Membrane protein go function two-way fusion reasoning method based on search enhancement and knowledge base
Patent Information
- Application Number
- CN202611142598.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-30
- Publication Date
- 2026-10-20
AI Technical Summary
在生物医学研究中，预测结果的可解释性和可追溯性至关重要，现有方法缺乏对预测结果来源的透明化展示，限制了预测结果在实际研究中的应用可信度
[0020]本发明实施例提供的技术方案带来的有益效果至少包括：
Smart Images
Figure FT_1 
Figure SMS_6 
Figure SMS_18
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics technology, and in particular to a dual-path fusion reasoning method for membrane protein GO function based on retrieval enhancement and knowledge base. Background Technology
[0002] Membrane proteins are a class of proteins that are embedded in or attached to biological membranes and play a crucial role in life activities such as cell signaling, substance transport, and energy conversion. Gene Ontology (GO) functional annotation is an important standard for describing the molecular function, biological processes, and cellular components of membrane proteins. Accurate GO functional prediction of membrane proteins is of great significance for understanding the biological mechanisms of membrane proteins and for drug target screening.
[0003] Existing methods for predicting membrane protein functions using GO (Go of Function) typically employ deep learning models. These models input the amino acid sequence and spatial structure information of the membrane protein into an end-to-end neural network, with the classification head directly outputting the predicted function probability. However, membrane protein functions exhibit high diversity, and some functional categories have few samples in the training set. Existing end-to-end models are insufficient in predicting functional categories with limited samples, and their parameters are fixed after training, making it difficult to adapt to newly discovered membrane protein function annotations.
[0004] Existing retrieval-enhanced protein function prediction methods typically employ single-modal features (such as amino acid sequences only) for similarity retrieval, failing to fully utilize the multimodal information of membrane proteins (including amino acid sequences, spatial structure information, and physicochemical characteristics). The function of membrane proteins is closely related to their spatial topology and physicochemical properties; single-modal feature retrieval struggles to accurately match functionally similar membrane proteins, thus limiting the reliability of the retrieval prediction results.
[0005] Existing methods for predicting membrane protein function typically rely solely on the output of parameterized models, failing to establish a correlation mechanism between model predictions and known knowledge bases. When the membrane protein to be predicted exhibits high similarity to existing membrane proteins in the knowledge base, current methods do not utilize experimentally validated functional annotations from the knowledge base to aid in reasoning, resulting in uninterpretable predictions and a lack of effective fusion between parameterized predictions and retrieval-based predictions.
[0006] Existing methods for predicting membrane protein function typically only provide functional labels and their probability values when outputting prediction results, without providing the basis for the prediction or the chain of evidence. In biomedical research, the interpretability and traceability of prediction results are crucial. The lack of transparency in disclosing the source of prediction results in existing methods limits the credibility of the prediction results in practical research.
[0007] Updating existing membrane protein knowledge bases typically requires retraining deep learning models, resulting in high computational costs and long update cycles. When new experimental validation annotations are released, existing methods struggle to quickly incorporate the new knowledge into the prediction system, leading to insufficient timeliness and scalability of the knowledge base. Summary of the Invention
[0008] In view of this, embodiments of the present invention provide a dual-path fusion reasoning method for membrane protein GO function based on retrieval enhancement and knowledge base.
[0009] A dual-path fusion inference method for membrane protein GO function based on retrieval enhancement and knowledge base is provided, including the following steps:
[0010] S1. Construct a membrane protein knowledge base: Input the amino acid sequence, spatial structure information and physicochemical features of each membrane protein sample in the training set into the HeteroMem model that has been trained and converged, extract the feature tensor output by the fusion layer of the HeteroMem model, perform average pooling on the feature tensor to obtain the multimodal semantic vector of the membrane protein sample.
[0011] Each membrane protein sample is compiled into a knowledge base entry by combining its protein identifier, functional tag set, and multimodal semantic vector, and then stored.
