A method and system for joint optimization of polypeptide sequence, structure, and function
Patent Information
- Application Number
- CN202611329048.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-03
AI Technical Summary
Existing peptide design methods suffer from problems such as error accumulation, target fragmentation, lack of characterization sharing, post-processing of structural information, single-target limitations, and poor mobility, resulting in low optimization efficiency and difficulty in simultaneously achieving multiple functional indicators.
A protein language model is used for the joint optimization of peptide sequence, structure and function. The structure prediction module and function prediction module are coupled through a cross-attention mechanism to construct a differentiable computational graph. Gradient-driven optimization is performed using sequence latent representation. Combined with multi-objective functions and constraint terms, the joint optimization of sequence, structure and function is achieved.
It improves the predictive consistency and reliability of peptide design, reduces error accumulation, achieves synergistic optimization of activity and druggability, improves optimization efficiency and the yield of effective candidates, and supports the application of a unified framework for various design tasks.
Smart Images
Figure CN123033558A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and computational biology, and in particular to a method and system for joint optimization of peptide sequences, structures, and functions based on a protein language model, as well as a computer-readable storage medium and electronic device. Background Technology
[0002] Peptides and proteins (collectively referred to as peptides in this article) have wide applications in disease treatment, in vitro diagnostics, biocatalysis, and functional materials due to their high specificity, low toxicity, and controllability. With the development of artificial intelligence technology, de novo peptide design using machine learning has gradually become a research hotspot.
[0003] Existing AI-based peptide design methods, while differing in specific routes, typically employ a serial pipeline approach, sharing the same architectural characteristic: a non-differentiable serial pipeline that first generates candidate sequences, then evaluates them, and finally selects the final candidate sequences. Specifically, existing pipelines can be broadly categorized into the following three types: The first type is the exhaustive approach. This method first fine-tunes the protein big language model using wild-type peptides in the target domain. Then, it sequentially changes the amino acid residues at each site of the wild-type peptide. The fine-tuned model provides probability scores for each amino acid residue at each site, and selects candidate single-point mutation sequences that meet the score criteria. Next, it performs three-dimensional structural modeling on the candidate sequences and selects the structure with the highest score. Then, it inputs the candidate sequences and receptor structures into molecular docking software to evaluate the binding energy and selects the amino acid with the lowest binding energy at each site. Finally, it outputs several sequences with the lowest binding energy according to the number of mutation sites set by the user and the required quantity.
[0004] The second type is the random sampling route; this type of method encodes the target functional polypeptide sequence into a two-dimensional feature matrix, uses it to train a functional polypeptide sequence characterization generation model, and then generates a variety of new two-dimensional feature matrices by adjusting the sampling method of noise, and then decodes them into new functional polypeptide sequences.
[0005] The third type is the predictive filtering approach. This type of method uses a peptide training sequence with target information and preset property information to fine-tune the parameters of the original peptide sequence prediction model. Then, the fine-tuned model is used to predict the sequence of the reference peptide sequence to obtain the predicted sequence, which is then combined to form a candidate sequence. The candidate sequence is then used to predict the structure to obtain the structural information. Finally, the candidate sequences are filtered according to preset screening conditions and structural information to obtain the target sequence set.
[0006] The search methods for the three types of routes mentioned above degenerate into exhaustive search, random sampling, and post-prediction filtering, respectively. The inventors discovered that these three degenerate forms share the same root cause: amino acid sequences are discrete symbol sequences. Although the structure predictor and function predictor are internally differentiable, their inputs are discrete sequences. Therefore, the "derivative of the function score with respect to the sequence" does not mathematically exist. Since the gradient does not exist, the downstream structure and function evaluation results cannot react back to the upstream sequence generation in the form of gradients. The evaluator can only perform scoring and ranking functions, and the optimization process can only degenerate into a search in discrete space.
[0007] Therefore, the above-mentioned serial pipeline method has at least the following problems: First, error accumulation: Sequence generation, structure prediction and function prediction are trained and inferred independently at each stage. The prediction error in the upstream stage will be passed down the pipeline and amplified, and there is no correction mechanism.
[0008] Secondly, the objectives are fragmented; the downstream structural rationality and functional objectives cannot be directly perceived during the sequence generation stage, and the generation and optimization of the sequence are difficult to be oriented towards the expected functions, and the entire design process is in an open-loop state.
