Methods, apparatus and media for constructing datasets on nanomaterial-protein interactions
By processing literature corpora, we extracted and filled in structured feature data on the interaction between nanomaterials and proteins, which solved the defects in dataset construction, achieved a more complete and consistent dataset, and improved the predictive ability of the prediction model.
Patent Information
- Application Number
- CN202510976790.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-16
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2045-07-16
AI Technical Summary
Existing technologies face challenges in constructing datasets on the interaction between nanomaterials and proteins, including deficiencies in data integrity, difficulties in data extraction, data bias, and heterogeneity in data quality. These issues result in incomplete datasets and insufficient predictive power.
By processing the literature corpus, extracting the original structured feature data, performing semantic alignment and missing feature imputation, constructing positive and negative samples, and using local and global imputation methods, combined with the principle of multi-source reliability, a structured feature dataset is generated.
It improves the completeness and consistency of the dataset, enhances the predictive ability for unseen nanomaterials and protein types, reduces information distortion, and improves the modeling ability of the model.
Smart Images

Figure CN120877849B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and biomaterials, and in particular to a method, apparatus and medium for constructing a dataset of interactions between nanomaterials and proteins. Background Technology
[0002] Research on the interaction between nanomaterials and proteins has significant scientific and applied implications, specifically in revealing the biological behavioral mechanisms of nanomaterials, guiding their design and application, and evaluating their safety and efficacy. For example, in the field of nanomedicine, many proteins serve as disease biomarkers or drug targets. A deeper understanding of the interaction mechanisms between nanomaterials and proteins can help achieve the enrichment of specific protein biomarkers or precise drug targeting, thereby improving diagnostic sensitivity and treatment effectiveness.
[0003] In traditional methods, the study of nanomaterial-protein interactions primarily relies on experimental verification. For example, researchers typically introduce nanomaterials into biological fluids containing various proteins (such as serum, plasma, or cerebrospinal fluid), then separate the formed nanomaterial-protein complexes using methods such as centrifugation or magnetic adsorption. Proteomics techniques are then employed to detect the abundance and composition of proteins adsorbed on the nanomaterial surface, thereby characterizing the interaction between the nanomaterial and protein. However, this approach has significant limitations: firstly, the experimental process is time-consuming and costly, making it difficult to apply under large-scale material screening or high-throughput conditions; secondly, different types or combinations of nanomaterials and proteins often require individual experimental verification, lacking flexibility and scalability.
[0004] In response, some improved existing technologies attempt to use artificial intelligence techniques, including machine learning, to predict the interactions between nanomaterials and proteins, in an effort to reduce reliance on experiments. Those skilled in the art have attempted to use various artificial intelligence techniques for prediction.
[0005] However, regardless of the method, both rely on constructing corresponding datasets. In existing technologies, the construction of relevant datasets still depends on experimental methods. While some related technical fields construct databases through literature review, for example, a Chinese patent discloses a method for constructing a protein interaction database based on biomedical literature, belonging to the field of natural language processing. The protein interaction database construction method includes: biomedical named entity recognition based on word representation methods; multi-stage protein name standardization integrating multiple effective methods; a fusion of multiple methods for name resolution; protein interaction extraction based on word representation; and an online system and protein interaction database providing a graphical user interface based on a B / S model. Using this invention, biomedical named entity recognition, standardization, name resolution, and automatic extraction of protein interactions can be performed on biomedical text to be parsed. However, when applying this to the field of nanomaterials and protein interactions, the following problems arise:
[0006] 1) Data completeness deficiencies: In protein-protein interaction studies, complete amino acid sequence information can be obtained from specialized databases; however, in nanomaterial-protein interaction studies, the required data not only includes relatively easy-to-obtain protein information but also a variety of indispensable but often missing characterization data, including:
[0007] The physicochemical properties of nanomaterials include core composition, nanomaterial concentration, nanomaterial particle size, shape category, nanomaterial surface modification, nanoparticle surface modification category, hydrodynamics in water, surface charge in water, polydispersity index in water, dispersion medium, hydrodynamic diameter in the dispersion medium, surface charge in the dispersion medium, and polydispersity index in the dispersion medium.
[0008] Protein source parameters, including protein source fluid, protein source species, and protein source concentration;
[0009] Interaction environment parameters include temperature, buffer composition, interaction time between nanomaterials and proteins, interaction temperature between nanomaterials and proteins, flow state of the nanomaterials and protein interaction system, flow rate of the nanomaterials and protein interaction system, and state of the nanomaterials and protein interaction system.
[0010] The parameters for separating complexes include the type of separation method, centrifugation speed, centrifugation time, centrifugation temperature, and number of centrifugation repetitions;
[0011] Proteomics analysis methods, including the range of detectable proteins, analytical methods, and instruments.
[0012] 2) Difficulty in data extraction: These key characterization data are often presented in unstructured form in the literature, possibly existing in different chapters (such as materials and methods, results, figure descriptions, etc.), lacking a unified expression standard. This dispersed and unstructured characteristic makes:
[0013] Automated extraction techniques struggle to accurately locate and identify relevant information.
[0014] Manual extraction requires specialized knowledge and is inefficient.
[0015] Data standardization faces significant challenges.
[0016] 3) Data bias issue: Existing research mainly uses proteomics technology to analyze nanomaterial-protein interactions, which has an inherent detection bias:
[0017] It can only effectively capture interaction data with strong affinity.
[0018] The detection sensitivity for weak or transient interactions is insufficient.
[0019] The reported data generally only includes data on compatible samples, leaving a large amount of data on incompatible samples missing.
[0020] 4) The calculation methods for the interaction strength between nanomaterials and the interaction vary depending on the available data: different studies use different proteomics analysis methods and reporting indicators, and different literatures may use different indicators such as absolute protein content, number of peptides, spectrum count, peptide count or peptide signal intensity to characterize the interaction strength.
[0021] 5) Data quality heterogeneity: Due to the aforementioned data heterogeneity, complexity and diversity of interaction strength calculation methods, the constructed datasets have significant quality heterogeneity issues. Experimental data from different sources differ in terms of methodological rigor and data reporting completeness. Summary of the Invention
[0022] The purpose of this invention is to provide a method, apparatus, and medium for constructing a dataset of interactions between nanomaterials and proteins in order to overcome the deficiencies of the prior art.
[0023] The objective of this invention can be achieved through the following technical solutions:
[0024] A method for constructing a dataset on nanomaterial-protein interactions includes:
[0025] The literature corpus is processed to extract one or more first information groups, wherein the first information group contains protein identifiers, amino acid sequences obtained based on protein identifiers, raw data on the interaction between nanomaterials and proteins, and raw structured feature data. The raw structured feature data includes at least nanomaterial identifiers and biological fluid identifiers, and the proteins corresponding to the protein identifiers are detected in the biological fluids corresponding to the biological fluid identifiers.
[0026] Semantic alignment is performed on all the original structured feature data in the first information group to obtain unfilled structured feature data, wherein the missing feature data is represented by the first identifier;
[0027] Fill in the missing features in all the structured feature data in the first information group to obtain the filled structured feature data;
[0028] The interaction strength between the nanomaterials and each detected protein in the first information group is calculated respectively, and the first type of sample is generated based on the calculated interaction strength, structured feature data, nanomaterial identifiers and protein identifiers.
