Protein prediction method and system for designing phase separation function of transcription factor

Abstract

The invention belongs to the field of artificial intelligence and biological calculation, and discloses a protein prediction method for designing a phase separation function of a transcription factor, and the method comprises the steps: carrying out the iterative optimization of a target protein sequence related to the transcription factor through employing a genetic algorithm, and obtaining a target protein sequence with the optimized phase separation function of the transcription factor; in the genetic algorithm, the fitness evaluation function of the target protein sequence is as follows: fitness = alpha * A-beta * B; according to the method, through an optimized fitness function, LLPS scores and pathogenicity scores of a target protein sequence are taken as core indexes of the fitness function, and a genetic algorithm is combined, so that accurate regulation and control of the transcription factor liquid-liquid phase separation (LLPS) capability are realized; meanwhile, the invention also provides a system for realizing the method.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and biological computing, and particularly relates to a protein prediction method and system for designing phase separation function of a transcription factor. BACKGROUND

[0002] Protein engineering is an important branch of biotechnology, aiming to modify the sequence and structure of natural proteins to endow them with new functions or improve their performance. In recent years, with the rapid development of artificial intelligence (AI) technology, AI-driven protein design has gradually become an important research direction in this field. AI technology, especially machine learning and deep learning algorithms, can process and analyze large amounts of protein sequence and structure data, thereby predicting the properties and functions of proteins and guiding the rational design of proteins. Through AI algorithms, researchers can quickly screen out protein mutants with potential functions, greatly improving the efficiency and success rate of protein engineering.

[0003] Transcription factors are key proteins that regulate gene expression in cells. They bind to specific sequences on DNA to regulate the transcriptional activity of genes. Recent studies have shown that liquid-liquid phase separation (LLPS) plays a crucial role in chromatin remodeling and gene expression regulation (Bojia et al. Cell, 2018). LLPS refers to the process by which proteins or polypeptides form membraneless organelles or condensates in cells through weak interactions. These condensates can enrich specific transcription factors and co-factors, thereby enhancing the efficiency and specificity of gene expression. For example, in some cases, transcription factors can form transcriptional condensates through phase separation, promoting the simultaneous activation of multiple genes. Therefore, regulating the phase separation ability of transcription factors can become a new strategy for precise regulation of gene expression, thereby affecting the function and behavior of cells.

[0004] However, existing technologies only focus on the regulation of transcription factor expression and do not involve the dynamic balance mechanism of LLPS: for example, excessive phase separation can lead to abnormal aggregation of condensates, causing abnormal function of transcription factors; at the level of protein engineering, traditional methods reduce the activity of transcription factors through site-directed mutagenesis, but have the following defects:

[0005] Dependence on "hot spot" sites (such as charged residue substitution) in IDR regions, which cannot systematically optimize the multivalent interaction network;

[0006] Adopting conservative substitution strategies (such as D / E→A), which leads to a cliff-like decline in phase separation ability and affects the function of transcription factors;

[0007] Lack of quantitative evaluation of the pathogenicity of mutants, which may introduce unexpected cytotoxicity.

[0008] The regulation of the phase separation function of the transcription factor can refer to the following technologies:

[0009] It is recorded in the article "Protein Phase Separation Series Technology Service: Phase Separation Formation Ability Prediction and Identification, and Phase Separation Function Verification" that protein phase mutant and mutant complement construction and imaging verification; in order to verify the direct influence of phase separation on protein function, it is usually necessary to destroy the protein phase formation ability and complement the phase formation ability after the function verification; after obtaining the information of the key phase site of the protein, the mutation of the key phase site of the protein is designed under the premise of avoiding the important functional domain of the protein, the phase separation formation ability is destroyed, and the phase separation formation ability is complemented by fusing and expressing short peptides (such as FUS IDR, hnRNPac1 IDR, polyG, etc.) of proteins known to have phase separation formation ability. Then, the related eukaryotic and prokaryotic expression vectors are constructed, and the phase separation ability of the phase mutant is verified in cells and in vitro. The scheme regulates the phase separation ability by using short peptides with known phase separation ability for complementation. The limitation of the scheme lies in that the exogenous short peptides used for complementation may not completely simulate the phase separation behavior of the endogenous protein, and the expression level is difficult to accurately control, which easily leads to non-physiological aggregation or phase change, thereby affecting the accuracy of the function verification; in addition, the phase separation of part of the proteins depends on multivalent interaction or post-translational modification, and single phase site mutation is not enough to completely destroy the phase separation ability, thereby limiting the application range of the strategy.

[0010] In summary, the existing technical system has the following technical bottlenecks:

[0011] Phase separation regulation lacks quantitative means, and safety and effectiveness are difficult to balance; protein engineering cycle is long, and cannot be quickly iterated and optimized. These problems seriously restrict the clinical application transformation of cell rejuvenation technology. Therefore, it is urgent to establish a reversible, controllable and precise phase separation regulation strategy, and to develop a new type of engineering platform that can dynamically adjust the LLPS intensity while maintaining the integrity of the function of the transcription factor, which has become a technical bottleneck that needs to be broken through in the field of cell therapy based on transcription factors. SUMMARY

[0012] In view of the deficiencies of the prior art, the purpose of the present application is to provide a protein prediction method for designing the phase separation function of a transcription factor, which optimizes the fitness function, takes the LLPS score and pathogenicity score of the target protein sequence as the core indicators of the fitness function, and realizes the precise regulation of the liquid-liquid phase separation (LLPS) ability of the transcription factor by combining a genetic algorithm;

[0013] Meanwhile, the present application also provides a system for realizing the method.

[0014] To achieve the purpose of the application, the following technical solutions are adopted:

[0015] A protein prediction method for designing phase separation function of transcription factors, which adopts genetic algorithm to iteratively optimize target protein sequences related to transcription factors, to obtain target protein sequences optimized for phase separation function of transcription factors;

[0016] In the genetic algorithm, the fitness evaluation function of the target protein sequence is:

[0017] Fitness = alpha * A - beta * B;

[0018] Where alpha and beta are weight coefficients, A is the LLPS score, and B is the pathogenicity score.

[0019] The pathogenicity control of the application adopts a double security mechanism: the first is at the mutation operation level, and the mutation type is selected from the pre-constructed benign mutation library; the second is at the fitness evaluation level, and the overall safety of the sequence is quantitatively evaluated through the pathogenicity score B. Even if each mutation of the sequence itself is benign, their combination may produce unexpected synergistic pathogenic effects. Therefore, the -beta * B term in the fitness function serves as a penalty term, ensuring that the algorithm can filter out sequences that meet the local benign mutation condition but have a high overall pathogenic risk score while optimizing the LLPS capability, thereby achieving a balance between functionality and safety.

[0020] Genetic Algorithm (Genetic Algorithm, GA) is designed according to the evolution law of organisms in nature, and is a computational model simulating the natural selection and genetic mechanism of Darwin's biological evolution process. It is a method of searching for optimal solution by simulating the natural evolution process.

[0021] The basic operation process of genetic algorithm is as follows: (1) initialization: set the evolution generation counter t=0, set the maximum evolution generation T, and randomly generate M individuals as the initial population P(0).

[0022] (2) Individual evaluation: calculate the fitness of each individual in the population P(t).

[0023] (3) Selection operation: apply the selection operator to the population. The purpose of selection is to directly inherit the optimized individuals to the next generation or produce new individuals through pairing and crossing to be inherited to the next generation. Selection operation is based on the fitness evaluation of individuals in the population.

