Macromolecule multi-attribute parallel generation method based on latent space decoupling

Abstract

This invention relates to the field of computational design technology for macromolecules, and discloses a parallel generation method for multiple macromolecular attributes based on latent space decoupling. The method includes: training a model to output a feature tensor; converting constraint vectors into direction vectors via an association matrix; calculating the projection components of the feature tensor onto the direction vectors and determining adjustment coefficients; correcting the feature tensor using an orthogonal complement projection matrix to separate shared gradient components; outputting an orthogonalized tensor; feeding it to a decoding layer to predict residue probability distributions to generate macromolecular candidate sequences; and dynamically adjusting the adjustment coefficients based on multi-objective reward feedback. This invention eliminates directional hedging of physicochemical property gradients, prevents manifold oscillations, avoids spatial conformational collapse caused by local residue mutations, and prevents sequence syntax collapse.

Description

Technical Field

[0001] This invention relates to a parallel generation method for multiple properties of macromolecules based on latent space decoupling, belonging to the field of macromolecular computational design technology. Background Technology

[0002] In the field of macromolecular design and information computing, pre-training macromolecular language models using large-scale sequence datasets is a mainstream technical approach. The underlying macromolecular language model captures the long-range topological distribution patterns of amino acid residue sequences through self-supervised probabilistic modeling, thereby generating synthetic backbone sequences that conform to the spatial folding constraints of natural proteins. This constitutes the physical basis for highly dynamic molecular optimization. However, when faced with parallel optimization design scenarios that require simultaneous constraints on multiple physicochemical indicators such as affinity activity, immunogenicity response, and development stability, existing technologies often convert multidimensional attributes into conditional vectors and combine them with sequence features. Since the basic structure shared by physicochemical attributes in the feature manifold is not considered, the gradients of the conditional constraints of each attribute are prone to directional conflict in the same dimension. This causes the transformer decoding network to generate continuous oscillations in the feature manifold when generating probability distributions for each residue. This can easily lead to chain damage, causing local residue mutations that lead to the collapse of the overall spatial conformation, and ultimately resulting in the loss of macromolecular specific activity.

[0003] To mitigate the aforementioned gradient conflicts, existing attempts often employ methods such as expanding the model parameter scale or constructing a back-end cascaded funnel screening process. However, expanding the feature space dimension cannot prevent nonlinear interference between attribute vectors within a high-dimensional manifold. Furthermore, serial funnel screening, lacking an in-situ feedback mechanism, easily misses potential global optimal solutions in the initial coarse screening stage, ultimately degenerating the entire optimization process into inefficient random trial and error within the full parameter space. This significantly increases computational costs and fails to decouple multi-dimensional attribute conflicts. Existing technologies face the following technical challenges: 1. Severe conflicting gradients of multiple attributes occur in the sequential feature flow path, lacking in-situ decoupling methods within the latent space; 2. Lack of synergy between multi-objective constraint mechanisms and the folding rules of the basic language model, making local mutations prone to disrupting folding syntax; 3. The screening process lacks global adaptive feedback, resulting in low search efficiency for generating multiple physical indices.

[0004] Therefore, the technical problem to be solved by this invention is how to decouple the mutual perturbation gradients of multidimensional attributes in the feature transfer path of macromolecules and achieve parallel control of multidimensional mutually exclusive materialization indicators while maintaining the grammatical constraints of the sequence folding of the macromolecular base language model. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A parallel generation method for multiple properties of macromolecules based on latent space decoupling, comprising the following steps:

[0006] Step S1: Use a large-scale unlabeled macromolecular sequence dataset to pre-train a transformer model and output a sequence latent space feature tensor that represents the prefix of the macromolecular sequence.

[0007] Step S2: Extract the multi-attribute constraint embedding vector transformed from external multi-dimensional physical property indicators; fuse the sequence latent space feature tensor with the multi-attribute constraint embedding vector and send it to the flow path; through the macromolecular prior physical property correlation matrix, uniquely map the multi-attribute constraint embedding vector to a set of physical property feature direction vectors in the latent space to define the gradient evolution trend of each physical property in the molecular manifold; call the attribute gradient projection algorithm to calculate the projection components of the sequence latent space feature tensor on each physical property feature direction vector; determine the adjustment coefficient of each physical property deviation based on the algebraic difference between the obtained initial macromolecular physical property features and the preset safety boundary threshold; use the orthogonal complementary space projection matrix to transform and correct the sequence latent space feature tensor to separate the common gradient components that trigger parameter hedging; adopt the linear combination calibration method to subtract the sum of the products of each physical property deviation adjustment coefficient and each physical property feature direction vector from the input sequence latent space feature tensor, and output the orthogonalized sequence feature tensor after parameter-level calibration.

[0008] Step S3: Input the orthogonalized sequence feature tensor into the transformer decoding layer to predict the probability distribution of amino acid residues at each position to generate macromolecular candidate sequences. Input the macromolecular candidate sequences into the multi-objective reward function for quantitative evaluation of physical property scores. Dynamically update the adjustment coefficients of each physical property deviation based on the obtained score feedback results.

[0009] Preferably, the pre-trained transformer model adopts a conditionally generated pre-trained transformer model; in step S1, the macromolecular sequence syntax and structural folding rules are obtained through self-supervised pre-training on a large-scale unlabeled macromolecular sequence dataset, and are used as the supervision boundary for generating macromolecular candidate sequences; external multidimensional physical property indicators include macromolecular affinity indicators, stability parameters, and immunogenicity levels; multi-attribute constraint condition embedding vectors are used as condition vectors to guide the parallel optimization of macromolecular candidate sequences.

[0010] Preferably, in step S2, uniquely mapping the multi-attribute constraint embedding vector to a set of physical property direction vectors in the latent space includes the following steps: Step S21, extracting the multi-attribute constraint embedding vector transformed from the externally input affinity index, stability performance parameter, and immunogenicity level; Step S22, mapping the multi-attribute constraint embedding vector through linear transformation of the macromolecular prior physical property correlation matrix to generate a set of physical property direction vectors in the latent space.

[0011] Preferably, in step S2, determining the adjustment coefficients for each property deviation includes the following steps: Step S23, calling the attribute gradient projection algorithm to calculate the projection components of the current sequence latent space feature tensor on each property feature direction vector; Step S24, calculating each property deviation adjustment coefficient based on the algebraic difference relationship between the obtained initial macromolecular property features and the preset safety boundary threshold, and increasing the corresponding property deviation adjustment coefficient when the corresponding property index exceeds the preset safety boundary threshold, in order to offset the offsetting of each property feature direction vector in the same latent space dimension.

[0012] Preferably, in step S2, the transformation correction of the sequence latent space feature tensor includes the following steps: step S25, using the orthogonal complement space projection matrix to perform an orthogonal basis transformation on the sequence latent space feature tensor to correct the feature tensor; step S26, separating the shared gradient component that causes parameter hedging from the sequence latent space feature tensor to block the decoding probability fluctuation caused by the conflict of multi-objective materialization index constraints.

[0013] Preferably, in step S3, predicting the probability distribution of amino acid residues at each position includes the following steps: Step S31, the orthogonalized sequence feature tensor separated from the common gradient component is used as the input parameter to control the distribution probability and sent to the next-level transformer decoding layer; Step S32, when predicting the probability distribution of amino acid residues at candidate positions, the decoding probability fluctuation caused by multidimensional property tag conflict is blocked, so that the generated macromolecular candidate sequence is stabilized within a fixed folding structure.

[0014] Preferably, based on the orthogonalized sequence feature tensor output in step S2, the index constraints carried by the multi-attribute constraint embedding vector are decoupled to mutually orthogonal latent space channels, so that the physicochemical properties of the finally generated macromolecular candidate sequences converge to the preset threshold range, and the adjusted orthogonalized manifold features conform to the macromolecular sequence syntax and structural folding rules trained from a large-scale unlabeled macromolecular sequence dataset.

[0015] Preferably, in step S3, dynamically updating the adjustment coefficients of each property deviation based on the score feedback results includes the following steps: Step S33, inputting the generated complete macromolecular candidate sequence into the multi-objective reward function to quantitatively evaluate the macromolecular property score; Step S34, dynamically adjusting the adjustment coefficients of each property deviation based on the score feedback results output by the multi-objective reward function, so that the global search path performs adaptive gradient navigation on the denoised multi-dimensional attribute terrain map.

