A cas protein intelligent comprehensive evaluation system and method

Abstract

This invention discloses an intelligent comprehensive evaluation system and method for Cas proteins, belonging to the interdisciplinary field of protein engineering and artificial intelligence. The evaluation system utilizes LLM (Limited Learning Model) to parse user natural language commands, extracting intent tags and constraint thresholds. It relies on a workflow dynamic distribution control engine, combined with a Model Context Protocol (MCP) to schedule underlying evaluation tool plugins. A short-circuit protection mechanism is set in the execution chain, automatically blocking subsequent high-computing-power plugin calls when the current screening result is zero. The evaluation system performs feature truncation and dimensionality reduction on the underlying high-dimensional complex data, generating a professional evaluation report from the LLM. Furthermore, a visualization workbench provides reverse configuration of evaluation parameters and full data display. This invention achieves fully automated scheduling and computational optimization from natural language to complex bioinformatics pipelines, significantly improving the efficiency and accuracy of Cas protein screening.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of protein engineering and artificial intelligence, and specifically relates to a smart comprehensive evaluation system and method for Cas proteins. Background Technology

[0002] The CRISPR-Cas system, as a revolutionary gene-editing tool, has demonstrated enormous application potential in biomedical research, agricultural breeding, and industrial biotechnology. Cas proteins are the core effector proteins of the CRISPR-Cas system, and their functional characteristics directly determine the efficiency, specificity, and applicability of gene editing. With the continuous advancement of gene-editing technology, traditional natural Cas proteins have gradually revealed limitations in practical clinical and scientific applications, such as being restricted by specific PAMs, having excessively large molecular sizes leading to difficulties in viral vector delivery, and exhibiting potential off-target effects. Therefore, mining novel Cas proteins from metagenomic data or rationally designing and evaluating them through protein engineering has become a core frontier issue in the current gene-editing field, breaking through patent barriers and technological bottlenecks.

[0003] The comprehensive evaluation of novel Cas proteins is a highly complex, multi-dimensional, and computationally intensive process. Currently, researchers typically need to conduct systematic verification at multiple levels, including sequence homology, characteristic domains, three-dimensional conformation, and functional activity. This requires the cascading use of a series of underlying bioinformatics computing nodes with different functions. However, existing Cas protein bioinformatics evaluation paradigms suffer from the following significant technical shortcomings and pain points, greatly limiting the efficiency of discovering and developing novel Cas proteins: (1) The tools are highly fragmented and the technical threshold is extremely high. The heterogeneous bioinformatics tools and deep learning models mentioned above are usually scattered across different computing platforms or open-source communities, with huge differences in the underlying operating environment. Researchers who lack programming skills find it difficult to independently complete the complex end-to-end evaluation process. (2) Data flow is fragmented and lacks cross-platform automated scheduling. The input and output formats of various underlying computing tools are incompatible with each other. Data transfer between upstream and downstream links relies heavily on manually written scripts for cleaning and format reshaping, which easily leads to data silos and slows down the screening efficiency; (3) The analysis process is rigid and lacks intention-based intelligent task orchestration. Most existing automated pipelines are based on rigid topologies with static code, which cannot understand unstructured natural language instructions and are difficult to realize intelligent processes such as dynamic distribution of computing power, on-demand single-point calling of tools, and automatic triggering of short-circuit protection; (4) High-dimensional results are difficult to interpret, and users cannot intuitively understand the results. The data generated by multimodal tools is huge and complex, and traditional solutions cannot perform effective feature dimensionality reduction and human-readable natural language summarization. In addition, there is a lack of a unified workbench to support the high degree of freedom configuration of underlying bioinformatics evaluation parameters and the full export and analysis of the original high-dimensional data.

[0004] In summary, there is an urgent need in this field for a novel evaluation architecture that can break down the barriers between heterogeneous bioinformatics tools, accurately understand natural language intent, and achieve intelligent scheduling across the entire process from sequence cleaning to structure and function, thereby completely reshaping the traditional, cumbersome bioinformatics analysis workflow. This would improve the success rate and efficiency of Cas protein discovery and design, and further advance CRISPR gene editing technology. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a smart comprehensive evaluation system and method for Cas proteins, which solves the problems in existing technologies.

[0006] The objective of this invention can be achieved through the following technical solutions: A smart comprehensive assessment method for Cas proteins includes the following steps: Acquire natural language commands input by the user and protein sequence data to be evaluated; The natural language instructions are input into a pre-configured large language model for intent parsing, and structured intent parameters are extracted. The intent parameters include: task execution intent labels and constraint thresholds. Based on the task execution intent tag, following the preset plug-in calling protocol, the corresponding independent evaluation tool plug-in is scheduled and activated from the tool library, and the execution chain of each independent evaluation tool plug-in is determined. The protein sequence data is input into the activated independent evaluation tool plugin and evaluated in conjunction with the constraint threshold. In the case where the execution chain includes multiple serially scheduled independent evaluation tool plugins, a short-circuit protection step is performed: obtain the set of reserved sequences output by the preceding independent evaluation tool plugin; if the number of sequences in the set of reserved sequences is zero, then intercept and cancel the call instructions to subsequent independent evaluation tool plugins in the execution chain. After the independent evaluation tool plugin is executed, the high-dimensional multimodal analysis data output is obtained. In the backend runtime environment, the high-dimensional multimodal analysis data is subjected to feature truncation and dimensionality reduction processing to extract the preset core quantitative feature values ​​and reshape them into flat structure data. The flattened data is fed back to the large language model to generate a global evaluation report based on natural language.

