A protein function prediction method and system based on hard maximum propagation

Abstract

This invention provides a protein function prediction method and system based on hard maximum propagation, relating to the field of bioinformatics. In the GO map construction stage, a hierarchical index construction step based on hierarchical depth is added to support efficient topological traversal. In the GO term score propagation stage, a hard maximum propagation algorithm replaces the traditional average propagation algorithm, stipulating that the corrected score of a parent term is equal to the maximum of the scores of all its child terms and its own original score, thus strictly adhering to the biological hierarchical logic of GO terms mathematically. In engineering implementation, a single-protein grouped sparse dictionary processing mechanism is innovatively introduced, decomposing global dense matrix operations into independent sparse dictionary operations for individual proteins. After propagation, lag score filtering is performed, discarding only results with final scores below a preset threshold, ensuring the effective transmission of weak signals in the map while compressing the output file size.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a method and system for predicting protein function based on hard maximum propagation. Background Technology

[0002] Proteins are the main carriers of life activities, and accurate prediction of their functions is crucial for understanding life mechanisms, revealing disease patterns, and drug development. Gene Ontology (GO) is an internationally recognized gene function annotation system. Its terminology (GO terminology) forms a directed acyclic graph (DAG) through relationships such as "is_a" and "part_of," providing a structured knowledge framework for computational prediction.

[0003] Currently, protein function prediction based on GO terminology follows a core process commonly used in the industry, including: (1) processing GO terminology data sources; (2) constructing GO maps; (3) extracting protein sequence features and making initial predictions; (4) GO terminology score propagation; and (5) outputting prediction results. The core differences among industry methods are mainly reflected in the feature extraction and initial prediction model in step (3), while score propagation, as a key step to ensure that the prediction results conform to the GO hierarchical logic, directly affects the final prediction accuracy due to its algorithmic performance.

[0004] Currently, the closest existing technical solution (such as the excellent solution in the CAFA 5 competition) uses the average propagation algorithm for score propagation. This algorithm updates the parent node's score by calculating the average score of all child nodes under the parent node. However, this method has a fundamental flaw: it violates the "True Path Rule" in biology. This rule states that if a protein possesses a specific sub-function (such as high confidence), it must also possess the corresponding parent function; that is, the confidence of the parent function should not be lower than that of the sub-function. For example, if the parent node initially scores 0.2, and its two child node scores are 0.9 and 0.3 respectively, after average propagation, the parent node's score is (0.2+0.9+0.3) / 3≈0.467. This artificially reduces the support strength of the high-confidence sub-function (0.9) for the parent function, leading to biological logic distortion and reducing the accuracy of prediction.

[0005] Furthermore, in engineering implementation, traditional propagation methods typically require constructing a globally dense matrix of size "number of proteins × number of GO terms". Given the scale of competitions like CAFA or real-world research involving hundreds of thousands of proteins and tens of thousands of GO terms, this matrix operation consumes enormous amounts of memory, easily leading to Out of Memory (OOM) errors. To avoid OOM, some methods truncate and filter low-scoring terms before propagation, but this disrupts the propagation chain in the GO graph, causing loss of topological information and affecting propagation effectiveness. Summary of the Invention

[0006] To overcome the aforementioned shortcomings of existing technologies, this invention provides a protein function prediction method and system based on hard maximum value propagation. This invention aims to replace and optimize key steps such as score propagation in a targeted manner without altering the overall framework of the industry-standard protein function prediction process.

[0007] A protein function prediction system based on hard maximum propagation, whose specific implementation includes the following modules, corresponds to the industry's five-step general process: OBO File Parsing Module: Used to read standard GO OBO files and extract GO terms and their hierarchical relationships. This module's implementation logic is consistent with existing technologies, ensuring compatibility with standard data sources.

[0008] The GO Graph Construction Module is used to construct a Directed Acyclic Graph (DAG) of Go based on extracted terms and relations. Compared with existing technologies, this module adds an optimization step: Based on the DAG construction, all terms are hierarchically categorized according to the depth of GO terms, and terms at the same level are batch-numbered, thereby constructing a GO graph with a hierarchical index. This optimization aims to avoid redundant calculations in subsequent graph traversal processes and improve the efficiency of large-scale data processing.