[0016] Preferably, adaptive gradient navigation utilizes a closed-loop adaptive feedback mechanism to transform the serial step-by-step selection process into parallel deterministic sequence generation within a high-dimensional latent space, enabling the optimization control of multidimensional physicochemical properties to be completed within the denoised multidimensional attribute topographic map. By deeply coupling the macromolecular sequence grammar and structural folding rules in the pre-training stage, the orthogonal correction mechanism of the sequence latent space feature tensor, and the terrain navigation strategy of evolutionary reinforcement learning, a multidimensional collaborative optimization path is constructed from the underlying grammatical boundary to high-dimensional feature correction and then to closed-loop feedback.

[0017] Compared with the prior art, the beneficial effects of the present invention are:

[0018] 1. By constructing multi-track orthogonal decomposition units in the latent space feature flow path of macromolecular sequences, the sequence latent space feature tensor and multi-attribute constraint embedding vector are extracted. The attribute constraints are converted into spatial feature direction vectors using the prior property correlation matrix, and clear physical boundaries are established between the transformer decoding layers. The attribute gradient projection operator calculates the discrete deviation adjustment coefficient of each attribute based on the projection component of the feature on the property feature direction vector, dynamically constructs a shared orthogonal complement space for non-conflicting attributes, strips the shared gradient components that cause parameter hedging, and retains only the independent feature components, thereby eliminating the directional hedging of the control gradients of different physicochemical properties in the same latent space dimension.

[0019] 2. By using the orthogonalized sequence feature tensor containing independent feature components as the control input to the lower-level transformer decoding layer, when predicting the probability distribution of amino acid residues at candidate positions, the decoding perturbation caused by the conflict between multiple objective physical property index labels is blocked, and the chain reaction of overall conformational degradation caused by local amino acid residue mutation is avoided. A collaborative regulation mechanism among features not only ensures that the model outputs macromolecular sequences that meet specific physical property thresholds, but also enables the adjusted orthogonalized manifold structure to smoothly fit the inherent legal sequence syntax and spatial folding rules in the pre-trained macromolecular language model, avoiding sequence syntax collapse when optimizing multiple mutually constraining physical property indices in parallel.

[0020] 3. By introducing an evolutionary reinforcement learning algorithm into the sequence generation flow path, the currently generated complete macromolecular sequence is input into a multi-objective reward function. The deviation adjustment coefficients of each attribute are dynamically adjusted based on the feedback results of the multi-objective reward function, enabling the global path optimization mechanism to navigate accurately directly in the denoised multi-dimensional attribute topographic map. This closed-loop adaptive path adjustment transforms the traditional serial funnel-style screening into parallel deterministic generation, solving the problem of low optimization efficiency caused by blind search when multi-dimensional physicochemical attributes are intertwined. This eliminates the need for trial and error in a huge full-parameter search space, thereby significantly reducing computational costs. Attached Figure Description

[0021] Figure 1This is a flowchart of the parallel generation process for multi-attribute macromolecular candidate sequences in this invention.

[0022] Figure 2 This is a schematic diagram of the parallel generation system for multi-attribute macromolecular candidate sequences of the present invention.

[0023] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0025] A parallel generation method for multiple properties of macromolecules based on latent space decoupling includes the following steps:

[0026] Step S1: Use a large-scale unlabeled macromolecular sequence dataset to pre-train a transformer model and output a sequence latent space feature tensor that represents the prefix of the macromolecular sequence.

[0027] Step S2: Extract the multi-attribute constraint embedding vector transformed from external multi-dimensional physical property indicators; fuse the sequence latent space feature tensor with the multi-attribute constraint embedding vector and send it to the flow path; through the macromolecular prior physical property correlation matrix, uniquely map the multi-attribute constraint embedding vector to a set of physical property feature direction vectors in the latent space to define the gradient evolution trend of each physical property in the molecular manifold; call the attribute gradient projection algorithm to calculate the projection components of the sequence latent space feature tensor on each physical property feature direction vector; determine the adjustment coefficient of each physical property deviation based on the algebraic difference between the obtained initial macromolecular physical property features and the preset safety boundary threshold; use the orthogonal complementary space projection matrix to transform and correct the sequence latent space feature tensor to separate the common gradient components that trigger parameter hedging; adopt the linear combination calibration method to subtract the sum of the products of each physical property deviation adjustment coefficient and each physical property feature direction vector from the input sequence latent space feature tensor, and output the orthogonalized sequence feature tensor after parameter-level calibration.

[0028] Step S3: Input the orthogonalized sequence feature tensor into the transformer decoding layer to predict the probability distribution of amino acid residues at each position to generate macromolecular candidate sequences. Input the macromolecular candidate sequences into the multi-objective reward function for quantitative evaluation of physical property scores. Dynamically update the adjustment coefficients of each physical property deviation based on the obtained score feedback results.

[0029] Preferably, the pre-trained transformer model adopts a conditionally generated pre-trained transformer model; in step S1, the macromolecular sequence syntax and structural folding rules are obtained through self-supervised pre-training on a large-scale unlabeled macromolecular sequence dataset, and are used as the supervision boundary for generating macromolecular candidate sequences; external multidimensional physical property indicators include macromolecular affinity indicators, stability parameters, and immunogenicity levels; multi-attribute constraint condition embedding vectors are used as condition vectors to guide the parallel optimization of macromolecular candidate sequences.

[0030] Preferably, in step S2, uniquely mapping the multi-attribute constraint embedding vector to a set of physical property direction vectors in the latent space includes the following steps: Step S21, extracting the multi-attribute constraint embedding vector transformed from the externally input affinity index, stability performance parameter, and immunogenicity level; Step S22, mapping the multi-attribute constraint embedding vector through linear transformation of the macromolecular prior physical property correlation matrix to generate a set of physical property direction vectors in the latent space.

[0031] Preferably, in step S2, determining the adjustment coefficients for each property deviation includes the following steps: Step S23, calling the attribute gradient projection algorithm to calculate the projection components of the current sequence latent space feature tensor on each property feature direction vector; Step S24, calculating each property deviation adjustment coefficient based on the algebraic difference relationship between the obtained initial macromolecular property features and the preset safety boundary threshold, and increasing the corresponding property deviation adjustment coefficient when the corresponding property index exceeds the preset safety boundary threshold, in order to offset the offsetting of each property feature direction vector in the same latent space dimension.

[0032] Preferably, in step S2, the transformation correction of the sequence latent space feature tensor includes the following steps: step S25, using the orthogonal complement space projection matrix to perform an orthogonal basis transformation on the sequence latent space feature tensor to correct the feature tensor; step S26, separating the shared gradient component that causes parameter hedging from the sequence latent space feature tensor to block the decoding probability fluctuation caused by the conflict of multi-objective materialization index constraints.

[0033] Preferably, in step S3, predicting the probability distribution of amino acid residues at each position includes the following steps: Step S31, the orthogonalized sequence feature tensor separated from the common gradient component is used as the input parameter to control the distribution probability and sent to the next-level transformer decoding layer; Step S32, when predicting the probability distribution of amino acid residues at candidate positions, the decoding probability fluctuation caused by multidimensional property tag conflict is blocked, so that the generated macromolecular candidate sequence is stabilized within a fixed folding structure.

[0034] Preferably, based on the orthogonalized sequence feature tensor output in step S2, the index constraints carried by the multi-attribute constraint embedding vector are decoupled to mutually orthogonal latent space channels, so that the physicochemical properties of the finally generated macromolecular candidate sequences converge to the preset threshold range, and the adjusted orthogonalized manifold features conform to the macromolecular sequence syntax and structural folding rules trained from a large-scale unlabeled macromolecular sequence dataset.

[0035] Preferably, in step S3, dynamically updating the adjustment coefficients of each property deviation based on the score feedback results includes the following steps: Step S33, inputting the generated complete macromolecular candidate sequence into the multi-objective reward function to quantitatively evaluate the macromolecular property score; Step S34, dynamically adjusting the adjustment coefficients of each property deviation based on the score feedback results output by the multi-objective reward function, so that the global search path performs adaptive gradient navigation on the denoised multi-dimensional attribute terrain map.

[0036] Preferably, adaptive gradient navigation utilizes a closed-loop adaptive feedback mechanism to transform the serial step-by-step selection process into parallel deterministic sequence generation within a high-dimensional latent space, enabling the optimization control of multidimensional physicochemical properties to be completed within the denoised multidimensional attribute topographic map. By deeply coupling the macromolecular sequence grammar and structural folding rules in the pre-training stage, the orthogonal correction mechanism of the sequence latent space feature tensor, and the terrain navigation strategy of evolutionary reinforcement learning, a multidimensional collaborative optimization path is constructed from the underlying grammatical boundary to high-dimensional feature correction and then to closed-loop feedback.

