A method, device and readable medium for analyzing rhizosphere response of cadmium and mercury pollution wetland based on multi-omics data correlation

Abstract

This invention relates to the field of data processing technology. It discloses a method, device, and readable medium for analyzing the rhizosphere response of cadmium and mercury-contaminated wetlands based on multi-omics data association. The method includes: locating candidate evidence fragments using a context matrix; introducing red oxidation state-optimal transport to repeatedly screen candidate evidence fragments to obtain a set of evidence chains; constructing a complete technical closed loop from multi-source heterogeneous data fusion to collaborative remediation and maintenance; achieving deep alignment between cross-omics evidence and geochemical thermodynamic conditions through red oxidation state-optimal transport; and introducing an interpretable causal sensitivity score to screen out environmentally dependent and vulnerable evidence, significantly improving the accuracy and robustness of key resistance element identification; simultaneously, based on adversarial synthesis of mechanism isomorphic entities and curriculum-based inoculation training, combined with process projection anchor immune regularization, the target model can maintain stable causal inference and prediction performance even under strong disturbance scenarios such as typhoons and land reclamation.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method, apparatus, and readable medium for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association. Background Technology

[0002] The multi-omics data association-based rhizosphere response analysis method for cadmium and mercury polluted wetlands refers to a technical strategy that utilizes multi-level molecular data, including genomics, transcriptomics, proteomics, metabolomics, and microbiome, to systematically integrate and correlate the response characteristics of the rhizosphere microenvironment of wetland plants under cadmium and mercury combined pollution conditions. This method analyzes changes in gene expression, protein function, metabolites, and community structure in rhizosphere soil, roots, and microbial communities under heavy metal stress to construct network relationships between pollutants and the rhizosphere microecology. This reveals the tolerance, translocation, accumulation, and detoxification mechanisms of wetland systems to cadmium and mercury pollution, providing molecular-level evidence for wetland pollution diagnosis and ecological restoration.

[0003] The existing technology has the following main shortcomings: Traditional multi-omics association analysis relies solely on statistical co-occurrence or semantic similarity, lacking the hard constraint of geochemical and thermodynamic feasibility, leading to a large amount of spurious correlation evidence interfering with the identification of key elements. At the same time, conventional machine learning models assume that the training and testing distributions are consistent, which leads to a decline in prediction performance when faced with distribution drift caused by external interventions such as typhoons, water replenishment, and land reclamation. They also ignore the benefit gap caused by the shift of anchor scenarios in actual engineering, making it difficult to guarantee long-term stable operation. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method, apparatus, and readable medium for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for analyzing the rhizosphere response of cadmium and mercury-contaminated wetlands based on multi-omics data association, the method comprising: S101: Acquire multi-omics and rhizosphere physicochemical data, preprocess the multi-omics and rhizosphere physicochemical data, and form a data cube; S102: Based on the data cube, perform lexicalization and ontology normalization, perform sparse attention long sequence encoding to form a context matrix, use the context matrix to locate candidate evidence fragments, introduce red oxidation state-optimal transmission to screen candidate evidence fragments multiple times, and obtain a set of evidence chains. S103: Construct a set of mechanism isomorphic objects, conduct quality audits on the set of mechanism isomorphic objects to obtain adversarial examples, train multiple types of models to obtain the target model, use the target model to perform anomaly detection and traceable explanation, and obtain a comprehensive report; S104: Establish a collaborative repair combinatorial space for multi-objective optimization to obtain the optimal combinatorial solution set. Simulate the optimal combinatorial solution set to select the most stable solution. Based on the most stable solution, design a small-scale inoculation experiment and perform online data reinjection. S105: Connect the target model to the digital twin platform for simulation and operation and maintenance closed loop.

[0006] Furthermore, the sparse attention long sequence encoding is specifically performed as follows: The lexicalized data cube is input into the sparse attention encoder, and modality embedding and hierarchical relative position encoding are injected during the encoding process to obtain the context matrix.

[0007] Furthermore, the step of locating candidate evidence fragments using a context matrix specifically involves: A multi-task scoring head is connected to the top layer of the sparse attention encoder to perform evidence strength regression, entity type classification and ontology edge prediction respectively, and obtain three types of scores. The three types of scores are weighted and fused to obtain the comprehensive evidence potential. Superimposed conditional random fields are used to smooth the boundary of the comprehensive evidence potential energy to obtain a score matrix and initial candidate evidence fragments. The initial candidate evidence fragments are hierarchically aggregated according to five granularities to obtain candidate evidence fragments; The generation of the candidate evidence fragments employs a two-layer strategy.

[0008] Furthermore, the introduction of red oxidation state-optimal transport to screen candidate evidence fragments multiple times includes: An initial candidate slice pool was obtained through preliminary screening using red oxidation state-optimal transport. For each candidate slice in the initial candidate slice pool, an interpretable causal sensitivity score was calculated. Obtain the original model score, weighted fuse the interpretable causal sensitivity score and the original model score to obtain the final score, sort them in descending order according to the final score to obtain the evidence chain set; Among them, key resistance elements were identified based on the evidence chain set.