[0007] Furthermore, after acquiring the protein sequence data to be evaluated, and before inputting the protein sequence data into the activated independent evaluation tool plugin, the method further includes: The file stream of the protein sequence data is matched and parsed using regular expressions to remove header annotation information and whitespace characters from the protein sequence data. The stripped data is cleaned of illegal characters and formatted, then converted into a standardized string array, which serves as the unified input format for each independent evaluation tool plugin in the execution chain.

[0008] Furthermore, the independent evaluation tool plugin includes a basic screening unit based on sequence characteristics, and the constraint threshold includes a sequence length parameter; the step of invoking the basic screening unit includes: The input protein sequence data is filtered according to the sequence length parameter, and sequences containing non-standard amino acids are removed. Identify and remove low-complexity sequences where the proportion of a single amino acid exceeds a preset composition threshold, wherein the preset composition threshold is included in the constraint threshold; Non-compliant sequences are removed according to the k-mer repetition rules: 1-mer can be repeated a maximum of 6 times consecutively, 2-mer can be repeated a maximum of 4 times consecutively, 3-mer to 6-mer can be repeated a maximum of 3 times consecutively each, and 7-mer can be repeated a maximum of 2 times consecutively.

[0009] Furthermore, the independent evaluation tool plugin also includes a domain identification unit based on Blast / HMM; the steps for invoking the domain identification unit include: Calculate the family coverage of the input sequence and retain candidate sequences that satisfy the query coverage, target coverage, and alignment score all being greater than the corresponding constraint thresholds; Based on the comparison of the built-in structural domain configuration file using the Hidden Markov Model, candidate sequences with both target HMM structural domain recognition coverage and HMMER score greater than the corresponding constraint threshold are retained.

[0010] Furthermore, the independent evaluation tool plugin also includes a PAM compatibility and specific function activity prediction unit; the step of invoking the PAM compatibility and specific function activity prediction unit includes: Deep learning was used to validate the prediction of PAM sequences for protein sequence identification using a validation model. Output the PAM probability distribution matrix containing the nucleotide probabilities at each position; Given the target gRNA sequence and the target DNA sequence, output the cleavage activity score and the probability distribution of editing efficiency.

[0011] Furthermore, the high-dimensional multimodal analysis data includes three-dimensional atomic coordinate data, and the core quantization feature value includes the average pLDDT score obtained by calculating the confidence level of the three-dimensional atomic coordinates.

[0012] Furthermore, the method also includes: The raw, high-dimensional multimodal analysis data without dimensionality reduction is sent to the front-end interactive interface for visualization rendering; Receive evaluation parameter configuration instructions based on feedback from the front-end interactive interface, and update the constraint threshold accordingly to re-trigger the evaluation process.

[0013] A Cas protein intelligent comprehensive assessment system, performing the above method, includes: The user input module is used to receive natural language evaluation instructions and protein sequence data files; The intent parsing and data preprocessing module is used to extract task execution intent tags and constraint thresholds from the natural language evaluation instructions, and to clean and format the received protein sequence data. The workflow dynamic scheduling module has a built-in intent-based workflow dynamic distribution control engine, which is used to schedule the evaluation units in the automated parallel evaluation toolset module according to the task execution intent tag and the model context protocol; and to execute a short-circuit protection mechanism: when there are multiple serially scheduled evaluation units and the number of sequences output by the previous evaluation unit is zero, the call to the subsequent evaluation unit is intercepted. The automated parallel evaluation toolset module consists of multiple evaluation units encapsulated as independent tool plugins. These evaluation units are used to perform evaluation calculations on the input protein sequence data in conjunction with the constraint thresholds. The summary and output layer module is used to perform feature truncation and dimensionality reduction on the high-dimensional multimodal analysis data generated by each evaluation unit, extract core quantitative feature values ​​and reshape them into flat structure data, and then send them back to the large language model to generate a global evaluation report. The professional visualization analysis workbench module is used to present raw, high-dimensional multimodal analysis data without dimensionality reduction and to receive reverse configuration instructions for evaluation parameters.

[0014] Furthermore, the evaluation unit includes: a basic screening unit based on sequence characteristics, a domain identification unit based on Blast / HMM, a three-dimensional spatial conformation folding prediction unit based on deep learning, a protein structure similarity search and function mining unit, and a PAM compatibility and specific functional activity prediction unit.

[0015] A computer device, comprising: a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the steps of the intelligent comprehensive evaluation method for Cas proteins as described above when executing the computer program.

