Function prediction method based on migration lifelong learning and three-level structure information

Abstract

The invention discloses a function prediction method based on migration lifelong learning and three-level structure information. Protein data is processed and divided into three GO subdata sets, and three-dimensional structure features, sequence features and PPI topological features are constructed; the features are integrated to form a complete protein interaction network; inputting into a GAT model, outputting enhanced graph sharing features, and constructing a Fisher information matrix at the same time; a GAT model and three classifiers are combined and trained, the classifiers calculate and output predicted values through a linear layer, final loss is obtained according to task specific loss, elastic weight and regularization loss, a prediction model is updated, and finally updated parameters and feature information are stored in a Fisher information matrix; and inputting the protein needing to be predicted into the prediction model, and outputting a prediction result. According to the method, the problem of disastrous forgetting in multi-task learning is solved, the model design is effectively simplified, the waste of computing resources is reduced, and the prediction efficiency and accuracy are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a function prediction method based on transfer lifelong learning and tertiary structure information. Background Art

[0002] The function prediction method based on transfer lifelong learning and tertiary structure information is an integrated lifelong learning framework that aims to achieve cross-task knowledge sharing and transfer and solve the problem of catastrophic forgetting by training and predicting the three sub-ontologies of Gene Ontology (GO) in sequence.

[0003] Traditional protein function prediction methods usually rely on sequence similarity, which is less reliable for proteins with little or no sequence similarity. Current methods for predicting protein function rely on multiple sources of information, such as sequence, interaction, protein structure or gene ontology data.

[0004] Patent application number: 202310026625.4, a process disclosed in a protein function prediction method based on multi-source heterogeneous information aggregation: 1. Collect PPI network topological features of different species from the STRING database, and use the "comprehensive score" to represent the interaction confidence between proteins to construct a PPI network. 2. Based on the proteins collected in step 1, use BLAST to construct an SSN network with a "similarity score", where the smaller the similarity score, the more similar the protein sequences are. 3. The structural domains, subcellular locations, and signaling pathways of proteins are used as node features to construct feature matrices in PPI and SSN networks. 4. Using the GraphMAE framework, through mask operations and GCN encoders, the potential representation of each protein node in the PPI network is generated, and the node features are restored through a re-masked decoder. 5. The pre-trained ESM-1b model is applied to extract protein sequence features and map them to the same vector space as the PPI and SSN networks through a fully connected layer. 6. PPI, SSN and sequence features are integrated through a fully connected classifier (with Sigmoid activation function) to perform protein function prediction, and consensus representation is used to improve prediction stability.

[0005] Patent application number: 202410170159.1, the disclosed process of the protein function prediction method based on multimodal multi-attention fusion features: 1. Use one-hot encoding to encode the amino acid sequence of the protein to obtain a matrix of size 1000×21, and after nonlinear mapping of the fully connected layer, obtain a feature matrix of size 1000×d, represented as a d-dimensional vector x. 2. Construct the position code p through trigonometric functions, and add it to the initial feature x of the amino acid to obtain the feature representation Xs containing position information; then splice the aggregated representation vector Vs to the head of Xs to obtain the initial feature X of the protein sequence. 3. Use the multi-head self-attention mechanism to extract the semantic features of each amino acid from the initial feature X of the protein sequence and aggregate them into a representation vector. 4. Extract the three-dimensional structural information of the protein through the structural information encoding module based on the contact graph to obtain the structural feature representation matrix Xg. 5. Design the SC cross-modal cross-attention fusion network to fuse the semantic features with the structural features Xg, and mine the deep information through the multi-head self-attention network, and finally aggregate them into the representation vector V. The final deep feature S and aggregation vector V are obtained by connecting three stages in series. 6. The aggregation vector V is passed into the label predictor to obtain the final prediction result.

[0006] The obvious drawback is that these methods do not pay attention to the problem of catastrophic forgetting, that is, learning a new task will lead to the loss of previously acquired knowledge. These methods ignore the potential for knowledge sharing and transfer between different protein functions, limiting the prediction accuracy and generalization ability of the model. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a function prediction method based on transfer lifelong learning and tertiary structure information, which solves the problem of catastrophic forgetting in multi-task learning, effectively simplifies model design, reduces the waste of computing resources, and significantly improves prediction efficiency and accuracy.

