid sequence information. During the search process, consistency checks are performed on the identifiers to ensure that the returned amino acid sequences are unique in origin and free from any manual editing or mutation annotations, retaining only the wild-type sequences in their natural reference state. The obtained wild-type amino acid sequences are stored in linear sequence form, and integrity checks are performed on sequence length, residue order, and character validity to prevent subsequent analysis biases caused by missing, duplicate, or illegal characters. This wild-type amino acid sequence serves as a unified reference framework for subsequent mutation localization, homology alignment, and residue interaction network construction, remaining constant throughout the analysis process to ensure that different mutation instances are aligned and compared within the same coordinate system.

[0022] After obtaining the wild-type amino acid sequence, the protein identifier is used as an index to retrieve all missense mutation records associated with the protein-coding region from a pre-selected public database of population variation. This public database aggregates genomic variation statistics from sequencing projects across different populations. Each missense mutation record includes the gene location corresponding to the mutation, the type of base change, and the amino acid substitution information mapped from it. During implementation, the retrieved mutation records are first screened, retaining only missense mutations that can uniquely map to the wild-type amino acid sequence. The amino acid position number and the type of amino acid resulting from each mutation are clearly recorded, and this information is combined to form structured mutation information entries. Simultaneously, the frequency data of each mutation in the reference population genome is extracted from the same database; this frequency data reflects the actual distribution of the mutation in the population. Subsequently, the mutation information entries are associated with the corresponding allele frequency data to construct a mutation set containing all missense mutations and its corresponding allele frequency dataset. Based on this, according to phase separation-related experimental results, the mutation set is labeled, and mutations with clearly defined phase separation phenotypes are marked as a subset of mutations with known phase separation characteristics.

[0023] In S2, a wild-type residue interaction network is constructed.

[0024] The wild-type amino acid sequence of the target protein, which has been acquired and verified for consistency, is used as the unique query sequence and input into a pre-set protein sequence database to perform a homology sequence retrieval operation. This database contains protein sequence information from different species or evolutionary branches and is stored using a unified sequence encoding rule. During the homology sequence retrieval process, protein sequences with significant sequence similarity to the wild-type amino acid sequence in both overall length and local fragments are selected from the database using sequence similarity comparison rules. Abnormal sequences with obvious deletions or redundant repetitions are removed, forming a set of structurally complete and representative homology protein sequences. Subsequently, a multiple sequence alignment algorithm is used to uniformly align the wild-type amino acid sequence with the homology protein sequence set. During the alignment process, the wild-type sequence is maintained as a reference coordinate, aligning the corresponding amino acid sites in each homology sequence. After alignment, an alignment result matrix containing all aligned sites is generated, where each column corresponds to one amino acid site in the wild-type sequence, and each row corresponds to one homology protein sequence. By statistically analyzing the alignment result matrix, the different amino acid types and their occurrence frequency at each position are summarized, and further converted into the frequency distribution of various amino acid types at that position, thus forming a multi-sequence alignment statistical result that can reflect the residue diversity characteristics of the target protein in the evolutionary dimension.

[0025] After obtaining the multiple sequence alignment (MSA) statistical results, for each amino acid site in the wild-type amino acid sequence, a sequence entropy value is calculated based on the frequency distribution of different amino acid types observed at that site in the homologous sequence set. This sequence entropy value quantifies the degree of conservation exhibited by that site during evolution; a more concentrated frequency distribution corresponds to a lower entropy value, and a more dispersed frequency distribution corresponds to a higher entropy value. After calculating the sequence entropy values ​​for all sites, the joint distribution between any two different amino acid sites in the MSA results is further statistically analyzed. Specifically, the number of times different amino acid type combinations at these two sites co-occur in the same homologous protein sequence is counted, and the corresponding co-occurrence frequency distribution is obtained. Using a mutual information calculation method, the co-occurrence frequency distribution between the two sites is compared with their independent frequency distributions to obtain a correlation strength value that reflects the degree of cooperative change of these two sites during evolution. After repeating the above correlation strength calculation operation for all site pairs in the wild-type amino acid sequence, each amino acid site is treated as a node in the network, and the correlation strength between corresponding site pairs is used as the weight of the connection edges between nodes to construct a complete wild-type residue interaction network. The residue interaction network is a graph structure that represents the interrelationships of different residues within the target protein in an evolutionary statistical sense, providing a structured network representation for subsequent identification of core interaction clusters and analysis of mutation effects.

