Method for constructing virtual screening model of traditional chinese medicine for herpes zoster related pain

Abstract

The application discloses a virtual screening model construction method for herpes zoster related pain traditional Chinese medicine, and belongs to the field of calculation methods for drug screening. The method constructs a compound data set containing four levels of medicinal material component spectrum, extract, effective component and active monomer; a disease target knowledge graph is constructed, and the targets are two-dimensionally labeled according to three barrier compartments of peripheral nerve ending, dorsal root ganglion and spinal cord central and three disease course phases of acute phase, subacute phase and post-herpetic neuralgia phase, so as to obtain a compartment attribution and phase weight matrix; a binding affinity prediction network is embedded and trained with the compartment identifier as a condition, and a multi-task network outputs a three-compartment accessibility probability vector; an affinity matrix is element-by-element gated according to compartment attribution, and a penalty factor that decreases with the accessibility of the spinal cord central compartment exceeding a permissible threshold is applied to the peripheral compartment target, and a compartment-phase analgesic efficiency tensor is obtained through phase weight weighting; and four levels are differentially aggregated and scored, and a priority sequence is output.

Description

Technical Field

[0001] This invention relates to the field of computer-aided drug screening technology, specifically to a method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain, which belongs to the interdisciplinary field of computational methods for drug screening and pharmacoinformatics. Background Technology

[0002] Postherpetic neuralgia (PHN) is a neuropathic pain triggered by the reactivation of the varicella-zoster virus. Pain lasting more than three months is termed postherpetic neuralgia. This disease has a high incidence and prolonged course in the elderly. Existing calcium channel modulators, tricyclic antidepressants, and opioids have limited efficacy and significant central nervous system side effects, highlighting the urgent need for innovative analgesics. Traditional Chinese medicine resources contain tens of thousands of active ingredients and naturally possess multi-component, multi-target effects, which highly align with the multi-pathway pathogenesis of neuropathic pain, making it an important source for the discovery of new drugs for treating neuropathic pain.

[0003] Traditional screening of traditional Chinese medicine (TCM) resources mainly relies on in vitro analgesic activity experiments to verify each compound individually. This approach is limited by throughput, time-consuming, and costly when dealing with a vast number of compounds. To improve efficiency, computer-based virtual screening technology has been introduced. Existing TCM virtual screening methods can be broadly categorized into two types: one is single-target virtual screening based on molecular docking, which calculates conformation and scores for a single protein structure. However, it lacks comprehensive support for multi-target synergistic analgesic pathways involved in neuropathic pain, such as the NOD-like receptor heat protein domain-associated protein 3 (NLRP3) inflammasome, transient receptor potential vanillic acid subtype 1 (TRPV1) receptor, voltage-gated sodium channel Nav1.7, and glial cell activation. The other type is multi-target association mining based on network pharmacology and knowledge graphs. This method linearly sums the effects of compounds on multiple targets with static weights into a scalar total score, and then superimposes a drug-likeness score, represented by blood-brain barrier permeability, as the ranking criterion for candidate resources.

[0004] The existing methods described above reveal a common, deep-seated flaw that determines the success or failure of screening when applied to shingles-related pain: their modeling of disease targets is neither spatial nor temporal.

[0005] First, in terms of spatial dimension, existing methods assume all targets are within the same accessibility space, thus treating blood-brain barrier permeability as a monotonically increasing factor in drug-likeness, with higher permeability prediction values ​​resulting in higher overall scores. However, the core analgesic targets for shingles-related pain are not located in the same compartment: TRPV1 and calcitonin gene-related peptide targets are mainly located in intradermal nerve fibers and free nerve endings, with pore capillaries on their outer side, and almost no tight junction barrier; Nav1.7 and Nav1.8 and satellite glial cell targets are mainly located in the dorsal root ganglia, with the blood-dorsal root ganglion barrier on their outer side, which has significantly higher permeability than the blood-brain barrier; while central sensitization-related targets such as P2X4 receptor, brain-derived neurotrophic factor, and potassium-chloride cotransporter 2 are located in the dorsal horn of the spinal cord, with the dense blood-spinal cord barrier on their outer side. The tightness of these three types of barriers increases sequentially, resulting in completely different directions of screening pressure on the physicochemical properties of compounds. By simply adding points to permeability, we can systematically eliminate compounds that are highly polar, have low central exposure, and are best suited to act on peripheral and dorsal root ganglion targets. This type of peripherally restricted analgesic strategy is the direction that has been repeatedly validated in the development of analgesic drugs in recent years.

[0006] Second, in terms of time, existing methods assume that the pathogenic weight of a target remains constant throughout the disease course, thus imposing a fixed set of weights on target combinations. However, the dominant pathological mechanism of shingles-related pain shifts with the course of the disease: in the acute phase, viral replication-induced activation of the NLRP3 inflammasome and release of interleukin-1β are dominant; in the subacute phase, peripheral sensitization occurs, with upregulation of sodium channel subtype expression; after entering the postherpetic neuralgia phase, continuous activation of spinal microglia and astrocytes, and the P2X4-brain-derived neurotrophic factor-potassium chloride cotransporter 2 cascade become the dominant drivers, while the inflammasome weight in the acute phase has significantly decreased. In other words, the pathogenic weight of different targets reverses in order throughout the disease course. Imposing fixed weights will misclassify compounds that are effective only in the acute phase as optimal throughout the entire course, and will also suppress candidate compounds that are truly effective for the sequelae period at the bottom of the ranking.

[0007] Third, because the above two dimensions are compressed, existing methods can only use content-weighted linear summation when aggregating the scores at the monomer level to the three higher levels of effective components of traditional Chinese medicine, classic extracts, and the overall composition spectrum of the medicinal material. Linear summation compresses the spatial-temporal coverage structure of the composition spectrum into a scalar, thereby erasing the most valuable characteristics of traditional Chinese medicine resources. The composition spectrum of a single medicinal material is naturally distributed in different polarity ranges. Highly polar components remain in the peripheral compartments, while lipid-soluble small molecules enter the central compartments, and the two are complementary in terms of compartmentality; at the same time, the targets of different components are also complementary in terms of disease progression. Once the complementary structure is erased, medicinal materials containing a single potent component will continuously outperform those with moderately potent components but complementary coverage, and the latter often have a larger effect size in whole animal models.

[0008] The above three points all point to the same consequence in practice: the consistency between the ranking given by virtual screening and subsequent in vitro analgesic activity experiments is low; the experimental sampling order determined based on this ranking cannot effectively reduce the number of trial and error attempts; and virtual screening fails to truly play its role in reducing the scale of experiments. Therefore, how to establish a mapping relationship between compound structural data and disease target knowledge that can simultaneously characterize the spatial partitioning of the barrier and the temporal migration of the disease course is a core technical problem that urgently needs to be solved in this field. Summary of the Invention

[0009] The core technical problem to be solved by this invention is: how to establish a quantitative mapping relationship between multi-level compound structure data of traditional Chinese medicine and knowledge of multi-target pathways of herpes zoster-related pain, which is simultaneously constrained by barrier spatial partitioning and disease course time migration, so that the candidate resource ranking output by virtual screening is consistent with the results of in vitro analgesic activity experiments, and thus determine the effective experimental sampling order.