[0029] Based on all the first information groups from the same literature corpus, the second type of sample is obtained by locally filling the parts with the same biological fluid identifier;
[0030] Based on all the first information groups from different literature corpora, and based on the parts with the same biological fluid identifiers, global filling was performed to obtain the third type of sample;
[0031] The first type of samples are labeled according to the strength of their interaction, and the second and third types of samples are used as negative samples to construct the dataset.
[0032] The strength of the interaction between the nanomaterial and the protein is represented by the relative abundance of proteins adsorbed on the surface of the nanomaterial.
[0033] The structured feature data includes:
[0034] The physicochemical properties of nanomaterials include nanomaterial identification, core composition, nanomaterial concentration, nanomaterial size parameters, shape category, nanomaterial surface modification, nanoparticle surface modification category, hydrodynamics in water, surface charge in water, polydispersity index in water, dispersion medium, hydrodynamic diameter in the dispersion medium, surface charge in the dispersion medium, and polydispersity index in the dispersion medium.
[0035] Protein source parameters, including biological fluid markers, protein source species, and protein source concentration;
[0036] Interaction environment parameters include temperature, buffer composition, interaction time between nanomaterials and proteins, interaction temperature between nanomaterials and proteins, flow state of the nanomaterials and protein interaction system, flow rate of the nanomaterials and protein interaction system, and state of the nanomaterials and protein interaction system.
[0037] The parameters for separating complexes include the type of separation method, centrifugation speed, centrifugation time, centrifugation temperature, and number of centrifugation repetitions;
[0038] Proteomics analysis methods, including the range of detectable proteins in proteomics, analytical methods, and instruments and equipment;
[0039] The protein identifier includes at least one of the following: protein name, gene name, and protein identifier.
[0040] The second type of sample is obtained by locally filling in all first information groups from the same corpus, based on the parts with the same biological fluid identifiers, including:
[0041] A first set is formed by extracting all first information groups from the same literature corpus and having the same biological fluid identifier, and a second set is formed by constructing all first information groups with the same structured feature data. The second sets are then paired with the first set based on the principle of the same biological fluid identifier.
[0042] The union of protein identifiers in all first information groups in the first set is used as the first protein set;
[0043] The union of protein identifiers in all first information groups in each second set is used as the second protein set corresponding to that second set;
[0044] For any first set, iterate through all paired second sets, match the protein identifiers in the second protein set corresponding to the second set with the protein identifiers in the first protein set corresponding to the first set, and fill the second protein set with protein identifiers that exist in the first protein set but not in the second protein set;
[0045] Based on the second set of proteins after filling, each filled protein is identified, combined with the structured feature data of the corresponding second set, and the interaction strength is padded with zeros to construct the second type of sample.
[0046] The process of globally filling in the third type of samples based on the common biological fluid identifiers of all first information groups from different literature corpora to obtain the third type of samples includes:
[0047] A third set is formed by extracting all first information groups from all literature corpora and having the same biological fluid identifier, and a fourth set is formed by constructing all first information groups with the same structured feature data.
[0048] Based on the principle of multi-source credibility, the third set is screened to obtain the fifth set, and each fourth set is paired with the fifth set based on the principle of the same biological fluid identifier.
[0049] The union of all protein identifiers in the first information group in the fifth set is used as the third protein set;
[0050] The union of protein identifiers in all first information groups in each fourth set is calculated as the fourth protein set corresponding to that fourth set;
[0051] For any fifth set, iterate through all the fourth sets that are paired with it, match the protein identifiers in the fourth protein set corresponding to the fourth set with the protein identifiers in the third protein set corresponding to the fifth set, and fill the fourth protein set with the protein identifiers that exist in the third protein set but not in the fourth protein set.
[0052] Based on the filled fourth protein set, each filled protein is labeled, combined with the structured feature data of the corresponding fourth set, and the interaction strength is padded with zeros to construct the third type of sample.
[0053] The multi-source trust principle is as follows:
[0054] If any protein appears in the same biological fluid in at least three independently sourced literature corpora, then it contains a first information set corresponding to the protein identifier.
[0055] During the process of imputing missing features in all structured feature data in the first information group:
[0056] The surface modification of the nanomaterials is filled with default values; missing data is considered as no surface modification.
[0057] The categories of nanoparticle surface modifications were filled using a large inference model.
[0058] The size parameters of the nanomaterials were filled using a weighted average.
[0059] The hydrodynamic diameter, surface charge, and polydispersity index of nanomaterials in water, as well as their hydrodynamic diameter, surface charge, and polydispersity index in the dispersion medium, were filled using a large case-based model.
[0060] The dispersion medium is filled with the default value; missing data is considered as water as the dispersion medium.
[0061] The concentration of the nanomaterials was filled using a mean value.
[0062] The concentrations of the protein sources were filled with mean values.
[0063] The interaction time and temperature between the nanomaterials and proteins are filled with the mode.
[0064] The flow state of the nanomaterial-protein interaction system is filled with default values; missing data is considered as static.
[0065] The flow rate of the nanomaterial-protein interaction system is filled with either a default value or a mean value. When the flow state of the nanomaterial-protein interaction system is stationary, missing data is considered as 0. When the flow state of the nanomaterial-protein interaction system is flowing, missing data is filled with the mean value.
[0066] The states of the nanomaterial-protein interaction system are filled with default values; missing data is considered as in vitro.
[0067] The separation method type is filled with default values; missing data is considered as using the centrifugation method.
[0068] The centrifugation rate is filled with the mean value. If the separation method type is not centrifugation, it is filled with a second identifier to indicate invalidity.
[0069] The centrifugation time, centrifugation temperature, and number of centrifugation repetitions are all filled with the mode. If the separation method type is not centrifugation, it is filled with a second identifier to indicate invalidity.
[0070] Samples are constructed based on unfilled structured feature data and filled structured feature data respectively. An unfilled dataset is constructed based on the first, second and third class samples obtained from the unfilled structured feature data, and a filled dataset is constructed based on the first, second and third class samples obtained from the filled structured feature data.
[0071] Each sample is assigned a weight, which is determined based on the accuracy and directness of the data.
[0072] A device for constructing a dataset of nanomaterial-protein interactions includes a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the method described above.
[0073] A storage medium having a program stored thereon, which, when executed, implements the method described above.
[0074] Compared with the prior art, the present invention has the following beneficial effects:
[0075] 1. Based on extracting information groups from literature corpora, in addition to constructing positive samples by imputing missing features in structured feature data, negative samples are also constructed through local imputation and global imputation. This approach allows for the extraction of more sample sets from limited literature corpora. As the amount of literature corpora increases and the number of overlapping biological fluid markers grows, the number of negative samples that can be amplified will increase explosively. Furthermore, imputing missing features in structured feature data can provide more complete structured feature data, which is more conducive to aligning effective information between different templates. This improves the predictive ability of the prediction model for unseen protein and nanomaterial types during training.
