Search

CN-122024849-A - Mixed expert lncRNA-protein interaction prediction method for gating and interaction routing

CN122024849ACN 122024849 ACN122024849 ACN 122024849ACN-122024849-A

Abstract

The invention provides a mixed expert lncRNA-protein interaction prediction method for gating and interaction routing, and relates to the field of biological big data analysis and deep learning cross application. The method integrates protein interaction network information, lncRNA and protein sequence information, and secondary structure and physicochemical property characteristics extracted by the sequence to construct multi-mode representation, and introduces learnable differential characteristic gating weight under a unified training target to adaptively inhibit redundancy and noise characteristics. The prediction layer adopts a mixed expert routing mechanism of gating and interaction driving to finish expert selection and weighted fusion to output interaction probability, and synchronously outputs gating weight, interaction weight and expert routing weight as contribution description and explanation basis. The method is suitable for large-scale interaction relation prediction, candidate interaction pair screening and experimental verification assistance.

Inventors

  • HU CHUNLING
  • DAI XIANGLONG
  • RAO JUNNAN
  • XU MIN

Assignees

  • 合肥大学

Dates

Publication Date
20260512
Application Date
20260126

Claims (10)

  1. 1. A mixed expert lncRNA-protein interaction prediction method of gating and interaction routing is characterized by firstly constructing multi-mode feature representation based on protein interaction network information (PPI), lncRNA and protein sequence information, secondary structure information and physicochemical property information, secondly introducing learnable differentiable feature gating weights into a deep learning framework, carrying out importance modeling on each mode and feature dimension thereof under interaction prediction supervision, carrying out low contribution feature adaptive suppression to obtain a screened feature representation, then constructing a cross-mode interaction coding network, mapping each screened mode feature to a unified representation space, learning interaction weights through interaction mechanisms to obtain fusion representation, finally constructing a mixed expert (MoE) prediction layer of gating and interaction routing, wherein the mixed expert (MoE) prediction layer comprises a routing network, a plurality of light-weight private sub-networks and a weighted fusion unit, the routing network takes the interaction weights between the feature gating weights and the modes as inputs, generates expert routing weight vectors aiming at each expert RNA-protein pair to be predicted, carries out weighted fusion on the output of the sub-networks to obtain interaction feature vector interaction vector prediction results, and the interaction vector interaction feature vector to be used for the key attribute vector interaction vector to be interpreted key attribute expert interaction vector interaction feature interaction vector to be relevant key feature interaction vector prediction key feature interaction vector to be used for the key interaction vector interaction attribute prediction.
  2. 2. The method for mixed expert lncRNA-protein interaction prediction for gating and interaction routing of claim 1, comprising the steps of: (1) The method comprises the steps of constructing a multi-modal feature representation, namely acquiring lncRNA sequences, protein sequences and protein interaction network (PPI) information, and extracting features based on the sequences and the network information to construct the multi-modal feature representation, wherein the multi-modal feature comprises the protein interaction network (PPI) feature, the sequence features of the lncRNA and the protein, and secondary structural features and physicochemical property features extracted from the lncRNA sequences and the protein sequences; (2) A differential feature gating selection module is constructed, namely a learnable gating weight with a value between 0 and 1 is distributed to each dimension of the multi-modal feature, and the multi-modal feature is subjected to progressive weighting by using the gating weight to obtain gating weighted feature representation; (3) The method comprises the steps of constructing a cross-modal interactive coding network, namely transforming each screened modal characteristic into a uniform-dimension modal representation through a linear mapping layer, calculating the relevance of any target mode to other source modes by adopting a scaling dot product attention, and carrying out Softmax normalization on the relevance corresponding to the same target mode to obtain an inter-modal interactive weight matrix; (4) The method comprises the steps of constructing a mixed expert (MoE) prediction layer of gating and interaction routing, carrying out statistics aggregation on feature gating weights and interaction weights among modes to obtain condition vectors, inputting fusion representations and the condition vectors into a routing network to obtain expert routing weights, respectively inputting the fusion representations into a plurality of parallel expert sub-networks to obtain expert interaction scores, carrying out weighted fusion on the expert interaction scores according to the expert routing weights to obtain total scores, obtaining interaction prediction results through Sigmoid, and synchronously outputting feature gating weights, interaction weight matrixes among modes and expert routing weight vectors as interpretation information.