[0012] A nearest neighbor index is constructed based on the multimodal semantic vectors of all knowledge base entries;
[0013] S2. Parametric prediction: The amino acid sequence, spatial structure information and physicochemical characteristics of the membrane protein to be predicted are input into the parameter-frozen HeteroMem model. The classification head of the HeteroMem model outputs the first functional prediction probability vector of the membrane protein to be predicted on each candidate functional label.
[0014] S3. Retrieval and Prediction: The same HeteroMem model as in step S2 is used to extract the query vector of the membrane protein to be predicted.
[0015] Retrieve the K knowledge base entries with the smallest distance to the query vector from the nearest neighbor index constructed in step S1;
[0016] The voting weight of each knowledge base entry is determined based on the distance between the multimodal semantic vector of each knowledge base entry and the query vector.
[0017] The functional label sets of K knowledge base entries are weighted and aggregated according to the voting weights to obtain the second functional prediction probability vector of the membrane protein to be predicted on each candidate functional label.
[0018] S4. Dual-path fusion: The first functional prediction probability vector and the second functional prediction probability vector are weighted and linearly fused according to the fusion coefficient, and the fusion result is used as the functional prediction result of the membrane protein to be predicted.
[0019] Beneficial effects
[0020] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:
[0021] This invention freezes the amino acid sequence, spatial structure information, and physicochemical characteristics of the membrane protein to be predicted into a HeteroMem model. The classification head outputs a first function prediction probability vector, while the same HeteroMem model is used to extract a query vector. The K knowledge base entries with the smallest distance to the query vector are retrieved from the nearest neighbor index. Voting weights are determined based on distance, and the function tag sets of the K knowledge base entries are weighted and aggregated to obtain a second function prediction probability vector. The first and second function prediction probability vectors are then linearly fused according to a fusion coefficient. This solves the technical problem of existing technologies that rely solely on parameterized model output and lack a mechanism for linking model prediction with the knowledge base. It achieves the technical effect of dual-path fusion of parameterized prediction and retrieval prediction, comprehensively utilizing the model's generalization ability and empirical evidence from the knowledge base.
[0022] This invention employs an inverse proportional distance strategy to calculate voting weights. For each candidate functional label, the voting weights of the reference membrane proteins containing that candidate functional label in the functional label set are accumulated and normalized to obtain the retrieval prediction probability. This solves the technical problem in the prior art that the retrieval prediction does not consider the similarity difference between the reference membrane protein and the membrane protein to be predicted. It achieves the technical effect of making the reference membrane protein with higher similarity have greater influence and improving the accuracy of retrieval prediction through the distance-weighted voting mechanism.
[0023] This invention solves the technical problems of existing technologies, such as the lack of interpretability and traceability of prediction results and the limitation of the credibility of prediction results in actual research, by outputting the protein identifiers, functional tag sets, and voting weights of K retrieved knowledge base entries as an interpretable evidence chain while outputting the functional prediction results. It achieves the technical effect of transparently displaying the basis for prediction through the evidence chain and improving the interpretability and credibility of prediction results. Attached Figure Description
[0024] Figure 1 A flowchart provided for an embodiment of this application. Detailed Implementation
[0025] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.
[0026] like Figure 1 As shown, the dual-path fusion inference method for membrane protein GO function based on retrieval enhancement and knowledge base includes the following steps:
[0027] S1. Construct a membrane protein knowledge base: Input the amino acid sequence, spatial structure information and physicochemical features of each membrane protein sample in the training set into the HeteroMem model that has been trained and converged, extract the feature tensor output by the fusion layer of the HeteroMem model, perform average pooling on the feature tensor to obtain the multimodal semantic vector of the membrane protein sample.
[0028] Each membrane protein sample is compiled into a knowledge base entry by combining its protein identifier, functional tag set, and multimodal semantic vector, and then stored.
[0029] A nearest neighbor index is constructed based on the multimodal semantic vectors of all knowledge base entries;
[0030] S2. Parametric prediction: The amino acid sequence, spatial structure information and physicochemical characteristics of the membrane protein to be predicted are input into the parameter-frozen HeteroMem model. The classification head of the HeteroMem model outputs the first functional prediction probability vector of the membrane protein to be predicted on each candidate functional label.