[0076] 2. The structured feature data covers the physicochemical properties of nanomaterials, protein source parameters, interaction environment parameters, complex separation parameters, and proteomics analysis methods. Different filling methods are designed for different features. Compared with the unfilled method, it is beneficial to align the effective information between different templates, thereby improving the predictive ability of the prediction model for unseen protein and nanomaterial types during training. Compared with the fixed filling method, it can more flexibly adapt to data features from different sources and types based on domain knowledge, reduce information distortion or noise introduction caused by a uniform filling strategy, and thus improve the model's ability to model heterogeneous data.
[0077] 3. For all first information groups from the same literature corpus, the first negative sample is obtained by calculating the union of the proteins detected in each type of biological fluid, thereby providing a basis for supplementing incompatible samples in the same literature corpus.
[0078] 4. For all the first information groups from different literature corpora, on the one hand, they are screened and filtered through a multi-source reliable method, and on the other hand, the union of the proteins detected in each type of biological fluid is calculated and the second negative sample is obtained by reverse expansion, so as to supplement the incompatibility samples with evidence in different literature corpora. Attached Figure Description
[0079] Figure 1 This is a schematic diagram of the main steps of the method of the present invention;
[0080] Figure 2 This is a schematic diagram illustrating the information extraction and processing process;
[0081] Figure 3 This is a schematic diagram of local filling, where (a) is the set partitioning process, (b) is the matching process, and (c) is the filling process;
[0082] Figure 4The diagram illustrates the global filling process, where (a) represents the set partitioning process, (b) represents the matching process, and (c) represents the filling process.
[0083] Figure 5 This is a schematic diagram of the original data;
[0084] Figure 6 This is a schematic diagram of the data distribution after transformation.
[0085] Figure 7 A diagram illustrating a populated dataset and an unpopulated dataset;
[0086] Figure 8 The dataset is an overview of the constructed dataset, in which (a) are the physicochemical properties of nanomaterials, (b) are the parameters of protein source and interaction environment, (c) are the parameters of complex separation, (d) are the proteomics analysis methods, and (e) are the collected proteins and their corresponding sample counts, where Accession is the protein identifier and Accessionindex is the protein identifier sequence number.
[0087] Figure 9 A schematic diagram of the statistical results for filling missing data, where (a) shows the number of missing values and (b) shows the proportion of missing values for different features.
[0088] Figure 10 This is a performance illustration of a model trained on a constructed dataset. Detailed Implementation
[0089] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0090] A method for constructing a dataset on nanomaterial-protein interactions, such as Figure 1 As shown, it includes:
[0091] Step S1: Process the literature corpus to extract one or more first information groups, including protein identifiers, amino acid sequences obtained based on protein identifiers, raw data on the interaction between nanomaterials and proteins, and raw structured feature data. The raw structured feature data includes at least nanomaterial identifiers and biological fluid identifiers. The proteins corresponding to the protein identifiers are detected in the biological fluids corresponding to the biological fluid identifiers. The extraction and processing procedures are as follows: Figure 2 As shown.
[0092] Nanomaterial identifiers are used to represent the type of nanomaterials, biological fluid identifiers are used to represent the type of biological fluids, and similarly, protein identifiers are used to represent the type of proteins.
[0093] In this embodiment, protein identification includes at least one of protein name, gene name, and protein identifier;
[0094] Structured feature data includes:
[0095] Structured feature data includes:
[0096] The physicochemical properties of nanomaterials include nanomaterial identification, core composition, nanomaterial concentration, nanomaterial size parameters, shape category, nanomaterial surface modification, nanoparticle surface modification category, hydrodynamics in water, surface charge in water, polydispersity index in water, dispersion medium, hydrodynamic diameter in the dispersion medium, surface charge in the dispersion medium, and polydispersity index in the dispersion medium.
[0097] Protein source parameters, including biological fluid markers, protein source species, and protein source concentration;
[0098] Interaction environment parameters include temperature, buffer composition, interaction time between nanomaterials and proteins, interaction temperature between nanomaterials and proteins, flow state of the nanomaterials and protein interaction system, flow rate of the nanomaterials and protein interaction system, and state of the nanomaterials and protein interaction system.
[0099] The parameters for separating complexes include the type of separation method, centrifugation speed, centrifugation time, centrifugation temperature, and number of centrifugation repetitions;
[0100] Proteomics analysis methods, including the range of detectable proteins in proteomics, analytical methods, and instruments and equipment;
[0101] In this embodiment, the above information is extracted from the literature corpus database by a large model driven by a preset prompt word template. The prompt word template includes at least two elements from background knowledge description, task definition, role assignment, parameter description, and structured output format requirements. In this embodiment, the prompt word template includes all elements from the background knowledge description, task definition, role assignment, parameter description, and structured output format requirements. The prompt word template is as follows:
[0102] ###Task###
[0103] You are an expert researcher tasked with extracting precise experimental data from a scientific paper. For each specified parameter, follow these steps:
[0104] 1.Understand the Context:Carefully analyze the parameter and itsdescription to ensure you extract the correct value from the text.
[0105] 2.Locate the Evidence:Identify the specific text or table in the PDFwhere the parameter's value is stated.Cross-check the context to confirmaccuracy.
[0106] 3.Report with Citation:Output the parameter,its corresponding value,and a verbatim or summarized reference to the part of the paper where it ismentioned.Ensure the evidence is accurate and explicitly stated in the PDF.Ifa parameter or value is not provided,clearly state it is"not available"(N / A).
[0107] ###Protein Corona Research Background###
[0108] When nanomaterials(NMs)enter biological systems,they interact withbiomolecules,particularly proteins,forming a"protein corona"on theirsurface.This corona significantly alters the physical and chemical propertiesof the nanomaterials,influencing their biological interactions,cellularuptake,toxicity,biodistribution,and overall functionality.The proteincomposition of the protein corona is affected by many parameters,such asNanomaterial Properties,Dispersion Medium Properties,Incubation Conditions,and Centrifugation Parameters.
[0109] ###Rules###
[0110] 1.Do not infer values or fabricate any information.
[0111] 2.If a value is presented in multiple formats(e.g.,a range or anaverage),provide the exact format used in the paper.
[0112] 3.For unclear cases,MUST PROVIDE THE RELEVENT EVIDENCE FROM THE TEXT!
[0113] 4.If a parameter has multiple values,list them separated bysemicolons.
[0114] ###Output Format###
[0115] #Do not output redundant content#
[0116] 1.**Parameter Name**:[Parameter Name]
[0117] -**Value**:[Extracted Value]
[0118] -**Evidence**:"[Excerpt or summary from the article text,includingtable / figure references ifapplicable]"
[0119] 2.**Parameter Name**:[Parameter Name]
[0120] -**Value**:[Extracted Value]
[0121] -**Evidence**:"[Excerpt or summary from the article text,includingtable / figure references ifapplicable]" ...
