Functional protein screening method based on large language model

Abstract

The application discloses a functional protein screening method based on a large language model, and belongs to the technical field of protein screening, and specifically comprises the following steps: obtaining a protein three-dimensional structure and function literature, converting the protein three-dimensional structure and function literature into a multi-modal feature vector by using an encoder, and constructing a field-specific vector database; after receiving a screening instruction, associated feature data matched in semantics is called by a vector retrieval algorithm; subsequently, the data is input into a large language model to perform biological mechanism reasoning, and an initial candidate protein sequence containing predicted activity is output; further, a molecular docking program is used to calculate physical binding energy, a deviation coefficient of the initial candidate protein sequence and the predicted activity is generated, and the deviation coefficient is used as a penalty term to construct a dynamic correction function to update a scoring weight; finally, a secondary sorting is performed on a candidate set based on the updated weight, and a sequence ranked first is output as a target protein; the method effectively combines AI reasoning and physical verification, and improves screening accuracy.

Description

Technical Field

[0001] This invention relates to the field of protein screening technology, and more specifically to a functional protein screening method based on a large language model. Background Technology

[0002] Functional proteins, as core elements of biomedicine, synthetic biology, and industrial enzyme engineering, have always been a research hotspot in their screening and directed evolution. With the development of high-throughput sequencing technology and the exponential growth of protein structure databases, computational biology-based virtual screening techniques have gradually become a key tool in assisting experiments. Current mainstream screening techniques typically combine bioinformatics algorithms with computer-aided drug design (CADD) workflows, utilizing molecular docking, molecular dynamics simulations, and traditional deep learning models based on convolutional neural networks or recurrent neural networks to extract features and predict affinity from massive amounts of protein sequence or structural data. The aim is to leverage computing power to rapidly narrow down the experimental scope from a vast candidate library, thereby accelerating the development of novel enzymes, antibody drugs, or biomaterials.

[0003] However, while the aforementioned existing technologies have improved screening speed to some extent, current protein screening methods largely rely on predefined rules or static models, lacking adaptive and dynamic optimization capabilities. Once the screening logic and criteria are set, they are difficult to adjust based on real-time data, new discoveries, or specific R&D needs, leading to a rigid screening process and an inability to accurately match target proteins. This static approach often introduces biases when faced with complex and ever-changing biological information, reducing screening efficiency and accuracy, and increasing the cost and time consumption of repeated experiments. Summary of the Invention

[0004] The purpose of this invention is to provide a functional protein screening method based on a large language model, thereby solving the problems in the background art:

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for screening functional proteins based on a large language model includes the following steps:

[0007] S1: Obtain protein three-dimensional structure data and functional literature text, use an encoder to convert them into multimodal feature vectors and store them in a domain-specific vector database;

[0008] S2: Receives the target protein screening instruction and retrieves the associated feature data that semantically matches the instruction from the domain-specific vector database using a vector retrieval algorithm;

[0009] S3: Input the associated feature data into the large language model, perform biological mechanism inference, and output an initial set of candidate protein sequences containing predicted binding activity values;

[0010] S4: Input the initial set of candidate protein sequences into the molecular docking program, calculate the physical binding energy and generate the deviation coefficient between the physical binding energy and the predicted binding activity value;

[0011] S5: Introduce the bias coefficient as a penalty term into the sorting algorithm to construct a dynamic correction function, and use this dynamic correction function to update the scoring weights for the initial candidate protein sequence set;

[0012] S6: Perform a secondary sorting calculation on the initial candidate protein sequence set using the updated scoring weights, and output the sequence ranked first as the target protein.

[0013] As a further aspect of the present invention: in step S1, the process of obtaining protein three-dimensional structure data and functional literature text, converting them into multimodal feature vectors using an encoder, and storing them in a domain-specific vector database is as follows:

[0014] Extract protein atomic coordinate files and corresponding functional description literature from bioinformatics databases, filter out solvent molecule data in atomic coordinate files to generate pure structural data, and perform noise removal and cleaning on functional description literature to generate pure text data.

[0015] Pure structural data is input into a 3D geometric encoder to generate a high-dimensional structural feature vector, and pure text data is input into a semantic understanding encoder to generate a high-dimensional semantic feature vector. The two are then fused into a multimodal feature vector using a vector concatenation algorithm.

[0016] Each multimodal feature vector is assigned a unique index identifier. An inverted index algorithm is used to construct a retrieval tree structure for the vector space. The multimodal feature vectors and their index identifiers are then written into a domain-specific vector database for storage.

[0017] As a further aspect of the present invention: in step S2, the process of receiving the target protein screening instruction and retrieving the associated feature data that semantically matches the instruction from a domain-specific vector database using a vector retrieval algorithm is as follows:

[0018] The system receives a target protein screening instruction containing a description of the desired biological activity from the user, parses the text character sequence in the instruction, and performs noise reduction and normalization processing using a natural language processing algorithm.

[0019] The standardized text character sequence is input into the semantic understanding encoder, and instruction feature vectors with the same dimension as the feature space of the domain-specific vector database are generated by calculating the model's internal parameters.

[0020] The built-in search engine of the domain-specific vector database is invoked to calculate the semantic similarity between the instruction feature vector and all multimodal feature vectors stored in the database using the cosine similarity formula.

[0021] Based on semantic similarity scores, the multimodal feature vectors in the database are sorted in descending order. The original structural data and functional literature text mapped by the vectors at the top of the sorted sequence are extracted as associated feature data.

[0022] As a further aspect of the present invention: in step S3, the process of inputting the associated feature data into the large language model, performing biological mechanism inference, and outputting an initial set of candidate protein sequences containing predicted binding activity values ​​is as follows:

[0023] The protein three-dimensional structure data contained in the associated feature data is serialized into a one-dimensional text string, and then filled into a preset structured prompt word template along with the functional literature text to construct the inference input data.