[0016] The beneficial effects of this invention are: 1. This invention breaks through the limitations of traditional static hard-coded pipelines, and for the first time encapsulates heterogeneous bioinformatics software such as BLAST, HMMER, ESMFold, and Protein2PAM with deep learning models into a standard MCP independent tool plugin. Through dynamic invocation of the core control engine, it achieves seamless cross-platform scheduling of computing power and adaptive aggregation of multimodal data, completely eliminating the "data silos" between bioinformatics tools.

[0017] 2. This invention constructs an intent parsing hub with a large language model as its core. Researchers do not need to have complex programming skills or memorize lengthy command line parameters. They only need to input unstructured natural language commands, and the system can automatically complete intent extraction, constraint parameter generation, and routing and distribution of the underlying evaluation link, realizing truly zero-threshold intelligent interaction.

[0018] 3. This invention establishes a systematic evaluation loop encompassing basic sequence features (k-mer, amino acid composition), domain homology, three-dimensional spatial conformation (pLDDT), and functional activity (PAM matrix, editing efficiency). Furthermore, relying on a dynamic workflow distribution control engine, the system not only supports multi-link sequential execution but also allows decoupled calls for single functions and incorporates a short-circuit protection interception mechanism (stopping subsequent calculations when all current parameters are filtered), significantly saving hardware computing resources and time costs.

[0019] 4. To address the technical bottleneck of context overflow in large models caused by massive multimodal data, this invention performs feature dimensionality reduction and data interception on the backend to ensure the generation of accurate, human-readable summary reports. At the same time, it supports the export of the full format of the original high-dimensional analysis data without dimensionality reduction through a professional visual analysis workbench (Web frontend), and supports researchers to reverse-intervene and adjust the underlying screening thresholds, perfectly balancing the automation of AI and the precision requirements of scientific research. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the overall architecture and multi-link evaluation process of a novel intelligent comprehensive evaluation system for Cas proteins provided by the present invention; Figure 2 This is a flowchart of the workflow steps of a novel intelligent comprehensive evaluation method for Cas proteins provided by the present invention; Figure 3 This is a schematic diagram of the cross-platform scheduling logic of the intent-based workflow dynamic distribution and MCP tool provided by the present invention; Figure 4 This is a schematic diagram of the natural language interaction and intent extraction interface of the large language model provided by the present invention; Figure 5 This is a schematic diagram of the interface of the natural language global evaluation report generated by the summary and output layer module provided by the present invention; Figure 6 This is a schematic diagram of the highly flexible configuration interface for creating and evaluating analysis tasks in the professional visual analysis workbench module provided by this invention. Figure 7 A schematic diagram of the interface for displaying the underlying full analysis data and exporting high-dimensional raw data in the professional visualization analysis workbench module provided by this invention; Figure 8 This is a schematic diagram of the three-dimensional spatial conformation folding visualization results generated by the prediction of candidate Cas proteins provided by the present invention. Figure 9 This is a schematic diagram of the PAM probability distribution matrix for the prediction output of candidate Cas proteins provided by the present invention. Figure 10 A schematic diagram of the PAM feature visualization results of a specific Cas protein (such as SpCas9) provided by this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 like Figure 1 As shown, a novel intelligent comprehensive evaluation system for Cas proteins based on the fusion of a large language model and the MCP protocol includes: The user input module includes a dual-track parallel branch: one track is a natural language instruction input branch, used to receive unstructured evaluation instructions from the user; the other track is a biological sequence file input branch, used to receive files containing protein sequences. The intent parsing and data preprocessing module, with natural language interaction as its core, relies on the large language model to intelligently identify user needs and automatically extract evaluation parameters. It includes a large language model (LLM) intent parsing unit for extracting user intent and constraint parameters from natural language instructions, and a data processing and standardization unit for cleaning illegal characters and uniformly formatting and reshaping sequence files. The workflow dynamic scheduling module has a built-in intent-based workflow dynamic distribution control engine and a heterogeneous sequence data adaptive aggregation mechanism. This module uses the user intent parsed from the large language model as decision instructions, and innovatively combines the model context protocol to use the underlying heterogeneous bioinformatics computing nodes as external callable tools for the intelligent agent to perform automated, cross-platform dynamic calls and combinations. The automated parallel evaluation toolset module consists of multiple evaluation units encapsulated as independent MCP tool plugins. It is controlled by the control engine and activated and called on demand according to user intent. It relies on a built-in comprehensive database containing protein sequence and structural information such as Cas9 and Cas12 to evaluate the characteristics related to Cas proteins. Specifically, it includes five independent MCP tool plugins: a basic screening unit based on sequence characteristics, a domain identification unit based on Blast / HMM, a three-dimensional spatial conformation folding prediction unit based on deep learning, a protein structure similarity search and function mining unit, and a PAM compatibility and specific functional activity prediction unit. The summary and output layer module includes result feature extraction and LLM summary output layer, which is used to perform unified structured transformation and dimensionality reduction on the high-dimensional analysis data generated by each evaluation link, and generate professional evaluation reports from the large language model. The professional visualization analysis workbench module provides a web interface and communicates bidirectionally with the modules mentioned above. It is used to provide highly flexible configuration of evaluation parameters and full export of high-dimensional analysis data, effectively meeting the needs of refined operation in complex scientific research scenarios.