[0010] To address the aforementioned technical problems, this invention provides a method for constructing a virtual screening model for traditional Chinese medicine (TCM) for herpes zoster-related pain. Its core concept lies in: instead of treating disease targets as a flat set with constant weights in the same space, it reorganizes the targets onto a two-dimensional grid composed of three barrier compartments × three disease phases; correspondingly, the druggability prediction of compounds is transformed from a single barrier permeability scalar to a three-compartment reachability probability vector; this vector is then used to perform compartment matching gating on the binding affinity matrix. The key is that this gating is non-monotonic for peripheral compartment targets, meaning that a penalty is imposed when central compartment reachability exceeds an acceptable threshold; finally, through phase weighting, a compartment-phase analgesic efficacy tensor is obtained, and differentiated aggregate scoring is applied to the four compound levels accordingly.

[0011] This invention discloses a method for constructing a virtual screening model of traditional Chinese medicine for shingles-related pain, including:

[0012] S1. Construct a compound dataset containing four levels: overall component spectrum of Chinese medicinal materials, classic extracts of Chinese medicinal materials, effective components of Chinese medicinal materials, and active monomers of Chinese medicinal materials. Each compound record in the compound dataset carries a string representing the molecular structure of the compound.

[0013] S2. Construct a disease target knowledge graph, and label the target points in two dimensions according to the three barrier compartments of peripheral nerve terminal compartment, dorsal root ganglion compartment, and spinal cord central compartment, and the three disease stages of acute phase, subacute phase, and postherpetic neuralgia phase. Obtain the amino acid sequence of each target point from the protein sequence database to obtain the target compartment assignment, target amino acid sequence, and phase weight matrix normalized according to the phase column.

[0014] S3. Using the molecular diagram obtained from the molecular structure characterization string and the target amino acid sequence as input, and the target compartmentalization as a conditional embedding, train the affinity prediction network and output the compound-target affinity matrix.

[0015] S4. Train a multi-task barrier reachability network and output the reachability probability vector of the compound in the three barrier compartments;

[0016] S5. Based on the compartmental affiliation of the target, the compound-target affinity matrix is ​​gated element-by-element using the reachability probability vector; and for targets belonging to the peripheral nerve terminal compartment and the dorsal root ganglion compartment, a penalty factor is applied in the gating, the penalty factor decreasing as the reachability of the compound in the spinal cord central compartment exceeds an allowable threshold; then, the compartmental-temporal analgesic efficacy tensor is obtained by weighted summation using the temporal weight matrix.

[0017] S6. For the active monomer level of traditional Chinese medicine, a weighted power average score is calculated based on the compartment-phase analgesic efficacy tensor. For the other three levels, a score is calculated based on the content weighted probability sum of the compartment-phase analgesic efficacy tensor of each component and the matching degree between its compartment-phase coverage distribution and the target coverage distribution.

[0018] S7. Sort the four levels of traditional Chinese medicine resources according to the scores, and output a priority sequence for determining the sampling order of in vitro analgesic activity experiments.

[0019] The beneficial effects achieved by this invention and their formation mechanism are as follows.

[0020] Firstly, by transforming the monotonic scoring method that prioritizes higher barrier permeability into a non-monotonic gating method based on compartment matching, the rank correlation between virtual screening ranking and in vitro analgesic activity experimental results is improved. The experimental sampling order determined accordingly can then target active compounds at a higher position. The mechanism lies in the fact that the gating function has an extreme point in the central accessibility dimension, allowing peripherally selective compounds to obtain scores that match their target compartments, preventing them from being overwhelmed by highly lipid-soluble compounds in the ranking process.

[0021] Secondly, an unintended effect was that, although the central overexposure penalty term only applied to peripheral compartment targets, the predictive-experimental consistency of spinal central compartment targets was also improved. Without the penalty term, lipophilicity and passive diffusion constituted a shortcut feature, which the network tended to use to simultaneously interpret the accessibility labels of all three compartments, thus learning only coarse polarity correlations in the central compartments. After the penalty term was applied, this shortcut was explicitly suppressed in the peripheral compartment direction, forcing the network to learn true compartment-specific transport features. This improvement in representation ability propagated back along the shared backbone to the central compartment task head, forming a positive cross-task transfer.

[0022] Third, when the barrier order structure constraint and the phase weight column normalization are applied together, the model's prediction variance on small sample data such as active monomers of traditional Chinese medicine decreases by more than the sum of the decreases when either constraint is applied alone. The barrier order constraint compresses the three reachability outputs into an ordered cone of the probabilistic simplex, while the phase weight column normalization restricts the weights to the simplex. Together, they compress the nine degrees of freedom of the compartment-phase analgesic efficacy tensor to approximately four, which is equivalent to a strong structure regularization, thereby suppressing overfitting when labeled samples are scarce.

[0023] Fourth, after using the matching degree between compartmental-temporal coverage distribution and target coverage distribution for upper-level aggregation, a systematic misalignment occurs between the ranking of medicinal materials and the ranking of active monomers: medicinal materials containing a single potent monomer rank lower than those containing multiple moderately potent but compartmental-temporally complementary components. This misalignment cannot be mathematically generated by linear content weighted summation because the effect of linear operators on the component tensor is independent of the coverage distribution between components. This misalignment provides a calculable basis for the quantitative interpretation of the monarch-minister-adjuvant-guide combination in traditional Chinese medicine compound prescriptions.

[0024] Fifth, the experimental feedback is backfilled according to the compartment rather than globally, so that the dorsal root ganglion neuron patch-clamp data only updates the relevant parameters of the dorsal root ganglion compartment, and the spinal cord slice glial activation data only updates the relevant parameters of the spinal cord central compartment. This avoids the mutual contamination of errors between different compartment experimental systems, and the number of experimental batches required for model iteration convergence is reduced. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall architecture and data flow of the virtual screening model for traditional Chinese medicine for herpes zoster-related pain described in this invention;

[0026] Figure 2 The following is a comparison of the effects of the embodiments, where (a) is the ablation experiment results, (b) is the prediction-experiment consistency comparison by compartment, and (c) is the dominant compartment distribution of the top ten candidate resources at each level. Detailed Implementation

[0027] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. The following examples are used to explain the present invention, but are not intended to limit the scope of the present invention.

[0028] All steps of the method described in this embodiment are executed by a computer system. The hardware environment used is as follows: a 64-core x86-64 architecture CPU, 256GB of memory, four general-purpose computing accelerator cards with 80GB of video memory each, and a 20TB solid-state drive array for storage; the software environment is a 64-bit Linux operating system, Python 3.11 runtime, PyTorch 2.4 deep learning framework, RDKit cheminformatics toolkit, and graph database. The above hardware configuration is an example of implementing this method, and those skilled in the art can make equivalent substitutions according to the data scale.