[0024] The inference input data is fed into the large language model, which uses the biochemical rules stored in its internal parameters to deduce the arrangement logic of amino acid residues and generate natural language response text containing the predicted amino acid sequence.

[0025] An amino acid sequence and its corresponding generation probability value are extracted from the natural language response text using a text parsing algorithm. The generation probability value is then converted into a predicted binding activity value and combined to generate an initial set of candidate protein sequences.

[0026] As a further aspect of the present invention: the process of extracting amino acid sequences and corresponding generation probability values ​​from natural language response text using a text parsing algorithm, converting the generation probability values ​​into predicted binding activity values, and combining them to generate an initial candidate protein sequence set is as follows:

[0027] The system uses a pre-defined regular expression matching logic to traverse the natural language response text, identify and extract character sequences that conform to the standard amino acid encoding format, and simultaneously calls the model application programming interface to extract the conditional generation probability data corresponding to each word that constitutes the character sequence.

[0028] The extracted conditional generation probability data is subjected to average log-likelihood calculation to obtain the overall confidence level of the sequence. Based on the preset assumption that the confidence level is positively correlated with biological activity, the overall confidence level of the sequence is converted into a predicted binding activity value within a preset numerical range using the probability-activity mapping function.

[0029] The extracted character sequence is labeled as a candidate protein sequence. The candidate protein sequence is bound to the corresponding predicted binding activity value as a key-value pair data structure. This key-value pair data structure is then appended to the list object to generate the initial set of candidate protein sequences.

[0030] As a further aspect of the present invention: in step S4, the process of inputting the initial candidate protein sequence set into the molecular docking program, calculating the physical binding energy, and generating a deviation coefficient between the physical binding energy and the predicted binding activity value is as follows:

[0031] The initial candidate protein sequence set is traversed, and the linear amino acid sequence is converted into a three-dimensional atomic coordinate model using a protein structure prediction algorithm. Then, hydrogenation and charge distribution processing is performed on the three-dimensional atomic coordinate model to generate a structure file that conforms to the docking format.

[0032] The structure file is loaded into the molecular docking program, and the binding process between the three-dimensional atomic coordinate model and the target molecule is simulated under preset force field parameters. The complex conformation with the lowest system energy is locked by the conformation search algorithm. The intermolecular interaction energy of the complex conformation is calculated by calling the built-in scoring function of the program, and the physical binding energy is output. The physical binding energy is processed by the reverse normalization algorithm to construct a mapping relationship in which the lower the physical binding energy value, the higher the normalization score. The physical binding energy is mapped to the same numerical range as the predicted binding activity value to generate a physical verification score.

[0033] Read the predicted binding activity value corresponding to the sequence, calculate the absolute value of the difference between the physical verification score and the predicted binding activity value, and mark the absolute value of the difference as the bias coefficient of the sequence.

[0034] As a further aspect of the present invention: in step S5, the process of introducing the deviation coefficient as a penalty term into the sorting algorithm to construct a dynamic correction function, and using this dynamic correction function to update the scoring weights for the initial candidate protein sequence set, is as follows:

[0035] Read the deviation coefficients corresponding to each sequence in the initial candidate protein sequence set, use the reverse mapping algorithm to convert the deviation coefficients into penalty factors, and establish the negative correlation logic between the deviation coefficient values ​​and the penalty factor values.

[0036] The predicted combination of the sequence and the activity value is used as the basic weight, and the penalty factor is used as the correction variable to construct a dynamic correction function that includes the product structure of the basic weight and the correction variable.

[0037] The predictions of each sequence are combined with the activity value and the penalty factor and substituted into the dynamic correction function to calculate and generate a correction value that incorporates physical verification information, thereby completing the calibration calculation of the model prediction results.

[0038] The corrected value output by the calibration operation is defined as the updated scoring weight, and this scoring weight is associated with and stored in the initial candidate protein sequence set as the basis for secondary sorting.

[0039] As a further aspect of the present invention: in step S6, the process of performing a secondary sorting calculation on the initial candidate protein sequence set using the updated scoring weights, and outputting the sequence with the highest sorting as the target protein, is as follows:

[0040] Traverse the initial candidate protein sequence set, read the updated score weight generated in step five for each sequence in the set, and bind the score weight as a unique sorting key value to the corresponding initial candidate protein sequence.

[0041] The data sorting algorithm is invoked to perform a rearrangement operation on the initial candidate protein sequence set. Based on the score weight values ​​bound to each sequence, all sequences in the set are generated into an ordered sequence list in descending order.

[0042] The sequence data at the first index position in the ordered sequence list is locked, the amino acid sequence information of the data is extracted and marked as the target protein, and the target protein is output to the preset result storage path through the data interface.

[0043] The beneficial effects of this invention are:

[0044] This invention significantly improves the adaptability and generalization depth of the protein screening process by constructing a domain-specific vector database and combining it with the multimodal reasoning capabilities of a large language model. Unlike existing technologies that rely on rigid predefined rules, this application converts the three-dimensional structural data of proteins and functional literature texts into multimodal feature vectors, enabling the model to flexibly retrieve related data and perform deep biological mechanism reasoning based on real-time input screening instructions. This adaptive screening mode can effectively cope with complex and ever-changing biological information, accurately match specific R&D needs, and thus overcome the shortcomings of traditional static models that cannot dynamically adjust screening logic, greatly expanding the discovery space of target proteins.

[0045] This invention innovatively introduces a dynamic correction mechanism based on physical binding energy deviation, significantly improving the accuracy of screening results and the success rate of experiments. Addressing the issue of bias inherent in static methods, this application compares the physical binding energy calculated by molecular docking with the model's predicted activity, generating a deviation coefficient as a penalty term to dynamically correct the scoring weights of candidate sequences. This closed-loop optimization strategy combining AI inference and physical calibration can automatically identify and eliminate spurious sequences with high statistical probability but poor physical stability, effectively correcting model bias. This significantly reduces the trial-and-error costs and time consumption of subsequent wet experimental verification, achieving a dual improvement in screening efficiency and quality. Attached Figure Description

[0046] The invention will now be further described with reference to the accompanying drawings.