[0024] in: The biological sequence file input branch in the user input module supports multimodal reception of heterogeneous biological data: it supports receiving non-standard sequence files in formats including but not limited to FASTA and TXT, or text containing known protein names; in the data processing and standardization unit, redundant header annotation information and whitespace are automatically removed through regular expressions and file stream parsing technology, and the heterogeneous sequence data is uniformly formatted and reshaped into a system-standardized string array format for downstream modules to aggregate and call.

[0025] The operation mechanism of the intent parsing and data preprocessing module is as follows: After receiving natural language instructions, this invention limits the role and task of the large model through prompt words to shield against interference from irrelevant data, and configures a preset task intent enumeration set, a default fault tolerance mechanism for parameter boundaries, and a few-shot reference example. After parsing based on the above prompt words, the large language model is required to output machine-readable structured data (preferably in JSON format), specifically including task execution intent labels, sequence length thresholds, and preference parameters such as computing power scheduling, which serve as the underlying machine-readable constraint instructions for subsequent routing and distribution. For example, in a non-limiting specific implementation, the prompt words instruct the model to match intents from a preset set of analysis tasks. If the user inputs "perform a full analysis of the sequence, minimum length 100", the large model outputs a structured parameter object similar to {"task_intent": "Full_pipeline", "min_length": 100, "max_length": 9999, "use_gpu": true} based on preset rules and fault tolerance mechanisms. This mechanism effectively ensures the accuracy of downstream system scheduling and its anti-hallucination capability, but the specific implementation of this invention is not limited to the aforementioned specific prompt text and JSON key names.

[0026] The operation mechanism of the intent-based workflow dynamic distribution control engine in the workflow dynamic scheduling module is as follows: the task execution intent tags parsed from the large language model are used as decision instructions, and each evaluation unit in the automated parallel evaluation toolset module is regarded as an independently decoupled external tool library; the system supports targeted triggering and dynamic invocation of a single MCP tool plugin to perform local feature evaluation according to specific user intent; at the same time, it supports sequential chaining of independent MCP tool plugins under full-process instructions, and when the execution result of any preceding MCP tool plugin is a complete sequence filtering, the system automatically intercepts the call to subsequent MCP tool plugins and triggers a short-circuit protection mechanism to save system computing power.

[0027] The automated parallel evaluation toolset module consists of multiple evaluation units encapsulated as independent MCP tool plugins. It is controlled by the control engine and activated and called on demand according to user intent. It relies on a built-in comprehensive database containing protein sequence and structural information such as Cas9 and Cas12 to evaluate Cas protein-related features.

[0028] In this embodiment, ESMFold is selected as an example of a complex tool, which is encapsulated into the MCP structure prediction tool predict_structures_for_sequences. The input parameters of this tool include at least the protein sequence list (sequences), the inference device (device), whether to save a PDB file (save_pdb), and the output directory (output_dir). When the client calls the tool, it passes these parameters in the arguments field through the MCP tools / call request, thus replacing the sequence input, device selection, and file output controls in the native command line.

[0029] After receiving the request, the server maps the parameters to the underlying ESMFold inference process, completing model loading, batch structure prediction, PDB file saving, and result processing, and returns structured results. The returned results preferably include prediction results (`results`) and an output directory (`output_dir`), where `results` contains the prediction structure, confidence level, and an optional PDB file path. This method achieves the mapping and encapsulation of native command-line capabilities to the MCP standard tool protocol.

[0030] The system's built-in comprehensive database supports the benchmarking and verification of multimodal biometrics, specifically divided into: The Cas protein sequence database unit contains standardized sequence information from different known Cas protein families, which is used to provide benchmark data for homology alignment analysis tools that are packaged as standalone MCP tool plugins; The Cas protein feature domain database unit contains hidden Markov model configuration files for the Cas protein family, which are used to provide benchmark data for domain identification and feature identification for functional domain analysis tools that are packaged as standalone MCP tool plugins. The Cas protein structure database unit contains the three-dimensional spatial atomic coordinates and topological structure information of known Cas proteins, which is used to provide benchmark data for spatial structure and similarity search for structure similarity comparison tools that are packaged as standalone MCP tool plugins.

[0031] The sequence-characteristic-based basic screening unit, packaged as a standalone MCP tool plugin, is used for preliminary filtering and cleaning of protein sequences after data processing and normalization unit reshaping. It includes the following specific execution logic: The length screening subunit is used to screen protein sequences that meet the preset or LLM dynamic extraction length range, retaining sequences with a minimum length of 50 and a maximum length of 10,000 amino acids. The non-natural amino acid screening subunit is used to accurately identify and remove sequences containing non-standard amino acids, retaining only sequences consisting of 20 natural standard amino acids (ACDEFGHIKLMNPQRSTVWY). The sequence is composed of screening subunits, which are used to screen low-complexity sequences and to screen sequences in which the proportion of a single amino acid does not exceed a preset threshold; The k-mer repeat filtering subunit is used to remove sequences that do not meet the requirements according to the k-mer repeat rules. The specific rules include: 1-mer can be repeated a maximum of 6 times consecutively, 2-mer can be repeated a maximum of 4 times consecutively, 3-mer to 6-mer can be repeated a maximum of 3 times consecutively, and 7-mer can be repeated a maximum of 2 times consecutively.

