A protein function prediction system and method based on a protein-product multi-modal computational model

Abstract

The application belongs to the technical field of biological information calculation model, and particularly relates to a protein function prediction system and method based on a protein-product multi-modal calculation model, which comprises the following steps: obtaining a training sample set; extracting protein representation and product representation; obtaining protein local sequence features and product local structure features; obtaining protein-product joint representation through a cross-modal interaction module; training a multi-modal calculation model; in the prediction stage, combining the protein to be tested with candidate products in the training product library one by one and inputting the trained model for scoring, selecting the candidate product with the highest score, and obtaining the function classification result of the protein to be tested through a product-function label mapping relationship. The application can predict protein function by utilizing the correlation between protein sequence and product structure, and is suitable for scenarios such as protein function classification, enzyme function annotation, product category attribution and candidate functional protein screening which are difficult to be accurately recognized by traditional sequence similarity analysis.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics computing model technology, and relates to a protein function prediction system and method based on a protein-product multimodal computing model. Background Technology

[0002] Protein function prediction is a crucial task in bioinformatics, enzyme engineering, synthetic biology, and natural product discovery. Current prediction methods primarily rely on sequence homology analysis, conserved domain annotation, phylogenetic analysis, or sequence similarity networks, which are suitable for preliminary assessments of protein family classification and conserved functions. However, for product-directed proteins such as natural product biosynthetic enzymes, functional differences are often determined by a combination of factors, including local residues in the active pocket, substrate conformational constraints, and the stabilization mechanisms of reaction intermediates. Overall sequence similarity is insufficient to fully reflect the final product structure and catalytic outcome.

[0003] Terpene synthases can form highly diverse terpene skeletons through complex carbocation cyclization, migration, and rearrangement processes. These reactions are chemically spontaneous and involve complex pathways; even a single amino acid residue change in the active pocket can significantly alter the final product structure. Therefore, enzymes with high sequence similarity may produce different products, while enzymes with low sequence similarity may produce similar skeletons. This phenomenon makes it difficult for traditional sequence analysis methods to accurately predict terpene synthase function and limits the efficiency of discovering novel scaffold natural products. Therefore, traditional sequence analysis methods still suffer from insufficient accuracy in predicting terpene synthase product skeleton types, skeleton rearrangement capabilities, or other fine functional tags.

[0004] Machine learning and deep learning have become important techniques for protein function prediction. Existing methods often use protein sequences or structures as input, directly predicting functional labels after feature extraction or characterization learning, suitable for protein family classification, enzyme number annotation, or coarse-grained functional assessment. However, for product-directed proteins such as natural product biosynthetic enzymes, their functional categories are often determined by the structure, properties, or reaction outcomes of the catalytic product, such as product backbone type, product structural category, product bioactivity, product odor characteristics, or whether a specific backbone rearrangement reaction occurs. In this case, there is an important intermediate relationship between protein functional labels and product structures. If only protein information is used as model input, it is difficult to fully utilize the functional discrimination information contained in the product molecular structure, and it is also difficult to establish a correspondence between protein characterization and product characterization. Therefore, it is necessary to construct a protein-product multimodal computational model to improve the accuracy and applicability of fine-grained functional label prediction for unknown proteins. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of existing computational models that rely solely on protein sequence similarity or single-modal protein features for functional prediction, making it difficult to accurately identify low sequence similarity proteins, product-related functional tags, and protein-product associations. Therefore, this invention provides a protein function prediction system and method based on a protein-product multimodal computational model.

[0006] The objective of this invention can be achieved through the following technical solutions: One of the technical solutions of the present invention is to provide a protein function prediction system based on a protein-product multimodal computational model, comprising: The data acquisition module is used to acquire protein sequences, corresponding product molecular structure representations, and protein functional tags. The protein encoding module is used to generate an initial protein characterization based on the protein sequence; The product coding module is used to generate an initial characterization of the product based on the product molecular structure representation. The local feature extraction module is used to extract local protein features and local product features from the initial protein characterization and the initial product characterization, respectively. A cross-modal interaction module is used to fuse the local features of the protein and the local features of the product to obtain a joint protein-product characterization; The model training module is used to train a multimodal computation model based on product classification loss and the protein-product joint characterization alignment loss. The candidate scoring module is used to merge the protein to be tested with the candidate products in the data acquisition module one by one and input them into the multimodal calculation model to obtain the candidate score of each candidate product. The functional label determination module is used to determine the functional prediction result of the protein to be tested based on the candidate product with the highest candidate score and the preset product-functional label mapping relationship; The latent space analysis and results display module is used to perform dimensionality reduction analysis, cluster analysis and visualization of known proteins and test proteins based on the protein-product joint characterization, candidate scores and functional prediction results, and to display the position of the test protein in the latent representation space, its functional region, the ranking of candidate products, the predicted functional label and its relationship with the functional clusters of known proteins.

[0007] In some specific embodiments, in the data acquisition module, the product molecular structure is represented as the string SMILES, and the protein function tag is selected from any one or more of the following: backbone rearrangement tag and non-backbone rearrangement tag, product backbone type tag, product structure category tag, substrate preference category tag, and catalytic reaction category tag.

[0008] In some specific embodiments, the protein coding module processes the protein sequence obtained from the data acquisition module as follows: Remove non-standard amino acid characters, retain only the amino acid sequence in standard letter form, and truncate or padded the protein sequence to a preset maximum length to obtain the pre-processed protein sequence; The preprocessed protein sequence Input protein pre-trained encoder Initial protein characterization The protein pre-trained encoder uses the ESM-C protein language model, and the output initial protein representation dimension is ( , ,1152), of which For batch size, 1152 represents the length of the protein sequence and is the dimension of the fixed feature vector encoded by each amino acid residue in the ESM-C protein language model.

[0009] More preferably, the protein sequence is truncated or filled to a preset maximum length of 350 amino acid residues.