[0037] Example 1: In a computational materials science scenario for parallel design of macromolecular candidate sequences, the computing platform receives a cleaned, large-scale, unlabeled macromolecular sequence dataset. Each sequence in the dataset uses the amino acid residue sequence as the basic recording unit. The processor removes repetitive sequences, sequences containing non-standard residue markers, and sequences whose length does not meet the modeling requirements. The retained sequences are written into a training sample queue in memory. The pre-trained transformer model reads the training sample queue and performs self-supervised pre-training, outputting a sequence latent space feature tensor representing the macromolecular sequence prefix. This sequence latent space feature tensor serves as the input state for subsequent multi-attribute parallel generation. In this scenario, the design task simultaneously provides affinity indices, stability parameters, and immunogenicity levels. Serial funnel screening typically filters candidate sequences according to one attribute first, and then passes the remaining sequences to the next attribute for screening. When local amino acid residue adjustments improve affinity, stability parameters and immunogenicity levels may change inversely in subsequent screenings, causing candidate sequences retained in the previous stage to be removed in the later stage. The computing platform needs to process multiple physical property indices simultaneously in the same sequence generation flow path, rather than splitting multiple physical property indices into unrelated screening batches.

[0038] After receiving affinity indices, stability parameters, and immunogenicity levels, the processor converts the external multidimensional physical property indices into multi-attribute constraint embedding vectors. These vectors, along with the sequence latent space feature tensor, are then fed into the transfer path. The processor reads the macromolecular prior physical property correlation matrix and converts the multi-attribute constraint embedding vectors into a set of physical property direction vectors within the latent space. The macromolecular prior physical property correlation matrix is ​​established from the statistical relationships of macromolecular physical property features recorded in the training sample queue. Each item in the matrix represents the correlation strength between the corresponding physical property index and the sequence feature distribution. This allows the affinity indices, stability parameters, and immunogenicity levels to participate in correction within the transfer path of the sequence latent space feature tensor. The pre-trained transformer model then enters the candidate residue prediction process. During the process, the attribute gradient projection algorithm reads the latent space feature tensor of the current sequence and the direction vectors of each physical property feature, calculates the projection components of the latent space feature tensor of the current sequence onto each physical property feature direction vector, and determines the adjustment coefficient of each physical property deviation based on the algebraic difference between the initial macromolecular physical property features and the preset safety boundary threshold. The determination of this adjustment coefficient takes the physical property evaluation result corresponding to the current candidate sequence as input and the preset safety boundary threshold as the comparison benchmark. When a physical property index deviates from the preset safety boundary threshold, the processor increases the participation weight of the physical property feature direction vector corresponding to that physical property index in the correction process. When a physical property index returns to the vicinity of the preset safety boundary threshold, the processor decreases the participation weight of the physical property feature direction vector corresponding to that physical property index in the correction process.

[0039] In situations where improved affinity leads to changes in immunogenicity, the latent space multi-track orthogonal decomposition unit receives the projection components and property deviation adjustment coefficients output by the attribute gradient projection algorithm. It then uses the orthogonal complement projection matrix to transform and correct the sequence latent space feature tensor. The unit subtracts the offsetting components formed by each property deviation adjustment coefficient and each property feature direction vector from the input sequence latent space feature tensor, outputting a parameter-calibrated orthogonalized sequence feature tensor. This process separates the shared gradient components that trigger parameter offsetting from the sequence feature flow path, allowing the next-level transformer decoding layer to receive input states containing independent feature components. After the orthogonalized sequence feature tensor is written into the transformer decoding layer, the processor predicts the amino acid residue probability distribution at each candidate position based on the orthogonalized sequence feature tensor and generates macromolecular candidate sequences from the probability distribution. The macromolecular sequence syntax and structural folding rules formed during the pre-training stage... The orthogonalized sequence feature tensor serves as the supervisory boundary for candidate residue output, while the orthogonalized sequence feature tensor acts as the input parameter controlling the distribution probability. The two form a continuous interface in the same decoding process. The former restricts candidate sequences from entering the sequence space that conforms to the structural folding rules, while the latter suppresses decoding probability fluctuations caused by multidimensional property tag conflicts. After generating complete macromolecular candidate sequences, the processor inputs the macromolecular candidate sequences into a multi-objective reward function to evaluate the affinity index, stability parameters, and immunogenicity level. Based on the score feedback results, it dynamically updates the adjustment coefficients of each property deviation. The updated property deviation adjustment coefficients re-enter the latent space multi-track orthogonal decomposition unit to drive the generation of the orthogonalized sequence feature tensor in the next round. The candidate sequence evaluation results of the current round become the input conditions for the sequence latent space correction in the next round, forming a closed-loop path from candidate sequence generation, property score evaluation, adjustment coefficient update to orthogonalized feature regeneration.

[0040] In this closed-loop path, the macromolecular prior property correlation matrix provides the property feature direction vector for the multi-attribute constraint embedding vector. The property gradient projection algorithm gives the property deviation adjustment coefficient based on the projection components and the preset safety boundary threshold. The orthogonal complement space projection matrix transforms and corrects the sequence latent space feature tensor based on the property deviation adjustment coefficient. The evolutionary reinforcement learning algorithm adjusts the deviation adjustment direction of subsequent rounds based on the feedback results of the multi-objective reward function. The data structure output by the previous stage directly constitutes the data structure input by the next stage. The entire generation process does not involve a serial screening of affinity, stability parameters, and immunogenicity levels, but rather integrates these parameters within the same sequence latent space. In the spatial flow path, multiple physical property indices are constrained in parallel. After running the method, the macromolecular candidate sequences output by the computing platform retain the macromolecular sequence syntax and structural folding rules learned by the pre-trained transformer model. The affinity index, stability parameters, and immunogenicity level converge to the preset threshold range in the same candidate sequence generation process. The candidate sequence backtracking process caused by the destruction of the properties of the previous stage by the subsequent stage screening in the serial funnel screening is transformed into a projection correction and feedback update process in the latent space. The candidate sequence generation path is transformed from step-by-step trial and error in the full parameter space to adaptive gradient navigation in the corrected multidimensional attribute topographic map.

[0041] Example 2: In a computational materials science experimental scenario for parallel verification of multiple attributes of macromolecular candidate sequences, the computing platform uses non-redundant macromolecular sequences from publicly available protein sequence libraries as the source of unlabeled pre-training data and candidate sequences with completed property annotations as the source of verification data. The verification objects include affinity indices, stability parameters, immunogenicity levels, and exploitability thresholds. The experimental platform includes a processor capable of handling floating-point matrix operations, a runtime memory with a capacity of no less than 128GB, a solid-state memory for storing sequence samples and property evaluation results, and a data recording module for recording the property scores of candidate sequences. The processor reads a large-scale unlabeled macromolecular sequence dataset to train a pre-trained transformer model, and then reads the multi-attribute constraint embedding vectors and sequence latent space feature tensors to form the verification basis for the experimental and control groups under the same input conditions. The control group uses a conventional condition generation process, directly combining the multi-attribute constraint embedding vectors with sequence features and sending them to the transformer decoding layer. The experimental group uses a macromolecular multi-attribute parallel generation method based on latent space decoupling, converting the multi-attribute constraint embedding vectors into property feature direction vectors through the macromolecular prior property correlation matrix, and using an attribute gradient projection algorithm to calculate the sequence. The projection components of the latent space feature tensor onto the direction vectors of each physical property feature are used to determine the physical property deviation adjustment coefficient based on the algebraic difference between the initial macromolecular physical property features and the preset safety boundary threshold. The orthogonal complementary space projection matrix corrects the sequence latent space feature tensor according to the physical property deviation adjustment coefficient and outputs the orthogonalized sequence feature tensor. Both groups use the same training sample queue, the same external multidimensional physical property index, the same multi-objective reward function, and the same termination condition. The pre-training data used in the experiment is written into the training sample queue after sequence deduplication, abnormal residue marker removal, and length filtering. The validation data is extracted from the candidate sequence set. Each batch contains 1000 initial candidate sequences. The preset safety boundary thresholds for affinity, immunogenicity, stability, and exploitability are set at 8.7, 0.35, 0.72, and 0.68, respectively. These thresholds are derived from the quantile statistics of the labeled candidate sequences in the validation data. High-affinity candidate sequences are mainly distributed above 8.7, and low-immunogenicity candidate sequences are mainly distributed below 0.35. The stability and exploitability thresholds correspond to the lower limits for the validation set to enter subsequent experimental screening.