[0032] The Blast / HMM-based domain identification unit, encapsulated as an independent MCP tool plugin, executes the following logic: The homology comparison branch calls the sequence alignment analysis tool to calculate the family coverage of the query sequence in the built-in comprehensive database, and retains candidate sequences that simultaneously meet the requirements of query coverage, target coverage, and alignment score all being greater than the preset threshold. The domain identification branch calls the functional domain analysis tool, compares the built-in domain configuration file with the Hidden Markov Model, and retains candidate sequences whose target HMM domain identification coverage and HMMER score are both greater than the preset threshold.

[0033] The execution logic of the deep learning-based 3D spatial conformation folding prediction unit, which is encapsulated as an independent MCP tool plugin, is as follows: The three-dimensional spatial conformation is predicted directly from the amino acid sequence using the pre-trained deep learning prediction model ESMFold, and the atomic coordinate matrix and pLDDT confidence score are obtained. Sequences with an average pLDDT score lower than a preset threshold are removed.

[0034] The execution logic of the protein structure similarity search and function mining unit, which is encapsulated as an independent MCP tool plugin, is as follows: The Foldseek similarity comparison tool is used to perform spatial topological comparison between the preserved 3D structure and the Cas structure information in the built-in comprehensive database, outputting a structural similarity score and retaining sequences with a similarity greater than a preset threshold.

[0035] The execution logic of the PAM compatibility and specific function activity prediction unit, which is encapsulated as a standalone MCP tool plugin, is as follows: The deep learning validation model Protein2PAM is used to directly predict the PAM sequence of candidate proteins, outputting a PAM probability distribution matrix containing the nucleotide probability at each position and the most likely PAM sequence. Furthermore, given the target gRNA sequence and target DNA sequence, the model is validated and outputs the probability distribution of the predicted cleavage activity score and editing efficiency of the Cas protein candidate.

[0036] The operation mechanism of the result feature extraction and LLM summary output layer in the summary and output layer module is as follows: For the large and nested high-dimensional complex data files returned by the various independent MCP tool plugins at the bottom layer, dimensionality reduction is performed in the independent code environment of the system's backend; quantitative indicators used to characterize the effect of core technologies are forcibly extracted, including the number of filtered statistics, the mean confidence level, and the composition of core sequences. After being reshaped into flat structure data, it is sent back to the large language model to prevent context overflow when the large language model generates the summary and to ensure the accuracy of the global professional evaluation report.

[0037] The professional visualization analysis workbench module operates as follows: it provides a web-based visual interface, offers highly customizable evaluation parameters, and supports full export of high-dimensional analysis data after system analysis, effectively addressing the need for refined operations in complex research scenarios. The interface supports user input of single protein sequences or batch sequence files in multiple formats. Users can flexibly select the required functional modules according to their actual needs, realizing a modular evaluation process. The system supports user-defined screening threshold parameters for each module, such as length range, coverage threshold, and confidence score threshold, to meet the needs of different application scenarios.

[0038] The system employs a hierarchical result output mechanism, comprising underlying full-scale analysis data generated by various independent MCP tool plugins, and a natural language global evaluation report generated by the summary and output layer module. The professional visualization analysis workbench module is used to parse and display the underlying full-scale analysis data, specifically including: (1) The underlying output details of the basic sequence filtering include the total number of sequence inputs, the final retention statistics, and the structured log details of sequences that were blocked or removed due to length exceeding the limit or containing illegal amino acids; (2) Homology and domain identification underlying data, including the high-dimensional alignment result matrix of sequence alignment analysis tools (covering family coverage and alignment score), and the HMM feature domain identification mapping and missing domain details of functional domain analysis tools; (3) The original data of three-dimensional spatial configuration and similarity prediction includes the three-dimensional atomic coordinate file in PDB format generated by the deep learning model, the pLDDT confidence score distribution data, and the structural topological similarity score output by the similarity comparison tool; (4) Functional activity and PAM prediction feature data, including the PAM probability distribution matrix generated by the deep learning validation model, the most likely PAM sequence prediction results, and the evaluation indicators of PAM compatibility and theoretical editing efficiency of candidate proteins.