[0009] Third, representations are not shared; even ignoring the discreteness problem, the three stages in the serial pipeline each use independent representation forms, namely, the sequence generation model outputs the sequence or one-hot encoding, the structure predictor accepts the sequence or multiple sequence alignment information and outputs coordinates, and the function predictor accepts the structure or molecular descriptor; there are no shared intermediate variables among the three, so they cannot be pieced together into a single differentiable graph in the topology of the computation graph, and "joint optimization" cannot be achieved.
[0010] Fourth, structural information is only used for post-hoc filtering; in existing methods, structural modeling scores or structural confidence scores are only used to "select the structure with the highest score" or "filter according to screening criteria," that is, they play the role of post-hoc filters. In the vast space of discrete sequences, the pass rate of filtering is extremely low, and a large amount of computation is consumed in generating candidates that are subsequently discarded.
[0011] Fifth, the limitations of single targets: existing methods often use a single functional indicator (such as binding affinity to the target) as the optimization target, which makes it difficult to simultaneously take into account drug-like properties such as stability, membrane permeability, and immunogenicity, and easily produces candidates that are "highly active but not druggable"; moreover, there is generally a trade-off between activity and drug-likeness in biochemistry, and there is no single optimal solution. A single weighted target can only give a compromise point under a certain weight, and the selection of this weight often lacks objective basis.
[0012] Sixth, the optimization efficiency is low; the discrete sequence space of peptides grows exponentially with the sequence length, and the hit rate of search methods based on random mutation, exhaustive search or simple heuristics is low and the iteration cost is high.
[0013] Seventh, poor transferability; for different design tasks such as cyclic peptides, proteases, and polypeptide functional materials, different pipelines and models often need to be built separately, making it difficult to reuse them under a unified framework.
[0014] Therefore, how to jointly model the sequence, structure, and function of peptides within a unified framework, so that downstream structural and functional objectives can drive upstream sequence variables, and achieve efficient synergistic optimization under multi-objective constraints, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0015] To overcome the technical shortcomings of existing "serial pipeline" peptide design methods, this invention provides a method and system for joint optimization of peptide sequence, structure, and function. This joint optimization method and system can improve the consistency and reliability of predictions, and has high optimization efficiency, reducing the impact of error accumulation, target fragmentation, and single-target limitations on the final peptide sequence optimization results.
[0016] The present invention achieves the above objectives through the following technical solutions: This invention provides a method for joint optimization of peptide sequence, structure, and function, comprising the following steps: S1. Encode the peptide sequences in the peptide sequence population using a pre-trained protein language model to obtain a sequence latent representation; based on the sequence latent representation, predict the three-dimensional structural information of the peptide through a structure prediction module to obtain a structural representation and a structural confidence score; based on the sequence latent representation and the structural representation, after fusion through a cross-attention mechanism, predict the functional attributes of the peptide through a function prediction module to obtain a functional score; wherein, the functional attributes include at least one attribute characterizing activity and at least one attribute characterizing drug-likeness. S2. The sequence naturalness term, structural rationality term, and functional score term are weighted and fused to construct an objective function; wherein, the sequence naturalness term is calculated from the negative log-likelihood of the polypeptide sequence, the structural rationality term corresponds to the structural confidence score, and the functional score term corresponds to the functional score; S3. Using the sequence latent representation as the optimization variable, update the sequence latent representation in the latent space spanned by the sequence latent representation along the negative direction of the gradient of the objective function, and decode the updated sequence latent representation into an updated polypeptide sequence using a decoder. S4. Repeat steps S1-S3 until the preset termination condition is met, and output the optimized peptide sequence set.
[0017] In this invention, the protein language model structure prediction module and function prediction module are coupled to form a single differentiable computational graph, such that the structure representation, the structure confidence score, and the function score are all differentiable functions of the sequence latent representation.
[0018] The core concept of this invention lies in reconstructing the non-differentiable sequential pipeline in existing technologies into a single differentiable computational graph with the sequence latent representation of a protein language model as the sole optimization variable. This allows the gradients of structural rationality and multiple functional objectives to be backpropagated to the sequence variables, thereby directly driving sequence optimization in the continuous latent space formed by the sequence latent representation. It should be noted that the contribution of this invention does not lie in "considering sequence, structure, and function together"; rather, it lies in the specific technical implementation of sharing the same sequence latent representation and coupling them through a differentiable path, enabling downstream objectives to backpropagate and drive upstream variables.