[0029] like Figure 1 As shown, the model described in this invention consists of seven forward modules (S1 to S7) and a feedback closed-loop module (S8). The model input is multi-level compound structure data of traditional Chinese medicine (TCM), which is constructed into a four-level compound dataset via S1. S2 constructs a disease target knowledge graph and outputs the target compartmental affiliation and temporal weight matrix, serving as the disease knowledge foundation of the model. S3 and S4 output the compound-target affinity matrix and the compound barrier reachability vector, respectively. S5 couples these two into a compartmental-temporal analgesic efficacy tensor. S6 performs differentiated aggregation on the four levels. S7 outputs a priority sequence of candidate TCM resources. S8 uses the compartment-resolved experimental residuals to update the temporal weight matrix of S2 and the gating parameters of S5, forming a closed loop.

[0030] like Figure 1 As shown by the dashed arrows, the update signals of S8 flow back to nodes S2 and S5, but not to S3 and S4. This is because the network parameters of S3 and S4 are obtained by training with large-scale molecular binding data and barriers, and their statistical basis is much larger than the sample size of a single batch of analgesia experiments. If a small amount of experimental data is used to fill the gap, it will introduce bias.

[0031] S1. Construct four-tiered compound datasets. By integrating data from the Traditional Chinese Medicine Systems Pharmacology Database, the Traditional Chinese Medicine Chemical Constituent Database, and the unit's own component content determination data, four-tiered compound datasets were constructed. The Traditional Chinese Medicine Systems Pharmacology Database uses TCMSP (Traditional Chinese Medicine Systems Pharmacology Database, which includes 499 medicinal materials registered in the Chinese Pharmacopoeia, 29384 chemical components, and 3311 corresponding target sites). The Traditional Chinese Medicine Chemical Constituent Database uses one or more of TCMID, SymMap, ETCM, and HERB. Components and their mass fractions were obtained by combining the compiled values ​​from the above databases with the unit's actual measured values ​​using high-performance liquid chromatography-mass spectrometry (HPLC-MS / MS). In case of conflict, the actual measured values ​​prevail. This clarifies the source and verifiability of the data at this level.

[0032] Level L1 represents the overall composition spectrum of Chinese medicinal materials. Each record corresponds to a single medicinal material and includes a list of components and their mass fractions obtained from high-performance liquid chromatography-mass spectrometry (HPLC-MS) or literature compilation. Level L2 represents classic extracts of Chinese medicinal materials, corresponding to extraction fractions obtained under specific solvent and process conditions. The list of components and their mass fractions are re-determined based on the extraction fraction. Level L3 represents effective components of Chinese medicinal materials, corresponding to groups of similar structural components obtained through further separation and enrichment, such as total alkaloids, total saponins, and total flavonoids. Level L4 represents active monomers of Chinese medicinal materials. Each record corresponds to a specific small molecule compound, whose structure is represented by a standardized simplified linear molecular input string.

[0033] In this embodiment, the L1 level includes 612 medicinal materials, the L2 level includes 1843 extracts, the L3 level includes 927 active components, and the L4 level includes 34786 active monomers. All of these records are derived from publicly available databases such as TCMSP and our own measured data. The four levels are linked by a hierarchical relationship: each L1 record is associated with several L4 records and their corresponding mass fractions, and the same applies to L2 and L3 records.

[0034] Structural integrity checks were performed on each monomer at level L4, including seven checks: resolvability of the structural formula, compliance of valence bonds, consistency of perceived aroma, consistency of stereochemical labeling, absence of isolated fragments, molecular weight within a reasonable range, and inclusion of elements on the whitelist. Those failing these checks were not allowed to proceed to the next stage. As shown in Table 1, the active monomers of traditional Chinese medicine exhibit a wide range of polarities: paeoniflorin and baicalin have topological polar surface areas of 164.37 Ų and 187.12 Ų, respectively, and cLogP values ​​of −1.36 and 0.14, respectively, classifying them as typical high-polarity components that are difficult to penetrate dense barriers; while corydaline, ligustrazine, and capsaicin all have topological polar surface areas below 60 Ų, indicating strong central nervous system exposure capabilities. This range forms the material basis for subsequent compartmentalized modeling.

[0035] Table 1 Structural parameters of some active monomers of traditional Chinese medicine

[0036]

[0037] As shown in Table 1, the active monomers of traditional Chinese medicine exhibit a wide range in polarity: paeoniflorin and baicalin have topological polar surface areas of 164.37 Å, respectively. 2 With 187.12Å 2 cLogP were −1.36 and 0.14, respectively, indicating they are typical highly polar components that are difficult to penetrate dense barriers; while the topological polar surface areas of corydaline, tetramethylpyrazine, and capsaicin were all below 60 Å. 2 The central nervous system has a relatively strong exposure capability. This span is the material basis for subsequent compartment zoning modeling.

[0038] Using shingles-related pain as the disease node, a knowledge graph was constructed by searching target-disease association databases, pathway databases, protein-protein interaction databases, and literature databases. This graph included five types of nodes: disease nodes, target nodes, pathway nodes, cell type nodes, and anatomical location nodes. The target-disease association databases used were DisGeNET and OpenTargets, the pathway databases used were KEGG and Reactome, the protein-protein interaction database used was STRING, and the literature database used was PubMed. Changes in target expression or activation folds at different disease stages were obtained from publicly available time-series transcriptomics, protein quantification, and immunohistochemistry literature, thus clarifying the source and verifiability of the data in this step.

[0039] Simultaneously, for each target node in the knowledge graph, the standardized amino acid sequence of the target is retrieved from the UniProtKB / Swiss-Prot manual review sub-database of the UniProt protein sequence database using its unique protein identifier (UniProt accession number), and this sequence is written into the knowledge graph as the sequence attribute of the target node. Therefore, the target amino acid sequence input in S3 is not separately acquired data, but rather the target node attribute of the disease target knowledge graph described in this step. When multiple splice isoforms exist for the same target, the standardized sequence annotated by UniProtKB / Swiss-Prot is used, and the database version number and retrieval date are recorded in the notes to ensure verifiability and reproducibility. Edge types include target-disease association (with evidence strength attribute), target-pathway involvement, target-cell type expression (with expression abundance attribute), cell type-anatomical location attribution, and target-target upstream and downstream regulation.

[0040] The knowledge graph constructed in this embodiment contains 187 target nodes, among which core targets include NLRP3, interleukin 1β, caspase 1, TRPV1, transient receptor potential ankylosing spondylin 1, Nav1.7, Nav1.8, P2X4 receptor, P2X7 receptor, brain-derived neurotrophic factor, tropomyosin receptor kinase B, potassium-chloride cotransporter 2, Toll-like receptor 4, chemokine receptor CX3CR1, glial fibrillary acidic protein, and calcitonin gene-related peptide, etc. The compartmental distribution and disease course dominance of the above targets are determined based on published literature on the pathogenesis of postherpetic neuralgia.

[0041] Based on the anatomical location of the cell that primarily expresses the target and the type of its outer barrier, the target is classified into one of the three barrier compartments:

[0042] compartment It is a compartment of peripheral nerve endings, with pore capillaries on its outer side and no tight junction barrier; targets mainly expressed in intraepidermal nerve fibers, keratinocytes or free nerve endings belong to this category, such as TRPV1, calcitonin gene-related peptide and NLRP3 derived from keratinocytes.