[0024] (4) Cross operation: apply the crossover operator to the population. The core role in genetic algorithm is the crossover operator.

[0025] (5) Mutation operation: the mutation operator is applied to the population. That is, the gene value of some loci of the individual string in the population is changed. After selection, crossover and mutation operation, the next generation population P(t+1) is obtained.

[0026] (6) Termination condition judgment: if t=T, the individual with the maximum fitness obtained in the evolution process is output as the optimal solution, and the calculation is terminated.

[0027] Genetic operations include the following three basic genetic operators: selection; crossover; mutation.

[0028] The core innovation of the present application is that by optimizing the fitness function, the LLPS score and the pathogenicity score of the target protein sequence are used as the core indicators of the fitness function, and the genetic algorithm is combined to realize the precise regulation of the liquid-liquid phase separation (LLPS) ability of the transcription factor.

[0029] It should be noted that: the LLPS score and the LLPS ability are not absolutely equivalent concepts; the LLPS score refers to the distance between the phase separation function required by the research and development personnel and the phase separation function of the target protein sequence itself, the greater the distance, the lower the score; in some application scenarios, if the value of the LLPS ability is lower, the lower the LLPS score represents the higher the value of the LLPS ability, and vice versa, if the value of the LLPS ability is higher, the lower the LLPS score represents the lower the value of the LLPS ability. In more application scenarios, we need the LLPS ability of the target protein sequence to be at a specific value, and the higher the LLPS score represents the farther the LLPS ability deviates from the set value.

[0030] In the above protein prediction method, the method comprises the following steps:

[0031] Step 1: constructing a sequence set containing multiple sequence different target protein sequences;

[0032] Step 2: iterative optimization of the sequence set; each round of iterative optimization includes fitness evaluation, selection operation, crossover operation and mutation operation on the target protein sequence to obtain a new sequence set;

[0033] Step 3: output the target protein sequence with a high fitness score after the iterative optimization is completed.

[0034] More specifically, in step 1, in addition to the wild-type sequence, several known mutated or engineered initial sequences can also be included, and these sequences are combined to form part of the initial population; if the number of sequences possessed is insufficient to reach the set population size, a certain number of random mutations are made on the basis of the wild-type sequence to expand the population, and such random mutations should follow certain prerequisites, such as random mutations based on a CSV file of non-pathogenic mutation information, which records the amino acid species at different positions in the wild-type sequence, which amino acids can be replaced at this position, and such replaceable amino acids meet the requirements of benign mutations.

[0035] In the above protein prediction method, step 2 is specifically:

[0036] Step 21: fitness evaluation, using a fitness evaluation function to evaluate the fitness score of the target protein sequences in the sequence set;

[0037] Step 22: selection operation, selecting several target protein sequences with high fitness scores in the sequence set as the parent set;

[0038] Step 23: crossover operation, randomly selecting key region fragments of two target protein sequences in the parent set to perform crossover operation to generate offspring target protein sequences; the key region fragments are pre-selected sequence fragments in the protein intrinsic disordered region;

[0039] Step 24: mutation operation, randomly mutating the amino acids in the protein intrinsic disordered region of the offspring target protein sequences to obtain new target protein sequences;

[0040] The target protein sequences obtained in step 24 constitute a new sequence set.

[0041] As known in the art, the target protein sequence related to transcription factors includes functional domains and protein intrinsic disordered regions (IDRs); the functional domain is a functional unit that can exist independently in a protein molecule and can independently perform biological functions such as binding and catalysis, and has independent folding ability, functional self-consistency, and structural diversity; in contrast, the protein intrinsic disordered region refers to a region of random structure in a protein molecule, also known as an unordered fragment. Compared with the functional domain of the protein, the unordered region lacks a clear secondary structure (such as an alpha helix or a beta sheet) and exhibits high structural flexibility and diversity. Although it was once considered to be a useless part, with further research, people have gradually recognized the importance of protein intrinsic disordered regions. Although such protein sequences lack stable structures and are highly variable, with further research, people have found that the presence of protein intrinsic disordered regions makes it easier for proteins to form droplets, inducing phase transition generation and regulation.

[0042] In the present application, the intrinsic disordered region of the protein is obtained by data mining of protein databases and related literature, combined with sequence analysis and prediction tools, and comprehensive data information from different sources.

[0043] In the present application, the key region fragment is a certain fragment of the intrinsic disordered region of the protein, which is obtained by manual or specific algorithm screening in the specific operation process. These key region fragments can be pre-selected by the operator, and the genetic algorithm will focus more on these regions during mutation and crossover, thereby realizing the directional optimization of the target functional domain. These key region fragments are exactly known fragments that have an impact on the LLPS ability or fragments that the operator predicts should have an impact on the LLPS ability. The length of the key region is generally 3 to 30 amino acids. In a target protein sequence, there is one or more key regions. When performing crossover operation of the genetic algorithm, two target protein sequences are randomly selected from the parent set, and a crossover point is randomly selected within one or more predetermined key region fragments for sequence fragment exchange.

[0044] In the above protein prediction method, the probability of the target protein sequence being selected in step 23 is positively correlated with the fitness score;

[0045] In step 24, the distance between the amino acid in the intrinsic disordered region of the protein and the core functional region is negatively correlated with the mutation probability of the amino acid; the core functional region is one or more sequences in the target protein sequence outside the intrinsic disordered region of the protein, and the core functional region is pre-set.

[0046] In the above protein prediction method, the pathogenicity score is the sum of the pathogenicity scores of the mutation site and the mutation type; the mutation site and the mutation type are obtained by comparing the target protein sequence with the wild-type target protein sequence.

[0047] In the above protein prediction method, the target protein sequence in the sequence set in step 1 includes the wild-type target protein sequence, the target protein sequence obtained after mutation or modification, and the target protein sequence obtained after further mutation.

[0048] In the above protein prediction method, after step 3, step 4 is further included.

[0049] Step 4: function verification and optimization, dry experiment verification, wet experiment verification are performed on the target protein sequence obtained in step 3, and the parameter setting and fitness evaluation function of the genetic algorithm are optimized according to the verification result;

[0050] The dry experiment verification is used to verify the nuclear localization, structural stability and LLPS score of the target protein sequence.

[0051] The wet experiment verification is used for verifying cell viability, phase separation dynamics and phenotype of the target protein sequence.

[0052] Meanwhile, the application also discloses a system for optimizing transcription factor phase separation function, comprising the following modules.

[0053] The target recognition module is used for constructing a sequence set containing multiple target protein sequences with different sequences, selecting an intrinsic disordered region and a core functional region of the target protein sequence, and setting at least one sequence fragment in the intrinsic disordered region as a key region.

[0054] The prediction module is used for iteratively optimizing the sequence set and outputting target protein sequences with high LLPS scores after the iterative optimization; each round of iterative optimization comprises fitness evaluation, selection operation, crossover operation and mutation operation on the target protein sequence, so as to obtain a new sequence set.

[0055] In the fitness evaluation operation, the fitness evaluation function of the target protein sequence is as follows:

[0056] Fitness = α*A-β*B.

[0057] Wherein, α and β are weights; A is the LLPS score of the target protein sequence; and B is a pathogenicity score.

[0058] The function verification and optimization module is used for performing dry experiment verification and wet experiment verification on the target protein sequence output by the prediction module, and optimizing the parameter setting and the fitness evaluation function of the genetic algorithm according to the verification result.