[0009] Input file reading module: Used to read preliminary prediction result files generated by any initial prediction model (such as sequence alignment, machine learning, or deep learning models). This module is consistent with existing technology implementation logic, ensuring broad compatibility with various upstream prediction methods.

[0010] The Hard Maximum Propagation Module: This is the core difference between this invention and existing technologies. Based on the GO graph, this module propagates and corrects the predicted scores of sub-terms to their parent terms. Its specific technical implementation steps include: Graph traversal: Utilizing the constructed hierarchical GO graph index, starting from the lowest-level sub-terms, traverse upwards layer by layer to the top-level parent term, ensuring the correctness of the propagation order.

[0011] Hard maximum value calculation: For each parent term encountered, extract the confidence scores of all its direct child terms, and take the maximum value among all child term scores and the original score of the parent term as the final score of the parent term after propagation correction. The calculation formula is: Parent term corrected score = max(original score of parent term, score of child term 1, score of child term 2, …, score of child term n).

[0012] Batch Processing and Sparse Dictionary: To improve computational efficiency, this invention employs a batch processing mechanism based on protein grouping. The system groups proteins by their IDs. When processing each protein, only a sparse dictionary (scores = dict(zip(...))) needs to be maintained in memory, with GO term IDs as keys and scores as values, replacing the traditional global dense matrix. This reduces the memory complexity from O(N×M) to O(K), where N is the total number of proteins, M is the total number of GO terms, and K is the number of non-zero nodes predicted for a single protein, thus completely solving the memory overflow problem in large-scale data processing.

[0013] Lag score filtering: After the hard maximum propagation algorithm has been fully executed, the system filters out all prediction results with scores below a preset minimum threshold (e.g., 0.001). This "lag" strategy ensures the integrity of the propagation chain, allowing even weak signals (e.g., 0.002) to be effectively transmitted in the graph; filtering after propagation greatly reduces the size of the final output data while maintaining logical correctness.

[0014] The results output module is used to format and save the high-confidence prediction results obtained after hard-maximum propagation and hysteresis filtering into a standard output file. Compared with existing technologies that directly output all scores, this module only outputs meaningful high-confidence data, improving the reliability and usability of the results file.

[0015] Based on the above system, the present invention also provides a corresponding protein function prediction method.

[0016] The beneficial effects of this invention are: 1) Higher prediction accuracy: The hard maximum propagation algorithm strictly follows the biological hierarchical logic of GO (true path rules) at the mathematical level, avoiding the dilution of high-confidence information by the average propagation algorithm, and theoretically improving the reliability of the prediction results. 2) Significantly enhanced computational efficiency and robustness: Through a series of engineering optimizations such as hierarchical index traversal, single-protein sparse dictionary operations, and lag fraction filtering, the system can maintain extremely low memory usage and high computational performance when processing massive protein-GO data, avoiding the OOM risk of traditional methods and possessing the ability to process large-scale real-world datasets. 3) Good compatibility and ease of use: This invention follows industry-standard processes, and its input and output formats are compatible with existing technologies. It can be directly connected to various upstream initial prediction models, facilitating integration and application. Attached Figure Description

[0017] Figure 1 This is a flowchart of protein function prediction based on hard maximum propagation.

[0018] Figure 2 This is a comparison between our proposed solution and existing technologies. Detailed Implementation

[0019] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0020] OBO file parsing module Functional Description: This is the system's data source processing module, responsible for reading the standard format file (OBO file) of the Gene Ontology (GO) and extracting all functional terms and their hierarchical relationships.

[0021] In layman's terms: This module corresponds to the first step of the industry-standard process (GO terminology data source processing), and its implementation logic is consistent with existing technologies without any modifications, ensuring compatibility with GO standard OBO files.

[0022] GO Graph Construction Module Function Description: Based on the parsed GO terms and relations, construct a directed acyclic graph (DAG) structure to ensure that the hierarchical relationships between terms are clear and there are no circular references.

[0023] In layman's terms: This module corresponds to the second step of the industry-standard process (GO graph construction). The differences / modifications compared to existing technologies are: Existing technologies only construct conventional GO directed acyclic graphs without hierarchical sorting and indexing mechanisms; This invention, based on the construction of a DAG, adds a hierarchical sorting optimization mechanism—it hierarchically divides all terms according to the depth of GO terms, batch-numbers terms at the same level, and constructs a hierarchical GO graph index, avoiding redundant calculations in subsequent graph traversal processes and improving the efficiency of large-scale data processing.