[0026] In S3, the amino acid residues corresponding to the nodes with the most connection edges and sequence entropy values ​​below a preset threshold are identified in the wild-type residue interaction network. These residues, together with their directly interacting neighboring residues, are defined as phase-separated core interaction clusters.

[0027] Using the constructed wild-type residue interaction network as input, all nodes in the network are traversed one by one. Each node uniquely corresponds to an amino acid site in the wild-type amino acid sequence, and the connections between nodes are determined by the correlation strength calculated based on mutual information in the previous steps. During the traversal, the number of directly connected edges for each node is counted. This number reflects the breadth of statistical associations between the amino acid residue and other residues in the residue interaction network. The number of connections is recorded in integer form, without secondary weighting of edge weights, and is only used to measure the connection richness of the node in the network topology. Simultaneously, the sequence entropy value of the amino acid site corresponding to the node is retrieved from the aforementioned multiple sequence alignment and sequence entropy calculation results, and this sequence entropy value is associated with the number of connections of the node and stored to form a node-level structural attribute record table. After completing the attribute statistics of all nodes, sequence entropy screening conditions are set for selecting core nodes. These screening conditions are determined based on the distribution of all sequence entropy values. For example, according to the ranking of sequence entropy values ​​of all amino acid sites, the values ​​in the top 20% of the low-entropy range are selected as the entropy screening threshold to characterize highly conserved residue sets during evolution. Under this screening condition, all nodes are screened once, retaining the set of nodes with sequence entropy values ​​below the threshold. Subsequently, in this low-entropy node set, the number of connection edges is ranked, identifying the nodes with the highest number of connection edges. The determination of the highest number of connection edges uses a relative ranking method; for example, nodes with the highest number of connection edges in the set are selected as high-connectivity nodes that meet the criteria. Through this dual screening process, a set of nodes that simultaneously satisfies the characteristics of high conservatism and high connectivity is obtained. Each node in this set is considered a candidate core node with a key structural position in the wild-type residue interaction network.

[0028] After obtaining the set of candidate core nodes that meet the screening criteria, a neighborhood expansion process is further performed on each node in the set based on the wild-type residue interaction network. Specifically, for each candidate core node, all nodes in the residue interaction network with which it has a direct connection edge are retrieved. This direct connection edge relationship is limited to connections of length one in the network, excluding indirect connections through other nodes. The retrieved neighboring nodes correspond to amino acid sites that have a significant cooperative change relationship with the core node in an evolutionary statistical sense. For each candidate core node, it and all its corresponding direct neighbors are collected together, and the amino acid residues they represent are recorded in the same set. During this process, overlap between the neighboring node sets corresponding to different candidate core nodes is allowed. For duplicate amino acid residues, only one record is retained in the set to ensure the uniqueness of elements within the set. After completing the neighborhood expansion of all candidate core nodes, the resulting residue set contains a group of residues that are both highly conserved in the wild-type residue interaction network and tightly connected through direct interactions. This group of residues is defined as a phase-separated core interaction cluster, used to characterize important regions in the target protein that may participate in multivalent interactions under both statistical association and evolutionary constraints. During implementation, this phase-separated core interaction cluster is stored as a set of residue numbers and a clear site correspondence is established with the wild-type amino acid sequence. This ensures that during subsequent mutation instance construction and phase-separation characteristic evaluation, it is possible to accurately determine whether any missense mutation occurs within the phase-separated core interaction cluster.

[0029] In S4, the corresponding residue interaction network is reconstructed for each mutant sequence in the missense mutation set, and the difference between the sum of the shortest path lengths of all node pairs and the wild-type residue interaction network is calculated as the network topology perturbation score.