[0043] compartment It is a compartment of the dorsal root ganglion, with the blood-dorsal root ganglion barrier on the outside. This barrier has a partially porous structure, and its permeability is between that of the periphery and the center. Targets mainly expressed in the cell bodies of dorsal root ganglion neurons or satellite glial cells belong to this compartment, such as Nav1.7, Nav1.8, and interleukin-1β derived from satellite glial cells.

[0044] compartment It is a central compartment of the spinal cord, with the blood-spinal cord barrier on its outer side and a dense, tight junction. Targets mainly expressed in spinal dorsal horn neurons, microglia, or astrocytes belong to this compartment, such as P2X4 receptor, brain-derived neurotrophic factor, potassium-chloride cotransporter 2, and glial fibrillary acidic protein.

[0045] When the same target is expressed in two or more compartments, the target is split into multiple compartment instances based on the expression abundance attribute of the target-cell type expression edge in the knowledge graph. Each compartment instance is assigned a compartment affiliation and its evidence strength is allocated according to its abundance ratio. For example, NLRP3 is expressed in both keratinocytes and dorsal root ganglion satellite glial cells, and is therefore split into NLRP3@ With NLRP3@ Two instances. After splitting, the 187 target nodes were expanded into 246 compartment instances. Target nodes... The compartment belongs to .

[0046] The course of the disease is divided into three phases: The acute phase refers to the period from 0 to 30 days after the rash appears; The subacute phase refers to 31 to 90 days. This refers to the postherpetic neuralgia period, which lasts for more than 90 days.

[0047] For each target instance Extracting the strength of disease-related evidence from the knowledge graph Furthermore, the fold changes in expression levels or activation levels relative to healthy controls at different time points were extracted from temporal transcriptomics, protein quantification, and immunohistochemical literature data. Phase weight matrix Build in two steps using the following formula:

[0048] ,

[0049] ,

[0050] In the formula For phase set, The total number of target instances. This is the evidence sharpening index. The logarithmic transformation in the first equation is used to compress the dynamic range of fold changes, avoiding the dominance of weights by extreme fold values ​​in individual reports; its inner normalization characterizes the relative contours of the same target point across three time phases. Since the three time phases at the same target point have a common factor, an exponent is introduced. This adjusts the strength of evidence when comparing across targets. The second equation performs column normalization based on time phase, making... That is, the sum of the pathogenic weights of all targets within each time phase is always 1. It is a dimensionless quantity, and its value range is... A larger value indicates a greater pathogenic contribution of the target at that time phase. The time phase weight matrix constructed according to the above formula exhibits an order reversal characteristic: belonging to , The weights of the compartmental inflammasome target NLRP3 are determined by The monotonically decreasing value of 0.42 to 0.09; and attributable to The weights of P2X4 – brain-derived neurotrophic factor, a target for glial cell activation in compartments, are determined by… The 0.09 monotonically rose to The weight of potassium-chlorine cotransporter 2 was adjusted from 0.06 to 0.41. At time NLRP3 has a higher weight than P2X4, The time is lower than P2X4, and the two are in An intersection occurs nearby. This reversal of the order is a key constraint that distinguishes this invention from the fixed-weight scheme. If the average weight of the three phases is used instead, the intersection information is completely lost.

[0051] The compartmentalized affinity prediction network consists of five parts.

[0052] Molecular encoder. Converts compound molecular maps into atomic and substructure marker sequences, which are then embedded into an in-cell matrix. Multi-head self-attention Transformer encoder. This embodiment takes... Attention head count is 8, hidden dimension The encoder output is then attention-pooled to obtain the molecule representation vector. .

[0053] Target encoder. The amino acid sequence encoding of the target is fed into a one-dimensional dilated convolutional residual stacked network. This network consists of four residual blocks with dilation rates of 1, 2, 4, and 8, respectively. The kernel width is 7, and the number of channels is 384. The progressively increasing dilation rate allows the receptive field to grow exponentially with the number of layers, thus covering long-program sequence dependencies across domains without increasing the number of parameters. This is particularly necessary for channel proteins like Nav1.7, which have four homologous domains and a sequence length exceeding 1900 residues. The stacked output is then subjected to global max pooling to obtain the target representation vector. .

[0054] Compartment condition modulation layer. Compartmental assignment based on target location. Indexing yields the compartment embedding vector , ; via two layers of sensing machines Generate scaling and translation vectors, and perform a channel-wise affine transformation on the numerator representation vector: In the formula This represents element-wise product. This modulation causes the same compound to exhibit different representations when facing targets in different compartments, enabling the network to learn the relationship that the same structural feature has different weights in different compartments. This is key to injecting compartment information into affinity prediction.

[0055] Cross-attention layer and regression head. For query, with As keys and values, an interaction representation is obtained through multi-head cross-attention, and then a three-layer perceptron regressor head outputs a combined affinity prediction value:

[0056] In the formula This is the dissociation constant, expressed in mol / L. It is a dimensionless quantity, with a typical value range of 4 to 10. The larger the value, the stronger the bond.

[0057] The network is trained using a publicly available large-scale molecular binding dataset. In this embodiment, the training set contains 1,172,000 compound-target binding records, covering 6,300 targets. The loss function is mean squared error, the optimizer is AdamW, and the initial learning rate is... The batch size was 256, and the training ran for 60 rounds. (Data source description) The large-scale molecular binding dataset mentioned above comes from the publicly available binding affinity database BindingDB (2024 version, approximately 2.9 million experimentally determined binding data, approximately 1.3 million compounds, and thousands of targets) and ChEMBL, supplemented by the DAVIS and KIBA benchmark sets. In this example, 1,172,000 data points and 6,300 targets are used as a selection subset. The compound data at four levels come from the TCM Systems Pharmacology Database (TCMSP, which includes 499 Chinese medicinal materials, 29,384 components, and 3,311 targets from the Chinese Pharmacopoeia), as well as TCMID, SymMap, ETCM, HERB, and our unit's HPLC-MS measurements. The disease target knowledge graph comes from DisGeNET, OpenTargets, KEGG, Reactome, STRING, and PubMed.

[0058] To facilitate subsequent multiplication with the gating value, the affinity prediction value is normalized using a compression function:

[0059] In the formula As the center of affinity, take The corresponding dissociation constant is 1 μmol / L; For the scale parameter, take After normalization , is a dimensionless quantity.

[0060] The multi-task barrier reachability network takes a compound molecular graph and physicochemical descriptor vectors as input. The physicochemical descriptors include six terms: molecular weight, topological polar surface area, cLogP, number of hydrogen bond donors, number of hydrogen bond acceptors, and number of rotatable bonds, as listed in Table 1. The network consists of a shared Transformer backbone encoder and five task heads, with the shared backbone output sharing a common representation. .