[0009] Furthermore, the preliminary screening via red oxidation state-optimal transport specifically involves: For each candidate slice, functional lexical units and contextual lexical units are extracted, the red oxidation state-optimal transmission distance between functional lexical units and contextual lexical units is calculated, and the optimal transmission plan is approximated by the Sinkhorn iterative algorithm. Each candidate slice is decomposed into multiple subspaces, and an approximate nearest neighbor index is constructed in each subspace to obtain an initial candidate slice pool.

[0010] Furthermore, the specific steps for identifying key resistance elements based on the evidence chain set include: Edge weights are calculated based on the evidence chain set. Resistant microbial groups, core metabolic pathways, key metabolites, and environmental driving factors are regarded as nodes, and a co-occurrence network is constructed by combining the edge weights. By simulating the removal of a certain group or pathway node in the co-occurrence network, the change in the predicted value of cadmium and mercury speciation is observed, the causal contribution is quantified, and key resistance factors are identified.

[0011] Furthermore, the multiple models include: a goal-oriented robust fusion model, an interpretable student model, and a biomarker stress level assessment model; The steps for training a target-oriented robust fusion model include: The adversarial examples are sorted and arranged according to three-dimensional indicators. The vaccination priority score is calculated by combining the three-dimensional indicators. The training process is divided into three stages to obtain a phased vaccination training sequence. The phased vaccination training sequence is used to train the target-oriented robust fusion model. During each upgrade phase, a priority experience replay mechanism is adopted.

[0012] Furthermore, the specific steps for establishing a collaborative repair combinatorial space for multi-objective optimization are as follows: Based on key resistance factors, candidate plants and functional microorganisms with cadmium and mercury enrichment or transformation capabilities were screened to construct a synergistic remediation combination space. For each candidate scheme in the synergistic remediation combination space, a target-oriented robust fusion model was used to predict its cadmium removal rate, mercury removal rate and stress response under standard anchor conditions. A multi-objective optimization function is established, and a non-dominated sorting genetic algorithm is used to find the optimal combination solution set in the combinatorial space.

[0013] An electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the above-described method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association.

[0014] A computer-readable storage medium storing a computer program, which, when executed, implements the aforementioned method for analyzing the rhizosphere response of cadmium- and mercury-polluted wetlands based on multi-omics data association.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This application discloses a method, equipment, and readable medium for analyzing the rhizosphere response of cadmium and mercury-contaminated wetlands based on multi-omics data association. This method constructs a complete technical closed loop from multi-source heterogeneous data fusion to collaborative remediation and operation. It achieves deep alignment between cross-omics evidence and geochemical and thermodynamic conditions through red oxidation state-optimal transport, and introduces an interpretable causal sensitivity score to screen out environmentally dependent and vulnerable evidence, significantly improving the accuracy and robustness of key resistance element identification. Simultaneously, based on adversarial synthesis of mechanism isomorphic entities and curriculum-based inoculation training, combined with process projection anchor immune regularization, the target model can maintain stable causal inference and prediction performance even under strong disturbance scenarios such as typhoons and land reclamation. Combined with knowledge distillation of interpretable student models and anchor robustness multi-objective optimization strategies, this method ultimately achieves closed-loop management of dynamic early warning and remediation decisions through a digital twin platform, providing a solution with both deep mechanistic explanation and engineering feasibility for the accurate diagnosis and long-term governance of cadmium and mercury-contaminated wetlands. Attached Figure Description

[0016] Figure 1 A flowchart of a method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association provided by this invention; Figure 2 A schematic diagram of the structure of an electronic device provided by the present invention; Figure 3 This is a schematic diagram of the structure of a computer-readable storage medium provided by the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Please see Figure 1 As shown in the figure, this embodiment discloses a method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association. The method includes: S101: Acquire multi-omics and rhizosphere physicochemical data, preprocess the multi-omics and rhizosphere physicochemical data, and form a data cube; As a specific implementation method, the inputs in this embodiment are the Pearl River Estuary mangrove wetland ecological functional zoning map, historical typhoon and rainstorm records, and existing reclamation and water replenishment project ledgers preset by the project. First, the study area was cataloged, identifying three typical habitat types: the upstream estuary, the mid-section reclamation buffer zone, and the offshore control point. Three sampling profiles were established perpendicular to the shoreline in each habitat type, with one fixed sampling point at each high tide level, mean tide level, and low tide level, totaling 27 core sampling points. Simultaneously, one control point was established in the non-mangrove zone of the offshore beach, containing three parallel points. The latitude, longitude, tide level, and vegetation type of all sampling points were entered into a spatial database. A timeline of intervention events was simultaneously established, cataloging the dates and intensities of historical typhoons, rainstorms, water replenishment, pesticide application, and restoration projects.

[0019] Rolling sampling is conducted, covering both the dry and wet seasons, with two rounds of standard sampling performed each season. Intensive sampling is initiated within 72 hours of a typhoon or heavy rain event. Follow-up sampling is conducted 7 days before and after water replenishment, chemical treatment, or remediation projects. A unique batch identifier is generated for each sampling event, and the corresponding anchor variable vector is labeled, including: season type (dry / wet), tidal state (high / mid-tide / low tide), event type (no event / typhoon / heavy rain / water replenishment / chemical treatment / remediation), and tidal elevation and water temperature at the time of sampling.