[0039] Example 2 like Figure 2 As shown, based on the evaluation system in Example 1, this example also proposes a smart comprehensive evaluation method for Cas proteins, including the following steps: S1: Execute user input and preprocessing. Receive unstructured natural language instructions and biological sequence files (FASTA or TXT format) from the user through a dual-track parallel branch. The intent parsing and data preprocessing module extracts semantic features from the natural language and simultaneously cleans up illegal characters and reshapes strings in the sequence files. S2: The workflow dynamic scheduling module obtains the results of S1, performs adaptive aggregation of heterogeneous sequence data, and the intent-based workflow dynamic distribution control engine uses the intent extracted by LLM as decision instructions and establishes the automation tool invocation strategy in combination with the model context protocol. S3: Based on decision instructions, the workflow dynamic scheduling module dynamically activates one or more independent MCP tool plugins in the automated parallel evaluation toolset module across platforms, executing them sequentially or as needed: basic screening based on sequence characteristics, domain identification based on Blast / HMM, three-dimensional conformation folding prediction based on deep learning, protein structure similarity search and function mining, and PAM compatibility and specific functional activity prediction; during this process, the built-in comprehensive database is invoked for sequence, structure and characteristic domain alignment; S4: The summary and output layer module performs dimensionality reduction feature extraction on the high-dimensional data returned by each MCP tool plugin and inputs it into the large language model to generate a structured summary report; S5: The final evaluation report is presented through a professional visualization analysis workbench module, and the front end can receive highly flexible dynamic configuration of the underlying evaluation parameters. At the same time, it provides full visualization and export of high-dimensional raw analysis data, effectively making up for the need for refined operation in complex scientific research scenarios.

[0040] like Figure 3 As shown, the workflow dynamic scheduling module incorporates an intent-based workflow dynamic distribution control engine. This engine completely decouples the underlying bioinformatics computing nodes into five independent MCP tool plugins (basic screening unit, domain identification unit, 3D conformation folding prediction unit, similarity search unit, and PAM compatibility prediction unit). Based on the JSON intent parameters (such as task type and sequence length threshold) generated by the large language model, the engine sends call requests to the corresponding MCP tool plugins through a standardized interface.

[0041] In this execution architecture, the system is configured with a short-circuit protection mechanism. When a sequence flows through a front-end MCP tool plugin (such as a basic filtering unit), if all sequences in that batch are determined to be non-compliant and are intercepted and filtered, the front-end plugin will return an empty data stream signal to the control engine. In response to this signal, the engine immediately terminates the activation instructions for subsequent nodes (such as computationally expensive 3D conformational folding prediction units), thereby achieving dynamic adaptive optimization of computing resources.

[0042] Example 3 This embodiment details how the evaluation system proposed in Embodiment 1 achieves routing distribution and result dimensionality reduction for complex multimodal computations through natural language interaction; 1. Intent parsing and preliminary sequence screening: like Figure 4 As shown, the user sends the instruction "Please perform basic screening and predict the structure evaluation pLDDT for these sequences. The sequence length requirement is a maximum of 2000 and a minimum of 100", and uploads a file named "tset.txt" containing 5 unknown Cas protein sequences. The intent parsing unit extracts the constraint parameters (minimum length 100, maximum length 2000, task: basic screening, structure prediction) and passes these parameters to the control engine. The engine dynamically activates the "basic screening unit" MCP plugin and performs serial filtering on the 5 sequences according to the preset rules: (1) The length screening subunit intercepts a sequence whose length does not meet the requirements (e.g., 35aa) based on the parameters; (2) The non-natural amino acid screening subunit intercepts a sequence containing non-standard amino acid characters (e.g., "X"); (3) The sequence composition screening subunit intercepts a sequence whose single amino acid proportion exceeds the preset 25% threshold; (4) The k-mer repetition screening subunit intercepts a low-complexity sequence that triggers the continuous repetition rule (e.g., 8 consecutive identical amino acids).

[0043] Ultimately, only one compliant sequence was retained from this batch, generating the following: Figure 5 The "Sequence Base Length Filtering Analysis Results" shown will be further routed by the control engine to the "3D Spatial Conformation Folding Prediction Unit".

[0044] 2. Forced dimensionality reduction and overflow prevention in natural language report generation Once the 3D structure prediction is complete, the underlying MCP plugin generates a complex PDB file containing atomic space coordinates. To avoid context overflow in large language models due to massive coordinate system data, such as... Figure 5 As shown, the summary and output layer module truncates and reduces the dimensionality of the underlying data, extracting only the core numerical features (a total of 1 sequence folding prediction was completed; prediction time: 8.32s; average pLDDT score: 79.1). After receiving this flattened dictionary data, the large language model generates... Figure 5The "Cas93D Structure Prediction Report" is displayed, and a link to the visualization workbench is attached at the bottom of the interface, informing users that the 3D conformation of macromolecules needs to be viewed on the web, thus achieving separate presentation of data.

[0045] Example 4 This embodiment details how the professional visualization analysis workbench receives all the high-dimensional data from the system's underlying layers and supports highly flexible configuration. 1. Configuration of Web-based High-Degree-of-Freedom Analysis Tasks like Figure 6 As shown, users enter the web interface for creating a new analysis task. This interface provides text input and file upload components. In the analysis function configuration area, users can directly select the corresponding function modules of the underlying MCP plugin (including: PAM prediction, structure prediction, Foldseek alignment, BLAST filtering, and HMMER filtering). The system also provides an advanced parameter setting drop-down panel, allowing users to manually set the hard calculation thresholds of the underlying algorithms such as family coverage, E-value, and confidence score, and package these configuration parameters into a task flow, which is then directly sent to the control engine.