[0010] In some specific embodiments, the product encoding module processes the product molecular structure obtained from the data acquisition module as follows: The product structure encoder is used to segment words, and the SMILES string is truncated or padded to the preset maximum length to obtain the preprocessed product molecular structure. The molecular structure of the pretreated product Input product structure encoder Initial characterization of the product was obtained. , The product structure encoder uses the ChemBERTa-77M chemical language model to generate a contextual embedding representation of the product molecular structure.

[0011] As a better alternative, the SMILES string is truncated or padded to a preset maximum length of 128 tokens.

[0012] In some specific embodiments, in the local feature extraction module, the initial protein characterization obtained by the protein encoding module and the initial product characterization obtained by the product encoding module are respectively input into a multi-scale one-dimensional convolutional neural network to obtain local protein features. and local features of the product ,in , This represents the protein branch multi-scale convolution feature extraction function. For initial protein characterization, This represents the multi-scale convolution feature extraction function for product branches. This is the initial characterization of the product.

[0013] In some specific embodiments, the multi-scale one-dimensional convolutional neural network includes several convolutional branches with different kernel sizes, including 3, 5, and 7. Each convolutional branch includes at least one one-dimensional convolutional layer, a batch normalization layer, a ReLU activation function, and an adaptive max pooling layer.

[0014] The outputs of each convolution branch are concatenated and projected to a unified fusion dimension through a linear transformation, which is 1024 dimensions in the preferred implementation. Then, L2 normalization is performed to make protein characterization and product characterization located in a comparable shared latent space.

[0015] In some specific embodiments, the cross-modal interaction module processes the protein local features and product local features obtained from the local feature extraction module as follows: The protein local features and product local features are sequentially processed by a protein-product collaborative attention unit and a feature fusion unit. In the protein-product co-attention unit, First, the local features of the protein according to The product's local features according to Projecting to the same latent space yields the projected protein features. and the characteristics of the projected product ,in, This is a learnable weight matrix for protein features. Bias terms for protein characteristics; The learnable weight matrix for product features. This is a bias term for product characteristics; Next, multi-head scaled dot product attention is used to calculate the cross-modal interaction between proteins and products. For each attention head, the query matrix Q, bond matrix K, and value matrix V are obtained by linear transformation of the corresponding modal features, and are then processed according to... Calculate the attention output; where, The dimension of the key vector; Finally, the attention flow is output in two directions, with the first direction being the protein-to-product attention flow, based on protein features. As a query, product characteristics As bonds and values, protein characterization with enhanced product information is obtained: The second direction is the product-to-protein attention flow, based on product characteristics. As a query, protein characteristics As bonds and values, we obtain product characterization with enhanced protein information: ;in, Representation layer normalization and residual connections are used to stabilize the cross-modal feature update process; Protein characterization enhanced by collaborative attention and product characterization Protein branch representations were obtained using feature mapping networks. and product branching characterization and according to By concatenating along the feature dimensions, a concatenated representation is obtained. The splicing characterization Further input is a fusion network containing fully connected layers, ReLU activation function, and Dropout layers to extract nonlinear interaction information between protein and product features; simultaneously, learnable weight parameters ( Adaptive weighting of protein branching and product branching characterizations yields a weighted cross-modal characterization: ,

[0016] in, Furthermore, the contribution ratio of protein and product information in joint characterization is adaptively adjusted by a sigmoid function constraint; the cross-modal interaction features obtained by splicing and fusing are further... With the weighted cross-modal representation The fusion process was performed to obtain the final protein-product co-characterization. This is used for subsequent protein function prediction tasks.

[0017] Joint characterization of protein-product obtained from the cross-modal interaction module Input product classifier, according to Output product category logits Product category logits The product class probability is converted using the Softmax function; where, This is the weight matrix of the product classifier. This is the corresponding bias vector; The product category logits are further processed through a pre-constructed product-function label mapping matrix. Project to function label space ;in, For the function label logits, It is a fixed mapping matrix used to represent the correspondence between each product category and the functional label.

[0018] Therefore, the model does not rely directly on individual protein-to-label classification heads, but obtains protein function prediction results through the mapping relationship between product prediction results and product function labels.

[0019] In some specific implementations, the model training module employs a multi-objective loss function to simultaneously constrain product identification, functional label identification, and protein-product joint characterization alignment. The multi-objective loss function is: ;in, For product classification loss, To assist in supervising loss using functional labels, To monitor and compare losses, , , These are the loss weighting coefficients for product classification, functional label-assisted supervision, and supervision comparison, respectively. Product classification loss The cross-entropy loss is obtained using the cross-entropy loss method. in, The total number of functional label categories. This is a label for actual functionality. The product-functional label mapping matrix is ​​the first... Predicted probability of each functional label; The supervised comparison loss Used to bring together the joint characterization of proteins and products of the same functional class or experimental matching, and to bring together the joint characterization of proteins and products of different classes; for a batch of Each protein-product paired sample combines protein characterization and product characterization. One comparison sample, supervised comparison loss Represented as: ;in, Indicates the first Individual sample characteristics, Representation and Sample A set of positive samples with the same functional labels or matching relationships. Indicates except for samples The entire set of comparison samples outside of [the original text] Represents cosine similarity. This is the temperature coefficient.

[0020] In some specific implementations, during the training optimization process of the model training module, the training set is divided using hierarchical or grouped cross-validation to maintain the consistency of the distribution of different functional labels. The model parameters are preferably updated using the AdamW optimizer, and the initial learning rate is preferably 1×10⁻⁶. -5 The preferred weight decay coefficient is 1×10. -4 ; When the validation set loss no longer decreases within a preset number of epochs, the learning rate is dynamically reduced; when the validation loss continues to fail to improve after a preset number of training epochs, an early stopping mechanism is triggered, and the model parameters at the point of lowest validation loss are retained. This training strategy aims to reduce the risk of overfitting in small-sample multimodal training and improve the model's generalization ability.