[0059] In the above system, in the crossover operation, the probability of selecting the target protein sequence is positively correlated with the fitness score; in the mutation operation, the distance between an amino acid in the intrinsic disordered region and the core functional region is negatively correlated with the mutation probability of the amino acid; the core functional region is one or more sequences in the target protein sequence outside the intrinsic disordered region; the pathogenicity score is the sum of the pathogenicity scores of the mutation site and the mutation type; and the mutation site and the mutation type are obtained by comparing the target protein sequence with a wild-type target protein sequence.

[0060] In the above system, the dry experiment verification is used for verifying the nuclear localization, structural stability and LLPS score of the target protein sequence; and the wet experiment verification is used for verifying the cell viability, phase separation dynamics and phenotype of the target protein sequence.

[0061] Meanwhile, the application further discloses an electronic device, comprising at least one processor and a memory connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to make the at least one processor execute the method according to any one of the above.

[0062] The application further discloses a non-transient computer readable storage medium storing computer instructions, wherein the computer instructions are used to make a computer execute the method according to any one of the above.

[0063] The application further discloses a computer program product, comprising computer program instructions, wherein the computer program instructions are executed by a processor to implement the method according to any one of the above.

[0064] Compared with the prior art, the application has the following beneficial effects:

[0065] (1) The application realizes precise regulation of the liquid-liquid phase separation (LLPS) ability of a transcription factor by using an optimized fitness function, taking the LLPS score and pathogenicity score of a target protein sequence as core indexes of the fitness function, and combining a genetic algorithm.

[0066] (2) The application successfully develops a quantitative regulation system for the phase separation characteristics of a protein, balances the LLPS intensity and the stability of a functional domain, and solves the problem of cell identity loss caused by excessive phase separation.

[0067] (3) The application constructs an AI-based protein engineering closed-loop platform, breaks through the limitation of traditional directed evolution technology on the modification of an intrinsic disordered region (IDR), and realizes multi-objective collaborative optimization of the phase separation characteristics and functional integrity of a transcription factor.

[0068] In summary, through the optimization of the application, the design cycle is shortened to 3 weeks (the traditional directed evolution needs 6 months), the wet experiment consumption is reduced, the success rate of mutant screening is improved, in the iterative process, the genetic algorithm explores 10^5 mutant combinations every round of iteration, and the design space exploration efficiency is improved by 500 times. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 is a schematic diagram of the method of the application;

[0070] Figure 2 is a comparison diagram of the protein structures of the OCT4 mutant designed by the application and the wild type OCT4 at different angles;

[0071] Figure 3 is a plasmid map of the plasmid PCDNA3..4-CMV-GFP-OCT4;

[0072] Figure 4 Plasmid map of plasmid PCDNA3.4-CMV-GFP-OCT4mut;

[0073] Figure 5 Static phase separation IF map for identifying phase separation function of Example 1, wherein (A) OCT4 phase separation; (B) OCT4mut phase separation; (C) OCT4 and OCT4mut group coagulation body number statistics results;

[0074] Figure 6 FRAP results for phase separation function identification, wherein (A) OCT4-FRAP; (B) OCT4mut-FRAP; (C) OCT4 and OCT4mut phase separation FRAP statistics chart;

[0075] Figure 7 Plasmid map of plasmid PLVX-CMV-3Xflag-OCT4-IRES-ZsGreen;

[0076] Figure 8 Plasmid map of plasmid PLVX-CMV-3Xflag-OCT4mut-IRES-ZsGreen;

[0077] Figure 9 PLVX-CMV-3Xflag-OCT4-IRES-ZsGreen plasmid FLAG expression IF verification result map;

[0078] Figure 10 PLVX-CMV-3Xflag-OCT4mut-IRES-ZsGreen plasmid FLAG expression IF verification result map;

[0079] Figure 11 PLVX-CMV-3Xflag-OCT4-IRES-ZsGreen and PLVX-CMV-3Xflag-OCT4-IRES-ZsGreen plasmid FLAG and OCT4 / OCT4mut co-expression WB verification result map in SBKM57 cells;

[0080] Figure 12A List of reads distribution of SBKM57 cell OCT4 / OCT4mut;

[0081] Figure 12B Column chart of reads distribution of SBKM57 cell OCT4 / OCT4mut;

[0082] Figure 12C Sequencing peak chart of CUT&TAG experiment of SBKM57 cell OCT4.

[0083] Figure 12D Sequencing peak plot for CUT&TAG experiment for OCT4mut in SBKM57 cells;

[0084] Figure 12E PCA analysis result plot for OCT4 / OCT4mut in SBKM57 cells;

[0085] Figure 12F Gene functional region distribution plot for OCT4 / OCT4mut in SBKM57 cells;

[0086] Figure 13 Protein structure alignment plot for SOX2 mutant designed in the application and SOX2 wild type structure from different angles;

[0087] Figure 14A Wild type SOX2-WT phase separation FRAP identification result;

[0088] Figure 14B Wild type SOX2-WT phase separation FRAP statistical plot;

[0089] Figure 14C SOX2 mutant SOX2-MUT6 phase separation FRAP identification result;

[0090] Figure 14D SOX2 mutant SOX2-MUT6 phase separation FRAP statistical plot;

[0091] Figure 14E SOX2 mutant SOX2-MUT10 phase separation FRAP identification result;

[0092] Figure 14F SOX2 mutant SOX2-MUT10 phase separation FRAP statistical plot;

[0093] Figure 14G SOX2 mutant SOX2-MUT11 phase separation FRAP identification result;

[0094] Figure 14H SOX2 mutant SOX2-MUT11 phase separation FRAP statistical plot;

[0095] Figure 14I SOX2-WT and mutant phase separation FRAP summary statistical plot;

[0096] Figure 15 Framework plot for Example 3. DETAILED DESCRIPTION

[0097] The technical solutions of the present application will be further illustrated by specific embodiments. Those skilled in the art should understand that the embodiments are only to help understand the present application and should not be regarded as specific limitations of the present application.

[0098] Embodiment 1

[0099] As shown in FIG. 1, the embodiment of the present application provides a technical solution for regulating the phase separation characteristics of OCT4. The amino acid sequence of wild-type OCT4 is numbered as uniprot:Q01860.

[0100] The present application iteratively optimizes protein sequences based on genetic algorithm (GA). The core process includes population initialization, fitness evaluation, selection, crossover, mutation, and loop iteration. The main goal is to obtain sequence variants with excellent performance in LLPS characteristics under certain mutation and screening rules. Specific parameters can be flexibly adjusted according to experimental requirements and feedback corrected according to the verification results after iteration.

[0101] The parameters of the genetic algorithm for OCT4 protein are set as follows:

[0102] Population size (POPULATION_SIZE): 50-100;

[0103] Maximum iteration number (GENERATIONS): 10-30;

[0104] Mutation probability (MUTATION_RATE): 0.05-0.20;

[0105] Early stopping rounds (EARLY_STOPPING_ROUNDS): 3-5 (stop iteration if no better result is obtained for consecutive rounds);

[0106] Maximum allowed mismatch sites (MAX_MIS): 20-40 (maximum allowed amino acid difference compared to wild type sequence);

[0107] The method flow includes the following steps:

[0108] Step 1: Construct a sequence set containing multiple OCT4 protein sequences with different sequences;

[0109] OCT4 gene is also known as Oct3, POU5F1, Oct3 or Oct4, which is a member of POU transcription factor family. Human Oct4 gene is located on chromosome 6 (6p21.31), with a length of 16.40 kb, multiple transcription start sites, and transcribes different mRNA isoforms (Isoform), thereby translating into multiple proteins.