[0030] For the constructed missense mutation set, each missense mutation in the set was processed independently. Specifically, based on the amino acid substitution information corresponding to the missense mutation, a unique corresponding mutant amino acid sequence was generated based on the wild-type amino acid sequence, ensuring that the sequence only changes the amino acid type at the mutation site, while the positions of other residues remain consistent with the wild type. Subsequently, using this mutant amino acid sequence as the query sequence, a homology sequence search was performed in a pre-set protein sequence database, with the search strategy consistent with the wild-type sequence homology search. The retrieved homologous protein sequences and the corresponding mutant amino acid sequences were then subjected to multiple sequence alignment (MSA). During the alignment process, the mutant sequence was used as a reference coordinate to align the corresponding amino acid sites in each homologous sequence. After the alignment was completed, the distribution of amino acid types at each amino acid site in the mutant sequence within the homologous sequence set was statistically analyzed based on the alignment results, and the co-occurrence relationship of amino acid type combinations between any two different sites was further analyzed. Using statistical association calculation rules consistent with those for constructing wild-type residue interaction networks, the correlation strength between each pair of sites in the mutant sequence was calculated. Amino acid sites were used as nodes, and the corresponding correlation strengths were used as the weights of the connection edges between nodes to construct the mutant residue interaction network corresponding to that mutant. This method ensures that each missense mutation corresponds to a residue interaction network with the same structural form as the wild type, but whose connection weights reflect the changes in statistical association after the mutation.

[0031] After obtaining the wild-type residue interaction network and the corresponding mutant residue interaction network, a unified topological distance quantization calculation process is performed for each pair of networks. Specifically, the set of nodes and the set of edges in the network are first defined, where nodes correspond to amino acid sites, and the edge weights are determined by the aforementioned statistical correlation strength. Then, in each residue interaction network, all possible node pair combinations are traversed, and the shortest path length between any two different nodes is calculated. During the shortest path length calculation, the edge weights are used as the basis for the path cost, ensuring that the path length reflects the propagation distance of the statistical correlation strength between nodes in the network structure. After calculating the shortest path length for all node pairs in a single network, all the obtained shortest path length values ​​are summed to obtain the total shortest path length for that network. This calculation process is first performed once for the wild-type residue interaction network to obtain the total shortest path length of the wild-type network; subsequently, the same calculation process is repeated for each mutant residue interaction network to obtain the total shortest path length of the corresponding mutant network. During implementation, consistency should be maintained between wild-type networks and mutant networks in terms of node numbering, connection definition, and path calculation rules.

[0032] After calculating the sum of the shortest path lengths of the interaction networks between wild-type and each mutant residue, a process for determining the network topology perturbation score is performed for each missense mutation. The sum of the shortest path lengths of the mutant residue interaction networks corresponding to that mutation is numerically differiated from the sum of the shortest path lengths of the wild-type residue interaction networks; the resulting difference is defined as the network topology perturbation score for that missense mutation. This perturbation score is used to quantitatively describe the impact of a single missense mutation on the overall density of the residue interaction network structure. Based on this, for the subset of mutations identified as having known phase-separation characteristics, quantitative experimental data on their corresponding phase-separation phenotypes are further obtained. This experimental data originates from objective measurements of phase-separation-related phenotypes under uniform experimental conditions. Numerical annotation processing is performed on the above experimental data, converting the raw experimental readings into continuous numerical forms and mapping them to a fixed numerical range through normalization to eliminate the influence of different experimental units and value ranges. The processed continuous values ​​serve as supervisory labels for the phase separation characteristics affecting the score. Together with the corresponding network topology perturbation scores and other input features, they are used to train the graph neural network model, thereby ensuring that a clear and reproducible mapping relationship is established between the model output score and the actual phase separation phenotype.

[0033] In step S5, a graph neural network prediction model is constructed, which takes the network topology perturbation score, allele frequency data, and identification information of the phase-separated core interaction clusters corresponding to each mutation as input features, and outputs the phase-separation characteristic influence score of the mutation.