[0008] In order to solve the above technical problems, the present invention provides a function prediction method based on transfer lifelong learning and tertiary structure information, characterized in that it includes the following steps:

[0009] Step 1: protein data were processed and divided into three GO sub-datasets;

[0010] Step 2: construct protein three-dimensional structural features, construct protein sequence features, and construct PPI topological features of protein interaction relationships;

[0011] Step 3: Integrate the features into a graph structure to form a complete protein interaction network;

[0012] Step 4: Input the interaction network into the GAT model to output the enhanced graph sharing features, and construct and initialize the Fisher information matrix;

[0013] Step 5. Each GO sub-dataset is equipped with a classifier, and the classifier is trained to obtain a prediction model combining the GAT model and the three classifiers; wherein the classifier calculates the enhanced graph shared features of each task input through a linear layer and outputs a prediction value, calculates the task-specific loss according to the prediction value, and combines the task-specific loss with the elastic weight and regularization loss to obtain the final loss, updates the prediction model through the final loss, and stores the parameters and feature information of the prediction model in the Fisher information matrix; the elastic weight and regularization loss are calculated through the Fisher information matrix;

[0014] Step 6: Input the protein to be predicted into the prediction model and output the prediction result.

[0015] Furthermore, in step 1, protein datasets and GO were first generated from UniProtKB / Swiss-Prot, and then the protein datasets were divided by reading GO and protein datasets first, compiling relevant annotations and interaction information of molecular function sub-datasets, cellular component sub-datasets, and biological process sub-datasets, filtering out proteins that did not belong to GO, and generating three valid protein sub-datasets: molecular function sub-datasets, cellular component sub-datasets, and biological process sub-datasets.

[0016] Furthermore, in step 2, PDB files of different proteins with high-precision protein three-dimensional structures are first collected from the AlphaFold database, and then features related to the protein three-dimensional structure are extracted therefrom. The features include atomic coordinates, secondary structure information and the distance matrix between amino acid residues, and are input into the geometric vector perceptron GVP model to obtain a 20-dimensional binary vector for each protein, thereby constructing the protein three-dimensional structural features.

[0017] Furthermore, in step 2, the sequences of different proteins are extracted from the FASTA file and then input into the ESM-2 pre-trained model to obtain a 2560-dimensional sequence embedding for each protein to construct protein sequence features.

[0018] Furthermore, in step 2, the PPI network structures of different proteins were collected from the protein interaction relationship database STRING, and then the interactions with confidence scores less than 0.7 were filtered out, and finally the retained interactions were converted into STRING database identifiers using the mapping provided by UniProtKB to generate protein interaction relationships in order to construct the PPI topological features of protein interaction relationships.

[0019] Furthermore, in step 3, the three-dimensional structural features of the protein and the protein sequence features are spliced ​​to obtain a 2580-dimensional sequence embedding for each protein, which is used as a graph node in the protein interaction network. Each graph node represents a protein, and the protein interaction relationships constitute the edges in the network, thus forming a complete protein interaction network.

[0020] Furthermore, in step 4, the graph feature G is output by the number of layers of the GAT model con : G con =GATConv(G Relationship ,MLPBlock(G Structure )), where G Structure is a structural feature, G Relationship It is a relational characteristic;

[0021] And introduce the multi-head attention mechanism to output the enhanced graph shared feature G shared :

[0022] G shared =MultiHeadAttention(G con ).

[0023] Furthermore, the classifier uses a linear layer to generate the predicted value of the task output as Y task =Sigmoid(W task G shared ), W task is the weight matrix, G shared To enhance the graph sharing features;

[0024] Calculate the task-specific loss L by using the predicted value and the true value task , using the binary cross entropy loss function: as follows Among them, is y i The true label, p i is the probability of sample prediction, N is the number of samples;

[0025] Among them, the elastic weight combined regularization term is calculated through the Fisher information matrix, and the elastic weight combined regularization loss is as follows:

[0026]

[0027] Among them, λ ewc is the dynamically adjusted regularization coefficient, F ij is the Fisher information matrix, θ ij is the model parameter, θ ij old is the old model parameter, and the final loss is: L = L task +L ewc .