[0110] The functional domains and the intrinsically disordered regions of the proteins are mined from the protein database and related literatures, and the data information from different sources is integrated to lock the target segments (i.e. the intrinsically disordered regions of the proteins) of the OCT4 transcription factor to be optimized as 1-52, 88-114 and 300-360;

[0111] The sequence set of the present application also includes a number of known mutated or modified initial sequences in addition to the wild-type sequence, and these sequences are combined to form part of the initial sequence set. If the provided sequences are insufficient to achieve the set population size, a certain number of random mutations are made on the basis of the wild-type sequence to expand the population. The mutations are only allowed to be mutations from aspartic acid (D) and glutamic acid (E) to alanine (A) and are limited within the maximum number of mismatches MAX_MIS, and in the specific implementation, the mutation rate is controlled by MUTATION_RATE; at the same time, the mutated sequence is compared with the wild-type sequence, and if the number of mismatches exceeds 40, the variant is discarded.

[0112] Step 2: Iterative optimization of the sequence set; each round of iterative optimization includes fitness evaluation, selection operation, crossover operation and mutation operation of the target protein sequence to obtain a new sequence set;

[0113] More specifically, the following steps are included:

[0114] Step 21: Fitness evaluation, the fitness score of the target protein sequence in the sequence set is evaluated by using a fitness evaluation function;

[0115] The fitness evaluation function is:

[0116] Fitness = α * LLPS score output by Seq2Phase model - β * aggregate pathogenicity score;

[0117] Wherein, α and β are weight coefficients, between 0 and 1, which are configured according to experience or other ways in the initial stage, and can be corrected according to the iteration results in the subsequent iteration process;

[0118] aggregate pathogenicity score = the sum of alphamissense pathogenicity scores of all mutation sites and types;

[0119] Although the mutation operation of the present embodiment is limited to the known low-risk mutation of D / E→A, the potential impact of multiple mutation combinations on the overall stability and function of the protein can be evaluated by calculating the total pathogenicity score. In the fitness evaluation, this score is used as a key negative indicator to eliminate candidate sequences that may have potential cytotoxicity due to mutation accumulation.

[0120] AlphaMissense is a machine learning model developed by DeepMind based on AlphaFold, primarily designed to predict the pathogenicity of missense mutations. AlphaMissense analyzes protein structures and evolutionary conservation to predict the pathogenicity of missense mutations. It combines the protein structure prediction capabilities of AlphaFold2 with sequence language modeling techniques, enabling the classification of 89% of mutations without experimental data (benign / pathogenic).

[0121] To evaluate the potential performance of each sequence in terms of LLPS, this embodiment calls the external tool Seq2Phase and writes the sequences to be evaluated in batch to a temporary FASTA file through the run_seq2phase function, then calls the tool to calculate the LLPS-related score of each sequence, and finally parses the output results. If the external tool fails to run or cannot produce valid output, the score of the sequence is recorded as invalid or excluded.

[0122] Seq2Phase; a deep learning-based sequence function prediction tool that integrates protein sequence features and phase separation (LLPS)-related physical and chemical properties such as aromatic residue distribution, low complexity regions, and charge patterns to build a neural network model to predict the tendency of proteins to undergo liquid-liquid phase separation. Its input is a FASTA-formatted amino acid sequence, and its output is an LLPS score between 0 and 1, with higher values indicating stronger phase separation potential. The tool is trained and validated on multiple known phase separation protein datasets and has good generalization ability. Seq2Phase captures local sequence patterns through convolutional neural networks and models long-range dependencies between residues using long short-term memory networks (LSTM), further introducing an attention mechanism to weight the contributions of key functional regions. During model training, a multi-task learning strategy is used to jointly predict phase separation tendency and known functional domain annotations, improving prediction robustness. When running, the tool parses the input sequence into sliding window segments, extracts physicochemical features for each segment, and inputs them into the pre-trained model. Finally, the comprehensive score is output through a logistic regression layer. By standardizing the input sequence, removing non-standard amino acid symbols, and filling in missing items, the sequence is then divided into segments based on the pre-set window size. Each segment is converted into a high-dimensional vector by a feature encoder, containing 17-dimensional physicochemical features such as hydrophobicity, charge distribution, aromaticity, and secondary structure tendency. The encoded sequence matrix is then input into the convolutional and LSTM layers to extract local motifs and long-range interaction patterns, with the attention mechanism automatically focusing on key phase separation regions. Finally, the features from each layer are fused and output by a fully connected network to obtain the normalized LLPS score.

[0123] All evaluated sequences are ranked by LLPS score. Since this example assumes that "lower score means better performance", the top ranked sequences (e.g. top 50%) are kept as high priority candidates for the next round of genetic operations. Thus, low-score sequences are considered as better candidates for evolution, and are given higher probability to participate in offspring generation in the following selection, crossover and mutation operations. This selection strategy not only maintains population diversity, but also accelerates convergence to the target phenotype space. In each iteration, the high priority sequences are determined for parent pairing by roulette wheel selection, crossover operation exchanges sequence fragments with a preset probability to simulate the natural recombination process, and mutation randomly introduces single amino acid substitution or short fragment insertion to enhance exploration ability. All offspring sequences are re-scored by Seq2Phase to form a new generation of population, and the cycle is executed until the preset convergence threshold is reached or the maximum number of iterations is reached. The final output of the optimal sequence and its LLPS score is the evolution endpoint.

[0124] Step 22: Selection operation, select a number of target protein sequences with high fitness score in the sequence set as the parent set;

[0125] In the sorting result, a number of sequences with better fitness (i.e. LLPS score) are selected as the parent set for subsequent crossover operation.

[0126] Step 23: Crossover operation, randomly select key region fragments of two target protein sequences in the parent set for crossover operation to generate offspring target protein sequences; the key region fragments are sequence fragments pre-selected in the protein intrinsic disordered region;

[0127] In this embodiment, the crossover_sequence function is called, and the corresponding key region fragment of parent 2 is selected according to the 50% probability to replace the key region fragment of parent 1.

[0128] The crossover_sequence function: The crossover_sequence function randomly determines the crossover starting site within the predefined key region boundary based on the alignment position of the parent sequences, and determines whether to enable fragment exchange with a probability of 50%; if the crossover is triggered, the amino acid sequence of the corresponding interval of parent 2 is inserted into parent 1 to generate a hybrid offspring sequence; this process simulates the biological recombination mechanism, while preserving the integrity of the domain and introducing genetic diversity. When the function is executed, first, the key region boundaries of the parent sequences are aligned to ensure that the crossover range is limited within the disordered domain; then a starting site index is randomly generated within the boundary interval, and a uniform distribution random number is used to determine whether to trigger the exchange (p=0.5); if true, the fragment of parent 2 from the starting site to the end is replaced in the corresponding part of parent 1, otherwise the original sequence of parent 1 is retained; finally, the recombined offspring sequence is output. This operation maintains the stability of the overall sequence architecture while effectively introducing local variation, avoiding local optimization. The crossover offspring sequence retains the functional domain characteristics of the parent, and may produce new synergistic effects due to fragment recombination, improving the LLPS tendency. This process is repeated until a sufficient size of offspring population is generated for the next mutation operation.