[0019] In this invention, the sequence latent representation refers to the dense floating-point vector obtained by mapping discrete sequences to a high-dimensional continuous space through a multi-layer neural network in a protein language model. The properties of a polypeptide are determined by its primary sequence, which folds to form a specific three-dimensional spatial structure. This three-dimensional spatial structure further determines its functional properties such as target binding, catalytic activity, stability, membrane permeability, and immunogenicity. Through sequence latent representation, the sequence, structure, and function of the polypeptide are closely linked, exhibiting a highly nonlinear mapping relationship.
[0020] In some embodiments, in step S1, the polypeptide includes a protease with catalytic function.
[0021] In some implementations, in step S1, the polypeptide sequence population is obtained by random generation, template-based generation, or perturbation generation starting from known active molecules.
[0022] In step S1 of this invention, the structure prediction module can be a structure predictor, and the structure confidence score in the predicted polypeptide structure can be used to represent the confidence of the structure representation; the function prediction module can be a function predictor. The two modules are coupled to the protein language model through a cross-attention mechanism, so that the sequence latent representation, the structure representation, and the functional attributes are jointly represented in a unified latent space.
[0023] In a specific implementation, in step S1, a pre-trained protein language model is used to encode peptide sequences in a peptide sequence population to obtain a sequence latent representation; simultaneously, binding pocket information and / or epitope information of the peptide sequence target are acquired; based on the sequence latent representation, the pocket information, and / or the epitope information, the three-dimensional structural information of the peptide is predicted by a structure prediction module to obtain a structural representation and a structural confidence score; based on the sequence latent representation, the structural representation, the pocket information, and / or the epitope information, after fusion via a cross-attention mechanism, the functional attributes of the peptide are predicted by a function prediction module to obtain a functional score. Through this design, joint optimization can be performed under target conditions.
[0024] In some implementations, step S1, in the process of predicting the functional properties of the peptide, further includes an active learning step: evaluating the prediction uncertainty of the functional score, and experimentally measuring the peptide based on the functional score and the prediction uncertainty; then backfilling the experimental results to update the functional prediction module. This active learning-based design loop reduces experimental costs and continuously improves model accuracy.
[0025] Furthermore, the prediction uncertainty σ(x) of the functional score can be estimated using methods such as deep integration or Monte Carlo dropout, and expressed as the acquisition function a(x) = (x) + β · σ(x) (where Based on the predicted mean of the functional score (β is the coefficient for balancing exploration and utilization), a batch of candidate peptides are selected from the optimization results and delivered for experimental testing; the experimental test results are backfilled into the training data to update the functional prediction module, thereby forming a closed-loop iteration of "design, synthesis, and testing", reducing experimental costs and continuously improving prediction accuracy.
[0026] In step S1 of the present invention, the functional attribute needs to include at least one attribute characterizing activity and at least one attribute characterizing drug-likeness, so as to simultaneously meet the activity requirement and the drug-likeness requirement.
[0027] In some embodiments, in step S1, the property characterizing activity is at least one of binding affinity to the polypeptide sequence target and biological activity.
[0028] In some embodiments, in step S1, the property characterizing drug-likeness is at least one of thermal stability, protease stability, membrane permeability, immunogenicity, and solubility.
[0029] In some embodiments, in step S1, when the polypeptide is a protease with catalytic function, the functional property also includes catalytic activity.
[0030] In step S2 of the present invention, the objective function is a differentiable multi-objective function, which serves as the total loss function for subsequent latent space gradient changes; correspondingly, the optimized peptide sequence set output in step S5 is a peptide sequence set obtained by sampling on the Pareto front determined by the multi-objective function.
[0031] In step S2 of this invention, the sequence naturalness term is used to constrain the update trajectory of the sequence latent representation within the training distribution neighborhood of the protein language model, thereby suppressing the statistical regularity of the decoded polypeptide sequence deviating from the natural sequence. Specifically, gradient descent in the latent space pushes the optimization variable z into a region outside the training distribution of the protein language model. While the decoded sequence may obtain a high functional score, it is a false positive—foldable, non-synthetic, and unreproducible in experiments. By using the negative log-likelihood of the candidate sequence under the protein language model as the sequence naturalness term, the optimization trajectory is constrained to the vicinity of the natural sequence manifold.
[0032] In step S2 of this invention, the structural confidence score is used as a differentiable structural rationality term in the objective function and directly participates in gradient calculation, actively pushing the sequence towards the direction of "foldable and structurally reliable", that is, changing from rejection filtering to guided optimization; under the same conditions, the output ratio of effective candidates is significantly improved.