[0042] To cover the impact of scoring perturbations on the generated results, the data recording module retains the original scores output by the attribute evaluator in each batch and introduces perturbation samples from the validation set residual distribution. The perturbation amplitude of the affinity index is controlled within 0.12 score units, the perturbation amplitude of the stability parameter is controlled within 0.025, the perturbation amplitude of the immunogenicity level is controlled within 0.030, and the perturbation amplitude of the exploitability threshold is controlled within 0.020. The perturbed scores are used as the original input data in the generation process. The processor retains the perturbation samples, and the decoding probability fluctuations caused by multi-attribute conflicts are handled by the attribute gradient projection algorithm and the orthogonal complement space projection matrix. After the experiment starts, the control group directly predicts the amino groups at candidate positions based on the combined condition vector in each round of decoding. For the probability distribution of acid residues, the experimental group first calculated the projection components of the current sequence latent space feature tensor onto the physical property feature direction vector. Then, based on the deviation of each physical property index from the preset safety boundary threshold, the physical property deviation adjustment coefficient was adjusted. When the affinity index moved above the preset safety boundary threshold and the immunogenicity level increased synchronously, the physical property deviation adjustment coefficient corresponding to the immunogenicity level increased. The orthogonal complement projection matrix separated the common gradient component that triggered parameter hedging from the sequence latent space feature tensor. The processed orthogonalized sequence feature tensor entered the next-level transformer decoding layer. The mean values ​​of the original affinity index after perturbation for 10 batches were 8.42, 8.51, 8.36, 8.58, 8.47, 8.63, 8.39, and 8.55, respectively. The values ​​of 8.44 and 8.60, respectively, and the average original immunogenicity levels after perturbation were 0.386, 0.372, 0.401, 0.365, 0.394, 0.357, 0.408, 0.369, 0.391, and 0.361. Multiple batches simultaneously exhibited input states where both affinity indicators failed to reach the threshold and immunogenicity levels exceeded the threshold. The proportions of the control group simultaneously meeting the threshold were 9.8%, 11.1%, 8.7%, 12.4%, 10.2%, 13.6%, 8.3%, 12.0%, 9.6%, and 13.1%, respectively. The proportions of the experimental group simultaneously meeting the threshold were 24.6%, 25.2%, 22.9%, 26.1%, 24.1%, 27.3%, 22.4%, and 2... 5.7%, 23.8%, and 26.8%; within the same batch, the pass rate of the structure folding rule changed accordingly as the orthogonalized sequence feature tensor entered the decoding layer. The pass rates of the structure folding rule in the control group were 71.3%, 72.8%, 70.6%, 73.5%, 71.9%, 74.1%, 69.8%, 73.0%, 71.1%, and 73.7%, respectively, while those in the experimental group were 84.7%, 85.1%, 83.8%, 85.9%, 84.4%, 86.2%, 83.1%, 85.4%, 84.0%, and 86.0%, respectively. The above data correspond to the constraint effect of the macromolecular sequence syntax and structure folding rule formed in the pre-training stage on the candidate residue output process.

[0043] To observe the intermediate processing, the data recording module continuously recorded the generation results for five rounds in the fourth batch. The average original affinity index after perturbation in the fourth batch was 8.58, and the average original immunogenicity level after perturbation was 0.365. This batch was close to the affinity index threshold, but the immunogenicity level was still higher than the preset safety boundary threshold. In the generation of rounds 1 to 5, the average affinity index was 8.58, 8.66, 8.74, 8.82, and 8.86, respectively; the average immunogenicity level was 0.365, 0.356, 0.344, 0.338, and 0.336, respectively; the average stability performance parameter was 0.706, 0.718, 0.731, 0.739, and 0.741, respectively; and the immunogenicity level deviation adjustment coefficient was 0.18, 0.23, 0.21, 0.16, and 0, respectively. After orthogonalization, the proportions simultaneously satisfying the threshold were 14.2%, 18.7%, 22.9%, 25.4%, and 26.1%, respectively. Intermediate data from the fourth batch showed that as the generation round progressed from the first to the third round, the average affinity index increased from 8.58 to 8.74, the average immunogenicity level decreased from 0.365 to 0.344, the average stability parameter increased from 0.706 to 0.731, and the immunogenicity level deviation adjustment coefficient decreased after increasing to 0.23 in the second round. This change corresponds to the process of the immunogenicity level shifting from exceeding the preset safety boundary threshold to entering the threshold range. When the immunogenicity level deviates significantly, the processor increases the participation weight of the corresponding physical property direction vector, and decreases the participation weight after the immunogenicity level falls back, so that subsequent rounds reduce the suppression of the affinity index-related feature components.

[0044] To examine the processing results of the orthogonal complement projection matrix on the perturbation input, the experimental platform recorded the original input state, control group output state, and experimental group output state in the 7th batch. After perturbation, the average original affinity index was 8.39, the average original immunogenicity level was 0.408, the average stability performance parameter was 0.691, and the average exploitability index was 0.653, with a threshold satisfaction rate of 5.6%. In the control group output state, the average affinity index was 8.71, the average immunogenicity level was 0.381, the average stability performance parameter was 0.704, and the average exploitability index was 0.667, with a threshold satisfaction rate of 8.3%. In the experimental group output state, the average affinity index... The average value was 8.79, the average immunogenicity level was 0.343, the average stability performance parameter was 0.728, and the average exploitability index was 0.692. The proportion of all values ​​meeting the threshold was 22.4%. In the 7th batch, the average affinity index of the control group output state exceeded 8.7, but the average immunogenicity level was still higher than 0.35, and the average stability performance parameter was lower than 0.72. This indicates that directly combining conditional vectors can push individual attributes closer to the target, but there are still directional conflicts between multiple physical property indicators. In the experimental group output state, the affinity index, immunogenicity level, stability performance parameter, and exploitability index all entered the corresponding threshold range at the same time. This result corresponds to the separation processing of shared gradient components of the orthogonalized sequence feature tensor.

[0045] To observe changes near the boundary, the experimental platform, based on batch 4, varied the allowable deviation range near the preset safety boundary threshold for immunogenicity levels. When the allowable deviation range was 0.010, the mean affinity index was 8.76, the mean immunogenicity level was 0.342, and the mean stability parameter was 0.735, with a threshold satisfaction rate of 20.8%. When the allowable deviation range was 0.020, the corresponding values ​​were 8.82, 0.338, 0.739, and 25.4%. When the allowable deviation range was 0.030, the values ​​were 8.86, 0.336, 0.741, and 26.1%. When the allowable deviation range was 0.040, the values ​​were 8.88, 0.339, 0.736, and 26.4%. When the allowable deviation range was 0.050, the values ​​were 8.90, 0.346, 0.729, and 25.8%. When the allowable deviation is 0.060, the four data points are 8.93, 0.352, 0.719, and 25.1%, respectively. When the allowable deviation is 0.070, the four data points are 8.95, 0.361, 0.707, and 24.9%, respectively. These boundary data indicate that when the allowable deviation of the immunogenicity level is too small, the attribute gradient projection algorithm applies too large a correction weight to the direction vector of the corresponding physical property feature of the immunogenicity level, thus limiting the rate of improvement of the affinity index. When the allowable deviation is between 0.020 and 0.050, the affinity index, immunogenicity level, and stability performance parameters improve synchronously. After further expanding the allowable deviation, the average immunogenicity level rises back to above 0.352, and the average stability performance parameter drops to below 0.719. This change indicates that there is a coupling relationship between the allowable deviation and multi-attribute constraints, and the improvement of the affinity index alone cannot be taken as the endpoint of generation.