[0123] N.**Parameter Name**:[Parameter Name]
[0124] -**Value**:[Extracted Value or N / A]
[0125] -**Evidence**:"[Excerpt or summary from the article text,includingtable / figure references ifapplicable]"
[0126] #Nanomaterial Properties#
[0127] 1.Core:The main composition of the nanomaterial(e.g.,gold,silica,etc.).
[0128] 2.Core type:The specific type or chemical structure of the core(classify the core into seven classes:metal-based,metal oxide-based,polymer-based,lipid-based,carbon-based,core-shell,and other).
[0129] 3.Surface modification:Modifications applied to the surface of thenanomaterial(e.g.,functionalchemical groups).
[0130] 4.Modification type:The specific type of surface modification orfunctionalization(classify thesurface modification into three classes:neutral,cationic,and anionic).
[0131] 5.Shape:The geometric morphology of the nanomaterial(e.g.,spherical,rod-like,etc.).
[0132] 6.Primary size:The dry size of the nanomaterial core measured usingTransmission ElectronMicroscopy(TEM)or Scanning Electron Microscope(SEM),often representing the dry size.
[0133] 7.DLS size in water(nm):The hydrodynamic size of the nanomaterial inwater,determined usingDynamic Light Scattering(DLS).
[0134] 8.Zeta potential in water(mV):The surface charge of the nanomaterialin water,indicating stabilityand interactions.
[0135] 9.PdI in water:The polydispersity index in water,reflecting the sizedistribution uniformity(lowervalues indicate more uniform particles).
[0136] 10.DLS size in dispersion medium(nm):The hydrodynamic size of thenanomaterial in a specificdispersion medium.
[0137] 11.Zeta potential in dispersion medium(mV):The surface charge in thedispersion medium,reflecting particle stability under those conditions.
[0138] 12.PdI in dispersion medium:The polydispersity index in thedispersion medium,indicating sizeuniformity.
[0139] 13.Dispersion medium:The liquid medium in which the nanomaterials aredispersed(e.g.,water,buffer,serum).
[0140] 14.NM concentration:The concentration of the nanomaterial in thedispersion medium.
[0141] #Incubation Conditions#
[0142] 15.Incubation protein source:The origin of the protein used forincubation(e.g.,human plasma,human serum,fetal bovine serum,or mouse plasma).
[0143] 16.Protein source organism:The organism from which the protein isderived(e.g.,human,bovine,mouse).
[0144] 17.Protein source concentration:The concentration of plasma or serumin the incubation medium.
[0145] 18.Incubation culture:Specific culture conditions during incubation(e.g.,water,buffer,orDMEM).
[0146] 19.Incubation time(h):The duration of the incubation period.
[0147] 20.Incubation temperature(℃):The temperature at which the incubationoccurs,typically reflecting physiological or experimental conditions.
[0148] 21.Incubation flow condition:Whether the incubation is static orunder flow conditions(static or flow).
[0149] 22.Incubation flow speed:The speed of flow during incubation,relevantfor dynamic conditions.
[0150] 23.Incubation setting:Specifies whether the incubation occurs in invivo(within a living organism)or in vitro(in a controlled laboratoryenvironment)culture conditions.
[0151] #Separation Parameters#
[0152] 24.Separation method:The method used to separate protein corona(e.g.,centrifugation,Size exclusion chromatography or magnetic separation).
[0153] 25.Centrifugation speed:The relative centrifugal force applied duringcentrifugation,expressed in multiples of gravity or revolutions per minute.
[0154] 26.Centrifugation time(min):The duration of the centrifugation step.
[0155] 27.Centrifugation temperature(℃):The temperature duringcentrifugation,affecting nanomaterial stability and protein binding.
[0156] 28.Centrifugation repetitions:The number of centrifugation cyclesapplied to separate bound and unbound proteins.
[0157] ###Paper Text###
[0158] *Copy the text here.*
[0159] Step S2: Semantically align the original structured feature data in all the first information groups to obtain unfilled structured feature data, wherein the missing feature data is represented by the first identifier;
[0160] Specifically, since the wording may be inconsistent between different literature corpora, semantic alignment is required. Generally, the semantic alignment process can be achieved through terminology standardization mapping and expression standardization assisted by large models.
[0161] Step S3: Fill in the missing features in all the structured feature data in the first information group to obtain the filled structured feature data;
[0162] During the process of imputing missing features in all structured feature data in the first information group:
[0163] The surface modification of nanomaterials uses default values for filling; missing data is considered as no surface modification, and the specific fill value is None.
[0164] The classification of nanoparticle surface modifications is filled using a large inference model. Since this is a classification problem based on chemical knowledge, it requires a certain level of background knowledge and reasoning ability. Furthermore, surface modifications are highly diverse (hundreds of types), making manual filling time-consuming. Therefore, we first construct a vector database of domain knowledge, then perform a retrieval, providing the large inference model with context and background knowledge before allowing it to reason, which improves efficiency. The inference hints are as follows:
[0165] [Role]
[0166] You are a chemistry expert specializing in nanoparticle surfacemodifications.Your task is to provide the most commonly used chemical name orabbreviation for the given TEXT.
[0167] [TASK]
[0168] Based on the provided TEXT and relevant CONTEXT,return the widelyaccepted chemical name or abbreviation of the chemical of nanoparticlesurface modification.
[0169] [TEXT]
[0170] {term}
[0171] [OUTPUT FORMAT]
[0172] -Return only the chemical name or abbreviation.
[0173] -If a common abbreviation is available for the chemical,only theabbreviation for this chemical is returned.
[0174] -If there is a single chemical,return it as is.
[0175] -If multiple chemicals exist,return them separated by'@'(e.g.,A@B).
[0176] -Do not include any additional explanations or formatting.
[0177] [CONTEXT]
[0178] The context is retrieved via vector retrieval and may not berelevant.
[0179] {context}
[0180] The size parameters of nanomaterials are filled with weighted averages. Generally, for spherical nanomaterials, particle size is used as the identifier. Since it is related to five parameters: the core composition of the nanomaterial, the type of core chemical composition, the surface modification of the nanomaterial, the type of surface modification of the nanoparticles, and the shape of the nanomaterial, the available data are grouped and weighted averaged. The grouping is based on the above five parameters. The more detailed the grouping, the greater the weight.
[0181] The data are grouped from finest to coarsest as follows: The first group uses five features—core composition, core chemical composition category, surface modification, surface modification category, and shape—for the most detailed grouping, with a weight of 5. The second group is based on three features: core composition, core chemical composition category, and shape, with a weight of 3. The third group uses only two features: core chemical composition category and shape, with a weight of 2. The fourth group uses only one feature—core chemical composition category—for the coarsest grouping, with a weight of 1. The finer the grouping, the more complete and representative the feature information, and therefore the higher its weight.
[0182] The formula for weighted average is:
[0183]
[0184] in, ω represents the weighted average after filling. i μ represents the weight of the i-th level group. i Let V represent the mean of the i-th level group, and let V represent all groups with effective weights ω. i and effective mean μ i A set of grouped indexes.