[0033] In some implementations, in step S2, the objective function further includes a constraint term, which is weighted and fused with the sequence naturalness term, the structural rationality term, and the functional score term.
[0034] In a specific implementation, in step S2, the constraint term is at least one of the following: cyclization constraint term, active pocket geometry constraint term, non-natural amino acid constraint term, syntheticity constraint term, and novelty constraint term.
[0035] Wherein, the cyclization constraint term is used to constrain the spatial distance between cyclization sites; when the polypeptide is a cyclic peptide, the constraint term is a cyclization constraint term, which is used to constrain the cyclization conformation of the cyclic peptide under head-tail amide cyclization, disulfide bond cyclization, or staple bridge connection.
[0036] Wherein, when the polypeptide is a protease with catalytic function, the constraint term is an active pocket geometry constraint term, which is used to constrain the spatial configuration of the catalytic residues of the protease.
[0037] The novelty constraint term is used to penalize the similarity between the polypeptide and a pre-defined library of known molecules.
[0038] In some implementations, the objective function is expressed as: L(x) = λ s · L seq (x) + λ t · L struct (x) + Σ k λ k · L func,k (x) + Σ j μ j · C j (x) Where x represents the candidate polypeptide; L seq The sequence naturalness term characterizes the likelihood of candidate sequences under the protein language model, constraining the generated sequences to conform to the statistical regularities of natural proteins and possess syntheticity; L struct The structural rationality term measures the foldability and stability of the structure based on the structural confidence score and energy term; L func,k This is the k-th functional scoring item, used to optimize functional attributes such as binding affinity, activity, stability, membrane permeability, immunogenicity, and off-target toxicity; C j The j-th constraint term includes cyclization constraints, non-natural amino acid constraints, syntheticity constraints, and novelty constraints, etc.; λ s , λ t , λ k With μ j These are the corresponding weighting coefficients.
[0039] The unconventional aspect of this objective function lies not in its expression, but in the composition of its terms, the differentiability requirement, and its function. Existing "objective functions" (such as docking binding energy scores or screening threshold conditions) are essentially scoring functions used to rank or filter generated candidates; they do not need to be differentiable, and in fact, are not, thus belonging to scoring tables. The objective function of this invention must be differentiable with respect to the sequence's implicit representation z, because its function is not ranking, but rather generating update directions, thus belonging to loss functions.
[0040] In some implementations, in step S2, the gradient of the functional scoring item with respect to the sequence latent representation includes a direct path gradient and an indirect path gradient; the direct path gradient is directly fed back to the sequence latent representation by the functional prediction module, and the indirect path gradient is fed back to the sequence latent representation by the functional prediction module via the structural representation; the direct path gradient and the indirect path gradient are superimposed and used to update the sequence latent representation.
[0041] Among them, "the gradient of the functional score with respect to the sequence latent characterization" refers to the direction and magnitude of the change in the corresponding "functional score" when the "sequence latent characterization" of the peptide undergoes a small change.
[0042] In some implementations, step S2 involves normalizing each term based on the norm of its gradient in the objective function; here, each term refers to the sequence naturalness term, structural rationality term, and functional score term; if there are constraint terms, these are also included. This normalization process ensures that no single term dominates the update direction of the sequence latent representation.
[0043] In some implementations, step S3, updating the sequence latent representation along the negative direction of the gradient of the objective function, includes: using the sequence latent representation as the optimization variable z, and updating it according to z1 = z - η· L / z is updated, L is the objective function, and η is the step size; and at least one of momentum, adaptive step size, and gradient clipping is used in the update process.
[0044] Where, according to z1 = z - η· L / Gradient descent can change the space in which the optimization problem is located, allowing gradient information to be utilized, which significantly improves the optimization efficiency and hit rate in a large discrete sequence space.
[0045] In this process, after updating and optimizing the variables, the continuous latent representation is mapped back to the discrete amino acid sequence through the decoding head of the protein language model and combined with pass-through estimation. Alternatively, gradient-guided Gibbs sampling or evolutionary algorithms can be used to further refine the discrete sequence locally, thus balancing the efficiency of continuous differentiability optimization with the feasibility of discrete sequences.
[0046] In some implementations, step S3, after obtaining the updated polypeptide sequence, further includes discrete refinement: the updated polypeptide sequence is re-encoded into a sequence latent representation using the protein language model, a discrete local search is performed in the neighborhood of the sequence latent representation, and the search result is used as the sequence latent representation for the next iteration.