[0021] In the unknown protein inference stage, a candidate product library is constructed, preferably composed of SMILES products appearing in the training set. For the protein to be tested... Normalized protein characterization is obtained through a protein encoder and a convolution adapter. For each candidate product in the candidate product library Normalized product characterization is obtained through a product structure encoder and a convolution adapter. First, the characterization retrieval similarity between the protein to be tested and each candidate product is calculated: This similarity score is used to indicate how close the test protein is to the candidate product in a shared characterization space.

[0022] Subsequently, the protein to be tested is combined with each candidate product to form a protein-product combination, which is then input into a complete protein-product multimodal computational model. After passing through a collaborative attention module, a feature fusion module, and a product classifier, the model outputs the predicted probability of the candidate product's own category, which serves as a self-consistency score for the candidate product. This score measures how well the complete joint model supports the matching relationship between the test protein and candidate product combinations.

[0023] For each candidate product, a joint candidate score is obtained by combining its representation retrieval similarity and model self-consistency score: ,in and These are the weighting coefficients for retrieval similarity and self-consistency scores, respectively. The candidate product with the highest joint candidate score is determined as the predicted product of the protein to be tested.

[0024] Finally, based on the product-functional label mapping relationship established during the training phase, the predicted product is... Functional prediction results of the target protein: .in This represents the mapping function from product to functional label. Therefore, the functional label prediction result of this invention is always associated with the predicted product identity, providing a more explicit explanation of the product's origin compared to a single label classifier.

[0025] In the results analysis phase, this invention can also perform dimensionality reduction and density analysis on the high-dimensional joint characterization of training samples and test samples. Preferably, t-SNE is used to map the high-dimensional protein-product joint characterization to a two-dimensional space: .in, This represents the dimensionality reduction mapping function. Further, a kernel density estimation function is employed. Density estimation of sample distribution in two-dimensional latent space: It identifies functional cluster boundary regions, low-density regions, or candidate uncertain regions based on preset density thresholds. The results can be used to display the functional region to which the protein to be tested belongs, its relative relationship with known protein functional clusters, candidate product ranking, and functional label prediction results.

[0026] The second technical solution of the present invention provides a protein function prediction method based on a protein-product multimodal computational model. Based on the protein function prediction system based on the protein-product multimodal computational model described in one of the above technical solutions, the protein function prediction method includes the following steps: S1, Obtain a training sample set, wherein each training sample in the training sample set includes a protein sequence, a product molecular structure representation corresponding to the protein sequence, and a protein functional tag, wherein the protein functional tag is determined by at least one of the product molecular structure, product chemical category, product backbone type, product properties, product biological activity, product application attributes, or enzyme catalytic attributes. S2, Input the protein sequence into the protein coding module to obtain the initial protein characterization; S3, Input the product molecular structure representation into the product molecular coding module to obtain the initial characterization of the product; S4, Local feature extraction is performed on the initial characterization of the protein and the initial characterization of the product to obtain local features of the protein and local features of the product; S5, input the local features of the protein and the local features of the product into the cross-modal interaction module to obtain joint protein-product characterization; S6. Based on the protein-product joint characterization, output the candidate product classification results, and train the multimodal computation model by combining product classification loss and protein-product joint characterization alignment loss. S7, combine the protein sequence to be tested with the molecular structure representations of multiple candidate products in the training product library one by one, and input them into the trained multimodal computing model to obtain the candidate score corresponding to each candidate product. S8. Select the candidate product with the highest candidate score as the predicted product of the protein to be tested, and obtain the protein function prediction result of the protein to be tested according to the preset product-functional tag mapping relationship.

[0027] Compared with the prior art, the present invention has the following advantages: (1) This invention constructs protein coding functions and product encoding function This approach incorporates protein sequence information and product molecular structure information into the same computational model, overcoming the problem that traditional sequence similarity analysis and protein single-modal machine learning methods cannot explicitly utilize product structure information.

[0028] (2) This invention extracts local features of proteins and products through multi-scale one-dimensional convolution functions, and uses protein-product collaborative attention functions. Establishing cross-modal interaction relationships enables the model to capture the potential correspondence between local features of protein sequences and local features of product structures.

[0029] (3) The present invention adopts 、 and The multi-objective loss function, composed of these components, simultaneously constrains product identification, functional label identification, and protein-product characterization alignment during the same training process, thereby improving the model's generalization ability on small sample protein-product pairing data.

[0030] (4) This invention constructs a joint candidate score by characterizing retrieval similarity and candidate self-consistency score. It can sort candidate products of unknown proteins and further determine the functional tags of unknown proteins based on the product-functional tag mapping relationship, thereby improving the efficiency of candidate protein screening.

[0031] (5) This invention is applicable to protein function prediction tasks defined by product structure, product properties, or catalytic results, including but not limited to prediction of terpene synthase backbone rearrangement ability, product backbone type, substrate preference, product structure category, product activity category, and odor-related function prediction. It has shown good results in predicting the ability of terpene synthase to catalyze backbone rearrangement products. It can improve the efficiency of candidate protein screening, reduce the number of blind experimental verifications, and is suitable for functional protein mining from large-scale genomic resources.

[0032] (6) This invention can show the position of the protein to be tested in the known functional space through potential space dimensionality reduction and kernel density estimation analysis, and help identify candidates within functional clusters, boundary candidates and low-density region candidates, thereby improving the efficiency of candidate protein selection and new functional protein discovery. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the construction and functional prediction of a protein-product multimodal characterization learning model.

[0034] Figure 2 Figure A shows the comparison diagram between the BLAST sequence similarity network of DvVS and the protein-product joint potential space, where Figure A is the BLAST method and Figure B is the method of the present invention.

[0035] Figure 3 The joint latent space, KDE density, and cluster distribution of 38 known and 161 uncharacterized Type B BFTPS are plotted.

[0036] Figure 4 The main product structure and new skeleton type of candidate BFTPS.

[0037] Figure 5 This is a comparison diagram of the potential space of a protein single-modal ablation model. Detailed Implementation

[0038] 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.