[0034] The adjacency matrix of the constructed wild-type residue interaction network is used as the shared graph structure framework for all mutant instances. This adjacency matrix explicitly describes the statistical interaction relationships and connection strengths between amino acid residues in the target protein, and is used to fix the topological connection forms between residues. Based on this, a mutation instance graph corresponding to each mutation in the missense mutation set is constructed. In specific implementation, firstly, based on the amino acid site where the mutation occurs, the mutant node is located in the node index corresponding to the adjacency matrix. Only the mutant node is assigned the allele frequency data corresponding to the mutation to reflect the degree of occurrence of the mutation in the reference population. For nodes that have not undergone mutation, a preset default value is uniformly assigned. The default value adopts a fixed numerical form, such as setting it to zero, to indicate that the node does not carry mutation information in the current mutant instance. At the same time, based on the determination results of the aforementioned phase-separated core interaction cluster, a Boolean identifier is further assigned to the mutant node to clearly indicate whether the mutation occurs within the phase-separated core interaction cluster. The Boolean identifiers corresponding to nodes that have not undergone mutation also adopt a uniform default value. By employing the above method, while maintaining the graph structure, each mutation instance graph accurately reflects the site attributes and structural location attributes of the mutation at the node feature level. Furthermore, the network topology perturbation score pre-calculated for the mutation is stored as a global feature of the mutation instance graph. This global feature does not participate in inter-node propagation but is only used to describe the degree of perturbation caused by the mutation to the overall residue interaction network structure. Thus, the construction process of a mutation instance graph including a fixed adjacency matrix, mutation node features, and global perturbation features is completed, ensuring that each missense mutation corresponds to a graph structure input data with a unified structure but different features.

[0035] After constructing the mutation instance graph, the graph neural network model is further constructed and trained. The graph neural network model includes an input layer, several levels of graph feature propagation layers, a graph-level feature aggregation layer, and an output layer. The input layer receives node features, adjacency relationships, and global feature information from the mutation instance graph. During model training, a subset of mutations with phase-separation phenotypes identified in the mutation set is selected as the training sample set. Each training sample corresponds to a mutation instance graph and its corresponding phase-separation characteristic influence score label. At the start of training, the model first performs graph feature propagation on the mutation instance graph in each iteration, allowing node features to be aggregated multiple times under fixed adjacency constraints, thus forming a node-level representation that characterizes the mutation node and its structural environment. After completing a predetermined number of feature propagation layers, the graph-level feature aggregation operation integrates the representations of all nodes into a single graph-level representation vector, reflecting the overall residue interaction state under the mutation instance. Subsequently, this graph-level representation and the corresponding network topology perturbation score are input to the output layer to calculate the model's predicted score for the degree of phase-separation characteristic influence of the mutation. During training, the numerically quantified and normalized phase separation characteristic influence score is used as a supervision signal. Through multiple iterations, the internal connection weights of the model are continuously adjusted to gradually reduce the deviation between the model's output score and the true label. The training process terminates after reaching the preset number of iterations or the model convergence condition, ultimately yielding a graph neural network model for phase separation characteristic prediction. This training process enables the model to learn the intrinsic relationships between mutation sites, structural interaction environments, and overall network perturbations, and in subsequent applications, output stable and discriminative phase separation characteristic influence scores for unknown mutations.

[0036] In step S6, all mutations are sorted according to the phase separation characteristic impact score output by the graph neural network prediction model, generating a list of phase separation-related mutations arranged in descending order of potential impact.

[0037] The constructed mutation set is categorized, clearly distinguishing between the subset of mutations with experimentally labeled phase-separation characteristics and the remaining mutation sets without such labels. The unlabeled mutation set includes missense mutation instances for which phase-separation phenotypic data has not yet been obtained under current experimental conditions, but which possess complete mutation information, allele frequency data, and network topology perturbation scores. For this unlabeled mutation set, a pre-trained graph neural network prediction model with fixed parameters is invoked to perform prediction operations. For each unlabeled mutation, its corresponding mutation instance graph structure input data is read. This input data includes a fixed wild-type residue interaction network adjacency matrix, node features set for the mutation node, and the global network topology perturbation score corresponding to the mutation. The graph neural network model maintains the same feature propagation and convergence process as the training phase during the prediction phase, without updating the model's internal parameters. The model performs a complete forward inference process for each mutation instance graph, outputting a continuous numerical phase-separation characteristic impact score. This score is consistent with the label value range used in the training phase, quantifying the potential impact of the mutation on the phase-separation characteristics of the target protein. The predicted scores for all unlabeled mutations are recorded using unique identifiers for each mutation instance, forming a set of phase separation characteristic impact scores that corresponds one-to-one with the set of unlabeled mutations. This step enables a unified predictive evaluation of mutation instances lacking experimental phase separation phenotypic information, ensuring that each mutation receives a comparable numerical score.