[0061] The five task heads output the following respectively: accessibility of peripheral nerve endings in compartments. Dorsal root ganglion compartment accessibility Accessibility of spinal cord central compartments Neural tissue affinity With nervous system safety All five outputs are mapped to open intervals using the Sigmoid function. All of these are dimensionless quantities. Defined as a positive safety vector, meaning that the larger the value, the lower the risk of adverse reactions to the nervous system. The five task head labels originate from heterogeneous sources, but all use the same molecular structure and physicochemical properties of the compound as the sole input. The commonality lies in the fact that they all depend on the same set of physicochemical determinants (lipid solubility, polarity, molecular size, and charge) for cross-barrier and neural tissue distribution. Therefore, the commonality of this layer is represented by a shared backbone. The heterogeneity is absorbed by the independent shallow sub-networks of each task head and the homoscedastic uncertainty weighting (the e^(−uk) term is automatically weighted for tasks with high noise), avoiding the representation conflict caused by the forced sharing of the backbone by concepts at different levels. Neural tissue affinity α and tricompartmental accessibility ρ are mutually independent dimensions rather than redundant: ρ describes whether a certain compartment can be reached (compartmental specificity), and α describes the overall affinity tendency after reaching the neural tissue (global attribute). In the compartmental-phase analgesic efficacy tensor S=α·η·Σ(m:κ(pm)=c)W̄mτ·Anm·Gnm, α and safety η are used as global gating coefficients to uniformly weight each compartment component. Therefore, the purpose of α is clear and it is not redundant with ρ due to different action levels.

[0062] Barrier sequence constraints. The density of the three barriers is as follows: the peripheral pore capillary barrier is the least dense, the blood-dorsal root ganglion barrier is in the middle, and the blood-spinal cord barrier is the densest. Therefore, for the same compound, the accessibility of its three compartments should satisfy an ordered relationship. This anatomical fact is used as an inductive bias, applied with hinge loss:

[0063] ,

[0064] The loss only produces a non-zero value when the order relation is violated, and therefore does not introduce a bias for samples that satisfy the order relation.

[0065] Total loss across multiple tasks. The five task headers have different label sources, dimensions, and noise levels; a weighted average based on homoscedastic uncertainty is used. In the formula For the first The loss of a task leader, This is the learnable log-variance parameter of the task header. These are the constraint weights. In the formula... The item automatically reduces the weight of noisy tasks. This prevents the degradation of all weights approaching zero, thus eliminating the need for manual adjustment of the weight ratios among the five tasks.

[0066] Taking the monomers listed in Table 1 as an example, the reachability vector obtained from training in this embodiment is: paeoniflorin baicalin Berberine sinomenine Capsaicin Corydalis B Ligustrazine All samples satisfy It is evident that highly polar paeoniflorin, baicalin, and quaternary ammonium berberine have significantly lower accessibility to the central nervous system compartments, while low-polarity small molecule tetramethylpyrazine has high accessibility to all three compartments.

[0067] The compartment door control function is calculated using the following formula:

[0068] ,

[0069] In the formula This means taking the non-negative part, i.e. ; This refers to the permissible threshold for central nervous system exposure. Penalty intensity based on target compartment affiliation: The first factor in this formula characterizes whether the compound can reach the target compartment, and the second factor characterizes whether the compound over-enters the central compartment. For For the target site in the compartment, the compound needs to enter the central nervous system, therefore the penalty intensity is zero, and the gating degenerates into a monotonically increasing function of central accessibility; for , The target point in the compartment, the penalty intensity is taken as a positive value, and the gate is in... The dimensionality first increases and then decreases, exhibiting a non-monotonic pattern, with its peak point corresponding to the compound with the best peripheral selectivity. As a contrast, the dotted line represents the gating curve without a penalty term, which monotonically increases across the entire range, representing the pattern of "the higher the permeability, the better" adopted by existing methods.

[0070] The two curves are Below the allowable threshold The fact that the segments completely overlap indicates that the penalty does not change the score of low central exposure compounds, but only applies a decay to compounds that exceed the threshold by the square of the excess amount; the use of a square form instead of a nonlinear form is to make the penalty first-order continuous at the threshold and avoid abrupt changes in the ranking near the threshold.

[0071] Normalized affinity matrix With gate matrix Element-wise multiplication, then grouped according to target compartment affiliation, and weighted summed using a time-phase weight matrix. In the formula The compartmental-phase analgesic efficacy tensor has three dimensions: compound, barrier compartment, and disease phase. Because... When the sequence is normalized according to time, its sub-sums in a single compartment satisfy... ,and , , , All four fall within the open interval. Therefore It also falls within the open range. , is a dimensionless quantity, and the larger the value, the stronger the predicted analgesic efficacy of the compound in that compartment and at that time phase.

[0072] Taking capsaicin as an example: its target site TRPV1 is located in compartment, , ,Pick , The excess amount is The penalty factor is 1, and the gate value is... They went unpunished; while for those who acted similarly... compartmental target but The amount of tetramethylpyrazine exceeded was 0.31, and the penalty factor was [missing value]. Gating value The rating was lowered accordingly. This aligns with the fact that capsaicin topical preparations have been clinically validated in postherpetic neuralgia, and that ingredients with high central nervous system exposure are prone to drowsiness and dizziness.

[0073] S6 employs two different polymerization pathways for the L4 level and the L1 to L3 levels. Each compound in the monomer level directly holds its own tensor, and polymerization is performed using a weighted power-average method. In the formula , Grid point weights and , For powers of 1. When the expression degenerates into a linear weighted sum; when When the value is positive (less than 1), the formula is a concave function. It gives a score higher than the linear sum for compounds with balanced values ​​at all lattice points, and a score lower than the linear sum for compounds with high values ​​only at a single lattice point, thus demonstrating subadditivity. This embodiment takes... .

[0074] For each record at levels L1 to L3 Its component set is The mass fraction of each component is and First, calculate the efficacy of the composite at each grid point using content-weighted probabilities: In the formula The content elasticity index has a range of values. This formula is in probabilistic form, and its values ​​always fall within the open interval. Furthermore, it saturates as the number of components increases, thus avoiding the problem of simply accumulating a large number of weak components at the same grid point, which would lead to an inflated score. This is a distortion that cannot be avoided by linear content weighted summation.

[0075] Next, the performance of each grid point is normalized to a coverage distribution: Then calculate the coverage distribution and the target coverage distribution. Match degree: In the formula, the target coverage distribution It is obtained by normalizing the product of the proportion of patient pain burden in each phase and the proportion of pathological contribution in each compartment. , is a dimensionless quantity, and takes the value of 1 if and only if the coverage distribution of the composite is completely consistent with the target distribution. Relative entropy is used here instead of information entropy because information entropy only rewards the uniformity of coverage and includes grid points that do not need to be covered clinically in the reward; while relative entropy rewards the matching with the distribution of clinical needs, which is more accurate in direction.

[0076] Finally, the combined body hierarchical score is obtained based on a geometric trade-off between strength and complementarity: In the formula The complementarity preference coefficient. Larger sizes tend to favor resources that cover a wide range and are well-balanced. Smaller values ​​indicate a preference for highly efficient, single-point resources. In this embodiment, the L1 level is... L2 level L3 level L4 level directly makes The higher the level, the more components there are. The larger the value, the more the value of the higher level is reflected in the complementary structure of the component spectrum.