[0047] This invention also provides a joint optimization system for polypeptide sequence, structure, and function, which is used to implement the joint optimization method described above, and includes: The sequence characterization module is used to encode polypeptide sequences in a polypeptide sequence population using a pre-trained protein language model to obtain the sequence latent characterization. The structure prediction module is used to predict the three-dimensional structure information of the polypeptide based on the sequence hidden characterization, and to obtain the structure characterization and structure confidence score. The functional prediction module is used to predict the functional properties of the peptide based on the sequence latent representation and the structural representation after fusion through a cross-attention mechanism, and obtain a functional score. The joint optimization module is used to construct an objective function that integrates sequence naturalness, structural rationality, and functional score. The sequence latent representation is used as the optimization variable. The sequence latent representation is updated in the latent space spanned by the sequence latent representation along the negative direction of the gradient of the objective function. The updated peptide sequence is then decoded by the decoder. The output module is used to output an optimized set of peptide sequences after the preset termination conditions are met. The structure prediction module, the function prediction module, and the protein language model are coupled to form a single differentiable computational graph.
[0048] In this invention, the functional prediction module, the joint optimization module, and the output module work together to achieve synergistic optimization of peptide activity and druggability. The functional prediction module predicts both the activity-characterizing and druggability-characterizing attributes, which are then entered as parallel differentiable terms into the same objective function of the joint optimization module and differentiated with respect to the same optimization variable. Therefore, each update step follows the synthetic direction of the activity gradient and the druggability gradient, rather than "optimizing activity first and then remedying druggability." The output module outputs a set of peptide sequences sampled at the Pareto front determined by the objective function, explicitly providing a set of trade-offs between activity and druggability. Subsequent development selects from this set based on the route of administration, indications, and acceptable lower limit of stability, rather than a single compromise point determined by a subjectively chosen weight.
[0049] In some embodiments, the joint optimization system further includes an active learning module for evaluating the prediction uncertainty of the functional score, and for performing experimental measurements on the peptide based on the functional score and the prediction uncertainty, and then backfilling the results of the experimental measurements to update the functional prediction module.
[0050] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the joint optimization method as described above.
[0051] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the joint optimization method described above.
[0052] Compared with the prior art, the present invention has the following significant advantages: (1) By using the sequence latent representation as a common node, the sequence, structure and function are incorporated into a unified differentiable framework for end-to-end joint modeling, which eliminates the error accumulation in the serial pipeline and improves the consistency and reliability of the prediction; and provides a topological premise for the gradient to be back from the downstream target to the upstream variable. (2) By selecting attributes and coupling differentiable multi-objectives, the synergistic optimization of activity and druggability is achieved, which increases the proportion of druggable candidate molecules and avoids the problem of "high activity but not druggable". (3) The optimization guided by the gradient of the latent space spanned by the latent representation of the sequence, combined with the refinement of the discrete sequence, significantly improves the optimization efficiency and hit rate in the huge discrete sequence space. (4) By using structural confidence as a differentiable objective term instead of a post-hoc filter, the rejection screening is transformed into guided optimization, which improves the output ratio of effective candidates. (5) The unified framework can be transferred to various design tasks such as cyclic peptides, proteases and polypeptide functional materials, which is easy to reuse and thus builds a barrier at the technology platform level. Attached Figure Description
[0053] Figure 1 This is a schematic diagram of the overall process of a joint optimization method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of an architecture for the joint characterization and optimization of sequence, structure, and function provided in an embodiment of the present invention; Figure 3 This is a structural block diagram of a joint optimization system provided in an embodiment of the present invention. Detailed Implementation
[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0055] In the following embodiments, "peptide" refers to a molecule composed of several amino acid residues linked by peptide bonds, including but not limited to linear peptides, cyclic peptides, and proteases with catalytic functions; "protein language model" refers to a neural network model pre-trained on large-scale protein sequence data in a self-supervised manner (e.g., masked language modeling), which can provide an intermediate representation and sequence likelihood in the form of a continuous vector for the input amino acid sequence; "sequence latent representation" refers to the intermediate representation of sequence information output by the protein language model in the form of a continuous vector, which is also an optimization variable in this embodiment, denoted by z; "structure confidence score" refers to a measure of the confidence level of the predicted three-dimensional structure, such as residue-by-residue confidence and global confidence score.