[0038] After predicting the impact scores of phase separation characteristics on the unlabeled mutation set, the scoring results are further integrated. Specifically, the phase separation characteristic impact scores are retrieved from a subset of mutations with known phase separation characteristics. These scores are derived from the numerical and normalized processing of the aforementioned experimental phenotypic data and are consistent with the scores output by the graph neural network model in terms of numerical range and semantics. Subsequently, the scoring results of this known mutation subset are merged with the unlabeled mutation prediction score set obtained in step one to construct a complete score set covering all missense mutation instances in the mutation set. During the construction of the complete score set, each mutation instance is indexed using its unique identifier to ensure consistency and traceability between the score and mutation information. After the score set is constructed, all missense mutation instances are sorted according to their score values. The sorting rule uses a single deterministic order from high to low values, without introducing additional weights or grouping conditions. After sorting, the sorting results are output in the form of a structured list, which sequentially records the mutation identification information, the corresponding phase separation characteristic impact score, and the mutation's position number in the overall sort. This approach creates a complete, unified, and directly applicable list of phase-separation-related mutations that can be used for analysis and decision-making. This allows researchers to quickly identify mutation instances with high potential impact based on the scoring results, and to conduct subsequent experimental verification or functional studies accordingly.

[0039] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0040] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0041] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0042] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0043] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0044] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0045] If the aforementioned functions are implemented as software functional modules 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 application, in essence, or the part that contributes to the prior art, or a portion 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 application. 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.

[0046] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0047] In conclusion, the above description is only a preferred 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 within the protection scope of the present invention.

Claims

1. A method for predicting protein phase separation characteristics based on artificial intelligence, characterized in that, Includes the following steps: S1. Obtain the wild-type amino acid sequence of the target protein, and obtain the set of known missense mutations of the target protein and the corresponding allele frequency data based on the population variation database. S2. The wild-type amino acid sequence and homologous sequence are compared. The sequence entropy value of each site is calculated based on the distribution of amino acid types in the comparison results. At the same time, the wild-type residue interaction network is constructed based on the statistical correlation of amino acid types between sites. S3. Identify the amino acid residues corresponding to the nodes with the most connection edges and sequence entropy values ​​below a preset threshold in the wild-type residue interaction network, and define these residues together with their directly interacting neighboring residues as phase-separated core interaction clusters. S4. For each mutant sequence in the missense mutation set, reconstruct the corresponding residue interaction network and calculate the difference between the sum of the shortest path lengths of all node pairs and the wild-type residue interaction network as the network topology perturbation score. S5. Construct a graph neural network prediction model, taking the network topology perturbation score, allele frequency data, and identification information of the phase-separated core interaction clusters corresponding to each mutation as input features, and outputting the phase-separation characteristic influence score of the mutation. S6. Sort all mutations according to the phase separation characteristic impact score output by the graph neural network prediction model, and generate a list of phase separation-related mutations arranged in descending order of potential impact.

2. The method for predicting protein phase separation characteristics based on artificial intelligence according to claim 1, characterized in that, In step S1, obtaining the wild-type amino acid sequence of the target protein and obtaining the set of known missense mutations of the target protein and the corresponding allele frequency data based on the population variation database specifically includes: Based on the identifier of the target protein, the complete wild-type amino acid sequence is retrieved from a pre-set protein sequence database. All missense mutation records belonging to the protein are obtained from a public database containing information on the frequency of human genome variations. The specific amino acid position where the mutation occurs and the type of the mutated amino acid are extracted and integrated into mutation information. At the same time, the frequency data of this mutation information in the genome of the reference population are obtained. The extracted mutation information and occurrence frequency data are integrated to construct a mutation set and the corresponding allele frequency dataset, and a subset of mutations with known phase separation characteristics is identified in the mutation set.

