Biological sequence optimization degree prediction method, device, equipment, medium and product

Abstract

This application provides a method, apparatus, device, medium, and product for predicting the optimization level of biological sequences, relating to the field of artificial intelligence technology. The method includes: acquiring a set of biological sequences to be predicted, the set including: an original biological sequence and at least one candidate biological sequence, the candidate biological sequence being obtained by modifying the original biological sequence; inputting the set of biological sequences to be predicted into a prediction model, and predicting the candidate biological sequence based on the biological sequence evolutionary trajectory to obtain a prediction result. The prediction model is obtained by pre-training using the biological sequence evolutionary trajectory, where the evolutionary trajectory characterizes the gradual evolution of the biological sequence from low to high optimization level, and the prediction result is used to characterize the optimization level of the candidate biological sequence. This method improves the predictive ability of candidate sequence optimization level by introducing evolutionary trajectory information.

Description

Technical Field

[0001] This application relates to the field of bioinformatics, and in particular to methods, devices, electronic devices, computer-readable storage media, and computer program products for predicting the degree of optimization of biological sequences. Background Technology

[0002] Biological sequences are important carriers of life information, including protein sequences and nucleic acid sequences, whose structure and function are determined by sequence characteristics. In fields such as biomedicine, protein engineering, and synthetic biology, it is often necessary to optimize biological sequences to improve their specific functional properties, such as binding ability, stability, or expression efficiency.

[0003] In related technologies, methods based on structural simulation or machine learning are commonly used to predict the degree of optimization of biological sequences. Structural simulation methods predict function by constructing three-dimensional molecular structures and analyzing their interactions; while machine learning methods establish a mapping relationship between sequences and functional indicators by extracting sequence features or structural features.

[0004] However, the relevant methods all use a single biological sequence as the modeling object, lacking a systematic modeling of the evolutionary process of biological sequences, resulting in limited predictive ability of the optimization degree of candidate sequences. Summary of the Invention

[0005] The embodiments of this application aim to at least partially solve one of the aforementioned technical problems or at least provide a useful commercial option. In view of this, the embodiments of this application provide a method, apparatus, device, medium, and product for predicting the degree of optimization of biological sequences, which improves the predictive ability of candidate sequence optimization by introducing evolutionary trajectory information.

[0006] In a first aspect, embodiments of this application provide a method for predicting the optimization level of biological sequences. The method includes: acquiring a set of biological sequences to be predicted, the biological sequence set including: an original biological sequence and at least one candidate biological sequence, the candidate biological sequence being obtained by modifying the original biological sequence; inputting the set of biological sequences to be predicted into a prediction model to obtain a prediction result, wherein the prediction model is obtained by pre-training using the evolutionary trajectory of biological sequences, the evolutionary trajectory representing the gradual evolution process of biological sequences from low optimization level to high optimization level, and the prediction result being used to represent the optimization level of the candidate biological sequence.

[0007] In an exemplary embodiment, obtaining the prediction model by pre-training using biological sequence evolutionary trajectories includes: acquiring training data, the training data including: N evolutionary trajectories, each evolutionary trajectory including: multiple directed biological sequences, wherein the abundance of the starting biological sequence is lower than that of the ending biological sequence, the abundance being used to characterize the enrichment degree of the biological sequence in the population, and N being an integer not less than 1; and pre-training a machine learning model based on the training data to obtain the prediction model.

[0008] In an exemplary embodiment, obtaining N evolutionary trajectories includes: obtaining a standard biological sequence set; performing clustering processing on the standard biological sequence dataset based on the sequence characteristics of the biological sequences to obtain N clonal groups, wherein the sequence characteristics include sequence similarity, V(D)J gene information or clonal characteristics; and constructing phylogenetic relationships for each clonal group to obtain N evolutionary trajectories.

[0009] In an exemplary embodiment, the original biological sequence set is preprocessed to obtain a standard biological sequence set.

[0010] In an exemplary embodiment, the biological sequence is selected from antibodies. The preprocessing of the original antibody sequence set to obtain a standard antibody sequence set includes: completing the original antibody sequence set with missing N-terminus and / or C-terminus to obtain a complete antibody sequence set; aligning the complete antibody sequence set according to antibody sequence encoding rules to obtain an aligned antibody sequence set; and performing redundancy removal processing on the aligned antibody sequence set to obtain a standard antibody sequence set.

[0011] In an exemplary embodiment, the step of completing the original antibody sequence set missing at the N-terminus and / or C-terminus to obtain a complete antibody sequence set includes: annotating each original antibody sequence with gene categories in regions V and J to obtain sequence annotation results; performing similarity matching in an antibody sequence coding rule library based on the sequence annotation results to obtain a reference antibody; and filling in the original antibody sequences missing at the N-terminus and / or C-terminus based on the reference antibody to obtain a complete antibody sequence set.

[0012] In an exemplary embodiment, the reference antibody is selected from the antibody sequence with the highest homology in the matching results.

[0013] In an exemplary embodiment, the complete antibody sequence set is validated for consistency based on conserved region features.

[0014] In an exemplary embodiment, the step of constructing phylogenetic relationships for each clonal group to obtain N evolutionary trajectories includes: sorting the biological sequences in the nth clonal group according to their abundance, where n is an integer not less than 1 and not greater than N; taking the biological sequence with the lowest abundance as the evolutionary starting point and the biological sequence with the highest abundance as the evolutionary ending point, and performing directed connections on the biological sequences in the nth clonal group based on the similarity or mutation relationship of the biological sequences to obtain the nth evolutionary trajectory; and traversing the N clonal groups to obtain N evolutionary trajectories.

[0015] In an exemplary embodiment, the step of pre-training the machine learning model based on the training data to obtain a prediction model includes: for the nth evolutionary trajectory, taking any two biological sequences in the nth evolutionary trajectory as the starting point and the ending point, and constructing an evolutionary path sequence composed of multiple intermediate biological sequences by progressively applying state transition operations corresponding to single-point mutations; sequentially inputting each biological sequence in the evolutionary path sequence into a feature extraction network for encoding to obtain a corresponding sequence feature representation; based on the sequence feature representation, obtaining the forward action probability and backward action probability corresponding to each state through a forward action prediction network and a backward action prediction network, respectively; cumulatively calculating the forward action probability and backward action probability of each state in the evolutionary path sequence, and constructing a trajectory loss function using the reward value of the ending biological sequence as a supervision signal; and jointly training the feature extraction network, the forward action prediction network, and the backward action prediction network based on the trajectory loss function to obtain a prediction model.

[0016] In an exemplary embodiment, the step of pre-training the machine learning model based on the training data to obtain a prediction model includes: for the nth evolutionary trajectory, taking any two biological sequences in the nth evolutionary trajectory as the starting point and the ending point; inputting the structural information corresponding to the starting point biological sequence and the ending point biological sequence into a policy network model to obtain the probability distribution of each amino acid position in each sequence by the policy network; performing cumulative calculation on the probability distributions of the starting point biological sequence and the ending point biological sequence respectively to obtain the corresponding sequence log probability; inputting the structural information into a reference model to obtain the corresponding reference sequence log probability; calculating the reward deviation between the starting point biological sequence and the ending point biological sequence based on the log probability difference between the policy network and the reference model, and constructing a loss function based on the reward deviation; updating the parameters of the policy network based on the loss function to obtain the prediction model.

[0017] In an exemplary embodiment, the method further includes: determining a target biological sequence based on the prediction results, wherein determining the target biological sequence based on the prediction results includes: sorting candidate biological sequences based on the prediction results to determine the target biological sequence; or iteratively optimizing candidate biological sequences based on the prediction results, and determining the target biological sequence based on the prediction results of the optimized candidate biological sequences.

[0018] Secondly, embodiments of this application provide a biological sequence optimization degree prediction device, which includes: an acquisition module and a prediction module. The module includes an acquisition module for acquiring a set of biological sequences to be predicted, comprising: an original biological sequence and at least one candidate biological sequence, wherein the candidate biological sequence is obtained by modifying the original biological sequence; and a prediction module for inputting the set of biological sequences to be predicted into a prediction model to obtain a prediction result, wherein the prediction model is obtained by pre-training using the evolutionary trajectory of the biological sequence, the evolutionary trajectory representing the gradual evolution of the biological sequence from a low optimization level to a high optimization level, and the prediction result representing the optimization level of the candidate biological sequence.