[0077] The four rating levels have different scales, making direct comparisons meaningless. A robust standardization based on the median and the absolute deviation of the median is adopted. In the formula The median. The median absolute deviation is represented by a constant of 1.4826, ensuring that the standardization aligns with the standard deviation under a normal distribution. Robust statistics are used instead of the mean and standard deviation because the distribution of TCM resource scores exhibits a significant right-skewed long tail; a few extremely high-scoring resources can significantly inflate the mean and standard deviation, leading to standardization failure.

[0078] according to The four levels of traditional Chinese medicine resources are sorted in descending order, and four priority sequences are output. Each record includes nine components of its compartment-phase tensor, the dominant compartment, the dominant phase, and its corresponding level, which researchers use to determine the sampling order for subsequent in vitro analgesic activity experiments: the dominant compartment is... Candidates should be prioritized for testing on the skin or epidermal nerve endings, with the primary compartment being... The candidate was prioritized for the dorsal root ganglion neuron system experiment, with the main compartment being... Candidates are given priority for spinal cord slice system experiments.

[0079] Based on the aforementioned aggregate scoring, the synergistic effect of traditional Chinese medicine compound formulas was further quantified. The compounds were then... The tensor is expanded into a nine-dimensional vector by rows. Define the compartment-time complement index of the two components: In the formula For inner product, This is the Euclidean norm. Each component is non-negative, therefore , is a dimensionless quantity, and the larger the value, the more complementary the coverage of the two components on the compartment-temporal grid.

[0080] Compound The synergy index is calculated using the following formula: In the formula As the geometric mean of the effectiveness of component pairs, complementary but weak component pairs will not receive high synergy scores, thus constraining the synergy index to be a measure of both complementarity and effectiveness. , is a dimensionless quantity.

[0081] Compatibility constraints are applied according to the following rules: requirements for principal drug components The principal ingredient should be located within the top 20% of all ingredients in the compound formula; the assistant ingredient should be similar to the principal ingredient. Located in the middle section, meaning synergistic without duplication; adjuvant and guide drug components are required. Located in the top 20%, it undertakes a complementary and covering function. Taking a certain blood-activating and meridian-clearing compound as an example, the dominant compartment of its principal ingredient, corydaline, is... The dominant phase is The dominant compartment of the adjuvant component paeoniflorin is The dominant phase is The two The synergy index is 0.83. The value was 0.61; however, the adjuvant was replaced with the same dominant drug. , After taking ligustrazine, It dropped to 0.22. The value dropped to 0.34. This result provides a calculable basis for matching and rejection.

[0082] The priority sequence output by S7 is fed into three types of in vitro experimental systems, each corresponding to one of the three barrier compartments:

[0083] compartment Corresponding to the skin or epidermal nerve ending system, the measured indicators were the inhibition rate of calcitonin gene-related peptide release and the inhibition rate of TRPV1-mediated calcium flow; compartment For the corresponding dorsal root ganglion neuronal system, the measured index was the inhibition rate of Nav1.7 current recorded using whole-cell patch-clamp recording; compartment For the corresponding spinal cord slice system, the measured indicators were the inhibition rates of microglial ion calcium adaptor protein 1 and P2X4 receptor activation.

[0084] The residuals for each compartment are calculated separately, and the residuals for the three compartments are calculated independently and are not mixed: In the formula To be sent to the isolation room The corresponding batches of compounds in the experimental system To normalize to open intervals The experimentally determined values.

[0085] The temporal weight matrix is ​​updated using the exponential gradient method: In the formula This update method ensures that the weights remain positive and the phase sequence is automatically normalized over time. It eliminates the need for additional projection operations after the update, thus avoiding the numerical instability of the projected gradient method at the simplex boundary.

[0086] Gating parameters are adjusted according to the residuals of the peripheral compartments: when and Both values ​​are positive, indicating that the model systematically underestimates the effectiveness of the peripheral compartment compounds, and the penalty intensity should be reduced accordingly. Or raise the allowable threshold When both values ​​are negative, the values ​​are adjusted in the opposite direction. The updated model parameters are fed back to S2 and S5 and enter the next round of screening, accumulating round by round to form a disease-oriented virtual screening knowledge base for traditional Chinese medicine.

[0087] The skeletal distribution of active monomers in traditional Chinese medicine (TCM) differs significantly from that in general drug libraries, with the former being enriched with skeletal structures of natural products such as alkaloids, saponins, terpenes, anthraquinones, and coumarins. When the encoder pre-trained directly on the general molecular library is transferred to TCM monomers, the prediction variance is high during the cold start phase. Therefore, before training the molecular encoder in S3, a skeleton-aware contrastive learning pre-training is performed: using the Murcko skeleton as the dividing criterion, molecular pairs with the same skeleton but different substituents are treated as positive sample pairs, and molecular pairs with different skeleton categories are treated as negative sample pairs. The loss function is: In the formula For cosine similarity, For temperature coefficient, This is done in small batches. Simultaneously, the sampling probability of negative samples is reweighted based on the inverse of the sample size for each skeleton category to compensate for the low proportion of natural product skeletons in the general library.

[0088] The value ranges, value bases, adaptive adjustment rules, and failure degradation treatments for each parameter in this invention are listed in Table 2.

[0089] Table 2 Key Parameter Values

[0090]

[0091] Retrospective validation was performed using a self-built dataset of herpes zoster-related pain compounds and their activity. This dataset was compiled from publicly available literature and our unit's experimental records, containing 1246 traditional Chinese medicine monomers with reported in vitro analgesic activity data, of which 214 met the activity criteria in their respective compartment systems. Evaluation metrics included the Top-50 enrichment hit rate (the percentage of active compounds in the top 50 ranked compounds in the model output) and the Spearman rank correlation coefficient between the model's predicted values ​​and in vitro experimental measurements. Five-fold cross-validation was used, and the mean and standard deviation are reported.

[0092] The control group uses the existing traditional method, which ranks targets by statically weighted linear summation of multi-target affinity and by monotonic addition of blood-brain barrier permeability. The ablation protocol consists of three parts: ablation ① to remove the central overexposure penalty factor, i.e., to... Ablation ② removes the phase weight matrix, i.e., replaces it with the average weight of the three phases; Ablation ③ removes the coverage matching degree, i.e., lets .

[0093] like Figure 2 As shown in section (a), the Top-50 enrichment hit rate of the traditional control scheme was 0.32±0.041, ablation ① was 0.41±0.038, ablation ② was 0.46±0.035, ablation ③ was 0.50±0.033, and the complete model was 0.58±0.030. Among the three ablation methods, the reduction was largest in removing the central overexposure penalty factor, indicating that this penalty factor is the single technical feature that contributes the most to this invention. The reduction was smallest in ablation ③ because this mainly acts on the three upper levels from L1 to L3, while the evaluation object of this validation is the single level.