3. The method for predicting protein phase separation characteristics based on artificial intelligence according to claim 1, characterized in that, In S2, constructing the wild-type residue interaction network specifically includes: Using wild-type amino acid sequences as query sequences, homologous sequence searches are performed in protein sequence databases to obtain a set of homologous protein sequences. Multiple sequence alignment algorithms are then used to align the wild-type sequences with the homologous protein sequences, generating alignment results that characterize the frequency of different amino acid types at each amino acid site. For each amino acid site in the alignment results, the sequence entropy value of the site is calculated based on the frequency distribution of various amino acid types observed at the site. At the same time, the co-occurrence frequency of various amino acid type combinations between any two different amino acid sites in the alignment results is analyzed. The correlation strength between each pair of sites is calculated using mutual information, and a wild-type residue interaction network is constructed with sites as nodes and correlation strength as edges.

4. The method for predicting protein phase separation characteristics based on artificial intelligence according to claim 1, characterized in that, In step S3, the amino acid residues corresponding to the nodes with the most connection edges and sequence entropy values ​​below a preset threshold in the wild-type residue interaction network are identified, and these residues, together with their directly interacting neighboring residues, are defined as phase-separated core interaction clusters, specifically including: The number of connection edges corresponding to each node in the wild-type residue interaction network and the sequence entropy value of each amino acid site are obtained. Based on the preset sequence entropy value screening conditions, the nodes that simultaneously satisfy the condition of having the most connection edges and a sequence entropy value lower than the entropy value screening conditions are selected from all nodes. In the wild-type residue interaction network, identify all neighboring nodes that have direct connection edges with the selected nodes. Group the selected nodes and the amino acid residues represented by their corresponding neighboring nodes into a set, which is defined as the phase-separated core interaction cluster.

5. The method for predicting protein phase separation characteristics based on artificial intelligence according to claim 1, characterized in that, In step S4, for each mutant sequence in the missense mutation set, the corresponding residue interaction network is reconstructed, and the difference between the sum of the shortest path lengths of all node pairs and the wild-type residue interaction network is calculated as the network topology perturbation score, specifically including: For each mutant amino acid sequence in the missense mutation set, the homologous protein sequence is obtained based on the sequence and multiple sequence alignment is performed. The mutant residue interaction network corresponding to the mutant is constructed. For the wild-type residue interaction network and each mutant residue interaction network, the shortest path length between all possible node pairs in the network is calculated based on the weight of the connecting edges in the network, and the shortest path length of each network is summed to obtain the total shortest path length of each network. The difference between the sum of the shortest path lengths of the mutant residue interaction networks and the sum of the shortest path lengths of the wild-type residue interaction networks is used as the network topology perturbation score corresponding to the mutation.

6. The method for predicting protein phase separation characteristics based on artificial intelligence according to claim 1, characterized in that, In step S5, a graph neural network prediction model is constructed, using the network topology perturbation score, allele frequency data, and identification information of the phase-separated core interaction clusters corresponding to each mutation as input features. The output score of the phase-separation characteristic impact of the mutation specifically includes: The adjacency matrix of the wild-type residue interaction network is used as the mutation instance graph framework. The allele frequency data corresponding to each node and the Boolean identifier indicating whether the mutation falls into the phase-separated core interaction cluster are used as node features. The network topology perturbation score is used as the global feature. Together, they constitute the graph structure input of the model. A graph neural network model is constructed. The graph structure input is divided into subsets of mutations with phase separation characteristics as inputs. The graph neural network is trained under supervision. The model output represents the score of the influence of phase separation characteristics.

7. The method for predicting protein phase separation characteristics based on artificial intelligence according to claim 6, characterized in that, The phase separation characteristic influence score is obtained by numerically labeling the phase separation phenotype corresponding to the mutation subset in the graph neural network model training, and then normalizing the data to map it into a fixed numerical range.

8. The method for predicting protein phase separation characteristics based on artificial intelligence according to claim 1, characterized in that, In step S6, all mutations are sorted according to the phase separation characteristic impact score output by the graph neural network prediction model, and a list of phase separation-related mutations arranged in descending order of potential impact is generated, specifically including: The mutation set other than the mutation subset with known phase separation characteristics is input into the graph neural network prediction model to obtain the corresponding phase separation characteristic influence score set. The phase separation characteristic influence scores of the mutation subset with known phase separation characteristics are merged to form a complete score set. All missense mutation instances in all mutation sets are arranged in descending order of the values ​​in the complete score set to generate a structured list.

Images