[0056] Example 1 This embodiment provides a method for joint optimization of peptide sequence, structure, and function based on a protein language model. Figure 1 This is a schematic diagram of the overall process of the joint optimization method provided in this embodiment; Figure 2 This is a schematic diagram of the architecture for the joint characterization and optimization of sequence, structure, and function provided in this embodiment. For example... Figure 1 and Figure 2 As shown, the joint optimization method includes the following steps: Step s1: Obtain the initial peptide sequence population to be optimized. Depending on the design task, an initial population of candidate peptide sequences can be constructed using methods such as random generation, template-based generation, or perturbation generation starting from known active molecules, serving as the starting point for optimization.
[0057] Step s2: Encode the sequence latent representation Each polypeptide sequence in the population is encoded using a pre-trained protein language model to obtain the sequence latent representation z; The encoding method used in this step is known in the art; however, in this embodiment, the sequence latent representation z is not a static feature sent downstream, but an optimization variable for subsequent gradient descent.
[0058] Step s3: Predict the three-dimensional structure The sequence latent characterization z is input into the structure decoder to predict the three-dimensional structure information of the peptide, and the structure characterization and structure confidence score (including residue-by-residue confidence and global confidence) are obtained.
[0059] Step s4: Predict multiple functional attributes The sequence latent characterization z and the structural characterization are fused via a cross-attention mechanism and then input into the functional prediction head to jointly predict multiple functional attributes and obtain a functional score. Functional attributes include at least one activity attribute and at least one druggability attribute; the activity attribute is the binding affinity and / or biological activity to the target, and for protease design tasks, it also includes catalytic activity; the druggability attribute is at least one of thermal stability, protease stability, membrane permeability, immunogenicity, and solubility. In steps s2-s4, the structural decoder and the functional prediction head are coupled in parallel to the same sequence latent representation z, and then coupled with the protein language model via cross-attention. Together, they form a single differentiable computational graph. Therefore, the gradient of the functional score with respect to the sequence latent representation z has two paths: L func / z = ( L func / z) 直接 + ( Lfunc / h struct )·( h struct / z) 间接 Among them, L func For the function rating, h struct For structural representation; the gradients of the two paths are superimposed and used to update z.
[0060] Step s5: Construct a differentiable multi-objective function The objective function is a differentiable multi-objective function that integrates the sequence naturalness term, the structural rationality term, and the functional score term, and may include constraint terms, in the form of: L(x) = λ s · L seq (x) + λ t · L struct (x) + Σ k λ k · L func,k (x) + Σ j μ j · C j (x) Among them, the sequence naturalness term L seq The negative log-likelihood of the candidate sequence under the protein language model is taken to constrain the candidate sequence to conform to the statistical regularity of natural proteins, thereby improving foldability and syntheticability; the structural rationality term L struct Constructed based on structural confidence score and energy term; Functional score term L func,k Construct the constraint term C based on the functional scores obtained in step s4. j Including cyclization constraints, non-natural amino acid constraints, syntheticity constraints, and novelty constraints, λ s , λ t , λ k With μ j These are the corresponding weighting coefficients.
[0061] In this step, the parameters of the objective function are not arbitrarily replaceable. If the sequence naturalness term L is removed... seq Latent space optimization deviates from the natural sequence manifold, and the model likelihood and structural confidence of the decoded sequence decrease simultaneously, resulting in false positives of "high functional scores but incompressible and incomposable"; if the structural rationality term L is removed... structThe functional gradient pushes the sequence into the adversarial region of the structure predictor, resulting in inflated functional scores and low structure confidence. If the objective function is not differentiable, it degenerates into scoring and ranking, and the optimization degenerates into random search, with efficiency falling back to the level of the existing serial pipeline.
[0062] Step s6: Latent Space Gradient Optimization and Decoding In this step, the optimization variable is the sequence implicit representation z itself; the structural representation, structural confidence score, and functional score are all differentiable functions of z, and they are not directly modified but change with z. The specific process is as follows: Step s6.1, Calculation L / z, the gradient is backpropagated along the objective function, structural rationality term / functional score term / sequence naturalness term, each predictor head, cross attention, and optimization variable z; Step s6.2, Update: z1 = z η· L / z can employ momentum, adaptive step size, and gradient clipping; Step s6.3, Decoding: The updated optimized variable z1 is decoded into the updated polypeptide sequence by the decoder; Step s6.4, Discrete Refinement: Since continuous points in the latent space may not correspond to valid discrete sequences, after decoding, the obtained sequence needs to be re-encoded back into the latent space by the protein language model, and a discrete local search is performed in its neighborhood (for example, fixed-point replacement of several sites and comparison of objective function values). The search result is used as z for the next iteration, thereby ensuring consistency between continuous optimization and discrete implementation.