[0046] After the candidate sequences generated by the experimental group were screened by the computational platform, the candidate sequences that simultaneously met the thresholds in each batch entered subsequent experimental verification. The wet experiments used the same expression system and the same binding activity detection conditions. The data recording module only recorded whether the preset experimental hit criteria were met, without changing the model parameters in the generation process. In the 10 batches, the wet experiment hit rates of the control group were 9.6%, 10.4%, 8.9%, 11.5%, 9.8%, 12.7%, 8.4%, 11.1%, 9.2%, and 12.0%, respectively; the wet experiment hit rates of the experimental group were 22.8%, 23.1%, 21.3%, 24.2%, 22.4%, 25.1%, 20.6%, 23.8%, 21.9%, and 24.7%, respectively. The proportions of the experimental group that simultaneously met the thresholds after computational screening were 24.6%, 25.2%, 22.9%, 26.1%, 24.1%, 27.3%, 22.4%, 25.7%, 23.8%, and 26.8%, respectively. In the above experimental data, the control group... Multiple batches showed output states where affinity indicators met the standards, but immunogenicity levels or stability parameters did not meet the standards synchronously. Under the same original input and perturbation conditions, the experimental group formed a continuous processing chain through macromolecular prior property correlation matrix, attribute gradient projection algorithm, orthogonal complementary space projection matrix, and multi-objective reward function feedback update. This ensured that after the orthogonalized sequence feature tensor entered the transformer decoding layer, the probability distribution of candidate residues remained synchronously changing under multi-attribute constraints. After running under the above experimental conditions, the macromolecular candidate sequences output by the computing platform showed a synchronous convergence trend among affinity indicators, stability parameters, immunogenicity levels, and exploitability thresholds. The experimental group of 10 batches simultaneously met the thresholds with a proportion higher than 20%, the wet experiment hit rate was higher than 20%, and the structural folding rule pass rate remained between 83.1% and 86.2%. The candidate sequence generation process was transformed from step-by-step screening driven by single attributes to parallel correction and closed-loop update of the sequence latent space feature tensor under the participation of embedding vectors under multi-attribute constraints.

[0047] Example 3: This example combines Figures 1 to 2 This section describes a parallel generation method for multiple properties of macromolecules based on latent space decoupling, such as... Figure 1As shown, externally input multidimensional physical property indices, including affinity, stability, and immunogenicity, undergo transformation operations to generate multi-attribute constraint embedding vectors that represent the transformation of these indices. These vectors are then input into the macromolecular prior property correlation matrix via a flow path. Through a unique mapping to spatial feature directions and a fusion mapping process, a physical property feature direction vector is generated to define the gradient evolution trend in the molecular manifold. This vector is then provided as the input constraint direction to the attribute gradient projection algorithm. In this parallel flow path, a large-scale unlabeled macromolecular sequence dataset provides pre-training corpus for self-supervised learning. This pre-training corpus is input into the pre-trained transformer model to obtain sequence grammar and structural folding rules. The pre-trained transformer model outputs a sequence latent space feature tensor representing the prefix state of the macromolecular sequence. This latent space feature tensor is also provided as the initial input manifold to the attribute gradient projection algorithm. The attribute gradient projection algorithm combines the input constraint direction with the input... The initial manifold performs calculations of the projection components and property deviation adjustment coefficients, outputting the property deviation adjustment coefficients to the orthogonal complement projection matrix. The orthogonal complement projection matrix undergoes transformation correction and separates common gradient components. After decoupling and hedging extraction, an orthogonalized sequence feature tensor is output. This orthogonalized sequence feature tensor serves as an independent feature component after parameter-level calibration and is fed as a control parameter to the converter decoding layer. Based on the received control parameter, the converter decoding layer predicts the amino acid residue probability distribution at the predicted position and generates macromolecular candidate sequences through a verification sequence generation step. These macromolecular candidate sequences, as stable sequences with converged multidimensional property thresholds, are input to a multi-objective reward function through input quantization. The multi-objective reward function quantifies and evaluates the macromolecular multidimensional property scores and generates score feedback results. These score feedback results are transmitted to a closed-loop adaptive feedback mechanism to perform gradient navigation based on the score feedback results. Finally, the global optimization path is achieved by dynamically updating the coefficients, and the updated coefficients are cyclically fed back to the property gradient projection algorithm.

[0048] like Figure 2As shown, the storage and data server environment is configured with solid-state storage for storing sequence samples and property evaluation results, a training sample queue for writing retained sequences and training samples, and a task queue for sequentially writing to the macromolecule batch generation task. The hardware computing core node is configured with a processor for handling floating-point matrix operations and execution kernels, and a runtime memory with a capacity of no less than 128GB. The model and algorithm runtime environment deploys a pre-trained transformer model for the self-supervised pre-trained output sequence latent space feature tensor, an attribute gradient projection algorithm module for calculating projection components and determining property deviation adjustment coefficients, a latent space multi-track orthogonal decomposition unit for transforming and correcting features using orthogonal complement space projection matrix, a lower-level transformer decoding layer for inputting orthogonalized sequence feature tensors and predicting residue probability distributions, a multi-objective reward function module for quantifying and evaluating the property scores of macromolecule candidate sequences, and a storage... The system includes a data recording module for the correlation matrix, adjustment coefficients, and feedback results, as well as an attribute evaluator that outputs the original score and evaluation feedback. Each environment and module is connected via specific data pathways. The output of the training sample queue is connected to the input of the pre-trained transformer model for data input, and the output of the task queue is connected to the input of the attribute gradient projection algorithm module for task input. The pre-trained transformer model, attribute gradient projection algorithm module, latent space multi-track orthogonal decomposition unit, lower-level transformer decoding layer, multi-objective reward function module, and attribute evaluator are sequentially connected for feature transformation and evaluation. The output of the attribute evaluator is connected to a solid-state memory to write the original score and evaluation feedback back to the solid-state memory. The data recording module is connected to both the attribute gradient projection algorithm module and the attribute evaluator to enable the storage and interaction of the correlation matrix, adjustment coefficients, and feedback results.

[0049] Example 4: In a computational materials science application scenario for batch generation of macromolecular candidate sequences, the computing platform continuously receives multiple batches of macromolecular sequence samples. Each batch includes amino acid residue sequences, affinity indices, stability parameters, immunogenicity levels, and exploitability thresholds. The processor reads a large-scale unlabeled macromolecular sequence dataset from memory, removes repetitive sequences, sequences containing non-standard residue tags, and sequences exceeding the preset processing range, and writes the retained sequences into a training sample queue. The pre-trained transformer model takes the training sample queue as input and outputs a sequence latent space feature tensor representing the macromolecular sequence prefix. This sequence latent space feature tensor is then incorporated into the macromolecular candidate sequence. Sequence generation flow path; different batches of macromolecular candidate sequences have different local amino acid residue distributions, and the direction of change of affinity index, stability performance parameter and immunogenicity level in the sequence latent space is different. Adjustment of the same local residue may increase affinity index, while increasing immunogenicity level or decreasing stability performance parameter. When the processor directly combines the multi-attribute constraint embedding vector with sequence features, the transformer decoding layer is prone to receiving control information in opposite directions in the same latent space dimension during the prediction of the amino acid residue probability distribution of candidate positions. The output results are manifested as candidate sequence regression, decreased pass rate of structural folding rules and feedback fluctuation of multi-objective reward function.

[0050] Before initiating multi-attribute parallel generation, the computing platform establishes a macromolecular prior property correlation matrix. The processor reads affinity indices, stability parameters, immunogenicity levels, and exploitability thresholds from candidate sequences with completed property annotations. It establishes corresponding data records for the sequence latent space feature tensor of each candidate sequence and the aforementioned property indices. It statistically analyzes the synchronous changes of each feature channel in the sequence latent space feature tensor when each property index changes in batches. Feature channels with the same direction of change are recorded as co-directional correlation terms, feature channels with opposite directions of change are recorded as hedging correlation terms, and feature channels with a change magnitude lower than the minimum resolution capability of the data recording module are recorded as weak correlation terms. This yields the macromolecular prior property correlation matrix used to transform the multi-attribute constraint embedding vector; macromolecular prior property... After the affinity correlation matrix is ​​established, the processor reads the externally input affinity index, stability parameters, and immunogenicity level, converts the external multidimensional physical property index into a multi-attribute constraint embedding vector, and then converts the multi-attribute constraint embedding vector into a set of physical property feature direction vectors based on the macromolecular prior physical property correlation matrix. Each physical property feature direction vector corresponds to the main change direction of a physical property index in the sequence latent space feature tensor. The physical property feature direction vector corresponding to the affinity index is used to constrain the residue probability distribution related to binding activity, the physical property feature direction vector corresponding to the stability parameters is used to constrain the residue probability distribution related to structural folding rules, and the physical property feature direction vector corresponding to the immunogenicity level is used to constrain the residue probability distribution related to immunogenic fragments.