[0031] S3. Retrieval and Prediction: The same HeteroMem model as in step S2 is used to extract the query vector of the membrane protein to be predicted.
[0032] Retrieve the K knowledge base entries with the smallest distance to the query vector from the nearest neighbor index constructed in step S1;
[0033] The voting weight of each knowledge base entry is determined based on the distance between the multimodal semantic vector of each knowledge base entry and the query vector.
[0034] The functional label sets of K knowledge base entries are weighted and aggregated according to the voting weights to obtain the second functional prediction probability vector of the membrane protein to be predicted on each candidate functional label.
[0035] S4. Dual-path fusion: The first functional prediction probability vector and the second functional prediction probability vector are weighted and linearly fused according to the fusion coefficient, and the fusion result is used as the functional prediction result of the membrane protein to be predicted.
[0036] It should be noted that:
[0037] This method is designed for scenarios where functional labels in membrane protein function prediction have a long-tailed distribution. For functional labels that appear no more than 20 times in the training set, the pure parameterized model suffers from a memory bottleneck due to insufficient training samples.
[0038] This method integrates parameterized prediction pathways and retrieval prediction pathways, and dynamically supplements the parameterized model with explicit evidence of similar membrane proteins retrieved.
[0039] Fmax is the standard evaluation index for gene ontology function prediction, which is the maximum value of the harmonic mean of precision and recall under different classification thresholds.
[0040] Experimental results show that, on a few-sample functional label, the Fmax of this method is improved from 0.0030 in the pure parametric model to 0.1060, and the overall Fmax reaches 0.5503. The overall Fmax of the HeteroMem model is 0.3830, and the overall Fmax of the TALE+ method is 0.4142.
[0041] The dimensions of both the first and second function prediction probability vectors are equal to the total number of candidate function labels.
[0042] K is the number of nearest neighbor searches, which is an integer greater than or equal to 1;
[0043] The fusion coefficient ranges from [0,1].
[0044] As an optional embodiment, in step S1, the multimodal semantic vector is obtained by the following formula:
[0045]
[0046] in, Let be the multimodal semantic vector of the i-th membrane protein sample;
[0047] i is the sequence number of the membrane protein sample;
[0048] , , These represent the amino acid sequence, spatial structure information, and physicochemical characteristics of the i-th membrane protein sample, respectively.
[0049] Pooling represents the average pooling operation on the output feature tensor of the fusion layer;
[0050] 512 is the fixed dimension obtained by average pooling the output feature tensor of the HeteroMem model fusion layer;
[0051] The nearest neighbor index is constructed using the IndexFlatL2 algorithm from the FAISS vector index library;
[0052] The IndexFlatL2 algorithm uses Euclidean distance as a similarity metric to perform precise retrieval by comparing the multimodal semantic vectors of all knowledge base entries one by one.
[0053] It should be noted that:
[0054] The calculation process of multimodal semantic vectors is as follows: the HeteroMem model encodes and fuses amino acid sequences, spatial structure information and physicochemical features respectively, and then outputs feature tensors from the fusion layer;
[0055] The feature tensor is transformed into a dense vector with a fixed 512 dimensions by average pooling.
[0056] The IndexFlatL2 algorithm calculates the Euclidean distance and sorts all multimodal semantic vectors in the index during retrieval, achieving a recall rate of 100% and avoiding the omission of membrane protein samples similar to the membrane protein to be predicted by the approximate index.
[0057] As an optional embodiment, in step S1, the knowledge base entries are stored in the form of triples:
[0058]
[0059] in, For the i-th knowledge base entry;
[0060] This is the unique identifier for the i-th membrane protein sample in the UniProt database;
[0061] The set of gene ontology GO functional tags associated with the i-th membrane protein sample, wherein the set of functional tags is a set of gene ontology GO terms;
[0062] Let be the multimodal semantic vector of the i-th membrane protein sample;
[0063] During the training phase, the set of functional labels undergoes ancestor propagation according to the real path rules. That is, when a membrane protein sample is labeled as a certain GO term, all ancestor terms of that GO term in the GO hierarchy are labeled as positive samples.