[0185] The hydrodynamic diameter, surface charge, and polydispersity index of nanomaterials in water, as well as their hydrodynamic diameter, surface charge, and polydispersity index in the dispersion medium, are filled using a case-based large model. Here, considering the limited existing data and the influence of six parameters—the core composition of the nanomaterial, the category of the core chemical composition, the surface modification of the nanomaterial, the category of the surface modification of the nanoparticles, the shape of the nanomaterial, and the particle size of the nanomaterial—the data that are not missing are used as cases (the above six parameters and the hydrodynamic diameter). At the same time, the characteristic data of the six parameters for the missing data are given, so that the large model can estimate the missing hydrodynamic diameter based on the existing six parameters and the case of the hydrodynamic diameter. The same logic applies to surface charge and polydispersity index diameter in water. Furthermore, existing data on hydrodynamics in dispersion media is limited and influenced by six parameters: the core composition of the nanomaterial, the type of core chemical composition, surface modification of the nanomaterial, type of surface modification of the nanoparticles, shape of the nanomaterial, and particle size of the nanomaterial, as well as the dispersion medium. Using the available data as examples (the six parameters mentioned in this paragraph, the dispersion medium, and the hydrodynamic diameter in the dispersion medium), and providing six characteristic data points and the dispersion medium for missing data, the large model can estimate the missing hydrodynamic diameter based on the existing six parameters, the dispersion medium, and the corresponding examples of hydrodynamic diameter in the dispersion medium. Similarly, the same logic applies to surface charge and polydispersity index in dispersion media. Taking the large model filling of hydrodynamic diameter as an example, the prompts are as follows:
[0186] [Role]
[0187] You are a chemistry expert tasked with estimating hydrodynamicdiameters through systematic analysis. Validate consistency between theory and examples, and resolve conflicts by averaging theoretical reasoning and examples.
[0188] [BACKGROUND]
[0189] Estimate the hydrodynamic diameter through the following sequential adjustments:
[0190] 1. **TEM / SEM Size (nm)**
[0191] -Start with the TEM / SEM measured size as the baseline.
[0192] -This represents the dry nanoparticle size before hydration effects.
[0193] -The hydrodynamic diameter is typically larger than the TEM / SEMparticle size because itincludes the hydration layer,whereas TEM / SEM measuresthe dry particle's physical size.
[0194] -For Rod-like,Sheet-like,and Plate-like materials,the baseline sizeshould be the longestdimension among length,diameter,thickness,or width.
[0195] 2.**Core Contribution**
[0196] -Adjust based on core material properties:+
[0197] *Metal-based(e.g.,Au,Ag,Pt,Fe,Cu,etc.)
[0198] Increase by 10-20%due to high density and strong hydrationattraction.
[0199] Surface coatings(e.g.,PEGylation)may further increase hydrationeffects.
[0200] *Metal oxide-based(e.g.,Fe3O4,TiO2,ZnO,CeO2,etc.)
[0201] Increase by 5-15%depending on the oxide’s hydration affinity.
[0202] Hydrophilic oxides(e.g.,Fe3O4,TiO2)tend to have thicker hydrationlayers.
[0203] Hydrophobic oxides(e.g.,CeO2)show minimal hydration effects.
[0204] *Carbon-based(Graphene,CNTs,Carbon Quantum Dots,etc.)
[0205] Graphene&Non-functionalized CNTs→Increase by 5-10%,as they haveweak hydrationlayers due to hydrophobic surfaces.
[0206] Functionalized CNTs(-OH,-COOH,etc.)→Increase by 15-25%,ashydrophilic groupsattract more water molecules.
[0207] Carbon Quantum Dots(CQDs)→Increase by 10-20%,as their small sizeand surfacechemistry enhance hydration effects.
[0208] *Lipid-based(Liposomes,lipid nanoparticles,etc.)
[0209] Unilamellar liposomes→Increase by 5-15%,as they slightly swell inaqueous environments.
[0210] Multilamellar liposomes→Increase by 10-30%,due to additionalhydration layers betweenlipid bilayers.
[0211] Surface charge effect:
[0212] Charged liposomes(cationic / anionic)→Higher hydration,contributing toa larger increase.
[0213] Neutral liposomes(e.g.,PC,PE-based)→Minimal additional hydration.
[0214] *Polymer-based(PLGA,PEGylated nanoparticles,dendrimers,etc.)
[0215] No significant increase unless highly hydrophilic.
[0216] Hydration mainly depends on polymer composition and surfacemodifications.
[0217] *Core-shell(Hybrid structures,e.g.,Metal@Polymer,SiO2@Au,etc.)
[0218] Adjust based on the shell properties:
[0219] Hydrophilic shells(e.g.,PEG,SiO2)→Increase by 10-20%.
[0220] Hydrophobic shells(e.g.,PDMS,PMMA)→Minimal increase.
[0221] *Other(Zeolites,biomolecules,etc.)
[0222] Zeolites&porous structures→Increase by 10-25%,depending on wateradsorption capacity.
[0223] Protein-based nanoparticles→Increase by 15-30%,due to extensivehydration ofbiomolecular surfaces.
[0224] 3.**Surface Modification Effect**
[0225] The increase in hydrodynamic diameter can generally be categorized asfollows:
[0226] *Small Molecule or Monolayer Modifications(Increase:<2 nm or<10%)
[0227] Functional groups such as citrate,carboxyl(COOH),amine(NH2),thiol(SH),phosphate,andsulfonate introduce minimal steric hindrance,typicallyincreasing the hydrodynamic diameter by 0.5–2nm or up to 10%relative to thecore particle size.
[0228] *Polymer Coatings(Increase:2–10 nm or 10–50%)
[0229] Surface coatings with polymers such as polyethylene glycol(PEG),polyvinylpyrrolidone(PVP),dextran,or chitosan result in a moderate increasein diameter,typically ranging from 2 to 10 nmor up to 50%of the core size,depending on polymer molecular weight and grafting density.
[0230] *Biomolecular or Layered Coatings(Increase:10–50 nm or 50–200%)
[0231] Modifications involving proteins(e.g.,bovine serum albumin,transferrin),lipid bilayers(e.g.,liposomes,phospholipids),or multilayerpolyelectrolyte coatings can significantly increase thehydrodynamic diameterby 10–50 nm or 50–200%due to their extended conformations andhydrationeffects.
[0232] *Cellular Membrane and Large Biomolecule Functionalization(Increase:>50 nm or>200%)
[0233] Modifications with whole cell membranes(e.g.,red blood cellmembranes,leukocytemembranes)or large protein complexes can drasticallyincrease the hydrodynamic diameter by morethan 50 nm,often exceeding 200%ofthe original nanoparticle size due to the extensive biologicalinterface andhydration layers.
[0234] 4.**Shape Adjustment**
[0235] The shape of nanoparticles significantly influences theirhydrodynamic diameter,withdifferent shapes leading to varying increases fromthe dry diameter.General trends include:
[0236] *Spherical:Typically,spherical nanoparticles show the smallesthydrodynamic diameterincrease.The hydrodynamic diameter could increase by 20-40%compared to the dry size.
[0237] *Rod-like:Rod-shaped nanoparticles tend to have a larger increase inhydrodynamic diameter,generally around 40-60%more than spherical ones.