[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0040] Unless otherwise specified, the materials and processes described in the following embodiments or examples are conventional materials and processes used in the art to achieve the corresponding functions.

[0041] Example 1: First, in this embodiment, a bifunctional terpene synthase (BFTPS) from fungi is used as the protein function prediction target to construct a training dataset for bifunctional terpene synthase protein-product pairing.

[0042] BFTPS with clearly defined product structures and catalytic results were collected from published literature and experimental validation data. Each data point included at least the protein name, database accession number or data source, protein amino acid sequence, corresponding major terpene product name and molecular structure representation, and functional tag defined by the product skeleton and catalytic results.

[0043] In this embodiment, the product molecular structure is represented using the Simplified Molecular Linear Input Specification (SMILES) string. Protein functional tags are defined as skeletal-rearranging and non-skeletal-rearranging: when the major product generated by BFTPS undergoes a change in carbon skeleton linkage during the cyclization cascade, the protein is labeled as skeletal-rearranging BFTPS (SR-BFTPS); when the major product is formed by cyclization without the aforementioned carbon skeleton rearrangement, the protein is labeled as non-skeletal-rearranging BFTPS (NSR-BFTPS). In numerical encoding, SR-BFTPS is encoded as 1, and NSR-BFTPS is encoded as 0.

[0044] In this embodiment, a total of 75 BFTPS with experimentally validated cyclization products were collected, including 34 Type ABFTPS, 38 Type B BFTPS, and 3 atypical BFTPS. Among them, Type A BFTPS included 4 SR-BFTPS and 30 NSR-BFTPS; Type B BFTPS included 26 SR-BFTPS and 12 NSR-BFTPS. Since Type B BFTPS had a higher proportion of skeletal rearranged forms and a more direct correspondence between product skeletal differences and functional labels, 38 Type B BFTPS were selected as the main modeling objects in this embodiment.

[0045] Thirty-four protein-product pairing samples from these 38 Type B BFTPSs were used for model training, including 23 SR-BFTPSs and 11 NSR-BFTPSs. The remaining four BFTPSs were used for independent validation: FgGS, AbVS-1, NnNS, and CgDS, with the first three being SR-BFTPSs and CgDS being an NSR-BFTPS. The protein names, sequence numbers, major product SMILES strings, and functional tag information for these 38 BFTPSs are shown in Tables 1-1 and 1-2.

[0046] Table 1-1 Pairing data of 19 Type B BFTPS proteins and products

[0047] Table 1-2 Pairing data for 19 Type B BFTPS proteins and products

[0049] Then, in this embodiment, a protein-product multimodal computation model is used to train the model on 34 protein-product paired samples in the above-mentioned Type B BFTPS dataset. The other 4 BFTPS datasets are used for independent validation, namely FgGS, AbVS-1, NnNS and CgDS.

[0050] The protein-product multimodal computational model includes a protein sequence preprocessing module, a product SMILES preprocessing module, a protein language model encoding module, a chemical language model encoding module, a multi-branch one-dimensional convolutional local feature extraction module, a bidirectional multi-head collaborative attention module, a gating fusion module, and a product classification and functional label prediction module.

[0051] (a) Protein sequence preprocessing The standardization of protein amino acid sequences is performed according to the following steps: (1) Read the FASTA format sequence and convert it to uppercase characters; (2) Delete spaces, newlines, and tabs; (3) Using regular expressions or equivalent character checking programs, only 20 standard amino acid characters ACDEFGHIKLMNPQRSTVWY are retained; at the same time, B, J, O, U, X, Z and other non-standard characters are deleted; (4) Starting from the N-terminus, truncate the input sequence to a maximum of 350 amino acid residues; (5) For sequences with a length of less than 350, pad the C end with PAD markers to a uniform length and generate an attention mask so that the PAD position does not participate in subsequent feature calculation.

[0052] (ii) Pretreatment of product SMILES The processing of the product SMILES was completed using a word segmenter that is compatible with the ChemBERTa pre-trained model (ChemBERTa-77M-MLM).

[0053] The SMILES strings input to the model are pre-processed representations from the training data table. Subsequently, the ChemBERTa default tokenizer is used to segment the SMILES strings and automatically adds the special tokens required by the model, converting them into token IDs. The maximum token length is set to 128; SMILES sequences exceeding this length are truncated to the right, and any insufficient portion is padded with PAD tokens on the right, generating an attention mask. The final token IDs and attention masks are then fed into the ChemBERTa encoder as molecular-side inputs.

[0054] (III) Encoding of protein language models and chemical language models The standardized protein sequence processed in step (I) is input into the ESM-C protein language model, specifically using the ESMC_600M_202412 checkpoint, to obtain the context embedding representation of each amino acid position.

[0055] Input the normalized product SMILES from step (II) into the ChemBERTa-77M chemical language model to obtain the product structure context embedding representation at each token position.

[0056] Whether to freeze pre-training parameters, which hidden layer output to select, and whether to perform mixed precision calculation during model encoding should be filled in Table 2 according to the actual code.

[0057] Table 2 Specific parameters of the pre-trained encoder

[0058] (iv) Local feature extraction from multi-branch one-dimensional convolution The protein context embedding and product context embedding obtained in step (iii) are respectively input into a multi-branch one-dimensional convolutional neural network with the same structure and independent parameters.

[0059] Protein features refer to the local pattern representation of proteins obtained by extracting ESM-C positional embeddings through multi-branch one-dimensional convolution; product features refer to the local structural pattern representation of SMILES obtained by extracting ChemBERTa token-level embeddings through multi-branch one-dimensional convolution. The kernel sizes of each branch are 3, 5, and 7, used to capture local sequence or substructure information at different scales. Both the protein sequence branch and the SMILES branch adopt a multi-branch one-dimensional convolution structure. Each branch contains two sets of one-dimensional convolutions, batch normalization, and ReLU activation modules in sequence; the convolution operation uses padding="same" to keep the input position / token length unchanged. Subsequently, AdaptiveMaxPool1d(1) is executed at the end of each branch to compress the length dimension to 1 and convert it into a fixed-length vector through Flatten. After the outputs of each branch are concatenated along the channel dimension, they are mapped to a 1024-dimensional shared representation space through a fully connected projection layer and then L2 normalized. Specific parameters are shown in Table 3.