[0020] Meanwhile, rhizosphere sediments were collected from the root surface at a depth of 0 to 2 mm, gently scraped with a sterile bamboo knife and immediately placed into sterile cryovials; non-rhizosphere sediments were collected from sediments at a depth of at least 5 cm from the root surface. Root exudates were collected using a Rhizon ceramic rod (0.2 μm pore size) inserted into the rhizosphere, subjected to a vacuum of -20 kPa, and continuously microfiltration for 2 to 4 hours. The collected exudate was then aliquoted into brown glass bottles.

[0021] For iron film samples, healthy root segments of 2 to 5 cm in length were collected. First, the roots were gently rinsed with sterile seawater to remove any attached deposits. Then, they were immersed in pre-cooled DCB extraction solution (citric acid-sodium bicarbonate-potassium ferrocyanide system, concentrations of 0.03 mol / L, 1.0 mol / L, and 0.1 mol / L, respectively) and extracted for 20 minutes at a constant temperature of 25°C. The extract was used for subsequent metal and iron speciation analysis, while the root segment solids were retained for surface morphology characterization. All samples were transferred to a freezer at -80°C for freezing within 6 hours.

[0022] In-situ monitoring equipment was simultaneously deployed at all 30 sampling points to measure pH and redox potential of rhizosphere sediment profiles in 1-2 mm increments. Water temperature, salinity, and tidal level changes were continuously recorded using a CTD multi-parameter instrument and a pressure gauge. Pore water from surface sediments was collected using a portable air pump, and dissolved oxygen and sulfides were measured on-site. All in-situ monitoring data were automatically stored at 10-minute intervals via a data logger, forming a time series aligned with the sampling events. The monitoring period covered the complete tidal cycle, ensuring the capture of chemical gradient changes during tidal fluctuations.

[0023] Five omics experiments were conducted in parallel on rhizosphere sediment samples, namely DNA, RNA, proteomics, metabolomics, and microbial community.

[0024] The total iron, total cadmium and total mercury contents of the DCB extract were determined to obtain a table of total metal content and speciation concentrations, iron film thickness and crystallinity parameters, and the ratio of ferrous to ferric iron.

[0025] All sequencing data underwent quality assessment, removing adapter sequences and low-quality bases (reads with a quality value below 20 and lengths less than 50 base pairs were discarded); mass spectrometry data were extracted and aligned, removing noise peaks with a signal-to-noise ratio below 3. For protein identification results, peptides with an identification confidence level below 95% were filtered out.

[0026] It should be further noted that the preprocessing of multi-omics and rhizosphere physicochemical data, including missing value imputation and batch correction, is existing technology and will not be elaborated upon in this application.

[0027] The preprocessed multi-omics and rhizosphere physicochemical data are uniformly mapped to the domain ontology namespace. Specifically, genes and gene clusters are mapped to the National Center for Biotechnology Information (NCBI) database and genome classification database for taxonomic annotation; reaction steps are mapped to the Kyoto Encyclopedia of Genes and Genomes orthologs and metabolic pathway database for functional annotation; metabolites are mapped to chemical entities on the biological interest ontology; and environmental variables and sampling conditions are mapped to the environmental ontology. Hypergraph or hyperbolic embedding methods are used to establish cross-ontology synonym resolution and hierarchical associations, eliminating the fragmentation of names for the same entity across different omics platforms. Subsequently, genomic, transcriptomic, proteomic, metabolomic, microbiome, geochemical, and field text metadata are concatenated using unified sample identifiers and time stamps to obtain a data cube.

[0028] S102: Based on the data cube, perform lexicalization and ontology normalization, perform sparse attention long sequence encoding to form a context matrix, use the context matrix to locate candidate evidence fragments, introduce red oxidation state-optimal transmission to screen candidate evidence fragments multiple times, and obtain a set of evidence chains. For example, lexicalization and ontology normalization are performed on the basis of the data cube, including: Transform the modal data of the data cube into domain-specific sub-word elements; For genomic and metagenomic data, gene clusters and open reading frames are mapped to ortholog group numbers from the Kyoto Encyclopedia of Genomes and Genomes and enzyme committee numbers from the Metabolic Pathway Database, serving as functional sub-terms; for metabolomics data, metabolites are decomposed into substructural fragments and isotope peak type markers from the chemical entity biology interest ontology, serving as chemical sub-terms; for microbial community data, species annotations of metagenomic assembled genomes or amplicon sequence variants are mapped to phylogenetic levels from genomic classification databases or the National Center for Biotechnology Information (NCBI) taxonomy, serving as ecological sub-terms; for field text metadata, domain-specific vocabulary and word fragment models are used for segmentation to generate environmental descriptive sub-terms.

[0029] The sparse attention long sequence encoding process specifically involves: The lexicalized data cube is input into the sparse attention encoder. During the encoding process, modality type (six modality tags: genome, transcriptome, proteome, metabolome, microbiome, and text metadata) embedding and hierarchical relative position encoding (reaction step level, pathway level, and group level) are injected to obtain the context matrix.

[0030] The BigBird sparse attention mechanism is employed, and its attention mask construction follows the following structure: ; In the formula: This represents a fully connected attention mechanism for global lexical units across all lexical units. This represents the local sliding window attention of each word to its 128 left and 128 right neighbors. For each lexical unit, random attention is applied to 64 randomly selected lexical units.

[0031] The global word count is set to 10% of the sequence length.