[0047] Figure 1This is a schematic flowchart of a functional protein screening method based on a large language model according to the present invention. Detailed Implementation

[0048] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1 As shown, this invention is a method for screening functional proteins based on a large language model, comprising the following steps:

[0050] S1: Obtain protein three-dimensional structure data and functional literature text, use an encoder to convert them into multimodal feature vectors and store them in a domain-specific vector database;

[0051] S2: Receives the target protein screening instruction and retrieves the associated feature data that semantically matches the instruction from the domain-specific vector database using a vector retrieval algorithm;

[0052] S3: Input the associated feature data into the large language model, perform biological mechanism inference, and output an initial set of candidate protein sequences containing predicted binding activity values;

[0053] S4: Input the initial set of candidate protein sequences into the molecular docking program, calculate the physical binding energy and generate the deviation coefficient between the physical binding energy and the predicted binding activity value;

[0054] S5: Introduce the bias coefficient as a penalty term into the sorting algorithm to construct a dynamic correction function, and use this dynamic correction function to update the scoring weights for the initial candidate protein sequence set;

[0055] S6: Perform a secondary sorting calculation on the initial candidate protein sequence set using the updated scoring weights, and output the sequence ranked first as the target protein.

[0056] In one embodiment of the present invention, step S1, which involves obtaining protein three-dimensional structure data and functional literature text, converting them into multimodal feature vectors using an encoder, and storing them in a domain-specific vector database, is as follows:

[0057] In the first stage of the data preparation process, this embodiment first connects to a publicly available bioinformatics database or a locally stored protein data warehouse through a preset data interface to batch retrieve atomic coordinate files containing the three-dimensional spatial information of the target protein, along with associated academic literature or functional annotation text. For the obtained atomic coordinate files, the processing program traverses each line of data in the file, reads the recorded atom type markers and residue names, and identifies solvent molecules that do not belong to the protein backbone or side chain, such as water molecules or ionic ligands introduced during crystallization. Since these solvent molecules are in a dynamic state in the actual biological environment and interfere with the determination of the static intrinsic structural features of the protein itself, the processing program directly deletes the coordinate lines of these non-protein atoms according to a preset blacklist, retaining only the spatial three-dimensional coordinate data of carbon, hydrogen, oxygen, nitrogen, and other atoms in the amino acid residues, thereby generating pure structural data. At the same time, for the extracted functional description literature, this embodiment uses text cleaning logic from natural language processing technology to scan the text for special characters, HTML format tags, bracket markers in reference citations, and stop words without actual semantic meaning. The processing program locates these noisy characters through pattern matching and replaces them with spaces or removes them directly. At the same time, it converts all uppercase letters in the text to lowercase to reduce redundancy in the lexical space, ensuring that the remaining text contains only the core semantic information describing the biological function, catalytic mechanism, or binding site of the protein, thereby generating clean text data.

[0058] After completing the data cleaning and preprocessing operations described above, this embodiment begins the digital mapping and fusion of features, aiming to convert physical protein entities into high-dimensional numerical vectors that computers can understand. First, the clean structural data is input into a 3D geometric encoder. This encoder, built on the principles of graph neural networks or 3D convolutional neural networks, does not simply read the absolute coordinates of atoms, but calculates the relative distances, bond angles, and dihedral angles between atoms. By constructing a topological connectivity graph between atoms, it captures the local chemical environment and global spatial folding patterns of the protein. Through the nonlinear transformation of multi-layer neural networks, this complex geometric and topological information is compressed and mapped into a fixed-length one-dimensional numerical array, i.e., a high-dimensional structural feature vector. Simultaneously, clean text data is input into a semantic understanding encoder. This encoder, based on a large-scale pre-trained language model architecture, uses an attention mechanism to capture the contextual dependencies between words in the text. It converts discrete words describing protein functions into continuously distributed numerical representations in the semantic space, making semantically similar functional descriptions closer in the vector space, thereby generating a high-dimensional semantic feature vector. In order to construct a unified representation that includes both spatial structural information and biological functional information, this embodiment uses a vector concatenation algorithm to connect the high-dimensional structural feature vector and the high-dimensional semantic feature vector end to end in the numerical dimension. This fusion operation effectively aligns the physical conformation of the protein with its biological function in a multimodal manner, generating a multimodal feature vector that can comprehensively characterize the protein's properties.

[0059] After generating and fusing feature vectors, to support rapid retrieval and management of large-scale data, this embodiment requires the generated multimodal feature vectors to be stored in a domain-specific vector database in an orderly manner. First, the processing program generates a globally unique index identifier for each generated multimodal feature vector. This identifier is typically a long integer or hash string, which serves as metadata and establishes a mapping relationship with the original protein name, source ID, and other information, ensuring that the stored vector data can be accurately traced and located. Next, an inverted index algorithm or a tree-based approximate nearest neighbor search algorithm is used to construct the index structure of the vector space. Based on the distribution of multimodal feature vectors in the numerical space, the processing program uses cluster centroids or hyperplane segmentation techniques to divide the entire high-dimensional vector space into multiple levels of sub-regions, constructing a retrieval path structure similar to a multi-branch tree. In this way, the originally chaotic massive vector data is organized into a structure with hierarchical navigation capabilities, allowing for similarity searches without traversing all data in the database; a pruning search along the branches of the tree structure is sufficient. Finally, all multimodal feature vector data ontology and their corresponding index identifiers are written into the underlying physical storage medium in the form of binary files to complete the persistent storage of data, thereby establishing a domain-specific vector database with efficient read / write and semantic retrieval capabilities.