[0060] Table 3 Parameters of Multi-Branch One-Dimensional Convolutional Networks

[0061] (v) Two-way multi-head collaborative attention and gating fusion Let P be the protein feature matrix projected in step (iv), and M be the product feature matrix projected in step (iv). Each row of P corresponds to an amino acid position or a local position feature preserved by convolution, and each row of M corresponds to a SMILES token position or a local position feature preserved by convolution.

[0062] Bidirectional co-attention includes both protein-to-product and product-to-protein directions: (1) In the protein-to-product direction, P is used as query Q, M is used as key K and value V to obtain a protein representation with enhanced product information; (2) In the product to protein direction, M is used as query Q, P is used as key K and value V to obtain the product representation with enhanced protein information.

[0063] The collaborative attention mechanism employs eight attention heads, with a model dimension of 1024 and each attention head having a dimension of 128. The enhancement representations in both directions are residually connected and then concatenated before being input into the gated fusion layer. The gated fusion layer uses a sigmoid function to generate weight gating, enabling the model to adaptively adjust the contributions of protein and product information, resulting in a joint protein-product characterization. Specific parameters are shown in Table 4.

[0064] Table 4. Bidirectional collaborative attention and gating fusion parameters

[0065] (vi) Multi-objective loss function and training parameters The model training employs a multi-objective loss function, including product SMILES classification loss, function label-assisted classification loss, and supervised contrast loss. The product SMILES classification loss constrains the model to identify the true product category corresponding to the protein; the function label-assisted classification loss is calculated by projecting the product classification logits onto the SR / NSR label space via the SMILES-function label mapping matrix; the supervised contrast loss is used to narrow the distance between true matching protein-product pairs and samples of the same function category in the latent space, and to increase the distance between samples of different function categories. The total loss is expressed as L = λ1L_SMILES + λ2L_label + λ3L_SupCon, where λ1, λ2, and λ3 are filled in according to Table 5.

[0066] In a specific training parameter setting, the AdamW optimizer is used with an initial learning rate of 1×10⁻⁶. -5 The weight decay coefficient is 1×10 -4The batch size was 32, and the dropout was 0.4. Hierarchical five-fold cross-validation was used to ensure a relatively consistent ratio of SR-BFTPS and NSR-BFTPS across each fold. Dynamic learning rate adjustment and early stopping strategies were employed during training, and accuracy, precision, recall, and F1 score were recorded. Other parameters are shown in Table 5.

[0067] Table 5 Model Training Parameters

[0068] (vii) Model prediction performance and comparison with sequence similarity methods In this embodiment, the method of the present invention is compared with the traditional sequence similarity prediction method to verify the role of protein-product multimodal information in improving functional prediction.

[0069] Traditional sequence similarity comparisons employ the BLASTp nearest neighbor function for classification. Specifically, each BFTPS to be predicted is treated as a query and compared with all reported BFTPS in the Uniprot public database (https: / / www.uniprot.org / ). Matches are filtered based on E-value, sequence similarity, and coverage thresholds, and the nearest neighbor protein with the highest sequence consistency is selected, with its SR / NSR tag used as the label for the protein to be predicted. The prediction results based on BLAST nearest neighbor classification are shown in Tables 6-1 and 6-2, with an accuracy of 84.2% for the sequence similarity method.

[0070] Table 6-1 BFTPS Prediction Results Based on BLAST

[0071] Table 6-2 BFTPS Prediction Results Based on BLAST

[0073] When using the method of this invention, each protein to be tested is first paired one by one with candidate SMILES in the training product library. A joint protein-product representation is formed through a protein encoder, a product encoder, multi-branch convolution, bidirectional collaborative attention, and gating fusion. The product classifier outputs the candidate product category probability based on this joint representation and obtains the SR / NSR label probability through the SMILES-functional label mapping matrix. Therefore, the SR / NSR determination is based on the joint protein-product representation and the candidate product classification results, rather than solely on the protein single-modality representation.

[0074] The results of five-fold cross-validation show that the label prediction accuracy of the method of this invention in the Type B BFTPS functional prediction task is 92.5%-95.2%. The accuracy, precision, recall and F1 score of each fold should be supplemented according to Table 7 to form verifiable performance data.

[0075] Table 7. Training results of five-fold cross-validation

[0076] Example 2 This embodiment describes two characterization methods for the known variediene-producing enzyme DvVS. First, based on all known Type B BFTPS, an all-vs-all BLASTp is performed to construct a sequence similarity network (SSN), and then the network is configured according to a 1×10⁻⁶ pattern. -90 A threshold-based sequence similarity network was constructed; DvVS did not form a stable cluster with other variediene-producing enzymes in this network. Second, DvVS and its candidate product SMILES were input into the model constructed in Example 1, and protein-product co-characterization was extracted. This co-characterization was then combined with the co-characterization of other known proteins for t-SNE dimensionality reduction and clustering; DvVS was assigned to the variediene-related functional cluster. The complete processing procedure and quantitative results are available in [link to documentation]. Figure 2 .

[0077] Example 3 In this embodiment, the model trained in Example 1 was used for functional prediction of uncharacterized BFTPS. Candidate sequences were sourced from the Plant Endophytic Fungi Genome Database and the NCBI Protein Database. The Plant Endophytic Fungi Genome Database was a database built by the inventors' research group. The NCBI Protein Database was located at https: / / www.ncbi.nlm.nih.gov / protein / , and the search date was January 1, 2025.