[0028] Furthermore, in step 5, in multi-task learning, after each task is completed, the parameters and feature information of the prediction model are stored in the Fisher information matrix, and the Fisher information matrix is ​​updated by calculating the square gradient of the parameters in the Fisher information matrix, so as to transfer and retain the feature information between tasks, and the prediction model applies the knowledge from the previous task through the updated Fisher information matrix.

[0029] Furthermore, in step 5, at the final stage of predictive value output, the hierarchical structure of GO is utilized to ensure that the score of each GO term reflects not only the direct predictive value but also the score of the corresponding ancestral node;

[0030] For each GO term, first obtain the prediction score s t , and then the prediction score s of each GO term t Propagate to all its ancestor terms a, and the score of the ancestor term a is updated to the maximum value between the two:

[0031] s a =max(s a ,s t ), s a Give the prediction score for the ancestral term.

[0032] Beneficial effects of the present invention:

[0033] The integrated modeling framework of the present invention can train and predict the three sub-ontologies of GO in sequence, thereby simplifying the complexity of the traditional method that requires establishing an independent model for each sub-ontology; by introducing a lifelong learning mechanism, the catastrophic forgetting problem that occurs when the model learns a new task is successfully addressed, allowing the model to accumulate and retain knowledge across tasks, thereby improving the overall performance; by training in the same model, knowledge sharing and transfer between the three sub-ontologies are achieved, significantly improving the prediction accuracy and generalization ability; by performing multi-task training in a single model, the generalization ability of the model in multi-task scenarios is enhanced, enabling it to better cope with various prediction challenges.

[0034] Compared with the traditional independent modeling method, the model of the present invention reduces redundant design and wasted computing resources while improving prediction efficiency. It is not only an effective method to solve catastrophic forgetting in multi-task learning, but also demonstrates its strong potential. This method simplifies model design, reduces the waste of computing resources, and significantly improves prediction efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a step diagram of the prediction method of the present invention. DETAILED DESCRIPTION

[0036] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0037] Reference Figure 1 As shown, an embodiment of the function prediction method based on transfer lifelong learning and tertiary structure information of the present invention downloads the UniProtKB / Swiss-Prot2024_05 version protein data set and GO (GO), and filters the experimental function annotations with evidence codes EXP, IDA, IPI, IMP, IGI, IEP, TAS, IC, HTP, HDA, HMP, HGI and HEP to filter all proteins, and a total of 78,292 protein records are obtained; then the data set is divided, the GO and protein data sets are read, the relevant annotations and interaction information of CC, BP ​​and MF are compiled, and proteins that do not belong to GO are filtered out, and the processing is divided into three GO sub-datasets: molecular function sub-dataset (MF), cellular component sub-dataset (CC) and biological process sub-dataset (BP). In the CC data set, the number of terms is 2,836 and the number of proteins is 58,839. In the BP data set, the number of terms is 20,724 and the number of proteins is 60,145. In the MF dataset, the number of terms is 7,013 and the number of proteins is 44,798.

[0038] Then construct protein three-dimensional structural features and protein sequence features; collect PDB files of different proteins with high-precision three-dimensional protein structures from the AlphaFold database, and then extract features related to protein three-dimensional structure, including atomic coordinates, secondary structure information, and distance matrix between amino acid residues. Input the features into the GVP model to obtain a 20-dimensional binary vector for each protein, which is the protein three-dimensional structural feature, denoted as Feature_3D. Extract the sequences of different proteins from the FASTA file, input the sequences into the ESM-2 pre-trained model, and obtain a 2560-dimensional sequence embedding for each protein, which is the protein sequence feature, denoted as Feature_Seq. Combine the protein three-dimensional structural features and protein sequence features as nodes in the protein interaction network:

[0039] G Structure =concat(Feature_Seq,Feature_3D)

[0040] Each node represents a protein.

[0041] The PPI network structures of different proteins were collected from the protein interaction relationship database (STRING), and then the interactions with confidence scores less than 0.7 were filtered out. Finally, they were converted into STRING database identifiers using the mapping provided by UniProtKB to generate protein interaction relationships, represented as Feature_PPI, which were used as edges of the protein interaction network:

[0042] G Relationship =Feature_PPI

[0043] Each edge represents that the two connected nodes have an interaction relationship, forming a protein interaction network.