[0060] In one embodiment of the present invention, step S2, which involves receiving a target protein screening instruction and retrieving semantically matching associated feature data from a domain-specific vector database using a vector retrieval algorithm, is as follows:

[0061] Upon receiving a screening request from the user, the processing module first opens a text input area on the user interface, waiting for the user to enter a specific natural language description. This description typically includes detailed information such as the specific biological activity, catalytic reaction type, or binding ability to specific target molecules that the user expects the target protein to possess. Once the user submits this text sequence containing professional biological terminology, the processing module immediately activates its built-in natural language processing algorithm to perform in-depth parsing and preprocessing of the original instruction. It scans each word in the character sequence like reading an article, using a pre-built biological stop word list to identify conjunctions, modifiers, and punctuation marks that, while conforming to grammatical rules, do not actually contribute to expressing the core biological meaning. The processing module removes these characters marked as noise from the sequence. At the same time, it uses lemmatization technology to restore various tense changes of verbs or plural forms of nouns to standard prototypical roots. For example, it maps various expressions of combining words to standard professional terms. Through this noise reduction and standardization process, the originally lengthy input instructions that may contain colloquial expressions are simplified into a set of clean text sequences that retain only the core semantic information and have a standardized format, thus eliminating interference factors for the next step of feature conversion.

[0062] After standardizing the instruction text, to enable the computer to understand the deeper semantics behind it, this embodiment inputs the processed clean text character sequence into a pre-trained semantic understanding encoder. This encoder is a mathematical model built on a deep neural network architecture, containing hundreds of millions of parameter weights. These parameters have learned complex contextual relationships between words through pre-training on large-scale biomedical literature data. When the text sequence enters the encoder, the model's attention mechanism layer calculates the contribution of each word in the sequence to the overall semantics and maps the discrete text symbols into a continuous high-dimensional numerical space through multi-layer matrix multiplication and non-linear activation function transformations. During this process, the encoder generates a numerical array with fixed dimensions, namely the instruction feature vector, based on the calculation logic of the model's internal parameters. The dimension of this vector is strictly set to be exactly the same as the dimension of the multimodal feature vectors stored in the domain-specific vector database. The principle behind this is that only when two vectors are in the same mathematical space and have the same dimensional coordinate system can their positional relationship in geometric space truly reflect their similarity in semantic content, thereby ensuring that the generated instruction feature vector can accurately represent the biological intent of the user's original input instruction.

[0063] Next, we move to the core step of vector retrieval. This embodiment utilizes the built-in retrieval engine of a domain-specific vector database to perform similarity comparison. The retrieval engine doesn't compare the magnitude of every specific numerical value in the vectors; instead, it uses the geometric principles inherent in the cosine similarity formula to measure the directional consistency between vectors. The retrieval engine iterates through every multimodal feature vector stored in the database, pairs it with the newly generated instruction feature vector, and calculates the cosine of the angle between these two vectors in multidimensional space. This calculation involves summing the products of the corresponding dimensional values ​​of the two vectors and normalizing the vector length. In principle, the closer the two vectors point in space (i.e., the smaller the angle between them), the closer the calculated cosine similarity value will be to 1. This indicates that the protein in the database highly matches the user's screening instructions in terms of function and structural features. Conversely, if the angle is large, the similarity value will be close to 0 or even negative, indicating that the two are completely unrelated. Through this geometric-based calculation method, the retrieval engine can quickly quantify the semantic affinity between instruction feature vectors and massive amounts of data in the database, thereby assigning each stored protein a quantitative score representing its degree of matching.

[0064] Finally, after obtaining the similarity scores of all candidate objects, the processing module needs to use these scores to lock in the most suitable candidate set. It performs a quick sorting operation on all multimodal feature vectors in the database based on the calculated semantic similarity values, rearranging them into an ordered queue in descending order of value. The processing module sets a truncation threshold or directly selects the top few positions of the queue as preferred targets, because these vectors at the top of the sequence have the highest similarity values, meaning they are most likely to meet the user's filtering requirements. After determining these preferred vectors, the processing module reads the unique index identifier carried in these vectors and uses this identifier to trace back to the underlying storage partition of the database, accurately extracting the original protein 3D structure data file mapped by the vector and the corresponding functional literature text. This process achieves a reverse mapping from abstract numerical vectors to specific biological entities. The extracted original structural data and functional literature text are packaged as associated feature data and passed to the next processing stage, thus completing the accurate retrieval and conversion from the user's vague natural language instructions to specific, highly matched protein entity data.

[0065] In one embodiment of the present invention, step S3, which involves inputting the associated feature data into a large language model, performing biological mechanism inference, and outputting an initial set of candidate protein sequences containing predicted binding activity values, is as follows:

[0066] In the data preparation and prompt word construction phase, this embodiment first initiates a data serialization process to address the format compatibility issue where three-dimensional spatial data cannot be directly read by large language models. The processing module reads the protein three-dimensional structure data from the associated feature data. This data typically consists of the spatial Cartesian coordinates of tens of thousands of atoms. Directly inputting this data into the model would lead to the curse of dimensionality and a lack of semantic association. Therefore, the processing module employs a serialization algorithm based on a structural alphabet. By calculating the dihedral angles between backbone atoms and the spatial distance matrix between residues, the continuous three-dimensional geometry is discretized and mapped into a one-dimensional text string composed of specific characters. These characters represent local secondary structural features of the protein, such as helices, folds, or turns. Simultaneously, the processing module reads functional literature text, extracting natural language descriptions of key residues at binding sites or catalytic reaction mechanisms. Then, it calls a preset structured prompt word template. This template is essentially a text frame containing specific slots, defining the model's role, task objectives, and the arrangement rules of the input data. The processing module fills the structure slots of the template with the serialized structure string and the context slots with the functional literature text. Through character concatenation, it integrates structural features and semantic logic into a complete inference input data with clear instruction meaning, thereby providing the large language model with composite input information that includes both physical geometric constraints and biological function guidance.