[0078] The candidate library construction process includes: (1) The HMMER method is used to control redundancy based on a sequence consistency threshold of 90%; (2) Retain sequences with a length in the range of 550-800 amino acid residues; (3) Multiple sequence alignment was performed using ClustalW to confirm the integrity of the N-terminal terpene cyclase domain and the C-terminal isopentenyl transferase domain; (4) The integrity of the conserved motifs of the terpene cyclase domain DDXXD and NSE, and the isopentenyl transferase domain DDXXD and DDXDD was examined, and 160 candidate sources from plant endophytic fungi and 198 candidate sources from NCBI were obtained, for a total of 358. (5) Using MEGA v6.0, the maximum likelihood method was used to construct a phylogenetic tree. The bootstrap was set to 2000. BFTPS located on the Type B branch were screened for subsequent analysis, and 161 uncharacterized Type B BFTPS were finally obtained.

[0079] During the prediction phase, each BFTPS protein sequence to be tested is combined one by one with the candidate product SMILES from the training product library. For each protein-candidate product combination, the model calculates: (1) Cosine similarity S_cos between the characterization of the protein to be tested and the characterization of the candidate product; (2) After inputting the protein and candidate products into the complete model, the product classifier outputs the softmax probability S_self of the candidate product category. The joint candidate score is defined as S_joint=αS_cos+(1-α)S_self, where the weight α is 0.62. The model selects the candidate product with the highest S_joint as the predicted product and determines the SR / NSR label according to the SMILES-functional label mapping corresponding to the predicted product.

[0080] Furthermore, a latent space analysis was performed on 161 uncharacterized Type B BFTPSs and 38 known Type B BFTPSs. Specifically, a joint protein-product representation vector was extracted for each sample; L2 normalization was applied to both the protein-side enhanced representation and the product-side representation, and a joint representation was constructed based on the weight coefficients learned from the model, followed by another L2 normalization. To enhance the visual separation of product categories, the joint representation was further superimposed with the predicted label center vector, the predicted label random projection embedding, and the predicted SMILES random projection embedding, and the final feature was again L2 normalized. Subsequently, t-SNE was used to reduce the high-dimensional features to two dimensions; the perplexity of t-SNE was set to 35, the learning rate to 200, the number of iterations to 3000, the early exaggeration to 8, and the random seed to 35; the initialization method used the default settings of the current version of scikit-learn. To demonstrate the local sample density, density modeling was performed in two-dimensional space using Gaussian kernel density estimation, with a bandwidth of 0.5, and normalized density was calculated on a 100 × 100 grid. Low-density boundary candidates are defined as proteins whose sample point normalized KDE density is less than 0.05.

[0081] In this embodiment, 199 samples formed 14 functionally related clusters, clusters 1-14, with 13 uncharacterized candidates located outside the main clusters. Among the 38 known Type B BFTPS, 26 were SR-BFTPS and 12 were NSR-BFTPS; these two types of samples exhibited regional separation in the latent space. Proteins generating the same or similar product backbones tended to be distributed in the same or adjacent regions. Clusters 6 and 14 were dark clusters lacking known experimental characterization. The known representative enzymes, predicted tags, and corresponding products of each cluster are shown in Table 8, and the complete latent space map is shown in [reference needed]. Figure 3 .

[0082] Table 8 Functional cluster information for clusters 1-14

[0083] Based on the predicted tags and potential spatial distributions from the aforementioned model, six candidate proteins were selected from 161 uncharacterized Type B BFTPS for experimental validation. Among these, the model predicted three SR-BFTPS and three NSR-BFTPS. Candidate selection comprehensively considered predicted tags, their cluster location, whether they are located at functional boundaries, whether they belong to dark clusters, their potential distance from known enzymes, and their potential for discovering new product scaffolds. The complete list of these six candidate terpene synthases, their prediction results, selection criteria, and experimental results are shown in Table 9. Figure 4 .

[0084] Table 9. Results of experimental verification of the active candidate BFTPS and product detection.

[0085] It should be noted in the table that... Screening criteria: PCC, product consistency clustering verification.

[0086] The selected candidate BFTPS genes were synthesized according to the original genome sequence. Gibson assembly was used to construct the target gene into the pUARA4 vector, with expression elements including the promoter PamyB, the terminator TamyB, and an ampicillin selection marker. After confirming the correct insertion fragment by colony PCR and Sanger sequencing, the expression vector was transformed into the Aspergillus oryzae NSAR1 heterologous expression host via protoplast transformation. The selection medium and conditions were marked with an arginine nutrient curve. At least six independent transformants were obtained for each candidate, and an empty vector host was set up as a negative control.

[0087] Each independent transformant was first cultured in 100 mL of MPY liquid medium (3% maltose, 1% peptone, 0.5% yeast extract, 0.1% ammonium sulfate, 0.15% methionine, 0.01% adenine) at 30°C and 220 rpm for 48 h. Then, it was transferred at a 2% inoculum to large 1 L / 3 L Erlenmeyer flasks of MPY fermentation medium and cultured at 30°C and 220 rpm for 7 days. After fermentation, the fermentation broth was centrifuged at 3,000 rpm to collect the cells, and then subjected to ultrasonic-assisted extraction with acetone for 30 min. The extract was centrifuged again at 3,000 rpm, and the supernatant was collected and concentrated under reduced pressure. The concentrate was redissolved in water and ethyl acetate for partition extraction, and the ethyl acetate phase was collected and evaporated to dryness under reduced pressure to obtain the crude extract. The crude extract was dissolved in ethyl acetate to a final concentration of 50 mg / mL and then analyzed by GC-MS. Two biological replicates were set up for each candidate.

[0088] Analysis was performed using a SHIMADZU GC-MS system equipped with a DB-5 ms capillary column (30.0 m × 0.25 mm, film thickness 0.25 μm). The column temperature was initially set at 60 °C, increased to 280 °C at 25 °C / min, and then increased to 310 °C at 10 °C / min. The mass spectrometry conditions were an ion source temperature of 230 °C and an electron energy of 0.4 eV. To improve the separation of structurally diverse terpenoid products generated by FsPS and AcPS, optimized GC-MS temperature programming conditions were further employed. The column temperature was initially set at 60 °C, first increased to 165 °C at 10 °C / min, then increased to 200 °C at 1.5 °C / min and held for 6 min; subsequently increased to 240 °C at 4 °C / min, and then increased to 310 °C at 12 °C / min and held for 5 min.