[0064] The membrane protein knowledge base is built on the MemGO-Bench dataset and only includes functional annotations of membrane proteins with experimental verification evidence.
[0065] It should be noted that:
[0066] Protein identifiers are unique identifiers from the UniProt database, used for tracing the origins and biological verification of retrieved reference membrane proteins;
[0067] Ancestor propagation ensures that the set of functional tags under the GO hierarchy remains hierarchically complete, and the retrieval and prediction pathway covers the ancestral functional terms of the membrane protein to be predicted during tag migration.
[0068] Only functional annotations with experimental verification evidence are included to avoid introducing errors into the retrieval evidence from computational prediction annotations.
[0069] As an optional embodiment, an incremental expansion step of the membrane protein knowledge base is also included:
[0070] When adding functional annotations for membrane proteins with experimentally validated evidence, the HeteroMem model will be frozen with the input parameters of the added membrane protein sample's amino acid sequence, spatial structure information, and physicochemical characteristics.
[0071] Extract the multimodal semantic vectors of the newly added membrane protein samples in the same manner as in step S1;
[0072] The protein identifiers, functional tag sets, and multimodal semantic vectors of newly added membrane protein samples will be used to form new knowledge base entries;
[0073] Insert the new knowledge base entry into the nearest neighbor index;
[0074] The HeteroMem model is not retrained during incremental expansion.
[0075] It should be noted that:
[0076] The incremental expansion is implemented by first collecting functional annotations of newly added membrane proteins with experimentally validated evidence.
[0077] The second step involves freezing the HeteroMem model with the amino acid sequence, spatial structure information, and physicochemical characteristics of the newly added membrane protein samples as input parameters, then outputting feature tensors through a fusion layer and performing average pooling to obtain the multimodal semantic vectors of the newly added membrane protein samples.
[0078] The third step is to perform ancestor propagation on the functional label set of the newly added membrane protein samples according to the real path rules;
[0079] The fourth step is to write the newly added knowledge base entries into the IndexFlatL2 index;
[0080] Knowledge base scaling analysis shows that the Fmax of the few-sample feature labels increases from 0.0857 to 0.1221 as the knowledge base size increases from 20% to 100%. The predictive performance on long-tail features can be improved simply by expanding the knowledge base without retraining the model.
[0081] As an optional embodiment, in step S2, the spatial structure information of the membrane protein to be predicted includes contact diagrams and isomorphic diagrams.
[0082] The process of processing input data in the parameter-frozen HeteroMem model includes:
[0083] The evolutionary characteristics of amino acid sequences were obtained by encoding them using the ESM-2 protein language model.
[0084] Spatial structural feature representations are obtained by processing the contact graph and heterogeneous graph structure through heterogeneous relation graph convolution;
[0085] The evolutionary and spatial structure features are modulated using physicochemical features through FiLM feature modulation.
[0086] The modulation result is fused through multimodal processing, and the first function prediction probability vector is output by the classification head.
[0087] It should be noted that:
[0088] The first function prediction probability vector is denoted as... ,satisfy ;
[0089] Where |GO| represents the total number of candidate feature tags;
[0090] The t-th component of the first functional prediction probability vector represents the prediction probability that the membrane protein to be predicted has the t-th candidate functional label, where t is the index of the candidate functional label and the value of t ranges from 1 to |GO|.
[0091] Parameter freezing refers to not updating the weights of the HeteroMem model during the prediction phase, but directly reusing the sequences, structures, and physicochemical statistical laws learned during the training phase.
[0092] The parameterized prediction pathway outputs high-confidence predictions for common functions that occur frequently in the training set, serving as the baseline for overall prediction.
[0093] As an optional embodiment, in step S3, the query vector and the multimodal semantic vector in step S1 are extracted using the same HeteroMem model and the same pooling method, and the query vector is obtained by the following formula:
[0094]
[0095] in, The query vector is the membrane protein to be predicted;
[0096] , , These are the amino acid sequence, spatial structure information, and physicochemical characteristics of the membrane protein to be predicted.