  3. 3. The method for predicting mixed expert lncRNA-protein interactions for gating and interaction routing of claim 2, wherein the step (1) is constructed as follows: Step S101, preprocessing sequence data and constructing multi-element characteristics Performing de-duplication, length standardization and character legality verification on the obtained lncRNA and protein sequences, ensuring that the sequences only contain effective basic groups/amino acid symbols, and taking the processed sequences as unified input for subsequent feature extraction, thereby ensuring sample input consistency and reducing the influence of noise on modeling; Constructing a protein interaction network (PPI) feature, namely constructing a corresponding PPI subgraph for each protein based on a public protein interaction database, mapping the PPI subgraph into a topological vector with a fixed dimension by adopting a graph embedding method, and describing the functional position and neighborhood relation of the protein in the interaction network so as to be convenient for alignment and fusion with other modal features; The sequence modal feature construction comprises the steps of carrying out k-mer frequency statistics and sequence coding on the lncRNA sequence and the protein sequence respectively, mapping a local sequence mode into a fixed length vector, and then splicing sequence vectors at two sides to obtain a sequence joint feature vector, wherein the sequence joint feature vector is used for controlling the dimension while keeping key fragment information and reducing high-dimensional sparse redundancy; the construction of secondary structure features, namely respectively obtaining the structure tag information of the lncRNA and the secondary structure conformation (alpha-helix, beta-sheet and coil) of the protein, and carrying out statistical coding on a structure result to form a fixed dimension vector; And the physicochemical property characteristic construction comprises the steps of respectively counting the hydrophobicity, polarity and charge physicochemical properties of the protein, the GC content and stability index of the lncRNA and forming a fixed dimension vector, and then splicing the physicochemical vectors at two sides to obtain a physicochemical combined characteristic vector so as to supplement the biochemical property information which is difficult to cover by the sequence and the structure.
  4. 4. The method for predicting the mixed expert lncRNA-protein interaction of gating and interaction routing according to claim 2 or 3, wherein the differentiable feature gating selection module constructed in the step (2) introduces a sparse constraint term in the training process to promote the gating weight to be sparse, the sparse constraint term adopts an L1 norm and forms a combined optimization target with the interaction prediction loss weighting, and the inference stage sets the feature dimension with the gating weight lower than a preset threshold to zero to realize screening.
  5. 5. The method for predicting mixed expert lncRNA-protein interactions for gating and interaction routing of claim 4, wherein the step (2) is constructed as follows: step S102, modeling and screening the differentiable importance of the multi-modal features The characteristic representation of each sample is uniformly represented as: (equation 1) D represents the total dimension of the multi-modal features to realize the joint modeling and subsequent importance evaluation basis of the multi-source features in the same space; Introducing a learnable feature gating weight for each dimension feature, wherein the gating weight is obtained by mapping learnable gating parameters through Sigmoid: (equation 2) Wherein, the For characterizing the relative importance of the ith dimension feature in the current prediction task; (equation 3) Weighted feature representation Is used as input to a subsequent interaction prediction model and for calculating an interaction prediction loss Introducing sparse constraint term in training process The sparse constraint term adopts an L1 norm; comprehensively considering the predicted performance and the characteristic compression requirement, and defining an overall optimization objective function as follows: (equation 4) Wherein, the The weight coefficient is used for balancing the prediction performance and the characteristic compression degree; During model training, feature gating weights Synchronously updating network parameters of the interaction prediction model through a gradient back propagation mechanism; after training is completed, preserving feature dimensions with feature gating weights higher than a preset threshold value to form a simplified multi-modal feature representation: (equation 5) Wherein, the Screening a threshold value for the characteristic; Reduced feature representation As the input of the subsequent multi-modal feature interactive modeling and fusion learning.
  6. 6. The method for predicting mixed expert lncRNA-protein interaction of gating and interaction routing according to claim 2 or 5, wherein in the cross-modal interaction coding network constructed in the step (3), softmax normalization is performed on source modal correlations corresponding to the same target modality to obtain inter-modal interaction weights, so that the sum of source modal weights for the same target modality is 1.