[0044] The protein interaction diagram is:

[0045] G=(G Structure ,G Relationship )

[0046] Input to the GAT model:

[0047] G con =GATConv(G Relationship ,MLPBlock(G Structure ))

[0048] Among them, the number of neural network layers is expanded, and the graph features enhanced by the multi-head attention mechanism are introduced:

[0049] G shared =MultiHeadAttention(G con )

[0050] Among them G Structure is a structural feature, G Relationship is the relationship feature, G shared It is an enhanced graph sharing feature.

[0051] Use a simple linear layer to generate task specific output.

[0052] Each GO sub-dataset is equipped with a classifier, which is trained to obtain a prediction model that combines the GAT model with three classifiers. After the prediction model is obtained, the protein to be predicted is input into the prediction model, and the prediction result is output, which is an end-to-end output mode.

[0053] During the training process of the prediction model, the classifier calculates the shared features of the enhanced graph input for each task through a linear layer and outputs a predicted value. The task-specific loss is calculated based on the predicted value, and the task-specific loss is combined with the elastic weight and regularized loss to obtain the final loss. The prediction model is updated through the final loss and the parameters and feature information of the prediction model are stored in the Fisher information matrix; the elastic weight and regularized loss are calculated through the Fisher information matrix.

[0054] The predicted value is Y task =Sigmoid(W task G shared )W task is the weight matrix, G shared To enhance the graph sharing features;

[0055] Calculate the task-specific loss L by using the predicted value and the true value task , using the binary cross entropy loss function: as follows Among them, is y i The true label, p i is the probability of sample prediction, N is the number of samples;

[0056] Among them, the elastic weight combined regularization term is calculated through the Fisher information matrix, and the elastic weight combined regularization loss is as follows:

[0057]

[0058] Among them, λ ewc is the dynamically adjusted regularization coefficient, F ij is the Fisher information matrix, θ ij is the model parameter, θ ij old is the old model parameter, and the final loss is: L = L task +L ewc .

[0059] In the final prediction stage, the hierarchical structure of GO is used to ensure that the score of each GO term reflects not only its direct prediction value, but also the scores of its ancestor nodes. The regularization term is incorporated by calculating the elastic weight through the Fisher information matrix to reduce the impact of training different GO sub-datasets. For each GO term t, its prediction score s is first obtained. t Then, t Propagate the prediction score of each GO term t to all its ancestor terms a. The score of the ancestor term is updated to the maximum value between the two:

[0060] s a =max(s a ,s t )

[0061] Among them, s a Score the predictions for the ancestral terms, used to get more comprehensive predictions.

[0062] In multi-task learning, the prediction model must effectively integrate and apply knowledge from previous tasks. This requires episodic memory and a mechanism for updating the knowledge base. After learning each task, the parameters and feature information of the prediction model are stored in the Fisher information matrix. The Fisher information matrix is ​​updated by calculating the square gradient of the parameters in the Fisher information matrix, thereby effectively transferring and retaining feature information between tasks. By accumulating the square gradients of each parameter, saving the parameter state of the prediction model for the current task, and normalizing the Fisher information, it is ensured that the prediction model can utilize previous learning results in new task training, improve catastrophic forgetting, and improve the performance and robustness of multi-task learning. The updated formula is:

[0063] F new =F old +g 2

[0064] Where g represents the gradient, F old Represents the Fisher information matrix of the previous task. The normalized formula is:

[0065]

[0066] Where N represents the total number of training data.

[0067] During training, the present invention sets the learning rate to 1e-5 and uses the Adam optimizer for optimization. In order to improve the convergence of the prediction model in the later stage of training, the present invention uses the MultiStepLR scheduler to reduce the learning rate to 0.1 times the original in the 20th, 40th, 60th, and 80th rounds, respectively. At the same time, the EarlyStopping mechanism is set. When the performance of the prediction model on the validation set has not been significantly improved within 15 consecutive rounds, the training is stopped in advance to prevent overfitting. The batch size of the training is set to 8, and a total of 512 rounds of training are performed. All experiments are run on a server equipped with a 3090 graphics card and 40.0G video memory.