[0019] Thirdly, embodiments of this application provide an electronic device, which includes: a processor and a memory; the aforementioned memory is used to store a computer program; the aforementioned processor is used to execute the aforementioned computer program to implement the biological sequence optimization degree prediction method as provided in the first aspect.

[0020] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer instructions or programs that, when executed on a computer, cause the biological sequence optimization degree prediction method provided in the first aspect to be performed.

[0021] Fifthly, embodiments of this application provide a computer program product including computer instructions that, when some or all of the computer instructions are executed on a computer, cause the biological sequence optimization degree prediction method provided in the first aspect to be executed.

[0022] In a sixth aspect, embodiments of this application provide a computer program that, when run on a computer, causes the computer to execute the biological sequence optimization degree prediction method provided in the first aspect.

[0023] In summary, unlike related technologies that rely solely on static modeling based on a single biological sequence, the solution provided in this application's embodiments trains the prediction model by introducing biological sequence evolution trajectory data, thereby achieving dynamic modeling of the biological sequence evolution process. Specifically, during training, the model learns the gradual evolutionary patterns of biological sequences from low-performance to high-performance states, enabling the model to not only characterize the static features of a single sequence but also capture the evolutionary trend of the sequence during optimization and its impact on performance.

[0024] Based on the above approach, when evaluating candidate biological sequences, the prediction model can combine the learned evolutionary laws to more accurately and comprehensively characterize the degree of optimization of the candidate sequence relative to the original sequence, thereby effectively improving the ability to characterize the degree of optimization of biological sequences and the accuracy of prediction results. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of a system architecture according to an embodiment of this application; Figure 2 This is a flowchart illustrating a biological sequence optimization degree prediction method according to an embodiment of this application; Figure 3 This is a schematic diagram of the prediction model construction process according to an embodiment of this application; Figure 4 This is a schematic diagram of the raw biological sequence preprocessing process according to one embodiment of this application; Figure 5 This is a schematic diagram of a biological sequence optimization degree prediction device according to an embodiment of this application; Figure 6 This is a schematic block diagram of an electronic device according to an embodiment of the present application. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein. In embodiments of this application, "B corresponding to A" means that B is associated with A. In one implementation, B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0029] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0030] Before introducing the technical solution of this application, the relevant knowledge of this application will be introduced below: In this application, a biological sequence refers to an information sequence formed by arranging basic biological units in a certain order, including amino acid sequences and nucleotide sequences. Amino acid sequences typically correspond to proteins or polypeptides, while nucleotide sequences typically correspond to DNA or RNA. In the embodiments of this application, the biological sequence is preferably a protein sequence, especially an antibody amino acid sequence. For example, the amino acid sequence of an antibody heavy or light chain can be used as a biological sequence input into the model for processing.

[0031] In this application, the original biological sequence refers to the initial sequence used as a reference benchmark before optimization design. This sequence is typically derived from natural samples, database records, or existing experimental results. In the embodiments of this application, the original biological sequence is used as a comparison benchmark to evaluate the optimization level of candidate biological sequences. For example, an initial antibody sequence is used as a template sequence for subsequent mutation to generate candidate sequences.

[0032] In this application, a candidate biological sequence refers to a sequence to be evaluated obtained by modifying the original biological sequence (e.g., mutation, insertion, or deletion). In the embodiments of this application, the candidate biological sequence can be generated by single-point mutation or multi-point mutation, such as a new sequence obtained by replacing a certain amino acid site in an antibody sequence.

[0033] In this application, the degree of optimization of a biological sequence refers to a quantitative representation of the improvement in target performance of a candidate biological sequence relative to the original biological sequence. This performance may include, but is not limited to, binding ability, stability, expression efficiency, or in vivo survival advantage. In the embodiments of this application, the degree of optimization is typically represented by a score output by the model; for example, a higher numerical value indicates a greater degree to which the candidate sequence outperforms the original sequence in overall performance.

[0034] In this application, the prediction model refers to a computational model trained using machine learning methods for performance evaluation of biological sequences. In the embodiments of this application, the prediction model is obtained by learning from biological sequence evolutionary trajectory data; its input is a biological sequence or sequence pair, and its output is a prediction result characterizing the degree of optimization. For example, a sequence coding model based on a deep neural network can be used as the prediction model.

[0035] In this application, an evolutionary trajectory refers to a sequence path formed by connecting multiple biological sequences in a certain direction, used to characterize the process of a biological sequence gradually evolving from a low-performance state to a high-performance state. In the embodiments of this application, the evolutionary trajectory includes at least a starting biological sequence and an ending biological sequence, and may include multiple intermediate biological sequences. For example, in the same antibody family, a low-abundance antibody sequence is used as the starting point, a high-abundance antibody sequence as the ending point, and they are connected through mutation relationships to form a directed path.

[0036] In this application, abundance refers to the frequency or relative proportion of a biological sequence observed in a sample or dataset, used to characterize the degree of enrichment of the sequence in a population. In the embodiments of this application, abundance is used to approximate the dominance of a sequence; for example, in antibody sequencing data, high abundance sequences typically correspond to antibody sequences that have a strong competitive advantage in the immunization process.

[0037] In this application, a clonal cluster refers to a set of biological sequences originating from the same ancestral cell, which typically share similar structural features and genetic origins. In the embodiments of this application, clonal clusters can be classified based on sequence similarity, V(D)J gene information, or CDR region characteristics. For example, antibody sequences with the same V gene and similar CDR3 regions can be grouped into the same clonal cluster.

[0038] In this application, V(D)J gene information refers to the source information of the variable region gene fragments constituting a biological sequence, including the specific types of V (Variable), D (Diversity), and J (Joining) gene fragments. In the embodiments of this application, this information is used to assist in the grouping of biological sequences. For example, in antibody sequences, sequences with the same V and J gene sources are more likely to belong to the same clonal group.

[0039] In this application, sequence features refer to information extracted from biological sequences to characterize their properties, including but not limited to sequence similarity, structural features, physicochemical properties, or gene origin information. In embodiments of this application, sequence features are used to group or model biological sequences, such as clustering antibody sequences based on sequence similarity.

[0040] In this application, a state transition operation refers to the operation of changing from one sequence state to another in the biological sequence space. In the embodiments of this application, the state transition operation is preferably a single-point mutation, that is, replacing an amino acid at a single position in the sequence. For example, mutating the 50th amino acid in an antibody sequence from A to G constitutes a state transition.

[0041] In this application, an evolutionary path sequence refers to a set of sequences formed by sequentially connecting a starting biological sequence, an intermediate biological sequence, and a ending biological sequence according to state transition operations. In the embodiments of this application, this path reflects the process of gradual sequence optimization, such as the gradual evolution from a low-abundance antibody sequence to a high-abundance antibody sequence through continuous single-point mutations.

[0042] In this application, a feature extraction network refers to a neural network model used to convert an input biological sequence into a high-dimensional feature representation. In the embodiments of this application, the network may be a pre-trained protein language model or other sequence encoding model used to extract semantic or structural features of the sequence.

[0043] In this application, the forward action probability refers to the probability of transitioning from the current biological sequence state to the next state, and the backward action probability refers to the probability of regressing from the current state to the previous state. In the embodiments of this application, the forward action corresponds to the sequence optimization direction (from low performance to high performance), and the backward action corresponds to the reverse change, which are used to jointly constrain the probability distribution of the model learning the evolutionary path.

[0044] In this application, the reward value is a numerical indicator used to characterize the quality of a biological sequence or evolutionary path. In the embodiments of this application, the reward value can be determined based on the abundance of the endpoint biological sequence or experimental measurement results. For example, a higher reward can be assigned to a high-abundance sequence to guide the model to learn a better path.

[0045] In this application, the policy network refers to the model to be optimized, used for probabilistic prediction of biological sequences; the reference model refers to a control model with fixed parameters, used to provide a baseline distribution. In the embodiments of this application, a training objective is constructed by comparing the probability output differences between the policy network and the reference model for the same sequence, so as to achieve the learning of the superiority / inferiority relationship of sequences.

[0046] In this application, reward bias refers to the difference in rewards between the endpoint biological sequence and the starting biological sequence under model evaluation, used to measure the model's ability to distinguish between superior and inferior sequences. In the embodiments of this application, by increasing the reward bias between the endpoint biological sequence and the starting biological sequence, the model is more inclined to assign higher scores to high-performance sequences.