[0063] Step s7, Iteration and Output Determine if the preset termination conditions are met (e.g., reaching a preset number of iterations, objective function convergence, or functional score meeting a preset threshold); if not, return to step s2 for recoding and continue iteration; if yes, output the optimized peptide sequence set. When the objective function is a multi-objective function, sample at its determined Pareto front to output a set of peptide sequences that achieve different trade-offs between activity and druggability.
[0064] Furthermore, this embodiment also includes an active learning step after step s4: using deep integration or Monte Carlo discard estimation of the prediction uncertainty σ(x) of the function score, selecting a batch of candidate delivery experiments based on the function score and σ(x), and backfilling the measurement results to update the function prediction module, thereby forming a closed loop of design-experiment-model update.
[0065] Example 2 This embodiment designs cyclic peptide conjugates targeting a specific target. Before step s2, the target binding pocket information is obtained as a condition, and in steps s3 and s4, this pocket condition is applied to the structure and function predictions through a cross-attention mechanism. A cyclization constraint term is added to the objective function in step s5 to constrain the spatial distance between cyclization sites. The activity attribute is the binding affinity to the target, and the druggability attribute is membrane permeability and protease stability. Other steps and processes are the same as in Example 1.
[0066] The output of this embodiment is a set of cyclic peptide sequences obtained by sampling at the activity and druggability Pareto front.
[0067] Example 3 This embodiment is designed for a protease with catalytic function. The polypeptide is a protease, and the functional properties in step s4 include catalytic activity; the objective function in step s5 includes catalytic activity and thermal stability terms. Other steps and processes are the same as in Example 1.
[0068] This embodiment demonstrates the portability of the unified framework of the present invention. There is no need to build a separate pipeline for this task; only the set of attributes predicted by the function prediction head and the corresponding items in the objective function need to be changed.
[0069] Example 4 This embodiment also provides a system for implementing any one of the methods in embodiments 1-3 above.
[0070] Figure 3 This is a block diagram of the joint optimization system provided in this embodiment. See also... Figure 3 The system includes: a sequence characterization module for encoding peptide sequences using a pre-trained protein language model to obtain a sequence latent characterization; a structure prediction module for predicting three-dimensional structural information based on the sequence latent characterization to obtain a structure characterization and a structure confidence score; a function prediction module for jointly predicting multiple functional attributes based on the sequence latent characterization and structure characterization to obtain a function score; a joint optimization module, which includes an objective function construction unit, a latent space optimization unit, and a decoding unit, for constructing a differentiable multi-objective objective function and optimizing the peptide in the latent space guided by gradients and decoding it into an updated peptide sequence; an active learning module for estimating the prediction uncertainty of the function score, selecting candidate peptides for experimental determination, and backfilling the results to update the function prediction module; and an output module for outputting an optimized set of peptide sequences when the termination condition is met.
[0071] This embodiment also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor; the processor and the memory are communicatively connected via a bus; when the processor executes the computer program, it implements the method described in any of the above embodiments. The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in any of the above embodiments; the computer-readable storage medium includes, but is not limited to, various media capable of storing program code such as read-only memory, random access memory, magnetic disks, or optical disks.
[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for joint optimization of polypeptide sequence, structure, and function, characterized in that, It includes the following steps: S1. Encode the polypeptide sequences in the polypeptide sequence population using a pre-trained protein language model to obtain the sequence latent representation. Based on the sequence latent characterization, the three-dimensional structural information of the polypeptide is predicted by the structure prediction module to obtain the structural characterization and structural confidence score. Based on the sequence latent characterization and the structural characterization, after fusion via a cross-attention mechanism, the functional properties of the peptide are predicted by the functional prediction module to obtain a functional score; wherein, the functional properties include at least one property characterizing activity and at least one property characterizing drug-likeness; S2. The sequence naturalness term, structural rationality term, and functional score term are weighted and fused to construct an objective function; wherein, the sequence naturalness term is calculated from the negative log-likelihood of the polypeptide sequence, the structural rationality term corresponds to the structural confidence score, and the functional score term corresponds to the functional score; S3. Using the sequence latent representation as the optimization variable, update the sequence latent representation in the latent space spanned by the sequence latent representation along the negative direction of the gradient of the objective function, and decode the updated sequence latent representation into an updated polypeptide sequence using a decoder. S4. Repeat steps S1-S3 until the preset termination condition is met, and output the optimized peptide sequence set.