  7. 7. The method for predicting mixed expert lncRNA-protein interactions for gating and interaction routing of claim 6, wherein the step (3) is constructed as follows: Step S103, input organization of the multi-mode features after screening Will be Modal division into Path input branch: (equation 6) The multi-path branches are used as the input of a subsequent interactive coding network; Step S104, intra-modality representation mapping and dimension alignment For each mode characteristic, respectively constructing intra-mode characteristic coding sub-network Representation extraction and dimension alignment are carried out on each modal feature, and the modal feature is mapped to a unified dimension Is a representation space of (a): (equation 7) Wherein the method comprises the steps of Is the first The subsequent cross-modal interaction is performed in the same representation space through unified dimension mapping; Step S105, cross-modal interaction weight learning and representation updating For any modality representation Calculating the interaction strength of the model with other modes, and updating the model according to the interaction strength to obtain the interactive representation The interactive update is expressed as; (equation 8) Wherein, the Representing a modality To the mode Is used for the strength of the steel sheet, The mapping function is used for aligning interaction information; from scaling dot product attention Calculating and carrying out Softmax normalization on source mode weights of fixed target modes to enable the source mode weights to meet the requirement that the sum of the source weights of each target mode is 1: (equation 9) Each mode fuses complementary information from other modes while retaining self discrimination information, so that updated representation containing cross-mode dependency relationship is obtained; Step S106, generating multi-layer interactive stack and fusion representation Stacking the interactive units of step S105 Layers to strengthen the cross-modal dependency and improve the fusion expression capability layer by layer, the first The layer output is used as the next layer input, and finally the deep interactive representation of each mode is obtained Then fusing the multi-mode interactive representation to obtain a unified fusion vector The fusion mode is splicing or weighted aggregation: (equation 10) Fusion vector And meanwhile, the interaction relation between the effective information of each mode and the modes is contained and is used as the final input representation of the interaction prediction.
  8. 8. The method for predicting the interaction between the expert lncRNA and the protein of the gating and interaction routing according to claim 2 or 7, wherein the routing network constructed in the step (4) comprises a two-layer fully connected network, a K-dimensional routing score is output, an expert routing weight vector is obtained through Softmax, and the expert routing weight vector is used for preserving two experts with the largest weight in a Top-2 mode and re-normalizing the two experts for weighted fusion of expert output.
  9. 9. The method for predicting mixed expert lncRNA-protein interactions for gating and interaction routing of claim 8, wherein the step (4) is constructed as follows: Step S107, moE prediction layer output and interpretation information generation of gating and interaction route In step S106, a fusion representation is obtained Based on the step S104, the feature gating weight obtained by learning is taken As characteristic contribution degree representation, the inter-modal interaction weight obtained by learning in step S105 is taken As cross-modal interaction weight And (3) with Respectively carrying out statistics and convergence to obtain condition vectors And (3) with The method is used for describing characteristic contribution distribution and modal interaction modes of the sample; constructing a MoE prediction layer of gating and interactive routing, wherein the MoE prediction layer comprises a routing network Lightweight private subnetworks in parallel with K The routing network takes the fusion representation and the condition vector as input and outputs the expert weight vector : (Equation 11) Wherein, the Represent the first Contribution of individual expert subnetworks to current sample prediction, and ; Each expert sub-network receives the fused representation And outputs an interaction score : (Equation 12) Weighting and fusing the expert outputs according to the expert weight vector to obtain a total score And obtaining the interaction prediction probability through Sigmoid function : (Equation 13) (Equation 14) Outputting the result of the prediction of the interaction between the lncRNA and the protein, and outputting the probability of the prediction of the interaction Synchronously outputting feature gating weights Inter-modality interaction weight Expert weight vector Wherein, the method comprises the steps of, For locating key features and their contribution, Is used for positioning the relation between the key modes and the interaction strength, The method is used for representing the contribution degree of the sample level decision path and the expert, so as to form an explanation evidence chain of characteristic contribution-modal interaction-decision path.
  10. 10. A computer readable storage medium having stored thereon a computer program for performing the method of any of claims 1 to 9 when run on a processor.