[0097] The distance is the Euclidean distance;
[0098] In the nearest neighbor index, the knowledge base entries are sorted in ascending order of Euclidean distance, and the membrane protein samples corresponding to the K knowledge base entries with the smallest distance are taken as reference membrane proteins.
[0099] It should be noted that:
[0100] The query vector and the multimodal semantic vector in the knowledge base are obtained by the same encoder and the same pooling method, and they are in the same embedding space;
[0101] Euclidean distance directly measures the similarity between the membrane protein to be predicted and each reference membrane protein in terms of sequence evolution information, spatial structural features, and physicochemical microenvironment modulation signals.
[0102] Functionally similar membrane proteins with low sequence homology are closer together in this embedding space, making them easier to retrieve and providing evidence for functional prediction in low homology scenarios.
[0103] As an optional embodiment, in step S3, the voting weight is calculated using an inverse proportional distance strategy:
[0104]
[0105] in, The voting weight for the j-th reference membrane protein;
[0106] Let be the Euclidean distance between the multimodal semantic vector of the j-th reference membrane protein and the query vector;
[0107] ε is a constant to prevent fractions with a denominator of zero from being greater than zero;
[0108] The value of j ranges from 1 to K;
[0109] The weighted aggregation specifically involves summing the voting weights of reference membrane proteins containing that candidate functional label in the functional label set for each candidate functional label.
[0110] The accumulated results are normalized to obtain the retrieval prediction probability of the candidate function tag;
[0111] The retrieval prediction probabilities of all candidate function tags constitute the second function prediction probability vector, denoted as . ,satisfy .
[0112] It should be noted that:
[0113] The calculation steps of the inverse proportional distance strategy are as follows: First, calculate the Euclidean distance between the query vector and the multimodal semantic vectors of K reference membrane proteins;
[0114] The second step is to add ε to each Euclidean distance and take the reciprocal to obtain the voting weight of each reference membrane protein.
[0115] The third step is to accumulate the voting weights of the reference membrane proteins that contain that candidate functional label for each candidate functional label.
[0116] The fourth step is to divide the sum of the candidate function labels by the sum of the K voting weights to complete the normalization.
[0117] The inverse proportional distance strategy does not include the hyperparameters to be tuned, and the reference membrane protein with a smaller Euclidean distance receives a higher voting weight;
[0118] With a nearest neighbor retrieval count of 2, the inverse proportional distance strategy outperforms the Gaussian kernel weighted strategy and the hard voting strategy in terms of prediction performance.
[0119] Weighted aggregation enables soft label migration, meaning that the functional labels of the reference membrane protein are partially migrated to the membrane protein to be predicted in a probabilistic manner according to the voting weights, rather than directly copying all labels to the membrane protein to be predicted.
[0120] As an optional embodiment, in step S4, the weighted linear fusion is performed by the following formula:
[0121]
[0122] in, The functional prediction results for the membrane protein to be predicted;
[0123] α is the fusion coefficient and α∈[0,1];
[0124] Predict the probability vector for the first function;
[0125] Predict the probability vector for the second function;
[0126] The fusion coefficient α is set to 0.9. This value is determined through fusion coefficient sensitivity analysis. Multiple candidate values of the fusion coefficient are traversed on the validation set, and the performance index of the functional prediction result under each candidate value is calculated. The candidate value corresponding to the optimal performance index is selected as the value of the fusion coefficient α.
[0127] The performance index is the Fmax index of gene ontology function prediction, and is calculated separately in three namespaces: molecular function, biological process, and cellular component.
[0128] It should be noted that:
[0129] The calculation steps for the fusion coefficient sensitivity analysis are as follows: First, select multiple candidate values in the interval [0,1].
[0130] The second step is to calculate the functional prediction results of each membrane protein sample in the validation set for each candidate value using the weighted linear fusion formula.
[0131] The third step is to calculate the Fmax index corresponding to each candidate value.
[0132] The fourth step is to compare the Fmax index corresponding to each candidate value and take the candidate value corresponding to the maximum value as the value of the fusion coefficient α.
[0133] Sensitivity analysis results show that the overall performance of the functional prediction results is optimal when α=0.9, at which point the weight of the parameterized prediction pathway is 0.9 and the weight of the retrieval prediction pathway is 0.1.