[0067] After constructing the inference input data, this embodiment feeds it into a large language model deployed on a high-performance computing cluster to initiate the generative inference process of biological mechanisms. This large language model, pre-trained with massive amounts of protein sequence and structural data, implicitly stores the rules of biochemical interactions between amino acid residues and the mapping logic between structure and function within its hundreds of millions of neural network parameters. Upon receiving the input data, the model's multi-head attention mechanism layer begins operation, calculating in parallel the association weights between structural characters and functional descriptive words in the input text, capturing deep-seated patterns such as what functions a particular structural model typically corresponds to. Based on these weights, the model performs probabilistic inference in the latent vector space, predicting which amino acid residues should be filled in the current structural framework to satisfy the specific biological activity described in the input. This process is not a simple retrieval but a creative generation based on biochemical rules. The model outputs results word by word, ultimately generating a coherent natural language response text. This text not only contains explanatory inference statements but, more importantly, one or more entirely new predicted amino acid sequences deduced by the model based on contextual logic.

[0068] The amino acid sequences and corresponding generation probability values ​​are extracted from the natural language response text using a text parsing algorithm. These generation probability values ​​are then converted into predicted binding activity values ​​and combined to generate an initial set of candidate protein sequences. The specific details are as follows:

[0069] In the deep mining phase of the natural language response text generated by the large language model, the processing module first initiates a text pattern recognition-based parsing process. Using pre-defined regular expression matching logic, it scans and traverses the entire text output by the model character by character. This matching logic incorporates a standard library of twenty natural amino acid abbreviations, enabling it to accurately identify consecutive string segments in the text that meet a specific length threshold and consist only of specific uppercase letters. These segments are identified as potential amino acid sequences. Simultaneously, the processing module goes beyond simply acquiring the text content of the sequence. It synchronously calls the large language model's application programming interface to trace back the generation process, delving into the model's underlying output layer to extract the conditional generation probability data corresponding to each lexical unit constituting the character sequence. The principle behind this operation is that when the model generates each amino acid character, it essentially calculates the probability distribution of all possible characters in the current context. By extracting the probability value corresponding to the character ultimately selected by the model, the processing module can quantify the model's certainty in generating that specific amino acid, thus assigning a precise numerical confidence label to each tiny sequence unit.

[0070] After acquiring the conditional generation probability data for each amino acid term constituting the sequence, the processing module needs to integrate these discrete probability values ​​into a comprehensive index that can evaluate the quality of the entire sequence. Therefore, it performs an average log-likelihood calculation. The processing module first converts each probability value between 0 and 1 into its logarithmic form. This is done to transform the multiplicative probability operation into an summation operation to avoid numerical underflow. Then, it sums the logarithmic values ​​at all positions and divides them by the total length of the sequence to calculate the standardized overall sequence confidence. Next, based on a pre-set biological hypothesis—that large language models assign higher generation probabilities to protein sequences that conform to natural evolutionary laws and have stable structures—the processing module establishes a positive correlation between confidence and biological activity. Based on this logic, the processing module calls a probability-to-activity mapping function. This function, through linear scaling or non-linear fitting, directly maps and projects the originally abstract probability confidence values ​​to the range of biological activity values ​​that the user is interested in, thus completing the cross-domain conversion from the statistical probability of the language model to the predictive combined with activity values ​​in the biochemical field.

[0071] After sequence extraction and activity value calculation are completed, the processing module begins constructing a standardized data storage structure to facilitate batch processing and indexing of this data in subsequent computer programs. First, the processing module formally marks the character sequence extracted through regular expression matching as a candidate protein sequence, treating it as an independent entity object. Then, it logically and strongly associates this sequence object with the calculated predicted activity value, encapsulating it in a key-value pair data structure. The amino acid sequence serves as the unique lookup key, and the activity value as the corresponding value. This structured encapsulation ensures a one-to-one correspondence between the data, preventing misalignment during batch processing. Finally, the processing module initializes an empty list object as a container, sequentially appending each encapsulated key-value pair data structure to this list. As the traversal process concludes, this list object gathers all possible protein sequences and their predicted activities, ultimately forming a complete initial set of candidate protein sequences, providing ample data preparation for the physical validation phase.

[0072] In one embodiment of the present invention, step S4, which involves inputting the initial candidate protein sequence set into the molecular docking program, calculating the physical binding energy, and generating a deviation coefficient between the physical binding energy and the predicted binding activity value, is as follows:

[0073] In the molecular-level physical verification phase, the processing module first initiates a loop logic to process each data item in the initial candidate protein sequence set. It reads linear amino acid sequence information consisting of only single-letter or three-letter codes and invokes a deep learning-based protein structure prediction algorithm. This algorithm infers the folding pattern of the peptide chain in three-dimensional space by analyzing the co-evolutionary relationships between amino acid residues and multiple sequence alignment information. It uses a neural network to predict the spatial relative coordinates of each atom and the torsion angle of the backbone, thereby reconstructing the one-dimensional text sequence into a three-dimensional atomic coordinate model with a physical geometry. Next, to meet the stringent requirements of molecular mechanics calculations for atomic integrity, the processing module performs a fine-tuning preprocessing operation on the generated coarse model. It completes the model by filling in missing hydrogen atoms according to the topological definition of the standard residue library. This step is necessary because models trained with X-ray crystallography data often lack hydrogen atom coordinates, and hydrogen atoms are crucial for forming hydrogen bonds. Simultaneously, the processing module calculates the protonation state of each amino acid residue based on the solution pH environment, and uses Gasteiger or similar charge calculation methods to distribute a portion of the charge to each atom, generating a structure file that contains complete atomic coordinates, element types, and charge distribution information, conforming to the docking program input standard.