[0051] To ensure a repeatable determination process for the property deviation adjustment coefficients, the processor reads the initial macromolecular properties and a preset safety boundary threshold. It compares the direction and magnitude of the difference between each property index and the preset safety boundary threshold, and determines the corresponding property deviation adjustment coefficient accordingly. When the affinity index is below the preset safety boundary threshold, the property deviation adjustment coefficient corresponding to the affinity index increases; when the immunogenicity level is above the preset safety boundary threshold, the property deviation adjustment coefficient corresponding to the immunogenicity level increases; when the stability parameter is below the preset safety boundary threshold, the property deviation adjustment coefficient corresponding to the stability parameter increases; when the corresponding property index returns to the range of the preset safety boundary threshold, the processor decreases the corresponding property deviation adjustment coefficient, reducing the suppression of other property direction vectors during the correction process. The preset safety boundary threshold adopts... The batch calibration method was used. The processor read the candidate sequence property evaluation results from the verification data that had been labeled with physical properties, and established sample groups according to whether the candidate sequence entered the subsequent experimental verification queue. The affinity index threshold was set to the lower limit value that could distinguish between high affinity candidate sequences and ordinary candidate sequences, the immunogenicity level threshold was set to the upper limit value that could distinguish between low immunogenicity candidate sequences and high immunogenicity candidate sequences, and the stability parameter threshold and exploitability threshold were set to the lower limit value that could enter the subsequent experimental verification queue. In the aforementioned verification data, the preset safety boundary threshold for affinity index was 8.7, the preset safety boundary threshold for immunogenicity level was 0.35, the preset safety boundary threshold for stability parameter was 0.72, and the exploitability threshold was 0.68. The above values ​​were written into the memory as the initialization thresholds of the batch computing platform.

[0052] During candidate sequence generation, the attribute gradient projection algorithm takes the current sequence latent space feature tensor and the direction vectors of each physical property feature as input. The processor sequentially reads the numerical distribution of each feature channel in the sequence latent space feature tensor and compares it with the direction of change of the physical property feature direction vector on the same feature channel. When the direction of change is consistent, the corresponding feature component is retained. When the direction of change is opposite and the corresponding physical property index exceeds the preset safety boundary threshold, it is marked as a shared gradient component. When the change amplitude is lower than the minimum resolution of the data recording module, it does not participate in this round of adjustment. The attribute gradient projection algorithm outputs the projection of the current sequence latent space feature tensor on each physical property feature direction vector. The image component and the property deviation adjustment coefficients involved in this round of correction are received by the latent space multi-track orthogonal decomposition unit. After receiving the projection component and the property deviation adjustment coefficients, the processor corrects the latent space feature tensor of the sequence according to the orthogonal complement space projection matrix. It separates the common gradient component that causes parameter offset from the input sequence latent space feature tensor, retains the independent feature components that do not have directional offset with each property index, and outputs the orthogonalized sequence feature tensor. This orthogonalized sequence feature tensor is used as the input of the next-level transformer decoding layer. After reading the orthogonalized sequence feature tensor, the transformer decoding layer predicts the probability distribution of amino acid residues at each candidate position and generates macromolecular candidate sequences.

[0053] In a batch containing 1000 initial candidate sequences, the average initial affinity index was 8.58, the average initial immunogenicity level was 0.365, and the average initial stability parameter was 0.706. The processor determined that the affinity index was close to the preset safety boundary threshold, but the immunogenicity level was higher than the preset safety boundary threshold. The property deviation adjustment coefficient corresponding to the immunogenicity level was 0.18 in the first round, increased to 0.23 in the second round, decreased to 0.21 in the third round, decreased to 0.16 in the fourth round, and decreased to 0.12 in the fifth round. Correspondingly, the average affinity index changed from 8.58 to 8.66, 8.74, 8.82, and 8.86, the average immunogenicity level changed from 0.365 to 0.356, 0.344, 0.338, and 0.336, and the average stability parameter changed from 0.706 to... The values ​​changed to 0.718, 0.731, 0.739, and 0.741, while the percentages that met the threshold changed from 14.2% to 18.7%, 22.9%, 25.4%, and 26.1% respectively. In the above batches, the immunogenicity level deviation adjustment coefficient first increased and then decreased, corresponding to the process of the immunogenicity level changing from exceeding the preset safety boundary threshold to entering the threshold range. When the immunogenicity level deviated significantly, the processor increased the participation weight of the corresponding physical property direction vector to separate the common gradient component related to the increase in immunogenicity level. When the immunogenicity level fell back, the processor reduced the corresponding physical property deviation adjustment coefficient, so that the independent feature components related to affinity index and stability performance parameter continued to participate in the prediction of candidate residue probability distribution. This processing process enabled the multi-attribute constraint embedding vector to enter the correction path of the sequence latent space feature tensor.

[0054] To define the allowable deviation range, the processor uses the residual distribution of immunogenicity levels in the validation data as input, and progressively calibrates the allowable deviation range near the preset safety boundary threshold. When the allowable deviation range is 0.010, the mean affinity index is 8.76, the mean immunogenicity level is 0.342, and the mean stability parameter is 0.735, while the percentage meeting the threshold is 20.8%. When the allowable deviation range is 0.020, the above data are 8.82, 0, and 0 respectively. 0.338, 0.739, and 25.4%; when the allowable deviation is 0.030, the above data are 8.86, 0.336, 0.741, and 26.1% respectively; when the allowable deviation is 0.040, the above data are 8.88, 0.339, 0.736, and 26.4% respectively; when the allowable deviation is 0.050, the above data are 8.90, 0.346, 0.729, and 25.8% respectively; when the allowable deviation is 0.06... At 0, the above data were 8.93, 0.352, 0.719, and 25.1% respectively; when the allowable deviation was 0.070, the above data were 8.95, 0.361, 0.707, and 24.9% respectively; during the calibration process, when the allowable deviation was below 0.020, the correction weight of the direction vector of the corresponding physical property characteristics of the immunogenicity level was too high, and the rate of increase of the affinity index was limited; when the allowable deviation was between 0.020 and 0.050, the affinity index, immunogenicity level, and stability performance parameter were all within the corresponding threshold range; when the allowable deviation was above 0.050, the average immunogenicity level approached or exceeded 0.35 again, and the average stability performance parameter dropped to around 0.72 or below 0.72. Based on this, the processor wrote the allowable deviation of the immunogenicity level of this batch into the calibration range of 0.020 to 0.050, and prioritized reading the allowable deviation within this range that simultaneously met the threshold ratio when starting subsequent batches.

[0055] While the computing platform continuously processes multiple batches, the data recording module saves the macromolecular prior property correlation matrix, preset safety boundary threshold, property deviation adjustment coefficient, orthogonalized sequence feature tensor summary, and multi-objective reward function feedback results for each batch. After processing each batch, the processor reads the candidate sequence property score output by that batch and compares it with the candidate sequence results entering the subsequent experimental verification queue. When the proportion of three consecutive batches simultaneously meeting the threshold decreases by more than 3 percentage points, the processor rereads the candidate sequence property evaluation results of the most recent 20 batches, updates the unidirectional correlation terms, hedging correlation terms, and weak correlation terms in the macromolecular prior property correlation matrix, and recalibrates the allowable deviation range near the preset safety boundary threshold. In the large-scale generation task, the computing platform writes the macromolecular candidate sequence generation task into the task queue in batches. Each batch carries external multidimensional property indicators, preset safety boundary thresholds, and training sample queue numbers. The processor reads the batch data according to the task queue order and writes the property deviation adjustment coefficient and multi-objective reward function feedback results generated in the current batch back to the memory. The next batch... Upon startup, the processor first reads the adjustment record of the previous batch, then reads the initial macromolecular physical property characteristics of the current batch. When the difference in the distribution of the initial macromolecular physical property characteristics between the two batches is lower than the batch difference threshold set by the data recording module, the processor uses the prior physical property correlation matrix of the macromolecules from the previous batch. When the difference in the distribution of the initial macromolecular physical property characteristics between the two batches reaches or exceeds the batch difference threshold, the processor updates the prior physical property correlation matrix of the macromolecules and calibrates the allowable deviation range. After adopting the above processing flow, a continuous data interface is formed between the sequence latent space feature tensor, the multi-attribute constraint embedding vector, the physical property direction vector, the physical property deviation adjustment coefficient, the orthogonalized sequence feature tensor, and the feedback results of the multi-objective reward function. The data output from the previous stage is written into the data structure that can be read by the next stage. The steps involving selection, setting, determination, and adjustment in the candidate sequence generation flow path all have corresponding inputs, processing actions, and output states. The computing platform performs parallel correction between affinity index, stability performance parameter, immunogenicity level, and exploitability threshold. The output macromolecular candidate sequences enter the candidate sequence set that conforms to the structural folding rules.