[0068] The GO sub-datasets were trained and tested in different orders to evaluate the effects of different parameter transfers on protein function and the forgetting problem of different tasks, and compared with representative existing advanced methods. The effectiveness of each module was analyzed through experiments. For each GO sub-dataset, it was divided into training set and test set according to the ratio of 90% and 10%, and the validation set was based on 10% of the training set. The GO sub-datasets were trained and tested in different orders to evaluate the effects of different parameter transfers on protein function and the forgetting problem of different tasks, and compared with representative existing advanced methods. The effectiveness of each module was analyzed through experiments.

[0069] The present invention evaluates the three sub-ontologies of GO and uses models that perform different tasks step by step. Specifically, MODEL_MF completes MF feature learning, MODEL_MF_CC completes MF and CC feature learning, and MODEL_MF_CC_BP completes MF, CC and BP feature learning. From Tables 2, 3, and 4, it can be seen that the model of the present invention is superior to the baseline method in all three evaluation indicators, and it can be clearly seen from Table 3 that after the model completes the MF and CC tasks, the performance of the CC task is significantly improved, which reflects the positive transfer effect between the MF task and the CC task.

[0070] For protein function evaluation, the present invention uses protein-centric evaluation indicators, maximum F value (Fmax), area under the regional accuracy-recall curve (AUPR) and category center area under the receiver operating characteristic curve (AUC), which are standardized by the Critical Assessment of Functional Annotation (CAFA) Challenge. See Tables 5 to 7:

[0071] Fmax is the maximum value of the F1 score, which measures the comprehensive performance of the classification model's accuracy and recall. The F1 score is the harmonic mean of the accuracy and recall.

[0072]

[0073] The average AUC (area under the curve) is an indicator for evaluating the ability of a classification model to distinguish between different categories. AUC measures the accuracy of the probability ranking predicted by the model.

[0074]

[0075] AUPR evaluates the performance of the model under all possible precision-recall (PR) curves. A high AUPR value indicates that the model performs well in precision and recall at each threshold.

[0076]

[0077] Table 1: Dataset

[0078]

[0079]

[0080] Table 2: MF comparison

[0081]

[0082] Table 3: CC comparison

[0083]

[0084]

[0085] Table 4: BP comparison

[0086]

[0087]

[0088] Table 5: MF ablation experiment

[0089]

[0091] Table 7 BP ablation experiment

[0092]

[0093] The prediction model of the present invention significantly mitigates catastrophic forgetting by effectively retaining and transferring knowledge across tasks. The excellent performance of the prediction model in various evaluation metrics is evident, especially after multi-task training. The prediction model is able to capture and integrate common features between MF, CC, and BP tasks, which not only improves prediction accuracy but also enhances its generalization ability. These findings emphasize the importance and effectiveness of the lifelong learning framework of the present invention in bioinformatics, especially for protein function prediction and gene ontology annotation.

[0094] The above embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Any equivalent substitution or change made by a person skilled in the art based on the present invention is within the protection scope of the present invention.

Claims

1. A function prediction method based on transfer lifelong learning and tertiary structure information, characterized in that: The following steps are involved: Step 1: protein data were processed and divided into three GO sub-datasets; Step 2: construct protein three-dimensional structural features, construct protein sequence features, and construct PPI topological features of protein interaction relationships; Step 3: Integrate the features into a graph structure to form a complete protein interaction network; Step 4: Input the interaction network into the GAT model to output the enhanced graph sharing features, and construct and initialize the Fisher information matrix; Step 5. Each GO sub-dataset is equipped with a classifier, and the classifier is trained to obtain a prediction model combining the GAT model and the three classifiers; wherein the classifier calculates the enhanced graph shared features of each task input through a linear layer and outputs a prediction value, calculates the task-specific loss according to the prediction value, and combines the task-specific loss with the elastic weight and regularization loss to obtain the final loss, updates the prediction model through the final loss, and stores the parameters and feature information of the prediction model in the Fisher information matrix; the elastic weight and regularization loss are calculated through the Fisher information matrix; Step 6: Input the protein to be predicted into the prediction model and output the prediction result.