[0094] like Figure 2 As shown in section (b), statistical predictions – experimental consistency are listed by compartment: The peripheral nerve terminal compartments increased from 0.38±0.050 to 0.71±0.036. The dorsal root ganglion compartments increased from 0.35±0.050 to 0.66±0.039. The spinal cord central compartment increased from 0.44±0.046 to 0.63±0.041. Notably, the central overexposure punishment factor only affected… , The target in the compartment was effective, but Compartmental consistency was also improved. This result exceeded conventional expectations, and the mechanism has been explained in the beneficial effects section: after the shortcut features were suppressed, the shared backbone learned more compartment-specific representations, which were then transferred to the central compartment task head via backpropagation.

[0095] like Figure 2As shown in section (c), the dominant compartment distribution of the top ten candidate resources in the four tiers shifts downwards with each tier. Compartmental offset: The dominant compartment in the L1 medicinal material level is These 3 These 4 There were 3 active monomers at the L4 level, and 5, 3, and 2 at the L4 level, respectively. This is consistent with the expectation that the components of Chinese medicinal materials are naturally distributed across compartments, and that monomers selectively enrich towards the periphery after screening using this method.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Equivalent substitutions or partial improvements made by those skilled in the art within the scope of the technology disclosed in this invention, such as replacing the Transformer molecular encoder with a graph neural network encoder, subdividing the three disease progression phases into more phases, or further subdividing the three barrier compartments anatomically, should all be covered within the protection scope of this invention.

Claims

1. A method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain, characterized in that, include: S1. Construct a compound dataset containing four levels: overall component spectrum of Chinese medicinal materials, classic extracts of Chinese medicinal materials, effective components of Chinese medicinal materials, and active monomers of Chinese medicinal materials. Each compound record in the compound dataset carries a string representing the molecular structure of the compound. S2. Construct a disease target knowledge graph, and label the target points in two dimensions according to the three barrier compartments of peripheral nerve terminal compartment, dorsal root ganglion compartment, and spinal cord central compartment, and the three disease stages of acute phase, subacute phase, and postherpetic neuralgia phase. Obtain the amino acid sequence of each target point from the protein sequence database to obtain the target compartment assignment, target amino acid sequence, and phase weight matrix normalized according to the phase column. S3. Using the molecular diagram obtained from the molecular structure characterization string and the target amino acid sequence as input, and the target compartmentalization as a conditional embedding, train the affinity prediction network and output the compound-target affinity matrix. S4. Train a multi-task barrier reachability network and output the reachability probability vector of the compound in the three barrier compartments; S5. Based on the compartmental affiliation of the target, the compound-target affinity matrix is ​​gated element-by-element using the reachability probability vector; and for targets belonging to the peripheral nerve terminal compartment and the dorsal root ganglion compartment, a penalty factor is applied in the gating, the penalty factor decreasing as the reachability of the compound in the spinal cord central compartment exceeds an allowable threshold; then, the compartmental-temporal analgesic efficacy tensor is obtained by weighted summation using the temporal weight matrix. S6. For the active monomers of traditional Chinese medicine, a weighted power average score is calculated according to the compartment-phase analgesic efficacy tensor. For the remaining three levels, scores are given based on the content-weighted probability sum of the compartment-phase analgesic efficacy tensor of each component and the matching degree between its compartment-phase coverage distribution and the target coverage distribution. S7. Sort the four levels of traditional Chinese medicine resources according to the matching degree score, and output the priority sequence used to determine the sampling order of in vitro analgesic activity experiments.

2. The method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain according to claim 1, characterized in that, The target compartment classification described in S2 is determined based on the anatomical location of the cells primarily expressing the target and the type of their outer barrier: targets primarily expressed in epidermal nerve fibers, keratinocytes, or free nerve endings are classified as peripheral nerve ending compartments; targets primarily expressed in dorsal root ganglion neuronal cell bodies or satellite glial cells are classified as dorsal root ganglion compartments; targets primarily expressed in dorsal horn neurons, microglia, or astrocytes of the spinal cord are classified as central spinal cord compartments; when the same target is expressed in more than two compartments, the target is split into multiple compartment instances and assigned compartment classifications according to the expression abundance attribute of the association edge between the target and cell type in the disease target knowledge graph.

3. The method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain according to claim 1, characterized in that, The phase weight matrix described in S2 is constructed in two steps using the following formula: ， ， In the formula, Indicates the target points before normalization. In time phase The intermediate phase weight is a dimensionless scalar. Represents the target after column normalization In time phase The phase weight is a dimensionless scalar with a value range of . And for each time phase, it satisfies ; The index representing the target compartment instance. ; This represents the total number of target compartment instances, and is a positive integer. Indicates the phases of the disease course. ; This represents the set consisting of the three disease stages. , , , The stages are, in order: acute phase, subacute phase, and postherpetic neuralgia phase. For sets The phase dummy of the summation; A target dummy for summing over the target points; Indicate target point The strength of disease-related evidence in the disease target knowledge graph is a dimensionless scalar with a value range of [value range missing]. ; The evidence sharpening index is a dimensionless scalar with a value range of [value range missing]. ; Indicate target point In time phase The fold change in expression level or activation level relative to healthy controls is a dimensionless scalar quantity, and ; The phase weight matrix represents the natural logarithm; and the phase weight matrix satisfies the order reversal constraint: the phase weight of at least one of the inflammatory body-type targets belonging to the peripheral nerve terminal compartment or the dorsal root ganglion compartment decreases sequentially in the three disease phases, and the phase weight of at least one of the glial cell activation-type targets belonging to the spinal cord central compartment increases sequentially in the three disease phases.

4. The method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain according to claim 1, characterized in that, The binding affinity prediction network described in S3 includes a molecular encoder, which is a multi-layer multi-head self-attention encoder that takes the atomic and substructure labeling sequence of the compound molecular map as input and outputs a molecular representation vector after pooling. The target encoder is a one-dimensional dilated convolutional residual stacked network with progressively increasing dilation rates. It takes the residue encoding of the target amino acid sequence as input and outputs a target representation vector after global pooling. ; a compartment condition modulation layer, whose compartment embedding vector is obtained from the compartment identifier index. Scaled vector generated by mapping With translation vector And perform a channel-by-channel affine transformation on the molecular representation vector. In the formula Represents element-wise product. For partition Embedded vector, , These are the scaling vector and translation vector of the compartment, respectively. For compounds Facing the compartment The molecular representation vector after affine transformation; and the cross-attention layer and multilayer perceptron regression head, to For query, with For both the key and value, the output is a predicted binding affinity calculated using the negative logarithmic dissociation constant, which is then normalized to an open interval using a compression function. This is then used to construct the compound-target affinity matrix.