[0024] Input file reading module Function Description: Responsible for reading the initial protein function prediction result files, which contain preliminary prediction scores for the functions that each protein may have.

[0025] In layman's terms: This module corresponds to the third step (protein sequence feature extraction and initial prediction) of the industry's common workflow. It is consistent with the existing technology implementation logic, without any changes, and is compatible with the initial prediction result format generated by all existing technologies (sequence alignment, machine learning, DeepGOPlus, etc.).

[0026] Hard maximize propagation module Functional Description: This is the core algorithm module of the system. It utilizes the hierarchical structure of the GO graph to propagate protein function prediction scores upwards, ensuring that the prediction score of the parent function is not lower than that of the child function.

[0027] In layman's terms: The specific technical implementation steps of this module are as follows: ① Graph traversal: Based on the index of the hierarchical GO graph, starting from the lowest-level sub-terms of the GO term, traverse upwards to the top-level parent term to ensure the order of propagation; ② Hard maximum calculation: For each parent term, traverse all its direct sub-terms, extract the confidence score of each sub-term, and use the maximum value as the confidence score of the parent term after propagation correction. If there are no sub-terms under the parent term, its initial confidence score is retained; ③ Batch processing: Perform batch hard maximum calculation on GO terms of the same level to reduce the computation time of single-term traversal; ④ Score filtering: Set a flexibly adjustable confidence score threshold to filter prediction results with scores below the threshold after propagation correction, and only retain high confidence scores.

[0028] This module corresponds to the fourth step (GO terminology score propagation) in the industry-standard process. It represents the core technology replacement and optimization of the existing technology in this invention. The differences / modifications between this invention and the existing technology are as follows: The existing technology uses an average propagation algorithm (to calculate the average score of subterms) and lacks batch processing and score filtering mechanisms. This invention replaces the average propagation algorithm with a hard maximum value propagation algorithm and adds two optimization mechanisms: batch processing (to improve efficiency) and score filtering (to improve the reliability of results). This is the core manifestation of the innovation of this invention.

[0029] This invention proposes a protein function prediction system based on Hard-Max propagation. During the topology propagation phase, the system employs a Hard-Max algorithm, ensuring that the score of a parent node is strictly equal to the maximum of its initial score and the scores of all its child nodes, i.e., the Score. parent =max(Score parent Score child This perfectly aligns with the real-world path rules in biology at the mathematical level.

[0030] In terms of engineering implementation, this invention introduces a batch processing mechanism of "single-protein grouped sparse dictionary". The system reduces the dimensionality of global dense matrix operations to independent sparse dictionary processing for each single protein by grouping the prediction results by protein ID (Groupby). When traversing the topological sort for score propagation, memory usage is only related to the number of effective labels for a single protein, thereby reducing space complexity by orders of magnitude and completely solving the memory overflow problem under large-scale prediction. At the same time, this invention adopts a "hysteresis score filtering" mechanism, setting a minimum threshold (e.g., 0.001), and removing redundant scores only after ensuring that the hard maximum value topological propagation is fully executed. This approach not only ensures the effective transmission of weak signals in the spectrum but also greatly compresses the data volume of the final result.

[0031] Result Output Module Function Description: Formats and saves the optimized prediction results as a standard output file for easy analysis and submission.

[0032] In layman's terms: This module corresponds to the fifth step (prediction result output) of the industry's common process. The difference / modification from the existing technology is that the existing technology only performs simple formatting on all scores after propagation; the present invention only formats the high-confidence data after hard maximum value propagation and score filtering, removes low-confidence invalid data, and improves the reliability of the output results.

[0033] Data flow and workflow, such as Figure 1 As shown.

[0034] 1. Data Input: The system obtains input data from the OBO file and the initial prediction result file. 2. Graph Construction: Parse the OBO file and construct a hierarchical GO graph. 3. Score Optimization: Optimize the initial prediction results using the GO graph. 4. Output Results: Save the optimized results as a standard format file. The above workflow corresponds perfectly to the industry-standard 5-step protein function prediction workflow. Only the corresponding steps have had their algorithms replaced and optimization mechanisms added. The workflow is seamless and can be directly connected to the initial prediction results of existing technologies.