[0047] In this application, the prediction result refers to the evaluation value or score output by the prediction model for the candidate biological sequence, which is used to characterize its degree of optimization. In the embodiments of this application, the prediction result may be a scalar value or a ranking result. For example, by scoring and ranking multiple candidate sequences, the target biological sequence with the best performance can be determined.

[0048] The following is through Figure 1 An exemplary system architecture of the embodiments of this application will be described.

[0049] Figure 1 This is a schematic diagram of a system architecture according to an embodiment of this application. The system architecture includes a user device 101, a data acquisition device 102, a training device 103, an execution device 104, a database 105, and a content library 106.

[0050] The data acquisition device 102 is used to read training samples from the content library 106 and store the read training samples in the database 105. The training samples involved in this embodiment may come from public databases (such as the Observed Antibody Space (OAS) database).

[0051] The training device 103 pre-trains the machine learning model based on the training samples maintained in the database 105, so that the trained machine learning model (i.e. the prediction model) can effectively predict the optimization degree of candidate biological sequences.

[0052] Additionally, refer to Figure 1The execution device 104 is equipped with an I / O interface 107 for data interaction with external devices. For example, it receives input commands sent by the user device 101 via the I / O interface; the input commands include a set of biological sequences to be predicted. The computing module 108 in the execution device 104 processes the input commands using a trained machine learning model to obtain output commands; the output commands include prediction results, which characterize the optimization degree of the candidate biological sequences, and are sent to the user device 101 via the I / O interface.

[0053] User equipment 101 may include mobile phones, tablets, laptops, handheld computers, mobile internet devices (MIDs), desktop computers, or other terminal devices with browser installation capabilities.

[0054] The execution device 104 can be a server.

[0055] For example, the server can be a rack server, blade server, tower server, or cabinet server, etc. The server can be a standalone test server or a test server cluster composed of multiple test servers.

[0056] In this embodiment, the execution device 104 is connected to the user equipment 101 via a network. The network can be an intranet, the Internet, Global System for Mobile Communications (GSM), Wideband Code Division Multiple Access (WCDMA), 4G, 5G, Bluetooth, Wi-Fi, or a voice communication network, or other wireless or wired networks.

[0057] It should be noted that, Figure 1 This is merely a schematic diagram of a system architecture provided in this application embodiment, and the positional relationships between the devices, components, modules, etc. shown in the figure do not constitute any limitation. In some embodiments, the data acquisition device 102, user device 101, training device 103, and execution device 104 may be the same device. The database 105 may be distributed across one server or multiple servers, and the content library 106 may be distributed across one server or multiple servers.

[0058] The technical solution of this application will be described in detail below: Figure 2 This is a flowchart of a biological sequence optimization degree prediction method according to an embodiment of this application, with reference to... Figure 2 The method may include: S210, Obtain a set of biological sequences to be predicted, the biological sequence set including: original biological sequences and at least one candidate biological sequence, the candidate biological sequence being obtained by modifying the original biological sequence; In an exemplary embodiment, the original biological sequence in this step can be derived from naturally known sequences, sequences included in databases, or functional sequences obtained through experimental screening, such as protein sequences (e.g., enzyme, antibody heavy or light chains), nucleic acid sequences, etc.; the candidate biological sequence is obtained by structural or sequence modification of the original biological sequence.

[0059] The aforementioned candidate biological sequences can be obtained in various ways. In one embodiment, site mutations, including single-point mutations, multi-point mutations, or regional mutations, can be performed on the original biological sequence to generate candidate sequences with different amino acid combinations. In another embodiment, functional regions in the original sequence can be replaced or rearranged based on sequence fragment recombination, such as locally replacing the complementarity-determining region (CDR) of an antibody, to construct a diverse set of candidate sequences. In a further embodiment, the original biological sequence can be guided by machine learning models or evolutionary information, for example, generating candidate sequences with more potential functions based on sequence conservation, structural constraints, or known functional sites. All of the above methods can introduce sequence diversity while maintaining the basic structural characteristics of the original sequence, thereby providing a sufficient candidate space for subsequent predictions.

[0060] In practical applications, such as antibody optimization or enzyme activity enhancement, it is often necessary to screen for optimal sequences with target performance from a large number of candidate sequences. By constructing a biological sequence set containing the original sequence and its derived candidate sequences in this step, it is possible to make unified predictions for sequences generated by different modification strategies while ensuring the consistency of sequence sources, thereby improving the comparability and reliability of prediction results. At the same time, systematically introducing diverse candidate sequences helps to cover a wider sequence space, increasing the probability of obtaining high-performance sequences and avoiding the limited search space problem caused by relying solely on limited mutations or empirical design in related technologies, thus improving the efficiency of subsequent prediction and screening.

[0061] S220, the set of biological sequences to be predicted is input into the prediction model to obtain the prediction result. The prediction model is obtained by pre-training using the evolutionary trajectory of biological sequences. The evolutionary trajectory represents the gradual evolution process of biological sequences from low optimization to high optimization. The prediction result is used to represent the optimization degree of the candidate biological sequences.

[0062] During the model training phase, this embodiment of the application obtains a large amount of biological sequence data from a biological sequence database. Through methods such as clustering, phylogenetic relationship construction, and sorting based on abundance information, multiple directed evolutionary trajectories that can reflect the gradual optimization of sequences are constructed. These evolutionary trajectories are used to pre-train the basic model, enabling the model to learn and master the inherent laws and constraints of biological sequences gradually evolving from low-optimization states to high-optimization states during the evolutionary process.

[0063] During the prediction phase, the prediction model has fully learned the aforementioned evolutionary laws. Therefore, when the original biological sequence and candidate biological sequences are input into the prediction model, the model does not need to input additional evolutionary trajectory data. Instead, based on the evolutionary knowledge learned during the pre-training phase, it evaluates the optimization degree of the candidate biological sequence and outputs the corresponding prediction result. This prediction result is used to characterize the optimization degree of the candidate biological sequence relative to the original biological sequence.

[0064] In the above manner, the embodiments of this application internalize the evolutionary information contained in the evolutionary trajectory of biological sequences into the prediction model in the form of model parameters, so that the model can comprehensively consider the rationality and optimization potential of the sequence in the evolutionary path during the actual prediction process, thereby effectively improving the accuracy and generalization ability of the prediction.

[0065] For example, the functional enhancement of antibody molecules is typically accompanied by a high-frequency somatic mutation process, manifested as a gradual evolution from an initial sequence to a high-affinity sequence. Therefore, this application constructs an evolutionary trajectory of antibody sequences and predicts candidate antibody sequences based on this trajectory. This effectively utilizes evolutionary information during antibody maturation, overcoming the shortcomings of related technologies that model only based on a single antibody sequence and ignore the evolutionary relationships between sequences. Furthermore, by structurally modeling the phylogenetic relationships in large-scale antibody sequence data, the model can incorporate evolutionary constraints during prediction, thereby improving the accuracy of predicting the optimization level of candidate antibody sequences.

[0066] For example, this applies to other types of biological sequences, such as enzyme protein sequences or other functional protein sequences. While specific functional indicators (such as catalytic activity, substrate specificity, or stability) may differ in relevant application scenarios, the essence is the same: a gradual optimization process of the sequence during evolution or screening. Therefore, by constructing evolutionary relationships or sequence change paths for corresponding biological sequences, evolutionary information between sequences can be incorporated into predictive models to evaluate candidate sequences. For instance, in enzyme optimization scenarios, sequence evolution paths can be constructed based on homologous sequence sets or directed evolutionary experimental data, and these paths can be used to predict the performance improvement trend of candidate enzyme sequences, thereby improving the efficiency of selecting optimal sequences.

[0067] In an exemplary embodiment, the prediction model is obtained by pre-training using biological sequence evolutionary trajectories, with reference to... Figure 3 ,include: S310, acquire training data; The aforementioned training data includes N (N is an integer not less than 1) evolutionary trajectories. Each trajectory is formed by the directed connection of multiple biological sequences in a predetermined direction, where the abundance of the starting biological sequence is lower than that of the ending biological sequence. This abundance characterizes the degree of enrichment of the corresponding biological sequence in the population, thereby reflecting the trend of sequence quality changes during evolution or selection. The evolutionary trajectory data constructed in this way can transform originally discrete single sequences into sequentially related sequence paths, thus explicitly depicting the gradual evolution of biological sequences from low fitness to high fitness.