[0089] For candidate proteins exhibiting a significant new product peak, a 50 L fermentation scale was performed. The crude extract was sequentially purified using silica gel column chromatography and preparative HPLC. The purified product was molecularly determined by HRMS and analyzed by 1D and 2D NMR (including...). 1 H, 13 The planar structure and relative configuration were determined by C, COSY, HSQC, HMBC, and NOESY; the absolute configuration was determined by ECD calculation / measurement comparison, optical rotation, and single-crystal X-ray diffraction.

[0090] The experimental results show that the model SR / NSR prediction labels of these six candidate BFTPSs are consistent with the analysis of the experimental products, as shown in Table 9. The NMR comparison tables of the experimental products and known compounds are shown in Tables 10-14.

[0091] Table 10 Compound 1 ( 1 H NMR (600 MHz),13 C NMR (150 MHz), pyridine- d 5) Comparison table with reported fusicoccadiene NMR data (400 MHz, CD2Cl2) in the literature

[0092] Table 11 Compound 2 ( 1 H NMR (600 MHz), 13 C NMR (150 MHz), pyridine- d 5) Comparison table with ophiobolin F NMR data (400 MHz, CD2Cl2) reported in the literature

[0093] Table 12 Compound 8 ( 1 H NMR (600 MHz), 13 C NMR (150 MHz), pyridine- d 5) Comparison table with reported mangicdiene NMR data (400 MHz, CDCl3) in the literature

[0094] Table 13 Compound 15 ( 1 H NMR (600 MHz), 13 C NMR (150 MHz), benzene- d 6) Comparison with reported spiroviolene NMR data (700 MHz, benzene- d 6) Comparison Table

[0095] Table 14 Compound 16 ( 1 H NMR (600 MHz), 13 C NMR (150 MHz), benzene- d 6) Compared with the variediene NMR data reported in the literature (700 MHz, benzene- d 6) Comparison Table

[0096] Comparative Example To verify the role of product structure information in the method of this invention, a protein single-modality control model was set up. This control model removed the ChemBERTa encoder, the product SMILES input, the product branch in the bidirectional collaborative attention, and the SMILES classification loss, while the remaining protein encoder, convolutional network, classifier, training set partitioning, optimizer, and training parameters remained consistent with the complete model.

[0097] The results showed that after removing the product structure information, some known BFTPS functions made incorrect judgments. For example... Figure 5 As shown in Table 15, it is known that NSR-BFTPS CgDS is assigned to the SR region; proteins that generate fusicoccadiene and ophiobolin F are incorrectly clustered in the same region; the aggregation relationship between variediene-producing enzymes is weakened; and some boundary candidates and isolated candidates that were clearly identifiable in the full model can no longer be effectively distinguished.

[0098] Table 15. Product types and prediction results of terpene synthases in each cluster.

[0099] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A protein function prediction system based on a protein-product multimodal computational model, characterized in that, include: The data acquisition module is used to acquire protein sequences, corresponding product molecular structure representations, and protein functional tags. The protein encoding module is used to generate an initial protein characterization based on the protein sequence; The product coding module is used to generate an initial characterization of the product based on the product molecular structure representation. The local feature extraction module is used to extract local protein features and local product features from the initial protein characterization and the initial product characterization, respectively. A cross-modal interaction module is used to fuse the local features of the protein and the local features of the product to obtain a joint protein-product characterization; The model training module is used to train a multimodal computation model based on product classification loss and the protein-product joint characterization alignment loss. The candidate scoring module is used to merge the protein to be tested with the candidate products in the data acquisition module one by one and input them into the multimodal calculation model to obtain the candidate score of each candidate product. The functional label determination module is used to determine the functional prediction result of the protein to be tested based on the candidate product with the highest candidate score and the preset product-functional label mapping relationship; The latent space analysis and results display module is used to perform dimensionality reduction analysis, cluster analysis and visualization of known proteins and test proteins based on the protein-product joint characterization, candidate scores and functional prediction results. It displays the position of the test protein in the latent representation space, its functional region, the ranking of candidate products, the predicted functional label and its relationship with the functional clusters of known proteins.

2. The protein function prediction system based on a protein-product multimodal computational model according to claim 1, characterized in that, In the data acquisition module, the product molecular structure is represented as the string SMILES, and the protein function tag is selected from any one or more of the following: backbone rearrangement tag and non-backbone rearrangement tag, product backbone type tag, product structure category tag, substrate preference category tag, and catalytic reaction category tag.

3. The protein function prediction system based on a protein-product multimodal computational model according to claim 1, characterized in that, In the protein coding module, the protein sequence obtained from the data acquisition module is processed as follows: Remove non-standard amino acid characters, retain only the amino acid sequence in standard letter form, and truncate or padded the protein sequence to a preset maximum length to obtain the pre-processed protein sequence; The preprocessed protein sequence Input protein pre-trained encoder Initial protein characterization The protein pre-trained encoder uses the ESM-C protein language model, and the output initial protein representation dimension is ( , ,1152), of which For batch size, 1152 represents the length of the protein sequence and the dimension of the fixed feature vector obtained by the ESM-C model encoding each amino acid residue.

4. The protein function prediction system based on a protein-product multimodal computational model according to claim 1, characterized in that, In the product encoding module, the product molecular structure obtained from the data acquisition module is processed as follows: The product structure encoder is used to segment words, and the SMILES string is truncated or padded to the preset maximum length to obtain the preprocessed product molecular structure. The molecular structure of the pretreated product Input product structure encoder Initial characterization of the product was obtained. , The product structure encoder uses the ChemBERTa-77M chemical language model to generate a contextual embedding representation of the product molecular structure.