[0129] Step 24: Mutation operation, random mutation on the offspring sequence generated by the crossover operation;

[0130] In this embodiment, it is preferred to continue the random mutation based on the benign mutation dictionary on the offspring sequence generated by the crossover;

[0131] The mutation operation of this embodiment follows two principles: First, to ensure functional stability, the mutation probability is associated with the distance of the amino acid site from the core functional region. Specifically, the amino acid located in the intrinsically disordered region (IDR) of the protein, the closer the sequence distance from the nearest core functional region (POU domain (138-212) and Homoeodomain (230-292)), the lower the probability of being selected as a mutation site, which aims to protect the structural stability near the core functional region. Second, to control the risk of pathogenesis, the selection of mutation type is limited. In this embodiment, only specific benign mutations are allowed, i.e., mutating aspartic acid (D) and glutamic acid (E) to alanine (A). All mutations must ensure that the total number of mismatches between the final sequence and the wild type does not exceed the preset MAX_MIS threshold.

[0132] The mutate_sequence function: The mutate_sequence function traverses the amino acid sites of the unordered region of the offspring sequence, mutates all sites, but the mutation direction is the non-pathogenic mutation tolerated by the site saved in the benign_mutation_dict; for each mutable site, the function randomly selects a replacement amino acid according to the predefined benign replacement set, and controls the overall variation degree by combining the mismatch threshold, to ensure that the total number of mutations of a single sequence does not exceed the limited value; this process uses uniform random sampling to maintain diversity while avoiding damage to key functional residues; each mutation preserves the original sequence skeleton, and only the non-pathogenic residues in the "benign_mutations_dict" are mutated, thereby maintaining structural stability and folding potential while enhancing sequence diversity. This function limits the mutation range to known tolerant sites, combined with dynamic mismatch number monitoring, to ensure that the variation is within the safety threshold; the counter is updated after each mutation until the preset upper limit or no replacement site is reached; the final output sequence retains the core characteristics of the parent while introducing moderate diversity, providing an evolutionally potential candidate pool for subsequent LLPS function optimization. After the mutation operation, the offspring sequence maintains the overall conformational stability, and the adjustment of the local amino acid composition effectively optimizes the charge distribution and hydrophobicity balance, further enhancing its liquid-liquid phase separation (LLPS) tendency; all generated sequences are filtered by quality, eliminating individuals with high polymerization risk or damaged key functional domains in the unordered region, and retaining candidates with high phase separation potential and structurally soluble proteins; the final output sequence set has both genetic diversity and functional conservation, providing a theoretical priority ranking basis for subsequent high-throughput experimental verification.

[0133] The target protein sequence obtained in step 24 constitutes a new sequence set.

[0134] Step 3: After the end of the iterative optimization, output the target protein sequence with the highest fitness score.

[0135] Each generation will perform the cycle operation of "fitness evaluation → selection → crossover → mutation → population update", and the best LLPS score and corresponding sequence of the current generation will be counted;

[0136] If there is no score better than the historical best for consecutive EARLY_STOPPING_ROUNDS rounds, the algorithm is determined to have converged and is stopped early, saving computing resources.

[0137] After the iteration is completed, the system will output a number of preferred sequences (such as the top 30) ranked, record their LLPS scores, pathogenicity scores, and mismatch numbers with the wild-type sequence, etc. These preferred sequences will be used as a candidate set for downstream functional verification;

[0138] Step 4: Functional Prediction and LLPS Validation

[0139] For mutants that have entered this stage, computer simulation or biophysical analysis methods are used to predict their functional performance, including but not limited to structural stability and aggregation tendency related to protein phase separation behavior; the experiment is divided into two parts: dry experiment and wet experiment.

[0140] The verification results of dry and wet experiments further provide feedback for adjusting the parameters of the genetic algorithm and the weight coefficients of the fitness function, forming a closed-loop optimization mechanism.

[0141] Dry experiment:

[0142] The optimized sequence (SEQ ID NO.1) had a phase separation ability score of (0.5), which was significantly lower than the wild type score (0.9). At the same time, the aggregate pathogenicity score was 4.41 and the nuclear localization score was 0.96.

[0143] SEQ ID NO.1:

[0144] MAGHLASDFAFSPPPGGGGDGPGGPAPGWVAPRTWLSFQGPPGGPGIGPGVGPGSAVWGIPCPPPYAFCGGMAYCGPQVGVGLVPQGGLATSQPAGAAGVGVASNSAGASPAPCTVTPGAVKLAKAKLAQNPAASQAIKALQKELEQFAKLLKQKRITLGYTQADVGLTLGVLFGKVFS QTTICRFEALQLSFKNMCKLRPLLQKWVEEADNNENLQEICKAETLVQARKRKRTSIENRVRGNLENLFLQCPKPTLQQISHIAQQLGLEKDVVRVWFCNRRQKGKRSSSAYAQRAAFAAAGSPFSGGPVSFPLAPGPHFGTPGYGSPHFTALYSSVPFPEGEAFPPVSVTTLGSPMHSN

[0145] like Figure 2The structure of the core region (POU domain (138-212) and Homoeodomain (230-292)) complex of the mutant and wild type nucleosomes, predicted by Alphafold3, is shown. The structural alignment RMSD = 3.31. Compared with the wild type, the mutant shows little change in the core region structure (RMSD = 3.31), the DNA binding position remains unchanged, and no new structures appear that bind to DNA; this indicates that the mutation does not disrupt its key DNA binding ability. Further molecular dynamics simulations show that the mutant's oligomerization ability is weakened in the early stages of droplet formation, leading to a decrease in the phase separation driving force, which is consistent with the lower LLPS score.

[0146] Figure 2 The structures of the OCT4 mutant designed for this invention were compared with those of the wild-type OCT4, and the structures were predicted using Alphafold3. Yellow represents the core region of the OCT4 mutant, red represents the core region of the wild-type OCT4, green represents DNA, and gray represents histones. The structural comparison results show that, compared to the wild-type, the mutant exhibits minimal structural changes in its core region (RMSD = 3.31), the DNA binding site remains unchanged, and no other new structures are observed.

[0147] Wet experiment:

[0148] 1. Identification of phase segregation function in OCT4 mutants

[0149] 1.1 Plasmid Synthesis

[0150] Genewiz (Suzhou Genewiz Biotechnology Co., Ltd.) synthesized the GFP-OCT4 plasmid fragment and the GFP-OCT4mut plasmid fragment and cloned them into the pCDNA3.4 vector to obtain plasmids PCDNA3.4-CMV-GFP-OCT4 and PCDNA3.4-CMV-GFP-OCT4mut. Figure 3 and Figure 4 After the DNA sequence was confirmed to be correct, subsequent experiments were conducted.

[0151] 1.2 Cell Culture

[0152] U-2 OS cells were cultured in a 37°C, 5% CO2 incubator using U-2 OS ready-to-use complete culture medium (CELLCOOK, CC2005M). When the cells reached 90% confluence, they were digested with 0.05% Trypsin-EDTA (Gibico, 25300054), counted, and cultured at a concentration of 1.0 × 10⁻⁶ cells / year. 5 The wells were seeded at a density of / wells in 35 mm glass-bottom confocal culture dishes (Cellvis, D35-20-1.5P), and plasmid transfection was performed the next day.