[0238] *Sheet-like:Sheet-like nanoparticles show a significant increase indiameter,typically 50-80%larger than their dry size.
[0239] *Complex:For complex or branched nanoparticles,the increase can rangefrom 50-100%compared to spherical nanoparticles,depending on the complexityof the structure.
[0240] *Plate-like:Plate-like nanoparticles often show an increase of about40-60%in hydrodynamicdiameter.
[0241] *Polyhedral:Polyhedral nanoparticles usually exhibit an increase ofabout 30-50%inhydrodynamic diameter.
[0242] 5.**Aggregation Impact**
[0243] -If aggregation occurs,apply further adjustments:
[0244] -Small clusters(~10 particles)→Increase hydrodynamic diameter by**several-fold**.
[0245] -Strong hydrophilic modifications**reduce**aggregation,leading tominimal change.
[0246] -Hydrophobic or unmodified surfaces**promote**aggregation,significantly increasinghydrodynamic size.
[0247] [TASK]
[0248] Estimate the hydrodynamic diameter based on the provided nanoparticleparameters andFEW-SHOT EXAMPLES(in CSV format).
[0249] [INPUT]
[0250] TEM / SEM Size(nm):{primary_size}
[0251] Core:{core}
[0252] Core Type:{core_type}
[0253] Surface Modification:{surface_modification}
[0254] Modification Type:{modification_type}
[0255] Shape:{shape}
[0256] [OUTPUT FORMAT]
[0257] Return only a numerical value.
[0258] Do not include units.
[0259] If a range is estimated,return the midpoint.
[0260] Do not include explanations or formatting.
[0261] [FEW-SHOT EXAMPLES(CSV format)]
[0262] {csv_content}"""
[0263] Among them, the values of the six parameters of nanomaterial core composition, core chemical composition category, nanomaterial surface modification, nanoparticle surface modification category, nanomaterial shape and nanomaterial particle size are replaced with the corresponding identifiers, such as nanomaterial particle size is replaced with {primary_size}, and non-missing data cases are replaced with csv format instead of {csv_content}.
[0264] The dispersion medium is filled with the default value; missing data is considered as water as the dispersion medium.
[0265] The concentration of the nanomaterials is filled with the mean value. Specifically, the mean value is calculated based on the specific nanomaterial and the type of nanomaterial. The concentration value range is relatively large, for example, it may be 1 mg / L to 10000 mg / L. There are many value choices to calculate the mean value.
[0266] The concentration of the protein source is filled with the mean. Specifically, the mean is calculated based on the specific protein source. Because the values are very diverse, it is necessary to combine the protein concentration and the nanomaterial concentration to calculate the mean.
[0267] The interaction time and temperature between nanomaterials and proteins were filled with the mode. For the interaction time between nanomaterials and proteins, since most studies use fixed values, such as 15 minutes, 30 minutes, 1 hour, 12 hours, etc., the mode was obtained based on the protein source. For the interaction temperature between nanomaterials and proteins, the choice of values is relatively limited, such as 4 degrees Celsius, 25 degrees Celsius or 37 degrees Celsius. The mode filling can be more realistic and effective.
[0268] The flow state of the nanomaterial and protein interaction system is filled with default values, and missing data is considered as static.
[0269] The flow rate of the nanomaterial and protein interaction system is filled with default values or average values. When the flow state of the nanomaterial and protein interaction system is stationary, the missing data is considered as 0. When the flow state of the nanomaterial and protein interaction system is flowing, the missing data is filled with the average value. Specifically, because the values are very diverse, it is necessary to consider the protein source to calculate the average value. Therefore, the average value corresponding to the protein source is used for filling.
[0270] The states of the nanomaterial and protein interaction system are filled with default values. Missing data are considered as in vitro data because the values are very diverse and the average value needs to be calculated considering the protein source.
[0271] The separation method type is filled with default values; missing data is considered as using the centrifugation method.
[0272] Centrifugation speeds are filled with the mean value; if the separation method type is not centrifugation, they are filled with a second identifier to indicate invalidity.
[0273] Centrifugation time, centrifugation temperature, and number of centrifugation repetitions are all filled with the mode. If the separation method type is not centrifugation, it is filled with a second identifier to indicate invalidity.
[0274] In addition, other features are as follows:
[0275] The core components of nanomaterials do not require filling; all literature corpora record the core chemical composition.
[0276] The categories of core chemical components are filled in manually. Specifically, values are assigned based on the core chemical components, according to a fixed set of categories.
[0277] The shapes of nanomaterials are based on artificially defined, fixed categories, and are filled according to the material description.
[0278] Protein sources, such as biological fluids from humans or laboratory animals, including serum, plasma, and cerebrospinal fluid, do not require filling, as all literature corpora will record them. Protein source species are classified based on artificially established mapping relationships. Nanomaterials and protein incubation environments also do not require filling.
[0279] Finally, the depth of proteomics analysis was grouped according to the number of proteins detected: 0-500, 501-1000, 1001-2000, 2001-3000, and >3000.
[0280] Step S4: Calculate the interaction strength between the nanomaterials and each detected protein in the first information group, and generate the first type of sample based on the calculated interaction strength, structured feature data, nanomaterial identifiers and protein identifiers;
[0281] In this embodiment, the interaction strength between nanomaterials and proteins is represented by the relative abundance of proteins adsorbed on the surface of nanomaterials.
[0282] Based on the data types reported in the literature corpus, and correspondingly using methods for calculating interaction strength, the methods for calculating interaction strength include at least one of the following:
[0283] Absolute protein content determination: Using standard peptides or proteins of known concentrations as internal standards, the absolute protein content (e.g., fmol / μg) is calculated by comparing the mass spectrometry response values of the target protein and the standard, providing more accurate quantitative results.
[0284] Spectral counting method: count the number of MS / MS spectra detected for each protein, and evaluate its relative abundance using a statistical model.
[0285] Peptide counting normalization: This method estimates the relative abundance of proteins by counting the number of unique peptides identified for each protein and normalizing the total number of peptides or the total number of proteins in the sample. It is simple, intuitive, and suitable for semi-quantitative analysis based on mass spectrometry data.
[0286] The iBAQ algorithm (intensity-based absolute quantification) is based on the summation of peptide signal intensities in mass spectrometry. It sums the integral intensities of all peptide ion currents for each protein and divides the sum by the theoretically detectable number of peptides for that protein, thus enabling quantitative comparison between proteins.
[0287] Label-Free Quantification (LFQ): This method compares the mass spectrometry signal intensities of the same peptide in different samples and uses a normalization strategy to achieve relative quantification of proteins under different conditions.
[0288] Step S5: Based on all the first information groups from the same literature corpus, and based on the parts with the same biological fluid identifier, perform local filling to obtain the second type of sample.
[0289] Considering that the experimental group will only list the detected proteins in the literature, the above method will result in too few negative samples in the first type of sample. Therefore, this application also expands the negative samples by local filling. The core idea is that the proteins contained in the same biological fluid can be considered to be similar or even the same. Therefore, if certain proteins are detected in the same biological fluid in other experimental groups, it is considered that the protein exists in the biological fluid. Conversely, the nanomaterials and other structured feature data corresponding to the experimental group in which the protein was not detected in the biological fluid can be considered as a negative sample.