[0134] When α=0 (pure retrieval) and when α=1 (pure model), the performance of the function prediction results is lower than that when α=0.9.
[0135] When the target function of the membrane protein to be predicted belongs to a common function that appears frequently in the training set, the first function prediction probability vector has high confidence and the parameterized prediction pathway is dominant.
[0136] When the target function is a rare function or a zero-sample function, the first function prediction probability vector approaches 0, and the prediction pathway is supplemented by explicit evidence of similar reference membrane proteins.
[0137] As an optional embodiment, the nearest neighbor retrieval number K is set to 2, and this value is determined by the TOPSIS multi-objective decision-making method, specifically including:
[0138] For each candidate K value in the candidate set {1,2,3,4,5,6,7,8,9,10,20,50}, calculate the overall function prediction performance index OverallFmax and the few-shot function prediction performance index Few-shotFmax on the validation set respectively.
[0139] Min-max normalization is performed on both indicators to obtain normalized indicators. and ;
[0140] Calculate the Euclidean distance from the point corresponding to each candidate K value to the ideal point (1,1):
[0141]
[0142] Mapping the Euclidean distance to the (0,1) interval for the overall score:
[0143]
[0144] The candidate K value with the highest comprehensive score is selected as the value of the nearest neighbor retrieval number K.
[0145] It should be noted that:
[0146] in, The value of OverallFmax corresponding to the candidate K value is the min-max normalized value;
[0147] The value of Few-shotFmax corresponding to the candidate K value is the min-max normalized value;
[0148] Let K be the Euclidean distance from the normalized index point corresponding to the candidate K value to the ideal point (1,1);
[0149] TS(K) is the overall score of candidate K values;
[0150] The maximum Euclidean distance from the normalized index point to the ideal point (1,1) within the interval (0,1) is normalized to the interval (0,1).
[0151] The steps for calculating the K value are as follows: First, perform retrieval prediction and dual-path fusion for each candidate K value in the candidate set to obtain the corresponding OverallFmax and Few-shotFmax.
[0152] The second step is to perform min-max normalization on both indicators respectively;
[0153] The third step is to calculate the Euclidean distance and overall score corresponding to each candidate K value according to the above formula.
[0154] The fourth step is to select the candidate K value with the highest overall score;
[0155] The TOPSIS multi-objective decision-making results show that when K=2, the comprehensive score TS is 0.833, which is the largest in the candidate set. At this time, the Few-shot Fmax reaches a peak of 0.1201, and the Overall Fmax decreases by 0.39% from the highest point. K=2 takes into account both overall robustness and specificity of few samples.
[0156] As an optional embodiment, step S5 and evidence output are also included:
[0157] While outputting the functional prediction results, the protein identifiers of the K knowledge base entries retrieved in step S3, the functional label set of each knowledge base entry, and the voting weights corresponding to each knowledge base entry are also output as interpretability evidence chains.
[0158] It should be noted that:
[0159] The interpretable chain of evidence records the basis for the functional prediction results. By outputting explicit evidence of similar reference membrane proteins, the prediction logic of the functional prediction results is made white-boxed.
[0160] The steps for evidence output are as follows: First, read the protein identifiers of the K knowledge base entries that the search prediction pathway hits;
[0161] The second step is to read the set of functional tags and corresponding voting weights for each knowledge base item.
[0162] The third step is to correlate and output the protein identifier, functional tag set, voting weights, and functional prediction results.
[0163] Taking the prediction of bovine MICOS complex subunit MIC26 with UniProt identifier Q148H0 as an example, the cross-species evidence chain consisting of mouse DNAJC11, human DNAJC11, and yeast MIC60 was retrieved for the prediction pathway hit. Based on this, the membrane protein was predicted to be located in the inner mitochondrial membrane with a confidence level of 0.7316.