[0035] The key points and protected aspects of this invention are all implemented through specific technical means, focusing on the key aspects of common industry processes. A comparison is made with the CAFA 5 competition-winning solution, which is most similar to this invention, as detailed below: A GO terminology score propagation method based on the hard maximum propagation algorithm The implementation scheme of this invention is as follows: In the GO term score propagation stage, the algorithm traverses from the bottom sub-terms to the top parent term. For each parent term, the maximum value of the confidence scores of all its direct sub-terms is taken as the confidence score of the parent term after propagation correction. The algorithm logic is: parent term score = max (sub-term 1 score, sub-term 2 score, ..., sub-term n score).

[0036] The existing implementation scheme is as follows: In the GO term score propagation stage, the average propagation algorithm is adopted. For each parent term, the arithmetic mean of the confidence scores of all its direct child terms is calculated as the confidence score of the parent term after propagation correction. The algorithm logic is: Parent term score = (Child term 1 score + Child term 2 score + ... + Child term n score) / n.

[0037] The specific differences between the two are: the core difference lies in the computational logic of the propagation algorithm. This invention uses "maximum value" while the existing technology uses "average value". The algorithm of this invention conforms to the hierarchical biological logic of GO terminology "if the child function exists, the parent function must exist", which fundamentally avoids the score dilution problem of the existing technology.

[0038] Hierarchical Go methods for constructing directed acyclic graphs The implementation scheme of this invention is as follows: In the GO graph construction stage, the extracted GO terms are layered according to the level depth, and the terms at the same level are batch numbered to construct a GO directed acyclic graph with a level index, thereby realizing fast traversal of the graph.

[0039] The existing implementation scheme is as follows: In the GO graph construction stage, only a regular GO directed acyclic graph without index is constructed, and the terms are not sorted hierarchically. The graph traversal requires enumerating the terms one by one.

[0040] The specific differences between the two are as follows: This invention adds hierarchical sorting and index construction techniques, which are absent in existing technologies. This improvement avoids repetitive calculations in graph traversal and enhances the efficiency of large-scale data processing.

[0041] A Go terminology score propagation optimization mechanism that integrates batch processing and score filtering The implementation scheme of the present invention is as follows: In the hard maximum value propagation stage, batch hard maximum value calculations are performed on GO terms at the same level to reduce the calculation time of a single term; after the propagation is completed, a confidence score threshold is set to filter out prediction results below the threshold and retain only high confidence data.

[0042] The existing implementation scheme is as follows: In the average propagation stage, the average value is calculated by traversing each term one by one without a batch processing mechanism; after propagation, there is no score filtering step, and all prediction results are directly output.

[0043] The specific differences between the two are as follows: This invention adds two technical means, batch processing and fractional filtering, which are not available in the prior art. Batch processing further improves the computational efficiency, and fractional filtering removes invalid data and improves the reliability of the prediction results.

[0044] Overall comparison conclusion: The core difference between this invention and the prior art is "core algorithm replacement + key link optimization". It does not change the overall process of protein function prediction commonly used in the industry, has strong technical compatibility, and all improvements are specific and implementable technical means, rather than simple functional enhancements. It provides a clear technical solution to solve the core defects of the prior art.

[0045] Average Propagation Algorithm: When processing GO (Gene Ontology) hierarchical structures, this algorithm typically averages (or weights averages) the predicted probabilities of child nodes with the parent node's own probability to update the parent node's final probability. Example: Suppose a protein has a parent functional node P with an initial predicted score of 0.2. This parent node has two child nodes: child node C1 with a predicted score of 0.9 and child node C2 with a predicted score of 0.3.

[0046] Average propagation calculation: P = (0.2 + 0.9 + 0.3) / 3 = 0.467.

[0047] Flaw: This violates the "True Path Rule" in biology. The True Path Rule states that if a protein has a specific sub-function (with a probability of 0.9), then it must also have the corresponding parent function (with a probability of at least 0.9). Average propagation lowers the parent node score to 0.467, leading to a logical error in biology.