[0046] 2. Full display of bottom-level filter details and structural comparison results like Figure 7 As shown, after the task workflow is completed, the visualization workbench renders the analysis results interface. The top displays the core quantitative indicators panel (including the number of retained sequences (1 / 5), average pLDDT of 91.2, and average PAM confidence level of 95.3%). The bottom lists the log details of the intercepted sequences and the output files of each filtering stage. The system supports users downloading the basic sequence filtering logs, BLAST sequence alignment details, and HMMER structure field annotation files (such as .domtblout format) in their original format locally in their entirety.

[0047] Example 5 To further demonstrate the core technical capabilities of the underlying MCP tool plugin of the evaluation system of this invention in complex multimodal bioinformatics prediction, this embodiment takes the classic SpCas9 protein as an example to explain in detail the underlying execution logic of three-dimensional conformational deduction and PAM compatibility prediction.

[0048] 1. Deep learning-based prediction of 3D spatial configuration folding When the system control engine receives the input SpCas9 protein sequence (1368 amino acids in length), it activates the "3D spatial conformation folding prediction unit." This unit directly calls the underlying pre-trained deep learning model (such as one based on the ESMFold architecture) to perform end-to-end 3D structure deduction without templates, successfully generating a PDB file containing complete atomic coordinates. Figure 8As shown, the system outputs the prediction result as a high-resolution three-dimensional spatial topological framework diagram of the SpCas9 protein. Simultaneously, the system extracts the pLDDT confidence score from the model output, calculating an average score of 91.2. This score is significantly higher than the structural stability threshold, demonstrating that the candidate protein possesses extremely high folding stability, and that the system has the capability to perform high-precision spatial conformation assessment of large molecular weight Cas proteins.

[0049] 2. PAM compatibility and depth prediction of specific functional activities PAM (Paramita Anomalies) are flanking sequence patterns essential for Cas proteins to recognize target DNA sequences and are a key indicator for assessing the functional potential of gene editing tools. In this embodiment, the system invokes the "PAM Compatibility and Specific Functional Activity Prediction Unit" and uses the Protein2PAM deep learning model to directly predict the PAM sequence of the aforementioned 1368 amino acids in the SpCas9 sequence.

[0050] The prediction results are supported by two core high-dimensional data: (1) the PAM probability distribution matrix (e.g. Figure 9 As shown): The model output contains a matrix containing the probabilities of nucleotides (A, C, G, T) at each position, accurately displaying the probability distribution of each position. Specifically, the matrix shows that the probability of nucleotide A at position 1 is 0.24110, the probability of nucleotide C is 0.23341, the probability of nucleotide G is 0.20896, and the probability of nucleotide T is 0.31653; while the probability of nucleotide G at position 2 is as high as 0.96293, and the probability of nucleotide G at position 3 is also as high as 0.95947. (2) Visualization and prediction results of PAM sequence features (e.g. Figure 10 As shown in the figure): Based on the above probability distribution matrix, the system further converts it into an intuitive SequenceLogo feature map. The height of the letter stacking in the figure clearly reflects the absolute recognition preference of the sequence for guanine at the 2nd and 3rd positions. The system finally outputs the most likely PAM sequence prediction result as "NGG" (where N represents any deoxyribonucleotide).

[0051] This deep learning-based prediction result is completely consistent with the known typical PAM sequence of the SpCas9 protein. This fully demonstrates that the deep learning evaluation module of this invention can accurately predict the PAM compatibility and editing target characteristics of Cas proteins even in the absence of expensive wet experimental data. This prediction data provides a crucial and reliable theoretical reference for subsequent optimization and experimental validation of the CRISPR gene editing system.

[0052] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.

[0053] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A smart comprehensive evaluation method for Cas proteins, characterized in that, Includes the following steps: Acquire natural language commands input by the user and protein sequence data to be evaluated; The natural language instructions are input into a pre-configured large language model for intent parsing, and structured intent parameters are extracted. The intent parameters include: task execution intent labels and constraint thresholds. Based on the task execution intent tag, following the preset plug-in calling protocol, the corresponding independent evaluation tool plug-in is scheduled and activated from the tool library, and the execution chain of each independent evaluation tool plug-in is determined. The protein sequence data is input into the activated independent evaluation tool plugin and evaluated in conjunction with the constraint threshold. In the case where the execution chain includes multiple serially scheduled independent evaluation tool plugins, a short-circuit protection step is performed: obtain the set of reserved sequences output by the preceding independent evaluation tool plugin; if the number of sequences in the set of reserved sequences is zero, then intercept and cancel the call instructions to subsequent independent evaluation tool plugins in the execution chain. After the independent evaluation tool plugin is executed, the high-dimensional multimodal analysis data output is obtained. In the backend runtime environment, the high-dimensional multimodal analysis data is subjected to feature truncation and dimensionality reduction processing to extract the preset core quantitative feature values ​​and reshape them into flat structure data. The flattened structure data is fed back to the large language model to generate a global evaluation report based on natural language.