[0164] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A dual-path fusion reasoning method for membrane protein GO function based on retrieval enhancement and knowledge base, characterized in that, Includes the following steps: S1. Construct a membrane protein knowledge base: Input the amino acid sequence, spatial structure information and physicochemical features of each membrane protein sample in the training set into the HeteroMem model that has been trained and converged, extract the feature tensor output by the fusion layer of the HeteroMem model, perform average pooling on the feature tensor to obtain the multimodal semantic vector of the membrane protein sample. Each membrane protein sample is compiled into a knowledge base entry by combining its protein identifier, functional tag set, and multimodal semantic vector, and then stored. A nearest neighbor index is constructed based on the multimodal semantic vectors of all knowledge base entries; S2. Parametric prediction: The amino acid sequence, spatial structure information and physicochemical characteristics of the membrane protein to be predicted are input into the parameter-frozen HeteroMem model. The classification head of the HeteroMem model outputs the first functional prediction probability vector of the membrane protein to be predicted on each candidate functional label. S3. Retrieval and Prediction: The same HeteroMem model as in step S2 is used to extract the query vector of the membrane protein to be predicted. Retrieve the K knowledge base entries with the smallest distance to the query vector from the nearest neighbor index constructed in step S1; The voting weight of each knowledge base entry is determined based on the distance between the multimodal semantic vector of each knowledge base entry and the query vector. The functional label sets of K knowledge base entries are weighted and aggregated according to the voting weights to obtain the second functional prediction probability vector of the membrane protein to be predicted on each candidate functional label. S4. Dual-path fusion: The first functional prediction probability vector and the second functional prediction probability vector are weighted and linearly fused according to the fusion coefficient, and the fusion result is used as the functional prediction result of the membrane protein to be predicted.
2. The method for predicting the dual-path fusion function of membrane proteins based on retrieval enhancement generation and knowledge base according to claim 1, characterized in that, In step S1, the multimodal semantic vector is obtained by the following formula: in, Let be the multimodal semantic vector of the i-th membrane protein sample; i is the sequence number of the membrane protein sample; , , These represent the amino acid sequence, spatial structure information, and physicochemical characteristics of the i-th membrane protein sample, respectively. Pooling represents the average pooling operation on the output feature tensor of the fusion layer; 512 is the fixed dimension obtained by average pooling the output feature tensor of the HeteroMem model fusion layer; The nearest neighbor index is constructed using the IndexFlatL2 algorithm from the FAISS vector index library; The IndexFlatL2 algorithm uses Euclidean distance as a similarity metric to perform precise retrieval by comparing the multimodal semantic vectors of all knowledge base entries one by one.
3. The method for predicting the dual-path fusion function of membrane proteins based on retrieval enhancement generation and knowledge base according to claim 2, characterized in that, In step S1, the knowledge base entries are stored in the form of triples: in, For the i-th knowledge base entry; This is the unique identifier for the i-th membrane protein sample in the UniProt database; The set of gene ontology GO functional tags associated with the i-th membrane protein sample, wherein the set of functional tags is a set of gene ontology GO terms; Let be the multimodal semantic vector of the i-th membrane protein sample; During the training phase, the set of functional labels undergoes ancestor propagation according to the real path rules. That is, when a membrane protein sample is labeled as a certain GO term, all ancestor terms of that GO term in the GO hierarchy are labeled as positive samples. The membrane protein knowledge base is built on the MemGO-Bench dataset and only includes functional annotations of membrane proteins with experimental verification evidence.
4. The method for predicting the dual-path fusion function of membrane proteins based on retrieval-enhanced generation and knowledge base according to claim 3, characterized in that, It also includes incremental expansion steps for the membrane protein knowledge base: When adding functional annotations for membrane proteins with experimentally validated evidence, the HeteroMem model will be frozen with the input parameters of the added membrane protein sample's amino acid sequence, spatial structure information, and physicochemical characteristics. Extract the multimodal semantic vectors of the newly added membrane protein samples in the same manner as in step S1; The protein identifiers, functional tag sets, and multimodal semantic vectors of newly added membrane protein samples will be used to form new knowledge base entries; Insert the new knowledge base entry into the nearest neighbor index; The HeteroMem model is not retrained during incremental expansion.