[0056] Example 5: In a scenario where a new macromolecular candidate sequence design task is integrated into a computing platform, the processor reads the training sample queue, candidate sequences with completed property annotations, affinity indices, stability parameters, immunogenicity levels, and exploitability thresholds corresponding to the task. Candidate sequences are divided into two categories based on whether they enter the subsequent experimental verification queue. The processor records, batch by batch, the unidirectional, inverse, and low-amplitude changes in each feature channel of the sequence latent space feature tensor as the property indices change. These records are written into the unidirectional correlation terms, hedging correlation terms, and weak correlation terms in the macromolecular prior property correlation matrix, respectively. Then, the lower limit values ​​of high-affinity candidate sequences, upper limit values ​​of low-immunogenicity candidate sequences, lower limit values ​​of stability parameters entering the subsequent experimental verification queue, and lower limit values ​​of exploitability thresholds entering the subsequent experimental verification queue are written into the preset safety boundary threshold and allowable deviation range corresponding to the task. The macromolecular prior property correlation matrix, preset safety boundary threshold, allowable deviation range, and training sample queue number are stored together in memory as initialization data read when the task generates macromolecular candidate sequences.

[0057] When the computing platform is deployed to a new batch environment, the processor extracts no less than 1,000 initial candidate sequences and runs three rounds of pre-generation process. The data recording module records the mean affinity index, the mean immunogenicity level, the mean stability performance parameter, the mean exploitability index, the pass rate of structural folding rules, and the proportion of simultaneous thresholds in each round. If the decrease in the proportion of simultaneous thresholds in three consecutive rounds exceeds 3 percentage points, the processor reads the property evaluation results of the candidate sequences in the most recent 20 batches and updates the macromolecular prior property correlation matrix. If the immunogenicity level is higher than the preset safety boundary threshold and the affinity index reaches the preset safety boundary threshold, the processor increases the property deviation adjustment coefficient corresponding to the immunogenicity level and regenerates the orthogonalized sequence feature tensor. If the stability performance parameter is lower than the preset safety boundary threshold, the processor increases the property deviation adjustment coefficient corresponding to the stability performance parameter. The calibrated orthogonalized sequence feature tensor enters the converter decoding layer and outputs the candidate residue probability distribution. Macromolecular candidate sequences that conform to the structural folding rules and simultaneously meet the thresholds are written into the candidate sequence set.

[0058] Example 6: In a computational materials science scenario for batch design of candidate sequences for new target macromolecules, the computing platform performs pre-deployment calibration before officially generating candidate sequences. The processor reads no fewer than 5000 candidate sequences that have completed property annotation, writing 70% of them into a calibration queue and 30% into a validation queue. Each candidate sequence retains the amino acid residue sequence, affinity index, stability parameters, immunogenicity level, exploitability threshold, and the corresponding sequence latent space feature tensor. The data recording module checks the above fields one by one. Candidate sequences missing any field are not included in the calibration queue. The sequence latent space feature tensor in the calibration queue is used to establish the macromolecule prior property correlation matrix. The property evaluation results in the validation queue are used to determine the preset safety boundary threshold, allowable deviation range, and batch difference threshold. The output of the pre-deployment calibration includes the training sample queue number, the macromolecule prior property correlation matrix, the preset safety boundary threshold, the allowable deviation range, the batch difference threshold, and a tool for recording the results of subsequent batch runs. Data recording table; When the processor establishes the macromolecular prior property correlation matrix, it sorts the candidate sequences in the calibration queue according to the direction of change of affinity index, stability parameter, and immunogenicity level, respectively. Then, it reads the sequence latent space feature tensor corresponding to each candidate sequence and compares the direction of change of each feature channel between adjacent candidate sequences. When a feature channel increases synchronously with the increase of affinity index without the increase of immunogenicity level, the feature channel is written into the same-direction correlation term corresponding to affinity index. When a feature channel increases with the increase of affinity index and is accompanied by an increase of immunogenicity level or a decrease of stability parameter, the feature channel is written into the hedging correlation term. When the change amplitude of a feature channel is lower than the minimum resolution of the data recording module in two consecutive sorting intervals, the feature channel is written into the weak correlation term. After all feature channels are written, the same-direction correlation term, the hedging correlation term, and the weak correlation term together form the macromolecular prior property correlation matrix corresponding to this batch. The matrix version number and the calibration queue number are written into the memory together.

[0059] The calibration of preset safety boundary thresholds uses the verification queue as input. The processor reads the candidate sequence set entering the subsequent experimental verification queue and extracts the distribution records of affinity index, stability parameter, immunogenicity level, and exploitability threshold. The preset safety boundary threshold for affinity index is taken as the lower limit record entering the subsequent experimental verification queue; the preset safety boundary threshold for stability parameter is taken as the lower limit record entering the subsequent experimental verification queue; the preset safety boundary threshold for exploitability threshold is taken as the lower limit record entering the subsequent experimental verification queue; and the preset safety boundary threshold for immunogenicity level is taken as the upper limit record entering the subsequent experimental verification queue. In one calibration batch, the processor reads the preset safety boundary threshold for affinity index as 8.7, the preset safety boundary threshold for stability parameter as 0.72, the preset safety boundary threshold for immunogenicity level as 0.35, and the preset safety boundary threshold for exploitability as 0.68. These thresholds are written into the initialization data table of this task. The calibration of the allowable deviation range adopts a step-by-step trial run method. The processor reads the sample from the verification queue... 1000 initial candidate sequences were extracted, and allowable deviation ranges of 0.010, 0.020, 0.030, 0.040, 0.050, 0.060, and 0.070 were written near the preset safety boundary threshold for immunogenicity. Five rounds of candidate sequence generation were run for each allowable deviation range. The data recording module recorded the mean values ​​of affinity index, immunogenicity level, stability performance parameter, and the proportion of simultaneous threshold satisfaction. When the allowable deviation range was 0.020 to 0.050, the mean value of affinity index remained between 8.82 and 8.90, the mean value of immunogenicity level remained between 0.338 and 0.346, the mean value of stability performance parameter remained between 0.729 and 0.739, and the proportion of simultaneous threshold satisfaction remained between 25.4% and 26.4%. The processor wrote 0.020 to 0.050 into the allowable deviation range of immunogenicity level and prioritized reading records with a high proportion of simultaneous threshold satisfaction within this range during formal generation.

[0060] Batch difference thresholds are used to handle variations in the distribution of candidate sequences across different batches. The processor reads the initial macromolecular property distribution and post-generation property evaluation results from the most recent 20 historical batches. It compares the current batch with historical batches batch-by-batch in terms of the mean values ​​of affinity, stability, immunogenicity, and exploitability. It also records the proportion of simultaneous threshold satisfaction after using the historical macromolecular prior property correlation matrix. When the difference is less than 3 percentage points and the proportion of simultaneous threshold satisfaction decreases by no more than 3 percentage points, the processor uses the macromolecular prior property correlation matrix from the previous batch. When the difference reaches or exceeds 3 percentage points, the processor reads the current batch's candidate sequence property evaluation results, updates the syn-correlation terms, hedging terms, and weak correlation terms, and recalibrates the allowable deviation range. During formal operation, the processor... In each round, the device reads the latent space feature tensor, property feature direction vector, preset safety boundary threshold, and property deviation adjustment coefficient of the current sequence. It compares the differences in affinity index, stability performance parameter, immunogenicity level, and exploitability threshold with respect to the preset safety boundary threshold, and determines the property feature direction vector that needs to increase its weight in this round. When the affinity index is lower than the preset safety boundary threshold, the property deviation adjustment coefficient corresponding to the affinity index is increased. When the stability performance parameter is lower than the preset safety boundary threshold, the property deviation adjustment coefficient corresponding to the stability performance parameter is increased. When the immunogenicity level is higher than the preset safety boundary threshold, the property deviation adjustment coefficient corresponding to the immunogenicity level is increased. When the exploitability index is lower than the exploitability threshold, the property deviation adjustment coefficient corresponding to the exploitability index is increased. After the property index returns to the corresponding threshold range, the corresponding property deviation adjustment coefficient is decreased.