[0153] 1.3 Cell transfection

[0154] According to the instructions, plasmids PCDNA3.4-CMV-GFP-OCT4 and PCDNA3.4-CMV-GFP-OCT4mut were transfected into U2OS cells using lipofectamine 3000 (Thermo Fisher).

[0155] The specific method is as follows: Add 1.0 μg of plasmid to 50 μL Opti-MEM and mix thoroughly; take another tube, add 50 μL Opti-MEM and 1.0 μL lipofectamine 3000, mix the two tubes thoroughly, incubate at room temperature for 15 minutes, and add to the corresponding cells prepared the day before (from step 1.2).

[0156] 1.4 Observation of Static Phase Separation

[0157] Forty-eight hours after transfection, cells were fixed with 4% paraformaldehyde (Beyotime, P0099), washed thoroughly with PBS, and then stained with DAPI (Beyotime, P0131) for nuclear staining and mounting. The morphology and aggregation of GFP in the nucleus were observed using a Nikon A1 confocal fluorescence microscope under 100x oil immersion. OCT4 existed in the nucleus as highly condensed droplets, indicating that OCT4 exhibits classic phase-separation characteristics; while OCT4mut was diffusely distributed, indicating that amino acid modification could disrupt its phase-separation characteristics, as shown in the results. Figure 5 As shown.

[0158] 1.5 Dynamic phase separation observation (FRAP, fluorescence recovery after bleaching technique)

[0159] Cells were transfected using the same method, and the GFP signal in the nuclei of live cells was dynamically observed under a Nikon A1 confocal fluorescence microscope. The GFP fluorescent spots were first quenched using a 2-second laser (excitation light 488nm, fluorescence intensity 21%), and then recorded every 4 seconds within a 70-second observation window. It was found that OCT4 exhibited rapid recovery, while the fluorescent spots of OCT4mut did not recover after quenching. The results are as follows: Figure 6 As shown;

[0160] 2. CUT & TAG identification of OCT4 mutant genome regulatory function

[0161] 2.1 CUT & TAG plasmid synthesis

[0162] Due to the lack of ChIP-level OCT4 and OCT4mut antibodies, PLVX-CMV-3Xflag-OCT4-IRES-ZsGreen was designed to investigate the binding site of the OCT4 mutant in the genome (plasmid map reference). Figure 7 ) and PLVX-CMV-3Xflag-OCT4mut-IRES-ZsGreen (plasmid map reference) Figure 8 The plasmid was used for FLAG-CUT&TAG and was synthesized by Genewiz (Suzhou Genewiz Biotechnology Co., Ltd.). After enzyme digestion and DNA sequencing were confirmed to be correct, subsequent experiments were carried out.

[0163] 2.2 Validation of FLAG expression via CUT & TAG plasmid transfection

[0164] U2OS cells were transfected with PLVX-3Xflag-OCT4-IRES-ZsGreen and PLVX-3Xflag-OCT4mut-IRES-ZsGreen plasmids using lipofectamine 3000 (Thermo Fisher); the method was the same as in 1.3.

[0165] 2.2.2 IF Validation of FLAG Expressions

[0166] Forty-eight hours after transfection, fixation was performed using 4% paraformaldehyde (Beyotime, P0099). After thorough washing with PBS, incubation with primary antibody was performed: the OCT4 group was co-incubated with OCT4 (Santa Cruz, SC-5279) and FLAG (sigma, F1804) antibodies, while the OCT4mut group was incubated with FLAG antibody. After secondary antibody staining, nuclear staining and mounting were performed using DAPI (Beyotime, P0131). Observation was conducted using a Nikon A1 confocal fluorescence microscope under 100x oil immersion: OCT4 and FLAG could co-localize in the nucleus, and FLAG expression in the OCT4 group showed phase separation characteristics. Figure 9 In contrast, the OCT4mut group exhibited a diffuse distribution without showing an aggregate morphology. Figure 10 ).

[0167] 2.3 FLAG-CUT & TAG Validation of OCT4 Mutant Gene Regulatory Function

[0168] 2.3.1 Plasmid transfection

[0169] Skin fibroblast SBKM57 cells were cultured in DMEM (Gibico, 10566-016) complete medium at 37°C in a 5% CO2 incubator. When the cells reached 90% confluence, they were digested with 0.05% Trypsin-EDTA, counted at 5.0 × 10⁵ cells / sample, and resuspended in electroporation buffer (Lonza, VPD1001). 6 μg of the corresponding plasmid (PLVX-CMV-3Xflag-OCT4-IRES-ZsGreen or PLVX-CMV-3Xflag-OCT4mut-IRES-ZsGreen) was added, and electroporation was performed using a Lonza-2B Nucleofector electroporator with the P22 program. After electroporation, the cells were cultured again at 37°C in a 5% CO2 incubator. Each experiment was repeated three times.

[0170] 2.3.2 Western blot verification of gene expression

[0171] After 48 hours of electroporation and cell culture, the cells were digested with 0.05% Trypsin-EDTA (Gibico, 25300054), and a portion of the cells were harvested for Western blotting to verify the co-expression of FLAG (sigma, F1804) and OCT4 / OCT4mut (Santa Cruz, SC-5279). The results are as follows: Figure 11 As shown.

[0172] 2.3.3 CUT & TAG Experiment

[0173] After centrifuging and collecting the remaining cells from the previous step, CUT&TAG library construction was performed (using Sigma-Aldrich (F1804) as the library construction antibody and Novizan (TD903) as the library construction kit). After the library passed the inspection, Illumina sequencing was performed. The sequencing results showed PCA clustering that there was a significant separation between the OCT4 mutant and the OCT4 group. Figure 12E Furthermore, the binding sites of OCT4 and OCT4mut groups on the genome changed from clustered to diffuse. Figures 12A to 12D ).

[0174] Further analysis showed that OCT4mut was significantly less enriched in the promoter and enhancer regions. Figure 12FThis study used an in vitro cell model to verify the impact of OCT4's phase-separating ability on its transcriptional regulatory function. It confirmed that OCT4 specifically accumulates in key regulatory regions through phase-separation, thereby stably binding to and activating the expression of downstream pluripotent genes. Conversely, OCT4mut, lacking phase-separating ability, exhibits random diffusion of its genome binding, resulting in significantly weakened transcriptional regulatory function. This suggests that phase separation is a crucial molecular basis for OCT4's normal function. This finding not only reveals the close relationship between OCT4's phase-separating characteristics and its transcriptional regulatory function but also provides a new perspective for understanding the molecular mechanisms of pluripotency maintenance.

[0175] The above-mentioned dry and wet experiments further demonstrate the accuracy of the method of the present invention. Using the method of the present invention, the design cycle of related proteins is shortened to 3 weeks (traditional directed evolution requires 6 months), the consumption of wet experiments is reduced, and the success rate of mutant screening is improved.

[0176] Example 2

[0177] This embodiment is mainly used for regulating the phase separation properties of SOX2 protein; its main methods are largely the same as those in Example 1.

[0178] SOX2 is an important transcription factor that plays a crucial role in embryonic development and cell differentiation. Belonging to the SOX gene family, it is widely distributed in various organisms and plays a vital role in maintaining stem cell pluripotency and regulating gene expression. SOX2 stands for "SRY-related HMG-box 2," named for its similarity to the SRY (sex-determining region Y) gene. The SOX2 protein contains a highly conserved HMG (high-mobility family) box domain that binds to DNA and regulates the transcription of target genes. It is highly expressed in embryonic stem cells, neural stem cells, and other pluripotent cells, and is one of the core factors maintaining the undifferentiated state of these cells.