[0074] Once the complete structural file is available, the processing module loads it into a specialized molecular docking program to conduct virtual screening experiments. Simultaneously, it reads in the pre-prepared structural data of the target molecule. The processing module defines a three-dimensional search grid box in the active site region of the target molecule and sets the force field parameters, including van der Waals radius, electrostatic interaction cutoff value, and solvation effect parameters. Under the physical constraints of the preset force field parameters, the docking program simulates the dynamic binding process between the candidate protein model (as ligand) and the target molecule (as acceptor). It uses conformational search algorithms such as genetic algorithms or Monte Carlo simulated annealing to continuously change the position, rotation angle, and flexible torsion state of the side chains of the candidate protein, calculating the current system energy after each positional change. The goal of the search algorithm is to find the state with the tightest binding and least repulsive force in the vast conformational space. After thousands of iterations and optimizations, the processing module finally locks in and outputs the complex conformation with the lowest total free energy of the system. This conformation represents the thermodynamically most stable binding mode.

[0075] After determining the optimal binding conformation, the processing module calls the program's built-in scoring function to quantify the strength of this bond. Based on physical principles, the scoring function weights and sums the van der Waals attraction, electrostatic attraction, hydrogen bond contribution energy, and desolvation penalty energy between all atomic pairs in the complex conformation. The calculated result is the intermolecular interaction energy, which is directly output as the physical binding energy. Since lower energy values ​​(usually negative) in physical thermodynamics represent more stable bonds, while the activity value predicted by the large language model in the preceding steps is usually higher (closer to 1) for better performance, there is a contradiction in the numerical direction. Therefore, the processing module uses a reverse normalization algorithm to process the physical binding energy. The algorithm first statistically analyzes the physical binding energy distribution of all sequences in the current set, determines the maximum and minimum value boundaries, and then constructs a linear or nonlinear mapping relationship so that data with lower original energy values ​​receive higher scores after mapping. Through this mathematical transformation, the negative energy data with varying original ranges are projected into a closed interval of 0 to 1 that is exactly the same as the predicted binding activity value, thereby generating a standardized physical verification score. This score directly reflects the level of binding affinity verified at the physical level.

[0076] Finally, the processing module performs crucial numerical comparison and bias calculation operations to quantify the consistency between the predictions of the large language model and the verification results of physical laws. The module first reads the predicted binding activity value corresponding to the currently processed sequence from memory. This value represents the theoretical activity derived from statistical laws. It then compares this value with the previously calculated physical verification score. The module performs a subtraction operation to calculate the difference between the two and takes the absolute value of the difference to eliminate directional influence. This absolute value is defined as the bias coefficient. This coefficient has clear physical and logical meaning. If the bias coefficient is very small, it indicates a high degree of agreement between the language model's prediction and the physical simulation results, indicating a very high degree of confidence in the sequence. Conversely, if the bias coefficient is large, it usually means that the language model may have misjudged, predicting a highly active but physically unstable sequence. The processing module marks the calculated bias coefficient in the sequence's data structure as a key indicator of the potential risk of the candidate protein, providing a quantitative penalty basis for the next step of weight adjustment and ranking.

[0077] In one embodiment of the present invention, in step S5, the process of introducing the deviation coefficient as a penalty term into the ranking algorithm to construct a dynamic correction function, and using this dynamic correction function to update the scoring weights for the initial candidate protein sequence set is as follows:

[0078] In the refinement stage of the initial candidate protein sequence set, the processing module first initiates an iterative reading program to traverse each candidate sequence object stored in memory and extract its associated bias coefficient. This coefficient is the difference value obtained in the previous stage by comparing physical calculations with model predictions. The processing module calls the built-in inverse mapping algorithm to process this bias coefficient. The core logic of this algorithm lies in establishing a numerical penalty mechanism: the larger the bias, the lower the reliability of the model prediction, and therefore a stronger penalty needs to be applied. The algorithm sets a baseline value, usually 1, and then uses linear subtraction or a non-linear exponential decay function to map the bias coefficient to a value between 0 and 1, which is defined as the penalty factor. Through this inverse mapping processing method, the processing module establishes a strict negative correlation between the bias coefficient value and the penalty factor value, ensuring that sequences whose physical verification results are significantly different from the model prediction results will receive a very small penalty factor, while sequences that are highly consistent with the model prediction results will retain a factor close to 1, thus providing a quantitative mathematical basis for the dynamic adjustment of weights.

[0079] After establishing the penalty factor, the processing module delves deeper into the data structure to extract the predicted combined activity value initially generated by the large language model. This value is defined as the base weight, representing the theoretical activity potential deduced solely from the perspective of biochemical language rules. Next, the processing module introduces the newly calculated penalty factor as a correction variable and begins constructing a dynamic correction function to adjust the final score. This function abandons simple addition and subtraction in its structural design, instead employing a product structure. The principle behind choosing the product structure lies in utilizing the gating effect of multiplication: if the penalty factor approaches zero due to severe physical conflict, then regardless of how high the base weight is, the final product will be forcibly lowered. In this way, the processing module constructs a mathematical expression that simultaneously reflects the creativity of the language model and the rigor of the physical model. This dynamic correction function treats the base weight as the operand and the correction variable as the operating coefficient, establishing a robust logical framework for subsequent numerical fusion operations.