[0061] In each round of correction, the attribute gradient projection algorithm outputs the projection components, shared gradient component labels, and property deviation adjustment coefficients involved in the correction. The latent space multi-track orthogonal decomposition unit reads these outputs and calls the orthogonal complement space projection matrix to correct the sequence latent space feature tensor. The processor separates feature components with opposite changing directions and corresponding property indices exceeding a preset safety boundary threshold from the sequence latent space feature tensor, retains feature components with consistent changing directions in the orthogonalized sequence feature tensor, and retains feature channels with changing amplitudes below the minimum resolution capability of the data recording module for the next round of recording. After the orthogonalized sequence feature tensor is written into the transformer decoding layer, the transformer decoding layer outputs the amino acid residue probability distribution at candidate positions. The data recording module saves the residue probability distribution, property score, and structural folding of this round. The rules are passed; when the computing platform experiences output fluctuations during continuous operation, the processor handles abnormal batches according to the online fault-tolerant procedure. If the threshold ratio decreases by more than 3 percentage points for three consecutive rounds, the processor pauses reading the current macromolecular prior property correlation matrix and instead reads the matrix versions from the most recent 20 batches that have small differences from the initial macromolecular property distribution of the current batch. It then runs three rounds of pre-generation process with 1000 initial candidate sequences. If the pass rate of the structure folding rule in the pre-generation process is less than 80%, the processor reverts to the deployment pre-calibration process and re-establishes the macromolecular prior property correlation matrix. If the pass rate of the structure folding rule reaches or exceeds 80%, the processor writes the matrix version into the current batch running record and continues to generate macromolecular candidate sequences.

[0062] In a batch of operations that has completed pre-deployment calibration, the computational platform reads the macromolecular prior property correlation matrix from records in the most recent 20 batches where the initial macromolecular property characteristic distribution difference is less than 3 percentage points. The results of three rounds of pre-generation before the formal generation show that the percentages simultaneously meeting the thresholds are 21.8%, 23.6%, and 24.1%, respectively, and the structure folding rule pass rates are 82.4%, 83.7%, and 84.2%, respectively. The processor retains the current matrix version and proceeds to the formal generation. After the fifth round of formal generation, the average affinity index is 8.84, the average immunogenicity level is 0.341, the average stability parameter is 0.736, and the average exploitability index is [missing value]. 0.689, while simultaneously satisfying the threshold ratio of 25.9%, the corresponding candidate sequence is written into the candidate sequence set; when the above-mentioned pre-deployment calibration and online fault-tolerant procedure are used, the training sample queue number, macromolecular prior property correlation matrix, preset safety boundary threshold, allowable deviation range, batch difference threshold, property deviation adjustment coefficient, orthogonalized sequence feature tensor, and multi-objective reward function feedback results are all formed into readable records in the memory. The processor completes the initialization before candidate sequence generation, feature correction during candidate sequence generation, and matrix update after candidate sequence generation according to the above records. The output macromolecular candidate sequence is retained in the candidate sequence set that conforms to the structural folding rules.

[0063] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A parallel generation method for multiple properties of macromolecules based on latent space decoupling, characterized in that, Includes the following steps: Step S1: Use a large-scale unlabeled macromolecular sequence dataset to pre-train a transformer model and output a sequence latent space feature tensor that represents the prefix of the macromolecular sequence. Step S2: Extract the multi-attribute constraint embedding vector transformed from external multi-dimensional physical property indicators; fuse the sequence latent space feature tensor with the multi-attribute constraint embedding vector and send it to the flow path; through the macromolecular prior physical property correlation matrix, uniquely map the multi-attribute constraint embedding vector to a set of physical property feature direction vectors in the latent space to define the gradient evolution trend of each physical property in the molecular manifold; call the attribute gradient projection algorithm to calculate the projection components of the sequence latent space feature tensor on each physical property feature direction vector; determine the adjustment coefficient of each physical property deviation based on the algebraic difference between the obtained initial macromolecular physical property features and the preset safety boundary threshold; use the orthogonal complementary space projection matrix to transform and correct the sequence latent space feature tensor to separate the common gradient components that trigger parameter hedging; adopt the linear combination calibration method to subtract the sum of the products of each physical property deviation adjustment coefficient and each physical property feature direction vector from the input sequence latent space feature tensor, and output the orthogonalized sequence feature tensor after parameter-level calibration. Step S3: Input the orthogonalized sequence feature tensor into the transformer decoding layer to predict the probability distribution of amino acid residues at each position to generate macromolecular candidate sequences. Input the macromolecular candidate sequences into the multi-objective reward function for quantitative evaluation of physical property scores. Dynamically update the adjustment coefficients of each physical property deviation based on the obtained score feedback results.

2. The parallel generation method for macromolecular multi-attribute based on latent space decoupling according to claim 1, characterized in that, The pre-trained transformer model adopts a conditional generation pre-trained transformer model; in step S1, the macromolecular sequence syntax and structural folding rules are obtained through self-supervised pre-training using a large-scale unlabeled macromolecular sequence dataset, and are used as the supervision boundary for generating macromolecular candidate sequences; external multidimensional physical property indicators include macromolecular affinity indicators, stability parameters, and immunogenicity levels; Multi-attribute constraint embedding vectors serve as condition vectors, guiding the parallel optimization of macromolecular candidate sequences.

3. The parallel generation method for macromolecular multi-attribute based on latent space decoupling according to claim 1, characterized in that, In step S2, uniquely mapping the multi-attribute constraint embedding vector to a set of physical property direction vectors in the latent space includes the following steps: Step S21, extracting the multi-attribute constraint embedding vector transformed from the externally input affinity index, stability performance parameter, and immunogenicity level; Step S22, mapping the multi-attribute constraint embedding vector through linear transformation of the macromolecular prior physical property correlation matrix to generate a set of physical property direction vectors in the latent space.

4. The parallel generation method for macromolecules based on latent space decoupling according to claim 1, characterized in that, In step S2, determining the adjustment coefficients for each property deviation includes the following steps: Step S23, calling the attribute gradient projection algorithm to calculate the projection components of the current sequence latent space feature tensor on each property feature direction vector; Step S24, based on the algebraic difference relationship between the obtained initial macromolecular property features and the preset safety boundary threshold, calculating each property deviation adjustment coefficient, and increasing the corresponding property deviation adjustment coefficient when the corresponding property index exceeds the preset safety boundary threshold, in order to offset the offsetting of each property feature direction vector in the same latent space dimension.

5. The parallel generation method for macromolecules based on latent space decoupling according to claim 1, characterized in that, In step S2, the transformation correction of the sequence latent space feature tensor includes the following steps: Step S25, using the orthogonal complement space projection matrix to perform orthogonal basis transformation on the sequence latent space feature tensor to correct the feature tensor; Step S26, separating the common gradient component that causes parameter hedging from the sequence latent space feature tensor to block the decoding probability fluctuation caused by the conflict of multi-objective materialization index constraints.

6. The parallel generation method for macromolecules based on latent space decoupling according to claim 1, characterized in that, In step S3, predicting the probability distribution of amino acid residues at each position includes the following steps: Step S31: The orthogonalized sequence feature tensor separated from the common gradient components is used as the input parameter to control the distribution probability and sent to the next-level transformer decoding layer; Step S32: When predicting the probability distribution of amino acid residues at candidate positions, the decoding probability fluctuation caused by multidimensional property tag conflict is blocked, so that the generated macromolecular candidate sequence is stabilized within a fixed folding structure.

7. The parallel generation method for macromolecules based on latent space decoupling according to claim 1, characterized in that, Based on the orthogonalized sequence feature tensor output in step S2, the index constraints carried by the multi-attribute constraint embedding vector are decoupled to mutually orthogonal latent space channels, so that the physicochemical properties of the finally generated macromolecular candidate sequences converge to the preset threshold range, and the adjusted orthogonalized manifold features conform to the macromolecular sequence syntax and structural folding rules trained from a large-scale unlabeled macromolecular sequence dataset.

8. The parallel generation method for macromolecules with multiple attributes based on latent space decoupling according to claim 1, characterized in that, In step S3, dynamically updating the adjustment coefficients of each property deviation based on the score feedback results includes the following steps: Step S33, inputting the generated complete macromolecular candidate sequence into the multi-objective reward function to quantitatively evaluate the macromolecular property score; Step S34, dynamically adjusting the adjustment coefficients of each property deviation based on the score feedback results output by the multi-objective reward function, so that the global search path can perform adaptive gradient navigation on the denoised multi-dimensional attribute terrain map.

9. The parallel generation method for macromolecules based on latent space decoupling according to claim 1, characterized in that, Adaptive gradient navigation utilizes a closed-loop adaptive feedback mechanism to transform the serial step-by-step selection process into parallel deterministic sequence generation within a high-dimensional latent space, enabling the optimization control of multidimensional physicochemical properties to be completed within the denoised multidimensional attribute topographic map. By deeply coupling the macromolecular sequence grammar and structural folding rules in the pre-training stage, the orthogonal correction mechanism of the sequence latent space feature tensor, and the terrain navigation strategy of evolutionary reinforcement learning, a multidimensional collaborative optimization path is constructed from the underlying grammatical boundary to high-dimensional feature correction and then to closed-loop feedback.

Images