2. The joint optimization method as described in claim 1, characterized in that, In step S1, the polypeptide includes a protease with catalytic function; The polypeptide sequence population can be obtained by random generation, template-based generation, or perturbation generation starting from known active molecules.
3. The joint optimization method as described in claim 2, characterized in that, A pre-trained protein language model is used to encode peptide sequences in a population of peptide sequences to obtain sequence latent representations; at the same time, binding pocket information and / or epitope information of peptide sequence targets are obtained. Based on the sequence hidden characterization, the pocket information and / or the epitope information, the three-dimensional structural information of the polypeptide is predicted by the structure prediction module to obtain the structural characterization and structural confidence score. Based on the sequence latent representation, the structural representation, the pocket information and / or the epitope information, the functional properties of the peptide are predicted by the functional prediction module after being fused through a cross-attention mechanism, and a functional score is obtained.
4. The joint optimization method as described in claim 1, characterized in that, In step S1, the process of predicting the functional properties of the polypeptide also includes an active learning step: evaluating the prediction uncertainty of the functional score, and performing experimental measurements on the polypeptide based on the functional score and the prediction uncertainty; and then backfilling the results of the experimental measurements to update the functional prediction module.
5. The joint optimization method as described in claim 1, characterized in that, In step S1, the property characterizing activity is at least one of binding affinity to the polypeptide sequence target and biological activity. The properties characterizing drug properties are at least one of thermal stability, protease stability, membrane permeability, immunogenicity, and solubility. When the polypeptide is a protease with catalytic function, the functional property also includes catalytic activity.
6. The joint optimization method as described in claim 1, characterized in that, Step S2 satisfies at least one of the following conditions: The objective function also includes a constraint term, which is weighted and fused with the sequence naturalness term, the structural rationality term, and the functional score term, and is at least one of the following: cyclization constraint term, active pocket geometry constraint term, non-natural amino acid constraint term, syntheticity constraint term, and novelty constraint term; The gradient of the functional scoring item with respect to the sequence latent representation includes the direct path gradient and the indirect path gradient; the direct path gradient is directly fed back to the sequence latent representation by the functional prediction module, and the indirect path gradient is fed back to the sequence latent representation by the functional prediction module via the structural representation. Based on the norm of the gradients of each term in the objective function, the terms are normalized.
7. The joint optimization method as described in claim 1, characterized in that, Step S3 satisfies at least one of the following conditions: The method of updating the sequence latent representation along the negative direction of the gradient of the objective function includes: using the sequence latent representation as the optimization variable z, and updating it according to z1 = z η· L / z is updated, L is the objective function, and η is the step size; and at least one of momentum, adaptive step size and gradient clipping is used in the update process. After obtaining the updated polypeptide sequence, the process further includes discrete refinement: the updated polypeptide sequence is re-encoded into a sequence latent representation using the protein language model, a discrete local search is performed in the neighborhood of the sequence latent representation, and the search result is used as the sequence latent representation for the next iteration.
8. A joint optimization system for polypeptide sequence, structure, and function, characterized in that, It includes: The sequence characterization module is used to encode polypeptide sequences in a polypeptide sequence population using a pre-trained protein language model to obtain the sequence latent characterization. The structure prediction module is used to predict the three-dimensional structure information of the polypeptide based on the sequence latent characterization, and obtain the structure characterization and structure confidence score. The functional prediction module is used to predict the functional properties of the peptide based on the sequence latent representation and the structural representation after fusion through a cross-attention mechanism, and obtain a functional score. The joint optimization module is used to construct an objective function that integrates sequence naturalness, structural rationality, and functional score. The sequence latent representation is used as the optimization variable. The sequence latent representation is updated in the latent space spanned by the sequence latent representation along the negative direction of the gradient of the objective function. The updated peptide sequence is then decoded by the decoder. The output module is used to output an optimized set of peptide sequences after the preset termination conditions are met. The structure prediction module, the function prediction module, and the protein language model are coupled to form a single differentiable computational graph.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the joint optimization method as described in any one of claims 1-7.
10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the joint optimization method as described in any one of claims 1-7.