[0080] Once the framework of the dynamic correction function is complete, the processing module executes the specific numerical substitution and calculation process. It simultaneously fills the function's variable slots with the unique predictions for each sequence, along with the activity value and corresponding penalty factor, and performs floating-point multiplication. During this calculation, the original prediction value, derived solely from the language model, is permeated and diluted by information from physical verification, generating a completely new value. This value integrates statistical predictions of sequence arrangement patterns with physical verification of intermolecular interactions; therefore, it is called the correction value incorporating physical verification information. This process is essentially a calibration operation of the model's prediction results. It effectively corrects erroneous predictions that, while having high textual probabilities, violate physical spatial constraints, bringing inflated scores back to levels consistent with physical reality. Simultaneously, it preserves the high scores of sequences that perform well in both dimensions, thereby greatly improving the authenticity and reliability of the screening results.

[0081] After completing the calibration operations for all sequences, the processing module formally defines the corrected output values ​​as the updated scoring weights. These new weights are no longer simple probabilities or energies, but a comprehensive evaluation metric. The processing module re-accesses the initial candidate protein sequence set via memory addressing, locates the corresponding sequence object, and stores the updated scoring weights as a new attribute field in the object's structure, or directly overwrites the original sorting criteria field. This storage operation not only updates the data but, more importantly, establishes a new evaluation standard, giving each sequence in the set a value label verified by physical laws. Once all data association and storage are complete, this comprehensively cleaned and corrected sequence set is ready to be directly passed as input data to subsequent modules, serving as the sole and authoritative basis for secondary sorting, ensuring that the final output target protein is the optimal solution selected through rigorous layers of screening.

[0082] In one embodiment of the present invention, step S6, which involves performing a secondary sorting calculation on the initial candidate protein sequence set using the updated scoring weights and outputting the sequence with the highest sorting score as the target protein, is as follows:

[0083] In the final selection and decision-making stage, the processing module first initiates a traversal reading program for the initial candidate protein sequence set. This program accesses each candidate object stored in the set one by one and accurately reads the updated score weights generated in the previous stage calibration calculation, which incorporate physical verification information. The principle behind this step is that although the object may still retain the original predicted activity values ​​or deviation coefficients from intermediate processes, only this updated score weight truly reflects the sequence's comprehensive performance in terms of biological rationality and physical stability. Therefore, it must be extracted as the sole evaluation criterion. The processing module establishes a temporary index mapping relationship for each sequence object in memory, explicitly assigning the extracted floating-point score weights as the sorting key value for that sequence object. This binding operation is equivalent to giving each candidate protein a final score label for comparison, ensuring that in the subsequent sorting calculation, the computer algorithm can accurately identify which numerical attribute should be used to determine the sequence's merits, thus introducing a quantifiable comparison benchmark to the unordered dataset.

[0084] After completing the binding operation of sorting key values, this embodiment calls an efficient data sorting algorithm to perform a rearrangement operation on the entire initial candidate protein sequence set. This algorithm adopts the logical architecture of quicksort or mergesort, adjusting the storage order of data in memory through multiple rounds of comparison and exchange operations. During execution, the algorithm continuously retrieves two sequence objects from the set and compares the values ​​of their bound score weights. According to a preset descending order rule, if the former's score weight is greater than the latter's, its position is moved forward; otherwise, it is moved backward. This comparison process is repeated within the set until all sequence objects find their correct relative position on the numerical axis, ultimately transforming the originally chaotic candidate set into a strictly ordered sequence list. In this generated ordered list, all sequences are arranged in descending order of score weight. This means that the sequences at the front of the list have the highest comprehensive score; that is, they are not only predicted by the language model to have high activity but also show extremely strong binding stability in physical simulations, thus achieving the goal of mathematically selecting the optimal solution from massive candidate data.

[0085] After the ordered sequence list is generated, the processing module immediately performs the final locking and output operations. It directly accesses the data unit at the first index of the list through a memory pointer. Because, according to the descending order, the sequence at index 0 must be the one with the highest score weight, which is the best-performing sequence in the entire screening process. The processing module reads the information stored within this data unit, removes intermediate calculation data such as score weights and bias coefficients, and extracts only the pure amino acid character sequence information, officially marking it as the target protein for this screening task. To enable users to view the result or for other applications to further utilize it, the processing module initiates the output process through a standardized data interface. It encapsulates the sequence text of the target protein and related key parameters into a universal format or file stream for network transmission and writes it to a preset result storage path. This path can be a specific folder on the local hard drive or a record line in a cloud database. This completes the entire process from complex computational reasoning to the final output of the effective product.

[0086] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for screening functional proteins based on a large language model, characterized in that, Includes the following steps: S1: Obtain protein three-dimensional structure data and functional literature text, use an encoder to convert them into multimodal feature vectors and store them in a domain-specific vector database; S2: Receives the target protein screening instruction and retrieves the associated feature data that semantically matches the instruction from the domain-specific vector database using a vector retrieval algorithm; S3: Input the associated feature data into the large language model, perform biological mechanism inference, and output an initial set of candidate protein sequences containing predicted binding activity values; the specific process is as follows: The protein three-dimensional structure data contained in the associated feature data is serialized into a one-dimensional text string, and then filled into a preset structured prompt word template along with the functional literature text to construct the inference input data. The inference input data is fed into the large language model, which uses the biochemical rules stored in its internal parameters to deduce the arrangement logic of amino acid residues and generate natural language response text containing the predicted amino acid sequence. The amino acid sequences and corresponding generation probability values ​​are extracted from the natural language response text using a text parsing algorithm. These generation probability values ​​are then converted into predicted binding activity values ​​and combined to generate an initial set of candidate protein sequences. The specific method is as follows: The system uses a pre-defined regular expression matching logic to traverse the natural language response text, identify and extract character sequences that conform to the standard amino acid encoding format, and simultaneously calls the model application programming interface to extract the conditional generation probability data corresponding to each word that constitutes the character sequence. The extracted conditional generation probability data is subjected to average log-likelihood calculation to obtain the overall confidence level of the sequence. Based on the preset assumption that the confidence level is positively correlated with biological activity, the overall confidence level of the sequence is converted into a predicted binding activity value within a preset numerical range using the probability-activity mapping function. The extracted character sequence is marked as a candidate protein sequence. The candidate protein sequence is bound to the corresponding predicted binding activity value as a key-value pair data structure. The key-value pair data structure is then appended to the list object to generate the initial candidate protein sequence set. S4: Input the initial set of candidate protein sequences into the molecular docking program, calculate the physical binding energy and generate the deviation coefficient between the physical binding energy and the predicted binding activity value; S5: Introduce the bias coefficient as a penalty term into the sorting algorithm to construct a dynamic correction function, and use this dynamic correction function to update the scoring weights for the initial candidate protein sequence set; S6: Perform a secondary sorting calculation on the initial candidate protein sequence set using the updated scoring weights, and output the sequence ranked first as the target protein.