[0068] In exemplary embodiments, the evolutionary trajectory can be constructed based on various methods, such as phylogenetic relationships, sequence similarity networks, and functional phenotypic data.

[0069] For example, constructing an evolutionary trajectory based on phylogenetic relationships includes: First, a standard biological sequence set is obtained, which can be derived from public databases or experimental sequencing data. This standard biological sequence set not only contains the sequences themselves, but also further includes sequence-related feature information, such as sequence similarity information, V(D)J gene annotation information, or clonal features reflecting origin relationships, thereby providing a foundation for subsequent clustering and trajectory construction.

[0070] Based on this, the standard biological sequence set is clustered according to the sequence characteristics of the biological sequences, dividing sequences with high similarity or origin correlation into multiple clonal groups. For example, in the antibody scenario, antibody sequences can be clustered by combining sequence similarity, V(D)J gene usage, and CDR region characteristics to identify clonal groups potentially originating from the same ancestral B cell. Through this clustering process, sequences within each clonal group exhibit strong homology and evolutionary correlation, thus providing a reasonable biological basis for subsequent phylogenetic relationship construction. Compared to existing methods that only model single sequences, this step introduces a clustering structure, organizing sequences into intrinsically related sets, which is beneficial for subsequent modeling of evolutionary relationships between sequences.

[0071] Furthermore, phylogenetic relationships between sequences are constructed within each clonal population to form corresponding evolutionary trajectories. In one implementation, connections between sequences can be established based on their similarity or mutation relationships, and the direction of these connections can be constrained by sequence abundance information. This means using low-abundance sequences as the evolutionary starting point and high-abundance sequences as the evolutionary endpoint, thereby constructing directional sequence paths. Specifically, sequences within the same clonal population can first be sorted according to their abundance values, and then directed connections can be established between adjacent or related sequences to form a path structure that gradually transitions from low-abundance sequences to high-abundance sequences, thus obtaining an evolutionary trajectory reflecting the gradual optimization process of the sequences. This evolutionary trajectory can characterize the process by which biological sequences transition from a low-fitness state to a high-fitness state during selection or natural evolution. By traversing all clonal populations, N evolutionary trajectories are obtained.

[0072] Through the aforementioned clustering and phylogenetic construction process, originally discrete biological sequences are transformed into trajectory structures with clear evolutionary relationships. This ensures that the training data contains not only information about the sequences themselves but also information about the evolutionary order and direction between sequences. In practical applications, such as antibody affinity optimization tasks, these evolutionary trajectories can reflect the maturation path of antibodies during immune selection, thus providing more constrained training signals for the prediction model. This allows the model to learn the directional patterns of sequence optimization, thereby improving its ability to predict the optimization degree of candidate sequences. Furthermore, this method is also applicable to other biological sequence types. For example, in the optimization of enzymes or other protein sequences, clustering can be performed based on sequence similarity or functional screening results, and similar evolutionary paths can be constructed. This enables unified modeling of the evolutionary processes of different types of biological sequences, demonstrating good versatility.

[0073] For example, constructing evolutionary trajectories based on networks of similarity between sequences includes: By treating each sequence in a standard biological sequence set as a node, connections are established based on the similarity or mutation distance between sequences, thus forming a sequence network structure. On this basis, pre-defined rules (such as a direction from low functional scores to high functional scores) can be used to orient the edges in the network, thereby extracting several directed paths as evolutionary trajectories. In this way, potential evolutionary paths can be mined from the overall sequence space without requiring a specific phylogenetic tree, thereby expanding the scope of evolutionary trajectory construction.

[0074] For example, evolutionary trajectories can also be constructed based on functional phenotypic data. In enzyme activity screening or antibody affinity assays, sequences can be sorted according to the functional assay results corresponding to different sequences, and connections between sequences can be established based on the direction of functional improvement, thereby constructing a trajectory of evolution from low-functional sequences to high-functional sequences. This approach, by directly introducing functional information, allows the evolutionary trajectory to reflect not only sequence changes but also the direction of functional optimization, thus enhancing the guidance for target performance during model training.

[0075] In an exemplary embodiment, the aforementioned standard biological sequence set is obtained by preprocessing the original biological sequence set. This preprocessing includes quality control, format standardization, redundancy removal, and feature normalization of the original biological sequences. Specifically, the original sequence data can first be quality-screened to remove low-quality sequences containing sequencing errors, abnormal lengths, or uncertain base / amino acid markers, thereby improving the reliability of subsequent analyses. Then, the retained sequences are format-standardized to ensure consistency in encoding methods, sequence orientation, and annotation formats across different sources, facilitating unified modeling. Based on this, highly repetitive or redundant sequences can be merged or deduplicated using sequence similarity comparison or clustering methods to reduce data redundancy and avoid model training bias. Furthermore, the sequences can be feature-standardized, such as by standardizing sequence length representation or extracting key region information, to form a standardized biological sequence set suitable for subsequent grouping, evolutionary trajectory construction, and model training. Through the above preprocessing steps, raw biological sequence data from complex sources and of varying quality can be transformed into a standardized biological sequence set with consistent structure, controllable quality, and representativeness. This provides a reliable data foundation for subsequent sequence feature-based grouping and evolutionary relationship modeling, and improves the stability and comparability of subsequent analysis results. Furthermore, this preprocessing method is universal and applicable to different types of biological sequence data, and can be further refined according to sequence type in specific application scenarios.

[0076] In an exemplary embodiment, reference is made to Figure 4 For antibody sequences, the original antibody sequence set (such as a spatial database of measured antibodies) is preprocessed to obtain a standard antibody sequence set, including: S410, completes the original antibody sequence set missing at the N-terminus and / or C-terminus to obtain a complete antibody sequence set; In actual sequencing data, due to factors such as sequencing depth or amplification preferences, some antibody sequences may have end deletions, affecting subsequent sequence alignment and feature extraction. To address this, missing regions can be filled in based on known conserved antibody sequences, homologous sequences in reference databases, or pre-defined antibody framework structures, resulting in a structurally complete set of antibody sequences. This step ensures a consistent structural basis among different sequences during subsequent processing, improving the accuracy of sequence comparisons.

[0077] In an exemplary embodiment, completion can be achieved as follows: For the original antibody sequences, the gene categories of regions V and J are identified and annotated to obtain the corresponding sequence annotation results. This annotation process can be based on established antibody gene annotation tools or alignment methods, enabling each antibody sequence to be classified into the corresponding V gene family and J gene family, thus providing a basis for subsequent matching.

[0078] After obtaining the sequence annotation results, similarity matching is performed in the antibody sequence encoding rule base based on the annotation information. In this embodiment, the antibody sequence encoding rule base preferably includes standard antibody gene reference sequences, such as the reference sequence set from the IMGT database. By comparing the original antibody sequence to be completed with the reference sequences of the corresponding V and J region families, candidate reference antibody sequences with high homology are screened, and the antibody sequence with the highest homology is preferably selected as the reference antibody.

[0079] Based on the reference antibody, the missing regions at the N-terminus and / or C-terminus of the original antibody sequence are filled. Specifically, based on the sequence information of the reference antibody in the corresponding region, the missing parts of the original sequence are inferred and completed, thereby restoring the complete structural framework of the antibody sequence. During the completion process, the structural partitioning information of the antibody sequence (such as the positional relationship between the frame region and the complementarity-determining region) can be combined to ensure that the completion result maintains structural rationality.

[0080] Preferably, after completion, the complete antibody sequence can be validated for consistency based on the characteristics of conserved regions in the antibody sequence. For example, conserved sites in the frame region can be compared to determine whether the completed sequence conforms to the typical characteristics of the antibody structure, thereby avoiding deviations introduced by mismatches or low-quality reference sequences. Sequences that fail the consistency check or cannot reliably match the reference antibody can be marked and removed in subsequent data processing.

[0081] The above-mentioned completion methods can address the issue of end-length deletions in the original antibody sequence, restoring its complete structural information and ensuring structural consistency among antibody sequences from different sources. This provides a reliable foundation for subsequent sequence alignment, clustering, and evolutionary trajectory construction, and improves overall data quality and analytical accuracy.

[0082] S420, according to antibody sequence encoding rules (such as IMGT, Kabat, or Chothia, etc.), the complete antibody sequence set is aligned to obtain an aligned antibody sequence set; In an exemplary embodiment, the frame region (FR) and complementarity-determining region (CDR) in the variable region can be located based on the structural partitioning characteristics of the antibody, and the sequences can be aligned according to a unified numbering system to make different antibody sequences comparable in key functional regions.