[0032] The method of locating candidate evidence fragments using a context matrix specifically includes: A multi-task word scoring head is connected to the top layer of the sparse attention encoder to perform evidence strength regression, entity type classification, and ontology edge prediction, respectively, to obtain three types of scores. The three types of scores are then weighted and fused to obtain the comprehensive evidence potential, as shown in the following formula: ; In the formula: For the context matrix of lexical units, the first item is... , For the weights and biases of the regression of evidence strength, For the Sigmoid activation function, in the second term , Weights and biases for entity type classification, in the third item , The weights and biases for the predicted edges of the ontology. This is the concatenation of adjacent word vectors. , , These are weighting coefficients, with values ​​of 0.5, 0.3, and 0.2 respectively. Superimposed conditional random fields are used to smooth the boundaries of the comprehensive evidence potential, resulting in a score matrix and initial candidate evidence fragments. The regularization coefficient of the conditional random field is 1×10⁻⁶. -2 ; The initial candidate evidence fragments were hierarchically aggregated according to five granularities to obtain candidate evidence fragments. The first type is the sentence segment level slice, which aggregates continuous high-scoring terms in text metadata into complete descriptive sentences. The second type is the functional cluster level slice, which aggregates terms belonging to the same gene cluster or the same protein complex. The third type is the pathway subgraph level slice, which aggregates adjacent high-scoring reaction steps and catalytic genes into subnetworks on the pathway maps of the Kyoto Encyclopedia of Genes and Genomes or the Metabolic Pathway Database. The fourth type is the co-expression module level slice, which aggregates gene groups with similar spatiotemporal expression patterns based on the co-expression relationships of transcriptome data. The fifth type is the community unit level slice, which aggregates terms of the same metagenomic assembly genome or the same functional amplicon sequence variant into microbial community units. The aggregation formula is as follows: ;

[0033] In the formula: The set of words contained in the slice. As the normalization factor, For graph attention network aggregation terms, G is the pathway subgraph or community co-occurrence subgraph corresponding to the slice.

[0034] The generation of the candidate evidence fragments adopts a two-layer strategy. First, the context vectors of each word in the candidate fragment are subjected to evidence score weighted average pooling. Then, if the candidate fragment involves a pathway or community structure, graph attention network aggregation is performed on the corresponding pathway graph or co-occurrence graph to introduce information from neighboring nodes.

[0035] The introduction of red oxidation state-optimal transport for multiple screening of candidate evidence fragments includes: An initial candidate slice pool was obtained through preliminary screening using red oxidation state-optimal transport. For each candidate slice in the initial candidate slice pool, an interpretable causal sensitivity score was calculated. ; In the formula: Given the anchor variable vector 'a', the fine-ranking model scores the credibility or support of this evidence slice, where A is the feasible region of the anchor variable. The gradient of the anchor variable vector measures the sensitivity of the support score to changes in environmental conditions. sup is the most unfavorable perturbation point in the feasible region of the anchor variable that maximizes the gradient norm.

[0036] Obtain the original model score, weighted fuse the interpretable causal sensitivity score and the original model score to obtain the final score, sort them in descending order according to the final score to obtain the evidence chain set; Among them, key resistance elements were identified based on the evidence chain set.

[0037] The preliminary screening using red oxidation state-optimal transport specifically involves: For each candidate slice, functional terms and context terms are extracted, and the red oxidation state-optimal transmission distance between functional terms and context terms is calculated. Then, the Sinkhorn iterative algorithm is used to approximate the optimal transmission plan. ; In the formula: For the transmission plan matrix, , For the i-th functional word in the candidate slice and the j-th conditional word in the context, For cross-modal semantic distance, These are characterization functions that map functional terms and environmental terms to redox potential states, respectively. The redox potential values ​​of functional terms are obtained from a database using the standard reaction potentials of the catalyzed terms, while the redox potential values ​​of environmental terms are obtained by normalizing the measured range of the microelectrode in the field. This represents the mapping value between the optimal pH for the functional reaction and the measured pH in the environment. The standard Gibbs free energy deviation or approximate barrier corresponding to the functional response set, under the given environmental context, is provided by a thermodynamic database and estimation model. The entropy regularization term of the transmission plan matrix is ​​used to accelerate the solution of the Sinckhorn iterative algorithm. The weight of the redox potential difference is 0.6. The weight for pH is 0.4. This is the baseline weight for the thermodynamic free energy deviation, with a value of 1.0. is the entropy regularization coefficient, with a value of 0.05. The Sinkhorn iterative algorithm has 50 iteration steps.

[0038] The red oxidation state-optimal transport distance not only measures the ontological semantic distance between functional terms and environmental context terms, but also introduces redox potential difference, acidity-base difference and standard Gibbs free energy deviation to ensure that the retrieved evidence slices are geochemically and thermodynamically feasible, thereby filtering out pseudo-related slices that are only semantically related but not supported by physicochemical conditions.

[0039] Each candidate slice is decomposed into three subspaces: functional space, chemical space, and ecological space. An approximate nearest neighbor index is constructed in each subspace to obtain an initial candidate slice pool.

[0040] The functional subspace index covers sections related to genes and pathways; the chemical subspace index covers sections related to metabolites and secretions; and the ecological subspace index covers sections related to microbial communities and environmental factors.