[0179] The differences from Example 1 are as follows:

[0180] The amino acid sequence length of the SOX2 protein is 317 aa; its functional domains are 39-118; and the protein's intrinsic disorder regions are (1, 43), (243, 266), (297, 317), etc.

[0181] By data mining of protein databases and related literature, combined with sequence analysis and prediction tools, and by integrating data information from different sources, the target regions for optimization of the SOX2 transcription factor were identified as 1-39 and 200-317.

[0182] In this embodiment, the mutation operation (or benign mutation) on the offspring sequence also follows two core principles:

[0183] First, a distance-based non-uniform mutation probability is employed. Amino acids in the protein's intrinsic disordered regions have a lower probability of mutation the closer they are to the core functional region (HMG box (41-109)), thus protecting the integrity of critical functional modules. Second, benign mutations based on prior knowledge are used. Specific mutation sites and replaceable amino acid types are strictly limited to a pre-loaded dictionary, "benign_mutations_dict," which contains information on known non-pathogenic mutations. In this way, the algorithm significantly reduces the risk of introducing harmful mutations while exploring the sequence space, and ensures that the total number of mismatches remains within the MAX_MIS limit.

[0184] An example of the dictionary mentioned above is a pre-prepared CSV file containing information on known non-pathogenic mutations at each point of the target protein. This file includes the following fields: ref_aa: reference amino acid (i.e., wild-type amino acid); position: the position of the amino acid in the sequence; alt_aa: alternative benign mutant amino acids.

[0185] In this embodiment, the main parameters are set as follows:

[0186] Population size (POPULATION_SIZE): 100-200;

[0187] Maximum number of iterations (GENERATIONS): 30-50;

[0188] Mutation probability (MUTATION_RATE): 0.05-0.10;

[0189] Truncation probability (TRUNCATION_RATE): 0.00-0.10 (Sequence truncation is not enabled in this embodiment);

[0190] Early Stopping Rounds: 3-5 (If no better result is obtained for several consecutive generations, the iteration will stop);

[0191] Maximum allowed number of mismatch sites (MAX_MIS): 10-30 (the maximum allowed number of amino acid differences between the sequence and the wild type).

[0192] In step 4, the results of the wet and dry test are as follows:

[0193] Dry experiment: For mutants that have entered this stage, computer simulation or biophysical analysis methods are used to predict their functional performance, including but not limited to structural stability and aggregation tendency related to protein phase separation behavior; the phase separation ability score of the optimized sequence is (0.3), which is significantly lower than the wild type score (0.8), while the aggregatepathogenicity score is 7.79 and the nuclear localization score is 0.95.

[0194] like Figure 13 The structure of the core region (HMG box (41-109) mutant and wild type nucleosome complex predicted by Alphafold3 is shown. The structure comparison RMSD = 0.155. The position of DNA binding has not changed, no other new structures have appeared, and there are no other positions where DNA binds.

[0195] Figure 13 The structures of the SOX2 mutant designed for this invention were compared with those of the wild-type SOX2, and the structures were predicted by Alphafold3. Yellow represents the core region (HMG-BOX) of the SOX2 mutant, red represents the core region (HMG-BOX) of the wild-type SOX2, green represents DNA, and gray represents histones. The structural comparison results show that the mutant exhibits minimal structural changes in its core region compared to the wild-type (RMSD = 0.155), the DNA binding site remains unchanged, and no other new structures are observed.

[0196] Wet assay: Identification of phase segregation function in SOX2 mutants

[0197] 1. Plasmid Synthesis

[0198] The GFP-SOX2 plasmid and the GFP-SOX2mut (mut1-mut11) plasmid were synthesized by Genewiz (Suzhou Genewiz Biotechnology Co., Ltd.) and cloned into the pCDNA3.4 vector. The PCDNA3.4-CMV-GFP-SOX2 and PCDNA3.4-CMV-GFP-SOX2mut (1-11) plasmids were synthesized. After the DNA sequences were identified by enzyme digestion and confirmed by sequencing, subsequent experiments were carried out.

[0199] 2. Cell Culture

[0200] U2OS cells were cultured in a 37°C, 5% CO2 incubator using ready-to-use U-2 OS complete culture medium (CELLCOOK, CC2005M). When the cells reached 90% confluence, they were digested with 0.05% Trypsin-EDTA (Gibico, 25300054), counted, and seeded at a density of 1.0 × 10⁵ cells / well in 35 mm glass-bottom confocal culture dishes (Cellvis, D35-20-1.5P). Plasmid transfection was performed the following day.

[0201] 3. Cell transfection

[0202] According to the instructions, PCDNA3.4-CMV-GFP-SOX2 and PCDNA3.4-CMV-GFP-SOX2mut(1-11) plasmids were transfected into U2OS cells using lipofectamine 3000 (Thermo Fisher). 1.0 μg of plasmid was added to 50 μL of Opti-MEM and mixed thoroughly. In another tube, 50 μL of Opti-MEM and 1.0 μL of lipofectamine 3000 were added, and the two tubes were mixed thoroughly and incubated at room temperature for 15 minutes before being added to the corresponding cells.

[0203] 4. Dynamic phase separation observation (FRAP, fluorescence recovery after bleaching technique)

[0204] Cells were transfected using the same method, and the GFP signal in the nuclei of live cells was dynamically observed under a Nikon A1 confocal fluorescence microscope. The GFP fluorescent spots were first quenched using a 2-second laser (excitation light 488nm, fluorescence intensity 21%), and then recorded every 4 seconds within a 60-second observation window. It was found that GFP-SOX2 exhibited rapid recovery, while the fluorescent spots of MUT6 / 10 / 11 in the GFP-SOX2mut mutant showed reduced recovery ability after quenching. The results are as follows: Figures 14A to 14I As shown.

[0205] This indicates that the phase separation kinetics of SOX2 are significantly altered after specific mutations, particularly at MUT6, MUT10, and MUT11 sites, which may play a crucial role in maintaining the dynamic equilibrium of protein-liquid phase separation. This further suggests that the HMG box region not only mediates DNA binding but may also participate in regulating the behavior of SOX2 during the aggregation of nuclear biomolecules. This discovery provides a new perspective on elucidating the spatial organization mechanism of SOX2 in transcriptional regulation and lays experimental groundwork for revealing the molecular basis of its abnormal aggregation in development and disease. Combined with mutant localization analysis, it is speculated that the HMG box domain may affect the function of the disordered region of SOX2 through conformational changes, thereby regulating phase separation potential. This dynamic characteristic may be closely related to the rapid assembly and disassembly of transcription factors in enhancer clusters, suggesting that phase separation may serve as a potential mechanism for regulating the spatiotemporal specificity of gene expression.

[0206] Example 3

[0207] refer to Figure 15 A system for optimizing the phase separation function of transcription factors, comprising the following modules:

[0208] Target identification module 1: Used to construct a sequence set containing multiple target protein sequences with different sequences, select the protein's intrinsic disordered region and core functional region, and set at least one sequence fragment in the protein's intrinsic disordered region as the key region;

[0209] Prediction Module 2: This module iteratively optimizes the sequence set and outputs the target protein sequences ranked highest by LLPS score after the optimization process. Each round of iterative optimization includes fitness evaluation, selection, crossover, and mutation operations on the target protein sequences to obtain a new sequence set.