[0083] In an exemplary embodiment, antibody sequences can also be aligned as a whole using a multiple sequence alignment method. This alignment process maps antibody sequences of different lengths or with insertion / deletion differences to a unified coordinate system, thereby providing a consistent data representation for subsequent sequence feature-based grouping and evolutionary trajectory construction.

[0084] S430, perform redundancy removal processing on the aligned antibody sequence set to obtain a standard antibody sequence set; In exemplary embodiments, highly repetitive or identical antibody sequences can be merged based on sequence similarity, or duplicate samples can be removed while maintaining sequence diversity, to avoid bias from certain high-frequency sequences during subsequent analysis or model training. This step reduces data redundancy while retaining representative sequence information, thereby improving data utilization efficiency.

[0085] Through the above-mentioned completion, alignment, and redundancy removal processes, the original antibody sequence data can be transformed into a standard antibody sequence set with complete structure, uniform representation, and controlled redundancy. This makes different antibody sequences comparable in terms of structure and functional regions, thus providing a reliable data foundation for subsequent sequence grouping, lineage relationship construction, and evolutionary trajectory modeling, and is conducive to improving the accuracy and stability of subsequent analysis results.

[0086] In an exemplary embodiment, the original enzyme protein sequence set can be preprocessed to obtain a standard enzyme protein sequence set, thereby providing a unified data foundation for subsequent sequence feature-based clustering and evolutionary trajectory construction. Specifically, since enzyme protein sequences from different sources may differ in sequencing integrity, annotation consistency, and functional region labeling, the original enzyme protein sequences are first subjected to quality screening and structural integrity assessment to remove low-quality sequences with obvious truncation, abnormal length, or containing uncertain amino acid markers, in order to ensure the reliability of subsequent analysis data.

[0087] After obtaining qualified enzyme protein sequences, domain identification and functional region annotation are performed. In one embodiment, catalytic and accessory domains in the enzyme protein sequence can be located and segmented based on known enzyme family databases or domain databases (e.g., Pfam or homologous enzyme structure annotation information), making different enzyme sequences comparable at the functional region level. This step maps the original sequence to a unified functional structural framework, thereby reducing analytical bias caused by differences in sequence length or structural defects.

[0088] After domain labeling, the enzyme protein sequences undergo sequence alignment to represent them in a unified coordinate system. In one embodiment, multiple sequence alignment methods can be used to globally or locally align homologous enzyme protein sequences, ensuring that key catalytic sites and conserved residues are positioned correspondingly in the sequence representation. In another embodiment, a structure-assisted alignment method can be used, combining three-dimensional structural information with sequence information to improve the consistent expression of key functional sites. This alignment process enables enzyme protein sequences from different sources to be comparable in structure and functional regions.

[0089] Furthermore, redundancy removal is performed on the aligned enzyme protein sequences to obtain a standard enzyme protein sequence set. In one implementation, highly repetitive or functionally consistent sequences can be clustered and merged based on sequence similarity thresholds to reduce data redundancy. In another implementation, functional annotation information can be used to representatively screen samples with the same catalytic function but highly similar sequences to retain representative sequences with structural diversity. This process reduces data size while ensuring information coverage, improving the efficiency of subsequent model training and evolutionary trajectory construction.

[0090] Through the above preprocessing steps, original enzyme protein sequences with complex origins and inconsistent structures can be transformed into a set of standard enzyme protein sequences with complete structures, aligned functional regions, and controlled redundancy. This provides high-quality input data for subsequent modeling, functional prediction, and optimization analysis based on sequence evolutionary relationships, and improves the model's ability to learn the laws of enzyme protein functional changes and its predictive stability.

[0091] S320, The machine learning model is pre-trained based on the training data to obtain a prediction model.

[0092] During training, the model uses sequence relationships within the evolutionary trajectory as supervisory signals, enabling it to learn the evolutionary patterns from low-abundance sequences to high-abundance sequences, thereby establishing a mapping relationship between sequence features and optimization directions. Because the training data incorporates explicit evolutionary direction constraints, compared to existing methods that model based solely on single sequences, this approach allows the model to explicitly capture sequence evolution trends during the learning process. Consequently, when predicting candidate biological sequences, it can not only assess their current characteristics but also determine their potential development direction within the evolutionary path, thus improving the predictive ability to discriminate the degree of sequence optimization.

[0093] In practical applications, such as antibody affinity optimization tasks, the prediction model trained based on evolutionary trajectories can more accurately identify candidate antibody sequences with potentially high affinity. Similarly, in the optimization of enzymes or other functional proteins, constructing corresponding evolutionary trajectories and training the model can effectively predict the functional improvement trends of candidate sequences. Therefore, this implementation method, by introducing evolutionary trajectory information into the model training process, achieves systematic modeling of the evolutionary laws of biological sequences, thereby improving the applicability and predictive performance of the prediction model in different biological sequence optimization tasks.

[0094] In exemplary embodiments, the training methods of the above-mentioned machine learning models include, but are not limited to, supervised learning based on regression heads, trajectory modeling based on Generative Flow Network (GFlowNet), and sequence ordering learning based on Direct Preference Optimization (DPO).

[0095] For example, pre-training based on trajectory modeling of generative flow networks includes: For the nth evolutionary trajectory, any two biological sequences with a clear sequential relationship are selected from this trajectory as the starting sequence and the ending sequence, and this pair of sequences is used as the basic constraint unit of a training sample. Based on this, by introducing a stepwise single-point mutation approach, the change process between the starting sequence and the ending sequence is refined. That is, the evolutionary trajectory, which originally only had an endpoint relationship, is expanded into an evolutionary path sequence composed of multiple continuous states, thus explicitly introducing intermediate state information of the stepwise evolution of the sequence at the model training level. During the path construction process, each state change can be constrained based on the edit distance between sequences or preset mutation rules, ensuring that each state transition corresponds to a change at the single amino acid or nucleotide level, thereby guaranteeing the biological rationality of the evolutionary path. In some implementations, the intermediate states can be generated by the shortest edit path or based on random perturbation combined with screening rules, enabling the model to cover diverse evolutionary path patterns and enhance its generalization ability to complex sequence spaces.

[0096] After obtaining the evolutionary path sequence, each biological sequence in the path is sequentially input into a feature extraction network for encoding to obtain a corresponding sequence feature representation. The feature extraction network converts discrete biological sequences into continuous vector representations, enabling it to simultaneously characterize the sequence's local structural features, conserved region information, and potential functional features, thereby providing a unified feature space for subsequent state transition modeling. In different implementations, the feature extraction network can employ an attention-based sequence model, or an encoding model based on graph or convolutional structures, to adapt to the data characteristics of different types of biological sequences.

[0097] After obtaining the sequence feature representation, the transition relationships between states are modeled using both forward and backward action prediction networks to output the probability of the current state evolving into a subsequent state and the probability of the current state backtracking to a previous state, respectively. By simultaneously introducing constraints in both the forward and backward directions, the model learns not only unidirectional sequence evolution paths during training but also the overall consistency of those paths, thereby reducing the risk of learning erroneous paths due to noise or local biases. In practical applications, this bidirectional modeling approach effectively enhances the model's robustness to complex evolutionary trajectories.

[0098] Furthermore, the probabilities of forward and reverse actions corresponding to each state in the evolutionary path sequence are cumulatively calculated, and an overall optimization objective is constructed by combining this with the reward information corresponding to the endpoint biological sequence. The reward value characterizes the quality of the endpoint sequence in terms of the target function; for example, in an antibody scenario, it corresponds to the degree of enhanced affinity, and in an enzyme protein scenario, it corresponds to the degree of improved catalytic activity or stability. By applying this reward information to the probabilistic learning process of the entire path, the model tends to learn evolutionary paths leading to high-reward sequences during training, thereby establishing a correlation between sequence changes and functional optimization.

[0099] After completing the above training process, the feature extraction network, forward action prediction network, and backward action prediction network are jointly optimized and updated to obtain the final prediction model. In this way, the model not only learns the static features of a single biological sequence but also the dynamic evolutionary patterns between sequences. This allows it to predict the possible evolutionary direction and optimization degree of the input candidate biological sequence during the inference stage, thereby achieving modeling and evaluation of the optimization path of biological sequences. In different application scenarios, such as antibody affinity optimization, enzyme activity enhancement, or other functional protein modification tasks, this model can consistently predict the potential optimization ability of sequences based on a unified evolutionary trajectory learning framework.