[0041] The process of identifying key resistance elements based on the evidence chain set includes the following steps: Edge weights are calculated based on the evidence chain set. Resistant microbial groups, core metabolic pathways, key metabolites, and environmental driving factors are regarded as nodes, and a co-occurrence network is constructed by combining the edge weights. By simulating the removal of a certain group or pathway node in the co-occurrence network, the change in the predicted value of cadmium and mercury speciation is observed, the causal contribution is quantified, and key resistance factors are identified.

[0042] It should be further explained that the S102 model is an integrated evidence retrieval engine, employing a six-level cascaded structure: The first layer is the input representation layer, which receives the data cube; the second layer is the sparse long sequence coding layer, which consists of 12 stacked sparse Transformer encoders. The core of each layer is a 16-head sparse self-attention mechanism, each head with a dimension of 64, which processes ultra-long sequences through three masking modes: global tokens, local sliding windows, and random attention.

[0043] The third layer is a word-level multi-task reasoning layer, deploying three task heads in parallel. The first is the evidence strength regression head: after linear projection and sigmoid activation, it outputs the evidence strength score of each word for the morphological transformation proposition of mercury cadmium. The second is the entity type classification head: after linear projection and Softmax activation, it outputs the probability distribution of whether the word belongs to a gene, metabolite, pathway step, environmental factor, or background description. The third is the ontology edge prediction head: it concatenates adjacent word vectors into a 2048-dimensional array, performs linear projection and sigmoid activation, and outputs the probability that there is an ontology-level association between adjacent words.

[0044] The fourth level is the evidence slice aggregation layer, which hierarchically aggregates candidate evidence fragments according to five granularity levels; the fifth level is the cross-modal alignment layer, which extracts functional terms and environmental context terms from each candidate evidence slice and calculates the red oxidation state-optimal transmission distance; the sixth level is the retrieval and ranking layer, which performs multiple screenings.

[0045] Finally, a candidate pool of biomarkers was constructed by integrating taxa, pathways, metabolites, root exudate components, and iron film characteristics with key resistance elements as the core.

[0046] S103: Construct a set of mechanism isomorphic objects, conduct quality audits on the set of mechanism isomorphic objects to obtain adversarial examples, train multiple types of models to obtain the target model, use the target model to perform anomaly detection and traceable explanation, and obtain a comprehensive report; In this embodiment, the mechanism of constructing the image set includes: A three-layer synthesizer is constructed to generate mechanistic isomorphs. The first layer is the physical process layer, which runs a differentiable geochemical-microbial dynamics coupled simulator to output the sensitivity matrices of iron film nucleation and stripping rates, Fe / S / N electron acceptor switching trajectories, and mercury methylation and demethylation rates to anchor variables. The second layer is the surface texture preservation layer, which trains a multi-omics texture encoder to measure the surface similarity of metagenomic K-mer distribution, transcript t-SNE structure density, metabolic fingerprint peak shape, and protein cluster annotation bag-of-words. During synthesis optimization, the texture distance between the mechanistic isomorph and the original sample is minimized, while the anchor-induced mechanism is maximized and reversed (e.g., iron film state flips from formation to destruction, redox potential flips from positive to negative, and root exudate flow flips from low to high). The third layer is the knowledge evidence layer, which calls a candidate slice set and constrains the generation direction by a domain expert knowledge graph. All samples must pass hard constraints of mass conservation, thermodynamic feasibility, and consistency of stable isotope flow.

[0047] The surface texture distance constraint formula is used for synthesis quality control: ; In the formula: The total texture preservation loss is represented by m, where m is the modality index. Let be the texture statistical feature extractor for the m-th modality; x is the original sample feature vector. The mechanism of synthesis is similar to that of the image. Denotes the Fraser distance of the m-th mode. Modal weights are assigned as follows: 0.3 for the genome, 0.25 for the transcriptome, 0.25 for the metabolome, and 0.2 for the proteome, with synthetic constraints. <0.1, and the mechanism inversion label must be the opposite of the original sample.

[0048] Among them, the domain expert knowledge graph refers to a machine-readable structured causal knowledge network jointly constructed by domain experts and documentary evidence for the rhizosphere biogeochemical processes of wetlands with combined cadmium and mercury pollution.

[0049] A three-layer quality audit was performed on the ensemble of isotopes. The first layer was a physical conservation audit, which checked the mass balance of carbon, sulfur, and iron to ensure that no elements were created or disappeared during the synthesis process. The second layer was a thermodynamic feasibility audit, which used a pH-redox potential phase diagram to verify whether the synthesized sample fell within the geochemical stability region of the corresponding habitat. The third layer was an isotope consistency audit, which verified whether the flow of organic carbon labeled with carbon-13, the reduction ratio of sulfate labeled with sulfur-34, and the iron film exchange flux labeled with iron-57 conformed to the measured isotope fractionation rules.

[0050] The various models include: a goal-oriented robust fusion model, an interpretable student model, and a biomarker stress level assessment model; The steps for training a target-oriented robust fusion model include: The adversarial examples are sorted and arranged according to three-dimensional indicators. The vaccination priority score is calculated by combining the three-dimensional indicators. The training process is divided into three stages to obtain a phased vaccination training sequence. The phased vaccination training sequence is used to train the target-oriented robust fusion model. The three-dimensional indicators are difficulty indicator, impact indicator, and cost indicator. The difficulty indicator is defined as the minimum amount of anchor disturbance required to cause the mechanism to reverse (e.g., if changing the pH by 0.3 units or the Eh by 50 millivolts triggers the reversal, the difficulty is high; if changing the pH by 2 units triggers the reversal, the difficulty is low), and the reciprocal is normalized. The impact index is defined as the absolute value of the prediction difference between the mechanism isomorphism and the original sample on the initial training version of the target-oriented robust fusion model; The cost indicators are weighted according to their feasibility in the field: uncontrollable natural events such as typhoons have high cost weights, while controllable engineering measures such as water replenishment and pesticide application have medium cost weights. This application does not impose any restrictions on this.