2. The function prediction method based on transfer lifelong learning and tertiary structure information as claimed in claim 1, characterized in that: In step 1, protein datasets and GO were first generated from UniProtKB / Swiss-Prot, and then the protein datasets were divided. The GO and protein datasets were read first, and the relevant annotations and interaction information of the molecular function sub-dataset, cellular component sub-dataset, and biological process sub-dataset were compiled. Proteins that did not belong to GO were filtered out, and three valid protein sub-datasets were generated: molecular function sub-dataset, cellular component sub-dataset, and biological process sub-dataset.

3. The function prediction method based on transfer lifelong learning and tertiary structure information as claimed in claim 1, characterized in that: In step 2, PDB files of different proteins with high-precision protein three-dimensional structures are first collected from the AlphaFold database, and then features related to the protein three-dimensional structure are extracted from them. The features include atomic coordinates, secondary structure information, and the distance matrix between amino acid residues. These features are input into the geometric vector perceptron GVP model to obtain a 20-dimensional binary vector for each protein, thus constructing the protein three-dimensional structure features.

4. The function prediction method based on transfer lifelong learning and tertiary structure information as claimed in claim 1, characterized in that: In step 2, the sequences of different proteins are extracted from the FASTA file and then input into the ESM-2 pre-trained model to obtain a 2560-dimensional sequence embedding for each protein to construct protein sequence features.

5. The function prediction method based on transfer lifelong learning and tertiary structure information as claimed in claim 1, characterized in that: In step 2, the PPI network structures of different proteins were collected from the protein interaction relationship database STRING, and then the interactions with confidence scores less than 0.7 were filtered out. Finally, the retained interactions were converted into STRING database identifiers using the mapping provided by UniProtKB to generate protein interaction relationships in order to construct the PPI topological features of protein interaction relationships.

6. The function prediction method based on transfer lifelong learning and tertiary structure information as claimed in claim 1, characterized in that: In step 3, the three-dimensional structural features of the protein and the protein sequence features are spliced ​​to obtain a 2580-dimensional sequence embedding for each protein, which is used as a graph node in the protein interaction network. Each graph node represents a protein, and the protein interaction relationship constitutes the edge in the network, thus forming a complete protein interaction network.

7. The function prediction method based on transfer lifelong learning and tertiary structure information as claimed in claim 1, characterized in that: In step 4, the graph feature G is output by the number of layers of the GAT model con : G con =GATConv(G Relationship ,MLPBlock(G Structure )), where G Structure is a structural feature, G Relationship It is a relational characteristic; And introduce the multi-head attention mechanism to output the enhanced graph shared feature G shared : G shared =MultiHeadAttention(G con )。 8. The function prediction method based on transfer lifelong learning and tertiary structure information as claimed in claim 1, characterized in that: The classifier uses a linear layer to generate the predicted value of the task output as Y task =Sigmoid(W task G shared ), W task is the weight matrix, G shared To enhance the graph sharing features; Calculate the task-specific loss L by using the predicted value and the true value task , using the binary cross entropy loss function: as follows Among them, is y i The true label, p i is the probability of sample prediction, N is the number of samples; Among them, the elastic weight combined regularization term is calculated through the Fisher information matrix, and the elastic weight combined regularization loss is as follows: Among them, λ ewc is the dynamically adjusted regularization coefficient, F ij is the Fisher information matrix, θ ij is the model parameter, θ ij old is the old model parameter, and the final loss is: L = L task +L ewc .

9. The function prediction method based on transfer lifelong learning and tertiary structure information as claimed in claim 1, characterized in that: In step 5, in multi-task learning, after each task is completed, the parameters and feature information of the prediction model are stored in the Fisher information matrix, and the Fisher information matrix is ​​updated by calculating the square gradient of the parameters in the Fisher information matrix, so as to transfer and retain the feature information between tasks, and the prediction model applies the knowledge from the previous task through the updated Fisher information matrix.

10. The function prediction method based on transfer lifelong learning and tertiary structure information as claimed in claim 1, characterized in that: In step 5, at the final stage of predictive value output, the hierarchical structure of GO is utilized to ensure that the score of each GO term reflects not only the direct predictive value but also the score of the corresponding ancestral node; For each GO term, first obtain the prediction score s t , and then the prediction score s of each GO term t Propagate to all its ancestor terms a, and the score of the ancestor term a is updated to the maximum value between the two: s a =max(s a ,s t ), s a Give the prediction score for the ancestral term.

Images