Description

Mixed expert lncRNA-protein interaction prediction method for gating and interaction routing Technical Field The invention relates to the field of biological big data analysis and deep learning cross application, in particular to a method for predicting lncRNA-protein interaction by combining differential feature selection, cross-modal interactive coding and mixed expert prediction of gating and interactive routing, namely a method for predicting lncRNA-protein interaction by using the mixed expert of gating and interactive routing. Background LncRNA-protein interactions (lncRNA-protein interaction, LPI) play an important role in biological processes such as gene expression regulation, cell differentiation, signal transduction, disease occurrence and development, and the like. Accurate prediction of lncRNA-protein interaction relationship is helpful for revealing complex regulation and control mechanism and provides support for research of disease mechanism, target spot discovery and drug research and development. Along with the development of high-throughput sequencing and histology technologies, the scale of related sequences and network data is rapidly increased, and the traditional experimental verification method has high cost and long period, so that efficient screening and prioritization of interaction relations are needed to be realized by means of a calculation method. From the technical route, the existing methods can be generally classified into two types, namely, a prediction method based on characteristic engineering and traditional machine learning, wherein information such as lncRNA sequences, protein structures, interaction networks and the like is generally subjected to statistical coding or manual feature construction, and then classification models are used for completing interaction discrimination. The method is relatively simple to realize, low in training cost and has a certain effect on small-scale data, but the performance of the method often depends on the quality of feature design, and the capability of describing complex correlations among multiple modes is limited. The other type is a prediction method based on deep learning, and the feature representation is automatically learned through a neural network to complete interaction prediction, so that the multi-source information can be more fully utilized, the nonlinear characterization capability is improved, and therefore, a good prediction result is obtained on a plurality of public data sets. However, the deep learning method generally has two challenges under a multi-mode scene, namely, firstly, the multi-source feature dimension is high, correlation and repeated information exist, the contribution degree difference of different modes is obvious, if an effective feature contribution evaluation and suppression mechanism is not available, redundant information is easy to introduce to influence training stability and generalization performance, secondly, complex dependency relationship exists among modes such as lncRNA sequences, protein structures, interaction network topology and the like, and if a simple splicing or shallow fusion mode is adopted, cross-mode association is difficult to fully describe, so that the expression capability of the model on the interaction mechanism is limited. In addition, LPI data often accompanies objective factors such as class imbalance, sample noise, and cross-dataset distribution differences, resulting in significant differences in the validity of different modality information among samples, thereby facilitating performance fluctuations of the model under different data sources and task settings. Therefore, a method for simultaneously realizing effective modeling of feature contribution and full characterization of cross-modal association under a unified learning framework and adaptively forming decision output according to sample features and interaction modes is needed to obtain an interaction prediction result which is more robust, more interpretable and more generalization capable. Disclosure of Invention The invention aims to solve the technical problems of effectively modeling and screening the contribution of each mode and the characteristic dimension thereof under a unified learning frame in the aspect of high-dimensional multi-mode characteristics composed of multi-source information such as protein interaction network information (PPI), lncRNA, protein sequences, secondary structures and physicochemical properties in the lncRNA-protein interaction prediction task, inhibiting the influence of redundancy and noise characteristics on training and generalization, and on the basis, explicitly describing the association between different mode representations to realize deep fusion expression of multi-mode information, and further, guiding the self-adaptive selection and weighted combination of the prediction process by utilizing the contribution modeling result and the interaction weight