[0051] During each upgrade phase, a priority experience replay mechanism is adopted. In the first phase (rounds 1-15), only 20% of the samples are low-difficulty single-anchor adversarial samples. In the second phase (rounds 16-35), 50% of the samples are mixed with medium-difficulty multi-anchor coupled adversarial samples. In the third phase (rounds 36-50), 80% of the samples are high-difficulty extreme event combination adversarial samples.

[0052] The formula for calculating vaccination priority is: ; In the formula: This represents the absolute value of the prediction difference between the mechanism-identical sample and the original sample on the initial training version of the target-oriented robust fusion model. To minimize anchor disturbance, As the event cost weight, The value is 0.01, intended to prevent division by zero.

[0053] The first phase selects samples with priority scores below the 40th percentile; the second phase selects samples from the 40th to 75th percentiles; and the third phase selects samples above the 75th percentile. The mixing ratio of adversarial samples to original samples in each phase is 20:80, 50:50, and 80:20, respectively.

[0054] The input is a candidate pool of markers, and the course is upgraded every 5 rounds during training. The coefficients λ and β are scheduled using a linearly increasing followed by cosine annealing strategy. The main objective function is: ; In the formula: x is the multi-omics fusion feature vector, which is composed of standardized omics-physicochemical features after screening by a biomarker candidate pool; a is the anchor variable vector; and y is the supervision label. This is a predictive model, containing a 12-layer Transformer encoder and a task-specific output header. As an adversarial scene synthesizer, it is parameterized by a three-layer synthesizer. To generate control variables, U is the feasible region constrained by the physics-process-texture triple constraints. For the supervised loss function, cross-entropy loss is used for classification tasks, and mean squared error loss is used for regression tasks. For non-process-supported projection operators, For the posterior covariance matrix of the anchor variable, , These are the process projection regularization coefficient and the anchor uncertainty regularization coefficient, respectively.

[0055] It needs to be further explained that, The course-based scheduling is adopted, with the value linearly increasing from 0.05 to 0.15 in rounds 1 to 5, and decreasing from 0.15 to 0.10 in rounds 6 to 50 using cosine annealing. The course-based scheduling is adopted. The value increases linearly from 0.05 to 0.20 in rounds 1 to 5, and decreases from 0.20 to 0.15 in rounds 6 to 50 using cosine annealing.

[0056] and A 6×6 diagonal matrix, with diagonal elements (variances) as follows: pH variance 0.25 (standard deviation 0.5), Eh variance 900 (standard deviation 30 mV), Fe... 2+ / Fe 3+ The variances of the ratio logarithms are 0.04 (standard deviation 0.2), secretion flow variance is 0.25 (standard deviation 0.5, normalized), tidal level variance is 0.11 (standard deviation 0.33), and seasonal variance is 0.25 (standard deviation 0.5). The off-diagonal elements are the covariances among the anchor variables, estimated from the measured covariances on site.

[0057] Optimizer: AdamW, initial learning rate 1×10-4 Batch size 16 (gradient accumulation 4 steps, equivalent batch 64), total training epochs 50.

[0058] Simultaneously, the robust, goal-oriented, and robust fusion model is transferred to an interpretable student model through knowledge distillation. The student model employs a single-layer Kolmogorov-Arnold Network, approximating the decision boundary of the teacher model with a combination of closed-form basis functions. During distillation, the student model not only learns hard labels but also fits the soft probability distribution of the teacher model's output, thus inheriting the teacher model's robust causal representation of anchor variables. After distillation, the student model can provide a lookup table of input variables, basis functions, and predicted outputs, achieving auditable interpretability at the function level.

[0059] In this embodiment, the distillation loss function is: ; In the formula: The cross-entropy loss between the student model and the real hard labels is used. The KL divergence between the soft labels of the student model and the objective-based robust fusion model. This is the Softmax function, where T is the distillation temperature and takes a value of 4. The soft label loss weight is set to 0.7. , These are the corresponding original output values; the number of basis functions for the student model is set to 32; the distillation rounds are 100, and the learning rate is 5×10⁻⁶. -4 .

[0060] Using the robust representations learned by the target-oriented robust fusion model on the biomarker candidate pool features as input, a biomarker stress level assessment model is constructed. A temperature scaling and ordination-preserving regression calibration module is added after the classifier output layer to ensure that the probability values ​​output by the model strictly reflect the true confidence level. Simultaneously, candidate slices and evidence chains are associated with each tiered decision, ensuring that each level of judgment output by the model can be traced back to specific microbial groups, metabolic pathways, root exudate indicators, and iron film characteristics.

[0061] The classification probability is: ; In the formula: For the output value of the i-th class in the output layer, The temperature is set to 1.5 for calibration. The grading thresholds are divided using equal-width probability partitioning: mild (p∈[0.50,0.75), moderate (p∈[0.75,0.90), severe (p∈[0.90,0.97), and extremely severe (p∈≥0.97)). When the maximum probability is below 0.50, the model outputs "uncertain" and triggers manual review.