[0100] For example, pre-training based on a sequence ranking learning method using Direct Preference Optimization (DPO) includes: For the nth evolutionary trajectory, any two biological sequences with sequential evolutionary relationships are selected from the evolutionary trajectory as the starting sequence and the ending sequence, and the starting sequence and the ending sequence are used as a pair of training samples to input into the model for joint modeling.

[0101] In one implementation, the structural information of the starting biological sequence and the ending biological sequence is first input into a policy network model. This structural information may include amino acid distribution features after sequence encoding, local structural features, or a high-dimensional embedding representation generated by a feature extraction network. The policy network model outputs a probability distribution for each amino acid position in the sequence, thereby forming a probabilistic model of the entire sequence generation or transformation process. After obtaining the probability distribution, the probability distributions of the starting and ending biological sequences at their respective sequence positions are cumulatively calculated to obtain the corresponding sequence log probability, thus transforming the overall sequence generation probability into an optimizable scalar representation.

[0102] Simultaneously, the same structural information is input into a reference model, which can be a base model with parameters frozen during the pre-training phase, used to provide a stable probability baseline. The corresponding log probability of the reference sequence is calculated through the reference model, thereby establishing a probability distribution difference constraint relationship between the policy network and the reference model.

[0103] Based on this, the reward bias between the starting biological sequence and the ending biological sequence is calculated based on the difference between the log probability output by the policy network and the log probability output by the reference model. This reward bias characterizes the degree of deviation of the current policy model from the reference model in predicting the direction of evolution or the trend of sequence optimization; essentially, it reflects the model's learning degree of "preferred sequence" priority in the evolutionary trajectory. In one implementation, this reward bias can be further used to construct a loss function, so that the model is simultaneously constrained by both sequence fitting ability and preference constraint ability during the optimization process.

[0104] Based on the loss function, the parameters of the policy network model are updated, allowing the model to gradually adjust its predictive ability for sequence position distribution. This makes the policy network more inclined to generate or assign higher probability weights to high-reward endpoint sequences during training. In this way, the model not only learns the static features of a single sequence but also learns the relative advantages and disadvantages of the evolution from the starting sequence to the endpoint sequence, thus implicitly encoding the evolutionary direction information of biological sequences in the parameter space.

[0105] In practical applications, such as antibody sequence optimization tasks, this training method can make the policy network more inclined to assign higher probabilities to antibody sequences with high abundance or high affinity, thereby enhancing the model's ability to model antibody maturation paths. In enzyme or other functional protein optimization tasks, the same can be done by introducing differences in functional superiority and inferiority as a source of reward bias, enabling the model to learn the evolutionary preference from low-functional sequences to high-functional sequences, thereby improving the predictive model's ability to identify the optimization direction of different biological sequences.

[0106] For example, pre-training based on supervised learning of the regression head includes: A regression head structure is introduced into the output layer of the model. The regression head structure is used to map the sequence representation obtained by the feature extraction network into continuous numerical output, thereby realizing the direct prediction of the target biological function indicators.

[0107] In antibody applications, the supervisory signal is preferably an antibody affinity index, such as binding strength, dissociation constant, or enrichment score—quantitative parameters that characterize the antibody's ability to bind to the antigen. In one implementation, the antibody sequence can be input into a prediction model, which then outputs a predicted affinity value. This predicted value is compared with the experimentally determined true affinity value to construct a regression loss function, thereby supervising the model's training. In this way, the model can gradually learn the nonlinear mapping relationship between the amino acid composition, structural region features, and affinity of the antibody sequence.

[0108] In one extended implementation, the regression head can be structurally adapted to different task requirements, such as employing a single-layer fully connected network, a nonlinear multilayer perceptron, or a regression module incorporating attention weighting mechanisms to enhance its ability to express complex sequence features. In some implementations, local region features of the sequence (such as CDR regions) and global features can be input into the regression head together, enabling the model to simultaneously focus on local key sites and overall structural information, thereby improving prediction accuracy.

[0109] In other biological sequence applications, such as enzyme protein or other functional protein optimization tasks, the monitoring signal corresponding to the regression head can be functional parameters such as catalytic activity, substrate conversion rate, stability index, or expression level. By mapping sequence features to corresponding functional values, the model can learn the influence of different amino acid mutations on functional changes, thereby achieving quantitative prediction of the direction of sequence optimization.

[0110] By employing the aforementioned supervised learning method based on regression heads, the predictive model can directly utilize experimentally measured functional data as supervisory signals during training, thereby establishing a continuous mapping relationship between biological sequences and functional phenotypes. Compared to methods that rely solely on sequence classification or relative ranking, this approach provides more refined numerical prediction capabilities, which is beneficial for improving the interpretability and prediction accuracy of the model in sequence optimization tasks.

[0111] In exemplary embodiments, for specific antibody affinity prediction or antibody sequence optimization tasks, after constructing a basic prediction model based on evolutionary trajectory training, the prediction model can be further fine-tuned to achieve rapid adaptation to specific downstream tasks. In this embodiment, the prediction model has typically learned general biological sequence evolution laws through large-scale evolutionary trajectory data; therefore, its main network parameters can characterize relatively general sequence evolution and optimization trends.

[0112] When adapting to downstream tasks, while keeping the main parameters of the pre-trained model essentially unchanged, a parameter-efficient fine-tuning (PEFT) mechanism is introduced to locally adjust the model. In a preferred embodiment, a low-rank adaptation (LoRA) method is used, introducing trainable low-rank parameter matrices into some linear transformation layers or attention calculation modules of the model. By incrementally modeling the original weight matrix in the form of low-rank decomposition, fine-grained adjustments to the model's output behavior are achieved.

[0113] Specifically, without altering the backbone structure and core parameters of the pre-trained model, only the newly added low-rank parameter matrix is ​​trained and updated. This allows the model to perform local feature reweighting and mapping adjustments for antibody affinity prediction or antibody sequence optimization tasks while maintaining its original evolutionary trajectory modeling capabilities. In this way, the model can be rapidly adapted to specific antibody data distributions and task objectives without requiring full parameter retraining, while consuming relatively few computational resources.

[0114] In antibody applications, this efficient parameter fine-tuning method allows the model to retain its overall understanding of antibody sequence evolution while focusing more on the characteristic differences of affinity-related key sites or complementarity-determining regions (CDRs), thereby improving sensitivity to changes in antibody binding ability and prediction accuracy. In other biological sequence applications, such as enzyme protein activity optimization or stability modification tasks, the same fine-tuning method can be used to enable the model to quickly adapt to different functional evaluation metrics without needing to rebuild or retrain the complete prediction model.

[0115] By using the above-mentioned efficient parameter fine-tuning methods, training computation costs and storage overhead can be significantly reduced, while effectively avoiding the degradation of the original evolutionary modeling capabilities that may be caused by full parameter fine-tuning. This enables efficient migration and stable adaptation of general biological sequence evolution prediction models to specific application tasks.

[0116] In an exemplary embodiment, the solution further includes: determining the target biological sequence based on the prediction results. Specifically, the prediction results are used to characterize the merits of each candidate biological sequence in terms of evolutionary trajectory or functional optimization dimension, such as the potential for affinity enhancement in antibody applications, or the degree of improvement in catalytic activity or stability in enzyme or other functional protein applications.

[0117] For example, determining the target biological sequence based on the prediction results includes: The candidate biological sequences are sorted based on the prediction results. Specifically, after inputting all candidate biological sequences into the prediction model, corresponding prediction scores or optimization indicators are obtained. The candidate sequences are then sorted from highest to lowest according to these scores, and the sequence with the highest score or above a preset threshold is selected as the target biological sequence. In this way, sequences with optimal or near-optimal characteristics can be quickly screened from a large-scale candidate sequence space, thereby improving sequence screening efficiency and reducing experimental verification costs.

[0118] For example, determining the target biological sequence based on the prediction results includes: Based on the prediction results, candidate biological sequences are iteratively optimized to gradually approach the target optimal sequence. In this approach, an initial candidate biological sequence can be used as a starting point. According to the optimization direction or score information output by the prediction model, the sequence is progressively modified, such as by introducing single-point mutations, local fragment replacements, or adjustments to conserved sites. After each round of optimization, the sequence is re-entered into the prediction model for evaluation, continuously updating the candidate sequence set. During multiple iterations, the prediction results output by the model guide the next optimization direction, causing the sequence to gradually evolve towards a higher prediction score, thus forming a closed-loop optimization process based on model feedback.