[0290] Based on this, such as Figure 3 As shown, step S5 specifically includes:
[0291] Obtain from the same corpus of literature, and such Figure 3As shown in (a), all first information groups with the same biological fluid identifier constitute a first set, and all first information groups with the same structured feature data are respectively constructed into a second set. Based on the principle of the same biological fluid identifier, each second set is paired with the first set. Specifically, each first set is generally paired with multiple second sets. The biological fluid identifier of the first information group in the first set is the same as that of the first information group in any of the second sets it is paired with. Furthermore, the union of all second sets is in fact the matching first set.
[0292] The union of protein identifiers in all first information groups in the first set is used as the first protein set;
[0293] The union of protein identifiers in all first information groups in each second set is used as the second protein set corresponding to that second set;
[0294] For any first set, iterate through all the second sets that are paired with it, such as... Figure 3 As shown in (b), the protein identifiers in the second protein set corresponding to the second set and the first protein set corresponding to the first set are matched, and the protein identifiers that exist in the first protein set but do not exist in the second protein set are filled into the second protein set;
[0295] Based on the filled second protein set, such as Figure 3 As shown in (c), each filled protein is labeled, combined with the structured feature data of the corresponding second set, and the interaction strength is padded with zeros to obtain the second information group, which is then filled into the second set. This record can correspond to a second type of sample.
[0296] Step S6: Based on all the first information groups from different literature corpora, and using the common parts of biological fluid identifiers, perform global filling to obtain the third type of samples.
[0297] Similarly, considering that the proteins contained in the same biological fluid are similar or even the same in different literature corpora or studies, negative samples can be further expanded by cross-literature corpora, that is, global filling.
[0298] Specifically, such as Figure 4 As shown, step S6 includes:
[0299] Obtain the corpus extracted from all literature, and such as Figure 4 As shown in (a), all first information groups with the same biological fluid identifier are formed into a third set, and all first information groups with the same structured feature data are respectively constructed into a fourth set;
[0300] Unlike local filling, global filling is based on the principle of multi-source reliability. It filters the third set to obtain the fifth set and pairs each fourth set with the fifth set based on the principle of the same biological fluid identifier. In fact, the fourth set that matches the fifth set also matches the third set corresponding to the fifth set. Similar to local filling, the third set is actually the union of all matching fourth sets. However, unlike local filling, since the fifth set is based on the result of filtering the third set, there may be some first information groups that are in a certain fourth set but not in the matching fifth set.
[0301] The union of all protein identifiers in the first information group in the fifth set is used as the third protein set;
[0302] The union of protein identifiers in all first information groups in each fourth set is calculated as the fourth protein set corresponding to that fourth set;
[0303] For any fifth set, iterate through all the fourth sets that are paired with it, such as... Figure 4 As shown in (b), the protein identifiers in the fourth protein set corresponding to the fourth set and the third protein set corresponding to the fifth set are matched, and the protein identifiers that exist in the third protein set but not in the fourth protein set are filled into the fourth protein set;
[0304] Based on the filled fourth protein set, such as Figure 4 As shown in (c), each filled protein is labeled, combined with the structured feature data of the corresponding fourth set, and the interaction strength is padded with zeros to construct the third type of sample.
[0305] Step S7: Label the first type of samples according to their interaction strength, and use the second and third types of samples as negative samples to construct the dataset.
[0306] Specifically, considering the possibility of classification and regression tasks, in one embodiment, for the classification task, a threshold of 0.001% is used to divide the samples into affinity and incompatibility, and binary classification modeling is performed. Therefore, in the first category of samples, the portion of the relative abundance of proteins adsorbed on the surface of the nanomaterial that is less than 0.001% is classified as negative samples, and the rest are classified as positive samples.
[0307] For regression tasks, such as Figure 5 As shown, the original data exhibits an extremely long-tailed distribution. Box-Cox transformation was used to standardize the relative abundance of proteins, and the results are as follows: Figure 6 As shown, it is then used for regression modeling.
[0308]
[0309] Wherein: λ is the transformation parameter. The optimal value of λ is usually estimated from the data through maximum likelihood estimation, aiming to make the transformed data as close as possible to the normal distribution. In this embodiment, the value of λ is 0.002999.
[0310] In this embodiment, as Figure 7 As shown, samples are constructed based on unfilled structured feature data and filled structured feature data respectively. An unfilled dataset is constructed based on the first, second and third class samples obtained from the unfilled structured feature data, and a filled dataset is constructed based on the first, second and third class samples obtained from the filled structured feature data.
[0311] In addition, in some other embodiments, each sample is configured with a weight, which is determined based on the accuracy and directness of the data.
[0312] Based on the above methods, the final dataset is summarized as follows: Figure 8 As shown, the missing data filling statistics are as follows: Figure 9 As shown, the constructed dataset includes 2.76 million samples, half of which are padded data and half are unpadded data.
[0313] also, Figure 10 The performance of the model trained on the dataset constructed in this invention is clearly demonstrated under different test conditions. Notably, the model trained solely on the padded dataset exhibits superior performance on the padded test set: in classification tasks, its accuracy is 0.8974, precision is 0.8361, recall is 0.8564, F1 score is 0.8461, and AUC score is 0.9604; while in regression tasks, the coefficient of determination (R²) is significantly higher. 2 The accuracy was 0.8698, the mean squared error (MSE) was 0.5544, and the mean absolute error (MAE) was 0.5221. However, when faced with test set samples containing missing data, the model's performance dropped significantly: in classification tasks, accuracy decreased to 0.5686, precision was only 0.4238, although recall remained high (0.8789), but the F1 score also decreased to 0.5718 and the AUC score to 0.7566; in regression tasks, the performance metrics deteriorated even more significantly, with R... 2The performance decreased to 0.3257, the MSE increased to 2.8655, and the MAE reached 1.2594. In contrast, the model trained on both padded and unpadded datasets showed stronger robustness. It not only maintained good performance on the padded test set—in the classification task, accuracy was 0.8953, precision was 0.8209, recall was 0.8704, F1 score was 0.8449, and AUC score was 0.9595; in the regression task, R... 2 The accuracy was 0.8752, the precision was 0.5303, and the recall was 0.8702, which is comparable to the model trained with only imputed data. More importantly, the model also performed well on the test set with missing data: in the classification task, the accuracy was 0.8949, the precision was 0.8201, the recall was 0.8702, the F1 score was 0.8444, and the AUC score was 0.9591; in the regression task, R... 2 The mean squared error (MSE) was 0.8728, the mean squared error (MSE) was 0.5407, and the mean squared error (MAE) was 0.5103, all maintaining a high level. Therefore, the dataset constructed using the method of this invention can be used to train a model with excellent predictive performance for interactions between nanomaterials and proteins. Furthermore, considering the common occurrence of missing feature data in practical applications, combining the filled dataset with the unfilled dataset for model training can effectively improve the model's generalization ability and stability in scenarios with missing data.