5. The method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain according to claim 1, characterized in that, The multi-task barrier reachability network described in S4 includes a shared backbone encoder and five task heads, which are for compound... Accessibility of peripheral nerve endings in separate compartments Dorsal root ganglion compartment accessibility Accessibility of spinal cord central compartments Neural tissue affinity With nervous system safety All five outputs fall within the open interval. During training, a barrier-order structure constraint hinge loss is applied, and a multi-task total loss weighted by homoscedasticity uncertainty is used. ， ， In the formula, Indicates a compound index. Indicates the first One compound; , , These are, in order, the peripheral nerve terminal compartment, the dorsal root ganglion compartment, and the spinal cord central compartment; Represents compounds In the next room The reachability probability is a dimensionless scalar. ; Represents compounds The neural tissue affinity of [something] is a dimensionless scalar. ; Represents compounds The neurological safety is defined as a positive safety vector, meaning that a larger value indicates a lower risk of adverse neurological reactions. It is a dimensionless scalar. ; This represents the loss of the barrier order structure constraint hinge, which is a dimensionless scalar that only occurs when the barrier order structure is violated. When the relation is ordered, the value is positive; This indicates taking the non-negative part; Indicates the total loss from multiple tasks; Indicates the task header index. ; Indicates the first The loss of one task leader; Indicates the first The learnable log-variance parameter of each task head is a dimensionless scalar. The weights for the task header are automatically derived from homoscedastic uncertainty. This represents the barrier order constraint weight, which is a dimensionless scalar with a value range of [value range missing]. .

6. The method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain according to claim 5, characterized in that, The gating and compartment-phase analgesic efficacy tensors described in S5 are calculated using the following formula: ， ， ， In the formula, Represents compounds Target The gate value is a dimensionless scalar; Indicates the first Example of a target compartment; Indicate target point The ownership of the compartment, ; Represents compounds At the target point The probability of accessibility to the assigned compartment; Represents compounds Accessibility probability in the spinal central compartment; This represents the permissible threshold for central nervous system exposure, a dimensionless scalar with a value range of [value range missing]. ; and Both indicate taking the non-negative part, that is ; This indicates the penalty intensity based on the target compartment affiliation value, when... Peripheral nerve terminal compartment or dorsal root ganglion compartment Take positive value ,when Spinal cord central compartment Take zero at time. The range of values ​​is ; Represents compounds In the next room Phase The analgesic efficacy component is a dimensionless scalar quantity. ; Represents the target point in the column normalized phase weight matrix. In time phase Element; The compound-target affinity matrix represents the compounds in the compound-target affinity matrix. Target Elements normalized by the compression function ; Indicates the classification of the compartment as Summing all target points.

7. The method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain according to claim 1, characterized in that, In S6, compounds at the active monomer level of traditional Chinese medicine are scored using a weighted power average, while records at the other three levels are scored using a content-weighted probability sum and coverage matching degree. ， ， ， ， ， In the formula, Indicates monomer The weighted power average score; This refers to the set consisting of the three barrier compartments. ; This represents the set consisting of the three disease stages; This represents the components of the compartment-phase analgesic efficacy tensor; Representing grid points The grid point weights are dimensionless scalars. ,and ; It represents the power, is a dimensionless scalar, and takes values ​​in the range of 1 to 2. ; This represents an index of a record in levels L1 to L3; Representing records The set of components, For the index of its components; , They are respectively for the assembly of compartments Phase set Dumb standard for summation; Indicates ingredients In Record The mass fraction in the figure is a dimensionless scalar. ,and ; The content elasticity index is a dimensionless scalar quantity with a value range of [value range missing]. ; Representing records At grid points Content-weighted probability and efficacy, ; Indicates multiplication; Representing records The compartmental-temporal coverage distribution is a dimensionless scalar; The target coverage distribution is obtained by normalizing the cross product of the proportion of patient pain burden at each stage of the disease course and the proportion of pathological contribution of each barrier compartment. Represents relative entropy; Represents an exponential function with the natural constant as its base; Representing records The coverage matching degree is a dimensionless scalar. If and only if the record When the coverage distribution is completely consistent with the target distribution, take ; Representing records Level rating; The complementarity preference coefficient is a dimensionless scalar with a value range of [value range missing]. The larger the value, the more it favors resources with broad and balanced coverage.

8. The method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain according to claim 7, characterized in that, This also includes traditional Chinese medicine compound prescriptions. Quantification of the synergistic analgesic effect of multiple components: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] The compartment-phase analgesic efficacy tensor is expanded into a nine-dimensional vector by rows. The compartment-time complement index of the two components and the synergistic index of the compound are calculated using the following formula: ， ， In the formula, Indicates a traditional Chinese medicine compound prescription; , Indicates compound The two components; Represents compounds The nine-dimensional vector obtained by expanding the compartment-phase analgesic efficacy tensor by rows; Represents the dot product of vectors; Denotes the Euclidean norm; Indicates ingredients and The complementary exponents of the compartments are dimensionless scalars. The larger the value, the more complementary the coverage of the two components on the compartment-temporal grid; Indicates compound The synergy index is a dimensionless scalar. ; Indicates ingredients In compound mass fraction in ; Indicates ingredients The score obtained by the weighted power average scoring formula For component pairs The geometric mean of efficiency; Indicates the use of compound Summing all disordered components; and applying compatibility constraints according to the following rule: the principal drug component is its synergistic index. Located in compound The first 20% of all ingredients; the adjuvant ingredient is its complementarity index with the principal ingredient. Located in compound The first 20% of all ingredients.

9. The method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain according to claim 1, characterized in that, It also includes S8: sending the priority sequence into three types of in vitro experimental systems that correspond one-to-one with the three barrier compartments, wherein the peripheral nerve ending compartment corresponds to the skin or epidermal nerve ending system, the dorsal root ganglion compartment corresponds to the dorsal root ganglion neuron system, and the spinal cord central compartment corresponds to the spinal cord slice system; calculating the residual between the experimental measurement value and the compartment-phase analgesic efficacy tensor independently for each compartment, and updating the phase weight matrix exponentially using the following formula: ， ， In the formula, Indicates compartment Phase The experimental residuals are dimensionless scalars; Indicates being sent to the isolation room The corresponding batches of compounds in the experimental system This represents the number of samples in this batch; Represents compounds In the next room Normalized to open intervals in the corresponding experimental system The experimentally measured values, superscript Indicates experimental determination; This represents the components of the compartment-phase analgesic efficacy tensor; Indicates the iteration round, which is a non-negative integer; , They represent the first , Phase weights of the wheel; This represents the update step size, which is a dimensionless scalar with a value range of [range missing]. ; Indicate target point The ownership of the compartment; The function represents an exponential function; and the penalty intensity of the penalty factor is adjusted according to the sign and magnitude of the residuals between the peripheral nerve terminal compartment and the dorsal root ganglion compartment. With the allowable threshold .

10. The method for constructing a virtual screening model of traditional Chinese medicine for herpes zoster-related pain according to claim 4, characterized in that, The molecular encoder is pre-trained using skeleton-aware contrastive learning before formal training: based on the Murcko skeleton, compound pairs with the same skeleton but different substituents are classified as positive sample pairs, and compound pairs with different skeleton categories are classified as negative sample pairs. The contrastive loss is calculated using the following formula: ， In the formula, This represents the comparative loss and is a dimensionless scalar. Indicates a small batch. Its sample size, For anchor sample index; , , These represent the numerator vectors for anchor samples, positive samples, and negative samples, respectively. A dummy standard for summing negative samples; Indicates cosine similarity; Represents the temperature coefficient, which is a dimensionless scalar and ; This represents an exponential function; and the sampling probability of negative samples is reweighted by the inverse of the sample size of the skeleton class.

Images