[0119] By using the above-mentioned target sequence determination methods based on sorting or iterative optimization, the prediction model can not only be used for static evaluation of candidate sequences, but also further guide the sequence design and optimization process. This enables an integrated sequence design process from "prediction" to "optimization" in application scenarios such as antibody affinity optimization, enzyme activity enhancement, or other functional protein modification, thereby improving the efficiency of obtaining target biological sequences and the accuracy of optimization.

[0120] The above text combined Figures 2 to 4 This paper describes an embodiment of the biological sequence optimization degree prediction method of this application. The following section combines... Figure 5 This application describes an embodiment of the apparatus.

[0121] Figure 5 This is a schematic block diagram of a biological sequence optimization degree prediction device 500 provided in an embodiment of this application, wherein the device 500 can be configured in an electronic device.

[0122] refer to Figure 5 The biological sequence optimization degree prediction device 500 includes: an acquisition module 510 and a prediction module 520.

[0123] The acquisition module 510 is used to acquire a set of biological sequences to be predicted, the biological sequence set including: original biological sequences and at least one candidate biological sequence, the candidate biological sequence being obtained by modifying the original biological sequence; the prediction module 520 is used to input the set of biological sequences to be predicted into a prediction model to obtain a prediction result, wherein the prediction model is obtained by pre-training using the evolutionary trajectory of biological sequences, the evolutionary trajectory representing the gradual evolution process of biological sequences from low optimization to high optimization, and the prediction result being used to represent the optimization degree of the candidate biological sequence.

[0124] In an exemplary embodiment, obtaining the prediction model by pre-training using biological sequence evolutionary trajectories includes: acquiring training data, the training data including: N evolutionary trajectories, each evolutionary trajectory including: multiple directed biological sequences, wherein the abundance of the starting biological sequence is lower than that of the ending biological sequence, the abundance being used to characterize the enrichment degree of the biological sequence in the population, and N being an integer not less than 1; and pre-training a machine learning model based on the training data to obtain the prediction model.

[0125] In an exemplary embodiment, obtaining N evolutionary trajectories includes: obtaining a standard biological sequence set; performing clustering processing on the standard biological sequence dataset based on the sequence characteristics of the biological sequences to obtain N clonal groups, wherein the sequence characteristics include sequence similarity, V(D)J gene information or clonal characteristics; and constructing phylogenetic relationships for each clonal group to obtain N evolutionary trajectories.

[0126] In an exemplary embodiment, the original biological sequence set is preprocessed to obtain a standard biological sequence set.

[0127] In an exemplary embodiment, the biological sequence is selected from antibodies. The preprocessing of the original antibody sequence set to obtain a standard antibody sequence set includes: completing the original antibody sequence set with missing N-terminus and / or C-terminus to obtain a complete antibody sequence set; aligning the complete antibody sequence set according to antibody sequence encoding rules to obtain an aligned antibody sequence set; and performing redundancy removal processing on the aligned antibody sequence set to obtain a standard antibody sequence set.

[0128] In an exemplary embodiment, the step of completing the original antibody sequence set missing at the N-terminus and / or C-terminus to obtain a complete antibody sequence set includes: annotating each original antibody sequence with gene categories in regions V and J to obtain sequence annotation results; performing similarity matching in an antibody sequence coding rule library based on the sequence annotation results to obtain a reference antibody; and filling in the original antibody sequences missing at the N-terminus and / or C-terminus based on the reference antibody to obtain a complete antibody sequence set.

[0129] In an exemplary embodiment, the reference antibody is selected from the antibody sequence with the highest homology in the matching results.

[0130] In an exemplary embodiment, the complete antibody sequence set is validated for consistency based on conserved region features.

[0131] In an exemplary embodiment, the step of constructing phylogenetic relationships for each clonal group to obtain N evolutionary trajectories includes: sorting the biological sequences in the nth clonal group according to their abundance, where n is an integer not less than 1 and not greater than N; taking the biological sequence with the lowest abundance as the evolutionary starting point and the biological sequence with the highest abundance as the evolutionary ending point, and performing directed connections on the biological sequences in the nth clonal group based on the similarity or mutation relationship of the biological sequences to obtain the nth evolutionary trajectory; and traversing the N clonal groups to obtain N evolutionary trajectories.

[0132] In an exemplary embodiment, the step of pre-training the machine learning model based on the training data to obtain a prediction model includes: for the nth evolutionary trajectory, taking any two biological sequences in the nth evolutionary trajectory as the starting point and the ending point, and constructing an evolutionary path sequence composed of multiple intermediate biological sequences by progressively applying state transition operations corresponding to single-point mutations; sequentially inputting each biological sequence in the evolutionary path sequence into a feature extraction network for encoding to obtain a corresponding sequence feature representation; based on the sequence feature representation, obtaining the forward action probability and backward action probability corresponding to each state through a forward action prediction network and a backward action prediction network, respectively; cumulatively calculating the forward action probability and backward action probability of each state in the evolutionary path sequence, and constructing a trajectory loss function using the reward value of the ending biological sequence as a supervision signal; and jointly training the feature extraction network, the forward action prediction network, and the backward action prediction network based on the trajectory loss function to obtain a prediction model.

[0133] In an exemplary embodiment, the step of pre-training the machine learning model based on the training data to obtain a prediction model includes: for the nth evolutionary trajectory, taking any two biological sequences in the nth evolutionary trajectory as the starting point and the ending point; inputting the structural information corresponding to the starting point biological sequence and the ending point biological sequence into a policy network model to obtain the probability distribution of each amino acid position in each sequence by the policy network; performing cumulative calculation on the probability distributions of the starting point biological sequence and the ending point biological sequence respectively to obtain the corresponding sequence log probability; inputting the structural information into a reference model to obtain the corresponding reference sequence log probability; calculating the reward deviation between the starting point biological sequence and the ending point biological sequence based on the log probability difference between the policy network and the reference model, and constructing a loss function based on the reward deviation; updating the parameters of the policy network based on the loss function to obtain the prediction model.

[0134] In an exemplary embodiment, the method further includes: determining a target biological sequence based on the prediction results, wherein determining the target biological sequence based on the prediction results includes: sorting candidate biological sequences based on the prediction results to determine the target biological sequence; or iteratively optimizing candidate biological sequences based on the prediction results, and determining the target biological sequence based on the prediction results of the optimized candidate biological sequences.

[0135] It should be understood that the device embodiments and method embodiments can correspond to each other, and similar descriptions can be found in the method embodiments. To avoid repetition, further details are omitted here. Specifically, Figure 5 The device 500 shown can perform Figure 2 The corresponding method embodiments, and the foregoing and other operations and / or functions of each module in the device 500 are respectively implemented to achieve Figure 2 For the sake of brevity, the corresponding processes in each method are not described in detail here.

[0136] The apparatus 500 of this application embodiment has been described above from the perspective of functional modules in conjunction with the accompanying drawings. It should be understood that this functional module can be implemented in hardware, in software instructions, or in a combination of hardware and software modules. Specifically, the steps of the method embodiments in this application can be completed by integrated logic circuits in the processor's hardware and / or by software instructions. The steps of the method disclosed in this application embodiment can be directly embodied as being executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. Optionally, the software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps in the above method embodiments.

[0137] Figure 6 This is a schematic block diagram of the electronic device 600 provided in an embodiment of this application. The electronic device 600 may be the training device or execution device described above, but is not limited thereto. Figure 6 As shown, the electronic device 600 may include: The system includes a memory 610 and a processor 620. The memory 610 stores a computer program 630 and transfers the computer program 630 to the processor 620. In other words, the processor 620 can retrieve and run the computer program 630 from the memory 610 to implement the methods described in the embodiments of this application.

[0138] For example, the processor 620 can be used to execute the steps in the above method according to the instructions in the computer program 630.

[0139] In some embodiments of this application, the processor 620 may include, but is not limited to: General-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0140] In some embodiments of this application, the memory 610 includes, but is not limited to: Volatile memory and / or non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0141] In some embodiments of this application, the computer program 630 may be divided into one or more modules, which are stored in the memory 610 and executed by the processor 620 to perform the method provided in this application. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 630 in the electronic device.