[0062] Finally, using continuous monitoring data from the past four weeks (28 days) as a historical window, the pollution indicators (total cadmium, total mercury, methylmercury fraction, and iron film thickness index) and stress levels for the next seven days are predicted on a rolling basis. During the prediction, anchor variables for the next seven days (such as tide forecasts and seasonal information) are input into the model as known conditions, and the output is the predicted value and prediction range. An early warning is triggered when the predicted value exceeds the preset ecological risk threshold, or when both the lower and upper limits of the prediction range cross the threshold boundaries simultaneously.

[0063] The rolling prediction equation is as follows: ; In the formula: This is a historical sequence of multi-omics and physicochemical characteristics over the past 28 days. The forecast step size is one week, and the rolling interval is one day. The anchor variable series is for the same period and the next 7 days (among the future anchor variables, the tide level is the astronomical tide forecast value, the season is a known constant, and the other environmental anchor variables are extrapolated from the moving average of the most recent 7 days).

[0064] The standardized residuals between measured values ​​and rolling forecasts are calculated daily. An anomaly is identified when the residual exceeds a specific quantile of the historical residual distribution. Once an anomaly is triggered, the key factors driving the anomaly are located, and a source tracing report is generated. The anomaly score is calculated using the following formula: ; In the formula: This is the measured value on day t. The standard deviation of the predicted residuals over the past 90 days; The anomaly detection threshold is 2.5. An abnormal alarm is triggered when the value is greater than 2.5.

[0065] S104: Establish a collaborative repair combinatorial space for multi-objective optimization to obtain the optimal combinatorial solution set. Simulate the optimal combinatorial solution set to select the most stable solution. Based on the most stable solution, design a small-scale inoculation experiment and perform online data reinjection. The specific steps for establishing a collaborative repair combination space for multi-objective optimization are as follows: Based on key resistance factors, candidate plants and functional microorganisms with cadmium and mercury enrichment or transformation capabilities were screened to construct a synergistic remediation combination space. For each candidate scheme in the synergistic remediation combination space, a target-oriented robust fusion model was used to predict its cadmium removal rate, mercury removal rate and stress response under standard anchor conditions. A multi-objective optimization function is established, and a non-dominated sorting genetic algorithm is used to find the optimal combination solution set in the combinatorial space.

[0066] The multi-objective comprehensive scoring function is: ; In the formula: c represents the collaborative repair combination scheme, encoded as a three-dimensional discrete vector. The value of 0 to 1 represents the predicted cadmium removal rate of this combination, output by the target-oriented robust fusion model under standard anchor conditions. The predicted mercury removal rate of this combination is represented by a value between 0 and 1. The anchor robustness score for this combination is set, ranging from 0 to 1, and is defined as the lowest predicted removal rate of this combination under four typical anchor scenarios: typhoon, rainstorm, land reclamation, and water replenishment. The unit area capitalized cost of this portfolio (ten thousand yuan per hectare) is normalized to the range of 0 to 1 using the min-max method. An ecological risk score, ranging from 0 to 1, is assigned to this combination based on an assessment of the risk of biological invasion by introduced microorganisms and the allelopathic effects of plants, and is determined manually. , , , , These are weighting coefficients, with values ​​of 0.2, 0.2, 0.2, 0.15, and 0.15, respectively.

[0067] The population size of the non-dominated sorting genetic algorithm is set to 100, with 200 iterations, a crossover probability of 0.9, and a mutation probability of 0.1.

[0068] For each candidate combination in the optimal solution set, the simulation was run continuously for 12 months in a simulation environment. Anchor variables were injected using Monte Carlo sampling according to seasonal tidal patterns and extreme times, and a total of 1000 independent simulation trajectories were executed. The time series of removal efficiency and stress level response of each combination under different anchor scenarios were recorded. The expected return of each combination in the whole scenario and the return gap under extreme events were calculated, and the combination that maintained an overall removal rate of not less than 80% of the benchmark value in more than 90% of the anchor scenarios was selected and marked as the most stable solution.

[0069] The formula for calculating the expected return across all scenarios is: ; In the formula: For the number of Monte Carlo simulations, For the anchor variable scenario in the s-th sampling, , The figures represent the removal rates of cadmium and mercury, respectively.

[0070] Profit gap under extreme events: ; in, .

[0071] S105: Connect the target model to the digital twin platform for simulation and operation and maintenance closed loop.

[0072] Example 2: Please see Figure 2 As shown, this embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the rhizosphere response analysis method for cadmium and mercury polluted wetlands based on multi-omics data association provided by the above methods.

[0073] Since the electronic device described in this embodiment is the one used to implement the rhizosphere response analysis method for cadmium and mercury polluted wetlands based on multi-omics data association in the embodiments of this application, those skilled in the art can understand the specific implementation and various variations of the electronic device in this embodiment based on the rhizosphere response analysis method for cadmium and mercury polluted wetlands based on multi-omics data association described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any electronic device used by those skilled in the art to implement the rhizosphere response analysis method for cadmium and mercury polluted wetlands based on multi-omics data association in the embodiments of this application falls within the scope of protection of this application.