5. The protein function prediction system based on a protein-product multimodal computational model according to claim 1, characterized in that, In the local feature extraction module, the initial protein representation obtained by the protein encoding module and the initial product representation obtained by the product encoding module are respectively input into a multi-scale one-dimensional convolutional neural network to obtain local protein features. and local features of the product ,in , This represents the protein branch multi-scale convolution feature extraction function. For initial protein characterization, This represents the multi-scale convolution feature extraction function for product branches. This is the initial characterization of the product.

6. The protein function prediction system based on a protein-product multimodal computational model according to claim 5, characterized in that, The multi-scale one-dimensional convolutional neural network includes several convolutional branches with different kernel sizes, including 3, 5, and 7. Each convolutional branch includes at least one one-dimensional convolutional layer, a batch normalization layer, a ReLU activation function, and an adaptive max pooling layer.

7. The protein function prediction system based on a protein-product multimodal computational model according to claim 1, characterized in that, In the cross-modal interaction module, the protein local features and product local features obtained from the local feature extraction module are processed as follows: The protein local features and product local features are sequentially processed by a protein-product collaborative attention unit and a feature fusion unit. In the protein-product co-attention unit, First, the local features of the protein according to The product's local features according to Projecting to the same latent space yields the projected protein features. and the characteristics of the projected product ,in, This is a learnable weight matrix for protein features. Bias terms for protein characteristics; The learnable weight matrix for product features. This is a bias term for product characteristics; Next, multi-head scaled dot product attention is used to calculate the cross-modal interaction between proteins and products. For each attention head, the query matrix Q, bond matrix K, and value matrix V are obtained by linear transformation of the corresponding modal features, and are then processed according to... Calculate the attention output; where, The dimension of the key vector; Finally, the attention flow is output in two directions, with the first direction being the protein-to-product attention flow, based on protein features. As a query, product characteristics As bonds and values, protein characterization with enhanced product information is obtained: The second direction is the product-to-protein attention flow, based on product characteristics. As a query, protein characteristics As bonds and values, we obtain product characterization with enhanced protein information: ;in, Representation layer normalization and residual connections are used to stabilize the cross-modal feature update process; Protein characterization enhanced by collaborative attention and product characterization Protein branch representations were obtained using feature mapping networks. and product branching characterization and according to By concatenating along the feature dimensions, a concatenated representation is obtained. The splicing characterization Further input is a fusion network containing fully connected layers, ReLU activation function, and Dropout layers to extract nonlinear interaction information between protein and product features; simultaneously, learnable weight parameters ( Adaptive weighting of protein branching and product branching characterizations yields a weighted cross-modal characterization: ， in, Furthermore, the contribution ratio of protein and product information in joint characterization is adaptively adjusted by a sigmoid function constraint; the cross-modal interaction features obtained by splicing and fusing are further... With the weighted cross-modal representation The fusion process was performed to obtain the final protein-product co-characterization. This is used for subsequent protein function prediction tasks.

8. The protein function prediction system based on a protein-product multimodal computational model according to claim 1, characterized in that, In the model training module, the protein-product joint characterization obtained from the cross-modal interaction module is performed. Input product classifier, according to Output product category logits Product category logits The product class probabilities are converted using the Softmax function; where, This is the weight matrix of the product classifier. This is the corresponding bias vector; The product category logits are further processed through a pre-constructed product-function label mapping matrix. Project to function label space ;in, For the function label logits, It is a fixed mapping matrix used to represent the correspondence between each product category and the functional label.

9. A protein function prediction system based on a protein-product multimodal computational model according to claim 1, characterized in that, In the model training module, a multi-objective loss function is used to simultaneously constrain product identification, functional label identification, and protein-product joint characterization alignment. The multi-objective loss function is: ;in, For product classification loss, To assist in supervising loss using functional labels, To monitor and compare losses, , , These are the loss weighting coefficients for product classification, functional label-assisted supervision, and supervision comparison, respectively. Product classification loss The cross-entropy loss is obtained using the cross-entropy loss method. in, The total number of functional label categories. This is a label for actual functionality. The product-functional label mapping matrix is ​​the first... Predicted probability of each functional label; The supervised comparison loss Used to bring together the joint characterization of proteins and products of the same functional class or experimental matching, and to bring together the joint characterization of proteins and products of different classes; for a batch of Each protein-product paired sample combines protein characterization and product characterization. One comparison sample, supervised comparison loss Represented as: ;in, Indicates the first Individual sample characteristics, Representation and Sample A set of positive samples with the same functional labels or matching relationships. Indicates except for samples The entire set of comparison samples outside of [the original text] Represents cosine similarity. This is the temperature coefficient.

10. A method for predicting protein function based on a protein-product multimodal computational model, characterized in that, A protein function prediction system based on a protein-product multimodal computational model according to any one of claims 1-9, wherein the protein function prediction method includes the following steps: S1, Obtain a training sample set, wherein each training sample in the training sample set includes a protein sequence, a product molecular structure representation corresponding to the protein sequence, and a protein functional tag, wherein the protein functional tag is determined by at least one of the product molecular structure, product chemical category, product backbone type, product properties, product biological activity, product application attributes, or enzyme catalytic attributes. S2, Input the protein sequence into the protein coding module to obtain the initial protein characterization; S3, Input the product molecular structure representation into the product molecular coding module to obtain the initial characterization of the product; S4, Local feature extraction is performed on the initial characterization of the protein and the initial characterization of the product to obtain local features of the protein and local features of the product; S5, input the local features of the protein and the local features of the product into the cross-modal interaction module to obtain joint protein-product characterization; S6. Based on the protein-product joint characterization, output the candidate product classification results, and train the multimodal computation model by combining product classification loss and protein-product joint characterization alignment loss. S7, combine the protein sequence to be tested with the molecular structure representations of multiple candidate products in the training product library one by one, and input them into the trained multimodal computing model to obtain the candidate score corresponding to each candidate product. S8. Select the candidate product with the highest candidate score as the predicted product of the protein to be tested, and obtain the protein function prediction result of the protein to be tested according to the preset product-functional tag mapping relationship.

Images