[0048] Hard-Max propagation: It strictly follows the True Path Rule P = max(P, C1, C2). In the example above, P = max(0.2, 0.9, 0.3) = 0.9.

[0049] The pain points of conventional methods (traditional matrix processing): Traditional propagation algorithms typically construct a massive dense matrix of protein-GO nodes. In large-scale predictions, there are often hundreds of thousands of proteins and over forty thousand GO nodes, and matrix multiplication can lead to extremely high memory consumption (OOM, Out of Memory). Forcibly truncating scores before propagation to save space would disrupt the propagation chain at the ontology level.

[0050] like Figure 2 As shown, the core technology lies in dictionary sparse propagation based on topological sorting + lag filtering: 1. Groupby Dictionary: Treat each protein as an independent batch. In memory, a tiny "sparse dictionary" is built only for the single protein being processed (scores = dict(zip(...))). This means that the memory complexity drops dramatically from O(N × M) to O(K) (where K is the number of non-zero nodes predicted for a single protein), completely eliminating memory overflow.

[0051] 2. Lagged Score Filtering: The code sets MIN_SCORE_THRESHOLD = 0.001, and the filtering operation is only performed after the topology propagation is completely finished. This ensures that even very weak prediction signals (such as 0.002) will not be prematurely discarded during propagation, as long as they can be transmitted to the parent node through the DAG graph; and redundant scores <0.001 are removed after propagation, greatly compressing the size of the final output file.

[0052] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A protein function prediction method based on hard maximum propagation, characterized in that, The method includes the following steps: GO graph construction steps: Read the term definition file of the Gene Ontology (GO), construct a directed acyclic graph (DAG) according to the hierarchical relationship between terms, and construct a hierarchical index for all terms according to the hierarchical depth of the terms in the DAG to generate a GO graph with hierarchical index. Score propagation correction step: Obtain the initial prediction score of the protein to be predicted for each GO term; for each parent term in the GO map, obtain the current prediction scores of all its direct child terms, calculate the maximum value among the current prediction scores of the parent term and all its direct child terms, and use the maximum value as the final prediction score of the parent term after propagation correction. The execution order of the fraction propagation correction step is based on the hierarchical index of the GO graph, starting from the lowest-level sub-terms and traversing upwards layer by layer to the highest-level parent term.

2. The protein function prediction method based on hard maximum propagation according to claim 1, characterized in that, In the fraction propagation correction step, a sparse dictionary operation based on protein grouping is used: Group the initial predicted scores by protein ID; For each protein, a sparse dictionary is built in memory with GO terms as keys and prediction scores as values; When processing a single protein, the fractional propagation correction operation of claim 1 is performed using the sparse dictionary and the hierarchical index of the GO graph.

3. The protein function prediction method based on hard maximum propagation according to claim 1, characterized in that, Its features are, The method further includes: Lag filtering step: After completing the score propagation correction step for all terms, a score threshold is set to filter out predictions with a final predicted score lower than the threshold.

4. A protein function prediction system based on hard maximum propagation, characterized in that, The system includes: The GO graph construction module is configured to read the term definition file of the Gene Ontology (GO), construct a directed acyclic graph (DAG) based on the hierarchical relationship between terms, and perform hierarchical and indexed construction of all terms according to the hierarchical depth of the terms in the DAG to generate a GO graph with hierarchical index. The hard maximum value propagation module is configured to perform the fraction propagation correction step as described in claim 1 or 2; The result output module is configured to output the predicted score after correction by the hard maximum value propagation module.

5. The protein function prediction system based on hard maximum propagation according to claim 4, characterized in that, The hard maximum value propagation module further includes: The grouping unit is configured to group the initial prediction scores by protein ID and create an independent sparse dictionary for each protein to store its initial prediction score. The traversal calculation unit is configured to use the hierarchical index generated by the GO graph construction module to traverse the terms in the GO graph in order from bottom to top, and perform a hard maximum value calculation operation on each parent term.

6. The protein function prediction system based on hard maximum propagation according to claim 4, characterized in that, The system also includes: The hysteresis filtering module, connected to the hard maximum propagation module, is configured to filter out prediction results whose final prediction scores are lower than a preset threshold after propagation is complete.

Images