[0074] Example 3: Please see Figure 3 As shown, this embodiment discloses a computer-readable storage medium, including a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the computer program, it implements the rhizosphere response analysis method for cadmium and mercury polluted wetlands based on multi-omics data association provided by the above methods.

[0075] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters, weights, and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0076] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0077] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0078] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0079] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only one method, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0080] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0081] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0082] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0083] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association, characterized in that, The method includes: S101: Acquire multi-omics and rhizosphere physicochemical data, preprocess the multi-omics and rhizosphere physicochemical data, and form a data cube; S102: Based on the data cube, perform lexicalization and ontology normalization, perform sparse attention long sequence encoding to form a context matrix, use the context matrix to locate candidate evidence fragments, introduce red oxidation state-optimal transmission to screen candidate evidence fragments multiple times, and obtain a set of evidence chains. S103: Construct a set of mechanism isomorphic objects, conduct quality audits on the set of mechanism isomorphic objects to obtain adversarial examples, train multiple types of models to obtain the target model, use the target model to perform anomaly detection and traceable explanation, and obtain a comprehensive report; S104: Establish a collaborative repair combinatorial space for multi-objective optimization to obtain the optimal combinatorial solution set. Simulate the optimal combinatorial solution set to select the most stable solution. Based on the most stable solution, design a small-scale inoculation experiment and perform online data reinjection. S105: Connect the target model to the digital twin platform for simulation and operation and maintenance closed loop.

2. The method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association as described in claim 1, characterized in that, The sparse attention long sequence encoding process specifically involves: The lexicalized data cube is input into the sparse attention encoder, and modality embedding and hierarchical relative position encoding are injected during the encoding process to obtain the context matrix.

3. The method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association as described in claim 2, characterized in that, The method of locating candidate evidence fragments using a context matrix specifically includes: A multi-task scoring head is connected to the top layer of the sparse attention encoder to perform evidence strength regression, entity type classification and ontology edge prediction respectively, and obtain three types of scores. The three types of scores are weighted and fused to obtain the comprehensive evidence potential. Superimposed conditional random fields are used to smooth the boundary of the comprehensive evidence potential energy to obtain a score matrix and initial candidate evidence fragments. The initial candidate evidence fragments are hierarchically aggregated according to five granularities to obtain candidate evidence fragments; The generation of the candidate evidence fragments employs a two-layer strategy.

4. The method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association as described in claim 1, characterized in that, The introduction of red oxidation state-optimal transport for multiple screening of candidate evidence fragments includes: An initial candidate slice pool was obtained through preliminary screening using red oxidation state-optimal transport. For each candidate slice in the initial candidate slice pool, an interpretable causal sensitivity score was calculated. Obtain the original model score, weighted fuse the interpretable causal sensitivity score and the original model score to obtain the final score, sort them in descending order according to the final score to obtain the evidence chain set; Among them, key resistance elements were identified based on the evidence chain set.

5. The method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association as described in claim 4, characterized in that, The preliminary screening using red oxidation state-optimal transport specifically involves: For each candidate slice, functional lexical units and contextual lexical units are extracted, the red oxidation state-optimal transmission distance between functional lexical units and contextual lexical units is calculated, and the optimal transmission plan is approximated by the Sinkhorn iterative algorithm. Each candidate slice is decomposed into multiple subspaces, and an approximate nearest neighbor index is constructed in each subspace to obtain an initial candidate slice pool.

6. The method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association according to claim 4, characterized in that, The process of identifying key resistance elements based on the evidence chain set includes the following steps: Edge weights are calculated based on the evidence chain set. Resistant microbial groups, core metabolic pathways, key metabolites, and environmental driving factors are regarded as nodes, and a co-occurrence network is constructed by combining the edge weights. By simulating the removal of a certain group or pathway node in the co-occurrence network, the change in the predicted value of cadmium and mercury speciation is observed, the causal contribution is quantified, and key resistance factors are identified.

7. The method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association as described in claim 6, characterized in that, The various models include: a goal-oriented robust fusion model, an interpretable student model, and a biomarker stress level assessment model; The steps for training a target-oriented robust fusion model include: The adversarial examples are sorted and arranged according to three-dimensional indicators. The vaccination priority score is calculated by combining the three-dimensional indicators. The training process is divided into three stages to obtain a phased vaccination training sequence. The phased vaccination training sequence is used to train the target-oriented robust fusion model. During each upgrade phase, a priority experience replay mechanism is adopted.

8. The method for analyzing the rhizosphere response of cadmium and mercury polluted wetlands based on multi-omics data association according to claim 7, characterized in that, The specific steps for establishing a collaborative repair combination space for multi-objective optimization are as follows: Based on key resistance factors, candidate plants and functional microorganisms with cadmium and mercury enrichment or transformation capabilities were screened to construct a synergistic remediation combination space. For each candidate scheme in the synergistic remediation combination space, a target-oriented robust fusion model was used to predict its cadmium removal rate, mercury removal rate and stress response under standard anchor conditions. A multi-objective optimization function is established, and a non-dominated sorting genetic algorithm is used to find the optimal combination solution set in the combinatorial space.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the rhizosphere response analysis method for cadmium and mercury polluted wetlands based on multi-omics data association as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the method for analyzing the rhizosphere response of cadmium- and mercury-polluted wetlands based on multi-omics data association as described in any one of claims 1 to 8.

Images