2. The method for screening functional proteins based on a large language model according to claim 1, characterized in that, In step S1, the process of acquiring protein three-dimensional structure data and functional literature text, converting them into multimodal feature vectors using an encoder, and storing them in a domain-specific vector database is as follows: Extract protein atomic coordinate files and corresponding functional description literature from bioinformatics databases, filter out solvent molecule data in atomic coordinate files to generate pure structural data, and perform noise removal and cleaning on functional description literature to generate pure text data. Pure structural data is input into a 3D geometric encoder to generate a high-dimensional structural feature vector, and pure text data is input into a semantic understanding encoder to generate a high-dimensional semantic feature vector. The two are then fused into a multimodal feature vector using a vector concatenation algorithm. A unique index identifier is assigned to each multimodal feature vector. An inverted index algorithm is used to construct a retrieval tree structure for the vector space. The multimodal feature vectors and their index identifiers are then written into a domain-specific vector database for storage.

3. The method for screening functional proteins based on a large language model according to claim 1, characterized in that, In step S2, the process of receiving the target protein screening instruction and retrieving the associated feature data that semantically matches the instruction from the domain-specific vector database using a vector retrieval algorithm is as follows: The system receives a target protein screening instruction containing a description of the desired biological activity from the user, parses the text character sequence in the instruction, and performs noise reduction and normalization processing using a natural language processing algorithm. The standardized text character sequence is input into the semantic understanding encoder, and instruction feature vectors with the same dimension as the feature space of the domain-specific vector database are generated by calculating the model's internal parameters. The built-in search engine of the domain-specific vector database is invoked to calculate the semantic similarity between the instruction feature vector and all multimodal feature vectors stored in the database using the cosine similarity formula. Based on semantic similarity scores, the multimodal feature vectors in the database are sorted in descending order. The original structural data and functional literature text mapped by the vectors at the top of the sorted sequence are extracted as associated feature data.

4. The method for screening functional proteins based on a large language model according to claim 1, characterized in that, In step S4, the process of inputting the initial candidate protein sequence set into the molecular docking program, calculating the physical binding energy, and generating the deviation coefficient between the physical binding energy and the predicted binding activity value is as follows: The initial candidate protein sequence set is traversed, and the linear amino acid sequence is converted into a three-dimensional atomic coordinate model using a protein structure prediction algorithm. Then, hydrogenation and charge distribution processing is performed on the three-dimensional atomic coordinate model to generate a structure file that conforms to the docking format. The structure file is loaded into the molecular docking program, and the binding process between the three-dimensional atomic coordinate model and the target molecule is simulated under preset force field parameters. The complex conformation with the lowest system energy is locked by the conformation search algorithm. The intermolecular interaction energy of the complex conformation is calculated by calling the built-in scoring function of the program, and the physical binding energy is output. The physical binding energy is processed by the reverse normalization algorithm to construct a mapping relationship in which the lower the physical binding energy value, the higher the normalization score. The physical binding energy is mapped to the same numerical range as the predicted binding activity value to generate a physical verification score. Read the predicted binding activity value corresponding to the sequence, calculate the absolute value of the difference between the physical verification score and the predicted binding activity value, and mark the absolute value of the difference as the bias coefficient of the sequence.

5. The method for screening functional proteins based on a large language model according to claim 1, characterized in that, In step S5, the process of introducing the deviation coefficient as a penalty term into the ranking algorithm to construct a dynamic correction function, and using this dynamic correction function to update the scoring weights for the initial candidate protein sequence set, is as follows: Read the deviation coefficients corresponding to each sequence in the initial candidate protein sequence set, use the reverse mapping algorithm to convert the deviation coefficients into penalty factors, and establish the negative correlation logic between the deviation coefficient values ​​and the penalty factor values. The predicted combination of the sequence and the activity value is used as the basic weight, and the penalty factor is used as the correction variable to construct a dynamic correction function that includes the product structure of the basic weight and the correction variable. The predictions of each sequence are combined with the activity value and the penalty factor and substituted into the dynamic correction function to calculate and generate a correction value that incorporates physical verification information, thereby completing the calibration calculation of the model prediction results. The corrected value output by the calibration operation is defined as the updated scoring weight, and this scoring weight is associated with and stored in the initial candidate protein sequence set as the basis for secondary sorting.

6. The method for screening functional proteins based on a large language model according to claim 1, characterized in that, In step S6, the process of performing a secondary sorting calculation on the initial candidate protein sequence set using the updated scoring weights and outputting the sequence with the highest sorting score as the target protein is as follows: Traverse the initial candidate protein sequence set, read the updated score weight generated in step five for each sequence in the set, and bind the score weight as a unique sorting key value to the corresponding initial candidate protein sequence. The data sorting algorithm is invoked to perform a rearrangement operation on the initial candidate protein sequence set. Based on the score weight values ​​bound to each sequence, all sequences in the set are generated into an ordered sequence list in descending order. The sequence data at the first index position in the ordered sequence list is locked, the amino acid sequence information of the data is extracted and marked as the target protein, and the target protein is output to the preset result storage path through the data interface.

Images