2. The intelligent comprehensive evaluation method for Cas proteins according to claim 1, characterized in that, After acquiring the protein sequence data to be evaluated, and before inputting the protein sequence data into the activated standalone evaluation tool plugin, the method further includes: The file stream of the protein sequence data is matched and parsed using regular expressions to remove header annotation information and whitespace characters from the protein sequence data. The stripped data is cleaned of illegal characters and formatted, then converted into a standardized string array, which serves as the unified input format for each independent evaluation tool plugin in the execution chain.

3. The intelligent comprehensive evaluation method for Cas proteins according to claim 1, characterized in that, The independent evaluation tool plugin includes a basic screening unit based on sequence characteristics, and the constraint threshold includes a sequence length parameter. The steps for invoking the basic filtering unit include: The input protein sequence data is filtered according to the sequence length parameter, and sequences containing non-standard amino acids are removed. Identify and remove low-complexity sequences where the proportion of a single amino acid exceeds a preset composition threshold, wherein the preset composition threshold is included in the constraint threshold; Non-compliant sequences are removed according to the k-mer repetition rules: 1-mer can be repeated a maximum of 6 times consecutively, 2-mer can be repeated a maximum of 4 times consecutively, 3-mer to 6-mer can be repeated a maximum of 3 times consecutively each, and 7-mer can be repeated a maximum of 2 times consecutively.

4. The intelligent comprehensive evaluation method for Cas proteins according to claim 1, characterized in that, The independent evaluation tool plugin also includes a domain identification unit based on Blast / HMM; the steps for invoking the domain identification unit include: Calculate the family coverage of the input sequence and retain candidate sequences that satisfy the query coverage, target coverage, and alignment score all being greater than the corresponding constraint thresholds; Based on the comparison of the built-in structural domain configuration file using the Hidden Markov Model, candidate sequences with both target HMM structural domain recognition coverage and HMMER score greater than the corresponding constraint threshold are retained.

5. The intelligent comprehensive evaluation method for Cas proteins according to claim 1, characterized in that, The independent evaluation tool plugin also includes a PAM compatibility and specific function activity prediction unit; The steps for invoking the PAM compatibility and specific function activity prediction unit include: Deep learning was used to validate the prediction of PAM sequences for protein sequence identification using a validation model. Output the PAM probability distribution matrix containing the nucleotide probabilities at each position; Given the target gRNA sequence and the target DNA sequence, output the cleavage activity score and the probability distribution of editing efficiency.

6. The intelligent comprehensive evaluation method for Cas proteins according to claim 1, characterized in that, The high-dimensional multimodal analysis data includes three-dimensional atomic coordinate data, and the core quantization feature value includes the average pLDDT score obtained by calculating the confidence level of the three-dimensional atomic coordinates.

7. The intelligent comprehensive evaluation method for Cas proteins according to claim 6, characterized in that, The method also includes: The raw, high-dimensional multimodal analysis data without dimensionality reduction is sent to the front-end interactive interface for visualization rendering; Receive evaluation parameter configuration instructions based on feedback from the front-end interactive interface, and update the constraint threshold accordingly to re-trigger the evaluation process.

8. A Cas protein intelligent comprehensive evaluation system, comprising the method described in any one of claims 1-7, characterized in that, include: The user input module is used to receive natural language evaluation instructions and protein sequence data files; The intent parsing and data preprocessing module is used to extract task execution intent tags and constraint thresholds from the natural language evaluation instructions, and to clean and format the received protein sequence data. The workflow dynamic scheduling module has a built-in intent-based workflow dynamic distribution control engine, which is used to schedule the evaluation units in the automated parallel evaluation toolset module according to the task execution intent tags and in combination with the model context protocol. And execute a short-circuit protection mechanism: when there are multiple evaluation units scheduled in sequence and the number of sequences output by the previous evaluation unit is zero, intercept the call to the subsequent evaluation unit; The automated parallel evaluation toolset module consists of multiple evaluation units encapsulated as independent tool plugins. These evaluation units are used to perform evaluation calculations on the input protein sequence data in conjunction with the constraint thresholds. The summary and output layer module is used to perform feature truncation and dimensionality reduction on the high-dimensional multimodal analysis data generated by each evaluation unit, extract core quantitative feature values ​​and reshape them into flat structure data, and then send them back to the large language model to generate a global evaluation report. The professional visualization analysis workbench module is used to present raw, high-dimensional multimodal analysis data without dimensionality reduction and to receive reverse configuration instructions for evaluation parameters.

9. The Cas protein intelligent comprehensive evaluation system according to claim 8, characterized in that, The evaluation unit includes: a basic screening unit based on sequence characteristics, a domain identification unit based on Blast / HMM, a three-dimensional spatial conformation folding prediction unit based on deep learning, a protein structure similarity search and function mining unit, and a PAM compatibility and specific functional activity prediction unit.

10. A computer device, characterized in that, include: Memory and processor; The memory is used to store computer programs; The processor is configured to implement the steps of the intelligent comprehensive evaluation method for Cas protein as described in any one of claims 1-7 when executing the computer program.

Images