5. The method for predicting the dual-path fusion function of membrane proteins based on retrieval enhancement generation and knowledge base according to claim 1, characterized in that, In step S2, the spatial structure information of the membrane protein to be predicted includes contact diagrams and isomorphic diagrams. The process of processing input data in the parameter-frozen HeteroMem model includes: The evolutionary characteristics of amino acid sequences were obtained by encoding them using the ESM-2 protein language model. Spatial structural feature representations are obtained by processing the contact graph and heterogeneous graph structure through heterogeneous relation graph convolution; The evolutionary and spatial structure features are modulated using physicochemical features through FiLM feature modulation. The modulation result is fused through multimodal processing, and the first function prediction probability vector is output by the classification head.
6. The method for predicting the dual-path fusion function of membrane proteins based on retrieval-enhanced generation and knowledge base according to claim 5, characterized in that, In step S3, the query vector and the multimodal semantic vector in step S1 are extracted using the same HeteroMem model and the same pooling method. The query vector is obtained using the following formula: in, The query vector is the membrane protein to be predicted; , , These are the amino acid sequence, spatial structure information, and physicochemical characteristics of the membrane protein to be predicted. The distance is the Euclidean distance; In the nearest neighbor index, the knowledge base entries are sorted in ascending order of Euclidean distance, and the membrane protein samples corresponding to the K knowledge base entries with the smallest distance are taken as reference membrane proteins.
7. The method for predicting the dual-path fusion function of membrane proteins based on retrieval-enhanced generation and knowledge base according to claim 6, characterized in that, In step S3, the voting weights are calculated using an inverse proportional distance strategy: in, The voting weight for the j-th reference membrane protein; Let be the Euclidean distance between the multimodal semantic vector of the j-th reference membrane protein and the query vector; ε is a constant to prevent fractions with a denominator of zero from being greater than zero; The value of j ranges from 1 to K; The weighted aggregation specifically involves summing the voting weights of reference membrane proteins containing that candidate functional label in the functional label set for each candidate functional label. The accumulated results are normalized to obtain the retrieval prediction probability of the candidate function tag; The retrieval prediction probabilities of all candidate function tags constitute the second function prediction probability vector, denoted as . ,satisfy .
8. The method for predicting the dual-path fusion function of membrane proteins based on retrieval enhancement generation and knowledge base according to claim 7, characterized in that, In step S4, the weighted linear fusion is performed using the following formula: in, The functional prediction results for the membrane protein to be predicted; α is the fusion coefficient and α∈[0,1]; Predict the probability vector for the first function; Predict the probability vector for the second function; The fusion coefficient α is set to 0.
9. This value is determined through fusion coefficient sensitivity analysis. Multiple candidate values of the fusion coefficient are traversed on the validation set, and the performance index of the functional prediction result under each candidate value is calculated. The candidate value corresponding to the optimal performance index is selected as the value of the fusion coefficient α. The performance index is the Fmax index of gene ontology function prediction, and is calculated separately in three namespaces: molecular function, biological process, and cellular component.
9. The method for predicting the dual-path fusion function of membrane proteins based on retrieval enhancement generation and knowledge base according to claim 8, characterized in that, The nearest neighbor retrieval number K is set to 2, and this value is determined by the TOPSIS multi-objective decision-making method, specifically including: For each candidate K value in the candidate set {1,2,3,4,5,6,7,8,9,10,20,50}, calculate the overall function prediction performance index OverallFmax and the few-shot function prediction performance index Few-shotFmax on the validation set respectively. Min-max normalization is performed on both indicators to obtain normalized indicators. and ; Calculate the Euclidean distance from the point corresponding to each candidate K value to the ideal point (1,1): Mapping the Euclidean distance to the (0,1) interval for the overall score: The candidate K value with the highest comprehensive score is selected as the value of the nearest neighbor retrieval number K.
10. The method for predicting the dual-path fusion function of membrane proteins based on retrieval-enhanced generation and knowledge base according to claim 9, characterized in that, It also includes step S5, evidence output: While outputting the functional prediction results, the protein identifiers of the K knowledge base entries retrieved in step S3, the functional label set of each knowledge base entry, and the voting weights corresponding to each knowledge base entry are also output as interpretability evidence chains.