[0210] In the fitness evaluation operation, the fitness evaluation function for the target protein sequence is:

[0211] Fitness = α*A - β*B;

[0212] Where α and β are weights; A is the LLPS score of the target protein sequence; and B is the pathogenicity score.

[0213] The prediction module 2 includes the following sub-modules:

[0214] Fitness assessment submodule 21: Used to evaluate the fitness score of target protein sequences in the sequence set using a fitness assessment function;

[0215] Selection Submodule 22: Used to select several target protein sequences with high fitness scores in the sequence set as the parent set;

[0216] Crossover submodule 23: used to randomly select key region segments of two target protein sequences from the parent set for crossover operation to generate offspring target protein sequences; the probability of the target protein sequence being selected is positively correlated with the fitness score;

[0217] Mutation submodule 24: used to randomly mutate amino acids in the intrinsic disordered region of the progeny target protein sequence to obtain a new target protein sequence and form a new sequence set. The distance between the amino acid in the intrinsic disordered region of the protein and the core functional region is negatively correlated with the mutation probability of that amino acid.

[0218] In the aforementioned system, the dry experimental verification is used to verify the nuclear localization, structural stability, and LLPS score of the target protein sequence; the wet experimental verification is used to verify the cell viability, phase separation dynamics, and phenotype of the target protein sequence.

[0219] Functional verification and optimization module 3: Used to perform dry and wet experimental verification of the target protein sequence output by the prediction module, and optimize the parameter settings and fitness evaluation function of the genetic algorithm based on the verification results.

[0220] The functional validation and optimization module validates the optimized target protein sequence using both dry and wet experiments. Based on the validation results, it adjusts the parameters of the genetic algorithm and the weighting coefficients α and β in the fitness function to balance the priorities of phase segregation capability and biosafety, further improving the functional performance and stability of the target protein in a living cell environment. The validation results are fed back to the prediction module, forming a closed-loop optimization cycle, significantly improving design efficiency and success rate. Through multiple rounds of closed-loop iteration, the system gradually converges to an optimized sequence that combines high LLPS tendency with low cytotoxicity, significantly enhancing the functional robustness and application potential of the target protein under complex physiological environments.

[0221] The applicant declares that the present invention is illustrated by the above embodiments, but the present invention is not limited to the above process steps, nor does it mean that the present invention must rely on the above process steps to be implemented. Those skilled in the art should understand that any improvements to the present invention, equivalent substitutions of the raw materials used in the present invention, additions of auxiliary components, and selection of specific methods, etc., all fall within the protection scope and disclosure scope of the present invention.

Claims

1. A method for predicting the phase separation function of transcription factors, characterized in that, Genetic algorithms were used to iteratively optimize the target protein sequences related to transcription factors, resulting in target protein sequences with optimized transcription factor phase separation function. In the genetic algorithm, the fitness evaluation function for the target protein sequence is: Fitness = α*A - β*B; Where α and β are weighting coefficients; A is the LLPS score; and B is the pathogenicity score.

2. The protein prediction method according to claim 1, characterized in that, The method includes the following steps: Step 1: Construct a sequence set containing multiple target protein sequences with different sequences; Step 2: Iteratively optimize the sequence set; each round of iterative optimization includes fitness evaluation, selection operation, crossover operation, and mutation operation of the target protein sequence to obtain a new sequence set; Step 3: After the iterative optimization is completed, output the target protein sequences with the highest fitness scores.

3. The protein prediction method according to claim 2, characterized in that, Step 2 specifically involves: Step 21: Fitness assessment, using a fitness assessment function to evaluate the fitness score of the target protein sequences in the sequence set; Step 22: Selection operation, select several target protein sequences with high fitness scores in the sequence set as the parent set; Step 23: Crossover operation, randomly select two key region fragments of the target protein sequences from the parent set and perform crossover operation to generate the offspring target protein sequence; the key region fragment is a pre-selected sequence fragment in the disordered region of the protein; Step 24: Mutation operation, randomly mutating the amino acids in the disordered region of the progeny target protein sequence to obtain a new target protein sequence; The target protein sequence obtained in step 24 constitutes a new sequence set.

4. The protein prediction method according to claim 3, characterized in that, In step 23, the probability of the target protein sequence being selected is positively correlated with the fitness score; In step 24, the distance between an amino acid in the intrinsic disorder region of the protein and the core functional region is negatively correlated with the mutation probability of that amino acid; the core functional region is one or more segments of the target protein sequence located outside the intrinsic disorder region of the protein, and the core functional region is predetermined.

5. The protein prediction method according to claim 1, characterized in that, The pathogenicity score is the sum of the pathogenicity scores of the mutation site and the mutation type; the mutation site and mutation type are obtained by comparing the target protein sequence with the wild-type target protein sequence.

6. The protein prediction method according to claim 2, characterized in that, The target protein sequences in the sequence set in step 1 include wild-type target protein sequences, target protein sequences obtained after mutation or modification, and target protein sequences obtained after further mutation.

7. The protein prediction method according to claim 2, characterized in that, After step 3 is completed, step 4 is also included; Step 4: Functional verification and optimization. The target protein sequence obtained in Step 3 is verified by dry and wet experiments, and the parameters of the genetic algorithm and the fitness evaluation function are optimized based on the verification results. The dry experimental verification is used to verify the nuclear localization, structural stability, and LLPS score of the target protein sequence; The wet assay was used to verify the cell viability, phase separation dynamics, and phenotype of the target protein sequence.

8. A system for optimizing the phase separation function of transcription factors, characterized in that, Includes the following modules: Target identification module: used to construct a sequence set containing multiple target protein sequences with different sequences, select the protein's intrinsic disordered region and core functional region, and set at least one sequence fragment in the protein's intrinsic disordered region as the key region; Prediction module: Used to iteratively optimize the sequence set and output the target protein sequences ranked higher by LLPS score after the iteration optimization is completed; each round of iteration optimization includes fitness evaluation, selection operation, crossover operation, and mutation operation of the target protein sequences to obtain a new sequence set; In the fitness evaluation operation, the fitness evaluation function for the target protein sequence is: Fitness = α*A - β*B; Where α and β are weights; A is the LLPS score of the target protein sequence; and B is the pathogenicity score. Functional validation and optimization module: Used to perform dry and wet experimental validation on the target protein sequence output by the prediction module, and optimize the parameter settings and fitness evaluation function of the genetic algorithm based on the validation results.

9. The system according to claim 8, characterized in that, In the crossover operation, the probability of the target protein sequence being selected is positively correlated with the fitness score; in the mutation operation, the distance between an amino acid in the intrinsic disorder region of the protein and the core functional region is negatively correlated with the mutation probability of that amino acid; the core functional region is one or more segments of the target protein sequence located outside the intrinsic disorder region of the protein; the pathogenicity score is the sum of the pathogenicity scores of the mutation site and the mutation type. The mutation site and mutation type are obtained by comparing the target protein sequence with the wild-type target protein sequence.

10. The system according to claim 8, characterized in that, The dry assay is used to verify the nuclear localization, structural stability, and LLPS score of the target protein sequence; the wet assay is used to verify the cell viability, phase separation dynamics, and phenotype of the target protein sequence.

11. An electronic device, characterized in that, The method includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the method of any one of claims 1-7.

12. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7.

13. A computer program product, characterized in that, It includes computer program instructions, wherein the computer program instructions, when executed by a processor, implement the method of any one of claims 1-7.

Images