[0142] like Figure 6 As shown, the electronic device 600 may further include: Transceiver 640, which can be connected to processor 620 or memory 610.

[0143] The processor 620 can control the transceiver 640 to communicate with other devices; specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 640 may include a transmitter and a receiver. The transceiver 640 may further include antennas, and the number of antennas may be one or more.

[0144] It should be understood that the various components in the electronic device 600 are connected through a bus system, which includes a data bus, a power bus, a control bus, and a status signal bus.

[0145] According to one aspect of this application, a computer-readable storage medium is provided that stores computer instructions or programs thereon, which, when executed by a computer, enable the computer to perform the methods of the above-described method embodiments. Alternatively, embodiments of this application also provide a computer program product containing instructions that, when executed by a computer, cause the computer to perform the methods of the above-described method embodiments.

[0146] According to another aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method described in the above-described method embodiments.

[0147] In other words, when implemented using software, it can be implemented wholly or partially in the form of a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0148] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0149] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0150] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs. For example, the functional modules in the various embodiments of this application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0151] The embodiments of this application will now be described in more detail, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. In the following embodiments, the methods for acquiring training data and training models involved are referred to the Detailed Description section, and will not be repeated here.

[0152] Example 1: Model Performance Validation in Antibody Affinity Prediction Task This embodiment validates the model trained based on the proposed scheme using publicly available affinity datasets such as SKEMPI-V2, AB-Bind, and IgDesign. The results are shown in Table 1. The results demonstrate that, on all three datasets, AbEvolve significantly improves the affinity prediction relevance compared to the baseline model.

[0153] Table 1

[0154] Note: Baseline refers to the base model, and AbEvolve refers to the prediction model obtained after training based on the scheme in this application.

[0155] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for predicting the degree of optimization of biological sequences, characterized in that, include: A set of biological sequences to be predicted is obtained, the set of biological sequences including: original biological sequences and at least one candidate biological sequence, the candidate biological sequence being obtained by modifying the original biological sequence; The set of biological sequences to be predicted is input into the prediction model to obtain the prediction result. The prediction model is obtained by pre-training using the evolutionary trajectory of biological sequences. The evolutionary trajectory represents the gradual evolution process of biological sequences from low optimization to high optimization. The prediction result is used to represent the optimization degree of the candidate biological sequences.

2. The method according to claim 1, characterized in that, The method of obtaining the prediction model through pre-training using biological sequence evolutionary trajectories includes: Acquire training data, which includes: N evolutionary trajectories, each evolutionary trajectory including: multiple biological sequences with directed connections, wherein the abundance of the starting biological sequence is lower than that of the ending biological sequence, the abundance is used to characterize the enrichment degree of the biological sequence in the population, and N is an integer not less than 1; The machine learning model is pre-trained based on the training data to obtain a prediction model.

3. The method according to claim 2, characterized in that, Obtain N evolutionary trajectories, including: Obtain a standard biological sequence set; Based on the sequence characteristics of biological sequences, the standard biological sequence dataset is divided into N clonal groups, wherein the sequence characteristics include sequence similarity, V(D)J gene information or clonal characteristics. For each clonal population, phylogenetic relationships were constructed, resulting in N evolutionary trajectories.

4. The method according to claim 3, characterized in that, The original biological sequence set was preprocessed to obtain the standard biological sequence set; Preferably, the biological sequence is selected from antibodies, wherein the original antibody sequence set is preprocessed to obtain a labeled antibody sequence set, including: The original antibody sequence set missing at the N-terminus and / or C-terminus is completed to obtain a complete antibody sequence set; According to the antibody sequence encoding rules, the complete antibody sequence set is aligned to obtain an aligned antibody sequence set; The aligned antibody sequence set is subjected to redundancy removal to obtain the standard antibody sequence set.

5. The method according to claim 4, characterized in that, The original antibody sequence set with N-terminal and / or C-terminal deletions is completed to obtain a complete antibody sequence set, including: Gene category annotations were performed on regions V and J of each original antibody sequence to obtain sequence annotation results; Based on the sequence annotation results, similarity matching is performed in the antibody sequence encoding rule base to obtain reference antibodies; Based on the reference antibody, the original antibody sequence missing at the N-terminus and / or C-terminus is filled to obtain a complete antibody sequence set; Preferably, the reference antibody is selected from the antibody sequence with the highest homology in the matching results; Preferably, the complete antibody sequence set is subjected to consistency verification based on conserved region features.

6. The method according to claim 3, characterized in that, The process of constructing phylogenetic relationships for each clonal population yields N evolutionary trajectories, including: Sort the biological sequences in the nth clonal group according to their abundance, where n is an integer not less than 1 and not greater than N; Using the biological sequence with the lowest abundance as the starting point of evolution and the biological sequence with the highest abundance as the ending point of evolution, directed connections are made to the biological sequences in the nth clonal group based on the similarity or mutation relationship of the biological sequences to obtain the nth evolutionary trajectory. By traversing N clone populations, N evolutionary trajectories are obtained.

7. The method according to any one of claims 2-6, characterized in that, The step of pre-training the machine learning model based on the training data to obtain a prediction model includes: For the nth evolutionary trajectory, taking any two biological sequences in the nth evolutionary trajectory as the starting point and the ending point, an evolutionary path sequence composed of multiple intermediate biological sequences is constructed by gradually applying state transition operations corresponding to single-point mutations; Each biological sequence in the evolutionary path sequence is sequentially input into a feature extraction network for encoding to obtain the corresponding sequence feature representation; Based on the sequence feature representation, the forward action probability and backward action probability corresponding to each state are obtained through the forward action prediction network and the backward action prediction network, respectively. The forward and backward action probabilities of each state in the evolutionary path sequence are cumulatively calculated, and the reward value of the endpoint biological sequence is used as a supervision signal to construct a trajectory loss function. Based on the trajectory loss function, the feature extraction network, the forward action prediction network, and the backward action prediction network are jointly trained to obtain the prediction model.

8. The method according to any one of claims 2-6, characterized in that, The step of pre-training the machine learning model based on the training data to obtain a prediction model includes: For the nth evolutionary trajectory, any two biological sequences in the nth evolutionary trajectory are taken as the starting point and the ending point; The structural information corresponding to the starting biological sequence and the ending biological sequence is input into the strategy network model to obtain the probability distribution of each amino acid position in each sequence for the strategy network. The probability distributions of the starting biological sequence and the ending biological sequence are cumulatively calculated to obtain the corresponding sequence log probabilities; The structural information is input into the reference model to obtain the corresponding reference sequence log probability; Based on the log probability difference between the policy network and the reference model, the reward deviation between the starting biological sequence and the ending biological sequence is calculated, and a loss function is constructed based on the reward deviation. The policy network parameters are updated based on the loss function to obtain the prediction model.

9. The method according to any one of claims 1-6, characterized in that, Further includes: Based on the prediction results, the target biological sequence was determined. The step of determining the target biological sequence based on the prediction results includes: Based on the prediction results, candidate biological sequences are sorted to determine the target biological sequence; or Based on the prediction results, the candidate biological sequences are iteratively optimized, and the target biological sequence is determined based on the prediction results of the optimized candidate biological sequences.

10. A biological sequence optimization degree prediction device, characterized in that, include: An acquisition module is used to acquire a set of biological sequences to be predicted, the biological sequence set including: original biological sequences and at least one candidate biological sequence, the candidate biological sequence being obtained by modifying the original biological sequence; The prediction module is used to input the set of biological sequences to be predicted into the prediction model to obtain the prediction result. The prediction model is obtained by pre-training using the evolutionary trajectory of biological sequences. The evolutionary trajectory represents the gradual evolution process of biological sequences from low optimization to high optimization. The prediction result is used to represent the optimization degree of the candidate biological sequences.

11. An electronic device, characterized in that, include: Processor and memory; The memory is used to store computer programs; The processor is configured to execute the computer program to implement the biological sequence optimization degree prediction method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions or programs that, when executed on a computer, cause the biological sequence optimization degree prediction method as described in any one of claims 1 to 9 to be performed.

13. A computer program product, characterized in that, The computer program product includes computer instructions that, when some or all of the computer instructions are run on a computer, cause the biological sequence optimization degree prediction method as described in any one of claims 1 to 9 to be executed.

Images