[0314] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
Claims
1. A method for constructing a dataset on the interaction between nanomaterials and proteins, characterized in that, include: The literature corpus is processed to extract one or more first information groups, wherein the first information group contains protein identifiers, amino acid sequences obtained based on protein identifiers, raw data on the interaction between nanomaterials and proteins, and raw structured feature data. The raw structured feature data includes at least nanomaterial identifiers and biological fluid identifiers, and the proteins corresponding to the protein identifiers are detected in the biological fluids corresponding to the biological fluid identifiers. Semantic alignment is performed on all the original structured feature data in the first information group to obtain unfilled structured feature data, wherein the missing feature data is represented by the first identifier; Fill in the missing features in all the structured feature data in the first information group to obtain the filled structured feature data; The interaction strength between the nanomaterials and each detected protein in the first information group is calculated respectively, and the first type of sample is generated based on the calculated interaction strength, structured feature data, nanomaterial identifiers and protein identifiers. Based on all the first information groups from the same literature corpus, the second type of sample is obtained by locally filling the parts with the same biological fluid identifier; Based on all the first information groups from different literature corpora, and based on the parts with the same biological fluid identifiers, global filling was performed to obtain the third type of sample; The first type of samples are labeled according to their interaction strength, and the second and third types of samples are used as negative samples to construct the dataset; The second type of sample is obtained by locally filling in all first information groups from the same corpus, based on the parts with the same biological fluid identifiers, including: A first set is formed by extracting all first information groups from the same literature corpus and having the same biological fluid identifier, and a second set is formed by constructing all first information groups with the same structured feature data. The second sets are then paired with the first set based on the principle of the same biological fluid identifier. The union of protein identifiers in all first information groups in the first set is used as the first protein set; The union of protein identifiers in all first information groups in each second set is used as the second protein set corresponding to that second set; For any first set, iterate through all paired second sets, match the protein identifiers in the second protein set corresponding to the second set with the protein identifiers in the first protein set corresponding to the first set, and fill the second protein set with protein identifiers that exist in the first protein set but not in the second protein set; Based on the second set of proteins after filling, each filled protein is identified, combined with the structured feature data of the corresponding second set, and the interaction strength is padded with zeros to construct the second type of sample; The process of globally filling in the third type of samples based on the common biological fluid identifiers of all first information groups from different literature corpora to obtain the third type of samples includes: A third set is formed by extracting all first information groups from all literature corpora and having the same biological fluid identifier, and a fourth set is formed by constructing all first information groups with the same structured feature data. Based on the principle of multi-source credibility, the third set is screened to obtain the fifth set, and each fourth set is paired with the fifth set based on the principle of the same biological fluid identifier. The union of all protein identifiers in the first information group in the fifth set is used as the third protein set; The union of protein identifiers in all first information groups in each fourth set is calculated as the fourth protein set corresponding to that fourth set; For any fifth set, iterate through all the fourth sets that are paired with it, match the protein identifiers in the fourth protein set corresponding to the fourth set with the protein identifiers in the third protein set corresponding to the fifth set, and fill the fourth protein set with the protein identifiers that exist in the third protein set but not in the fourth protein set. Based on the filled fourth protein set, each filled protein is identified, combined with the corresponding structured feature data of the fourth set, and the interaction strength is padded with zeros to construct the third type of sample.
2. The method for constructing a dataset of nanomaterial-protein interactions according to claim 1, characterized in that: The strength of the interaction between the nanomaterial and the protein is represented by the relative abundance of the protein adsorbed on the surface of the nanomaterial; The structured feature data includes: The physicochemical properties of nanomaterials include nanomaterial identification, core composition, nanomaterial concentration, nanomaterial size parameters, shape category, nanomaterial surface modification, nanoparticle surface modification category, hydrodynamics in water, surface charge in water, polydispersity index in water, dispersion medium, hydrodynamic diameter in the dispersion medium, surface charge in the dispersion medium, and polydispersity index in the dispersion medium. Protein source parameters, including biological fluid markers, protein source species, and protein source concentration; Interaction environment parameters include temperature, buffer composition, interaction time between nanomaterials and proteins, interaction temperature between nanomaterials and proteins, flow state of the nanomaterials and protein interaction system, flow rate of the nanomaterials and protein interaction system, and state of the nanomaterials and protein interaction system. The parameters for separating complexes include the type of separation method, centrifugation speed, centrifugation time, centrifugation temperature, and number of centrifugation repetitions; Proteomics analysis methods, including the range of detectable proteins in proteomics, analytical methods, and instruments and equipment; The protein identifier includes at least one of the following: protein name, gene name, and protein identifier.
3. The method for constructing a dataset of nanomaterial-protein interactions according to claim 1, characterized in that, The multi-source trust principle is as follows: If any protein appears in the same biological fluid in at least three independently sourced literature corpora, then it contains a first information set corresponding to the protein identifier.
4. The method for constructing a dataset of nanomaterial-protein interactions according to claim 2, characterized in that, During the process of imputing missing features in all structured feature data in the first information group: The surface modification of the nanomaterials is filled with default values; missing data is considered as no surface modification. The categories of nanoparticle surface modifications were filled using a large inference model. The size parameters of the nanomaterials were filled using a weighted average. The hydrodynamic diameter, surface charge, and polydispersity index of nanomaterials in water, as well as their hydrodynamic diameter, surface charge, and polydispersity index in the dispersion medium, were filled using a large case-based model. The dispersion medium is filled with the default value; missing data is considered as water as the dispersion medium. The concentration of the nanomaterials was filled using a mean value. The concentrations of the protein sources were filled with mean values. The interaction time and temperature between the nanomaterials and proteins are filled with the mode. The flow state of the nanomaterial-protein interaction system is filled with default values; missing data is considered as static. The flow rate of the nanomaterial-protein interaction system is filled with either a default value or a mean value. When the flow state of the nanomaterial-protein interaction system is stationary, missing data is considered as 0. When the flow state of the nanomaterial-protein interaction system is flowing, missing data is filled with the mean value. The states of the nanomaterial-protein interaction system are filled with default values; missing data is considered as in vitro. The separation method type is filled with default values; missing data is considered as using the centrifugation method. The centrifugation rate is filled with the mean value. If the separation method type is not centrifugation, it is filled with a second identifier to indicate invalidity. The centrifugation time, centrifugation temperature, and number of centrifugation repetitions are all filled with the mode. If the separation method type is not centrifugation, it is filled with a second identifier to indicate invalidity.
5. The method for constructing a dataset of nanomaterial-protein interactions according to claim 1, characterized in that, Samples are constructed based on unfilled structured feature data and filled structured feature data respectively. An unfilled dataset is constructed based on the first, second and third class samples obtained from the unfilled structured feature data, and a filled dataset is constructed based on the first, second and third class samples obtained from the filled structured feature data.
6. The method for constructing a dataset of nanomaterial-protein interactions according to claim 1, wherein each sample is assigned a weight, the weight being determined based on the accuracy and directness of the data.
7. A device for constructing a dataset of nanomaterial-protein interactions, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-6.
8. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements the method as described in any one of claims 1-6.
