AI-assisted microbial gene editing target prediction method and system

Abstract

The invention discloses an AI-assisted microbial gene editing target prediction method and system, and the method comprises the following specific steps: 1, fusing a gene sequence, a protein structure and metabolic flux data through a cross attention mechanism, and optimizing the prediction precision in combination with a generative adversarial network; step 2, performing full-process automation according to prediction, editing and verification in sequence by performing online monitoring according to a Raman spectrum, a CRISPR-Cas9 gene editing tool and a miniaturized fermentation chip; and step 3, dynamically recommending an optimal editing strategy and avoiding a toxic risk by combining a deep Q network with a metabolic network model. According to the method, intelligent design of the gamma-PGA high-yield strain is achieved through multi-modal data fusion, closed-loop feedback and reinforcement learning, compared with a traditional method, the prediction accuracy is greatly improved, the strain improvement period is shortened to 3 months, the yield reaches 62 g / L, and remarkable technical advantages and economic value are achieved.

Description

Technical Field

[0001] This invention relates to the field of microbial gene editing technology, and in particular to an AI-assisted method and system for predicting microbial gene editing targets. Background Art

[0002] γ-PGA is a natural amino acid polymer that is water-soluble, biodegradable, and non-toxic, and is widely used in medicine, food, environmental protection, and other fields.

[0003] Currently, microbial fermentation is the main method for synthesizing γ-PGA, but traditional strain improvement relies on experimental trial and error, which suffers from low efficiency, long cycle, and high cost. The combination of AI and gene editing provides a new path to solve this problem, but existing technologies have insufficient accuracy in predicting the editing effects of genes related to γ-PGA synthesis and lack systematic solutions. Summary of the Invention

[0004] This invention provides an AI-assisted method and system for predicting microbial gene editing targets to address the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: AI-assisted methods for predicting microbial gene editing targets include the following specific steps: The first step involves fusing gene sequence, protein structure, and metabolic flux data through a cross-attention mechanism, and then combining this with a generative adversarial network to optimize prediction accuracy. The second step involves automating the entire process of prediction, editing, and verification sequentially by using online Raman spectroscopy monitoring, CRISPR-Cas9 gene editing tools, and miniaturized fermentation chips. The third step involves using a deep Q-network combined with a metabolic network model to dynamically recommend the optimal editing strategy and mitigate toxicity risks.

[0006] As a further improvement to this technical solution: the cross-attention mechanism is introduced in the 3rd and 6th layers of the Transformer encoder, allowing gene sequence features to interact dynamically with protein structure features and metabolic flux features; the number of interaction layers is 2-6, the number of attention heads is 8, and the attention dimension is 64.

[0007] This invention also proposes an AI-assisted microbial gene editing target prediction system, including a multimodal data fusion and dynamic prediction model unit, a closed-loop feedback unit, and a reinforcement learning unit; The multimodal data fusion and dynamic prediction model unit is used to fuse gene sequence, protein structure, and metabolic flux data through a cross-attention mechanism, and combine it with a generative adversarial network to optimize prediction accuracy. The closed-loop feedback unit achieves full automation of the prediction, editing, and verification process by using online Raman spectroscopy monitoring, CRISPR-Cas9 gene editing tools, and miniaturized fermentation chips. The reinforcement learning unit uses a deep Q-network combined with a metabolic network model to dynamically recommend the optimal editing strategy and avoid toxicity risks.

[0008] As a further improvement to this technical solution: the miniaturized fermentation chip includes a microfluidic channel and a colony culture area. The microfluidic channel has a width of 100μm, a depth of 50μm, and a single channel volume of 1μL. The colony culture area is a circular chamber with a diameter of 200μm, and each chip can hold 1024 colonies.

[0009] As a further improvement to this technical solution: the gRNA of the CRISPR-Cas9 gene editing tool is designed using the CRISPRdirect tool, the PAM sequence is NGG, and the off-target rate is less than 0.1%.

[0010] As a further improvement to this technical solution: the state space of the deep Q network includes gene expression data, metabolite concentration data, and editing history data.

[0011] As a further improvement to this technical solution: the reward function of the deep Q-network is defined as: ; in, =1, =0.5, =2, toxicity risk is determined by the concentration threshold of toxic metabolites predicted by the metabolic network model.

[0012] As a further improvement to this technical solution: the metabolic network model calculates metabolic flux through constraint optimization, providing metabolic feasibility constraints for editing strategies; The constraints of the metabolic network model include substrate consumption rate, product formation rate, and toxic metabolite accumulation threshold.

[0013] As a further improvement to this technical solution: the system is applicable to the improvement of production strains for high-value-added biomaterials such as γ-polyglutamic acid and polyhydroxy fatty acid esters.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention achieves intelligent design of high-yield γ-PGA strains through multimodal data fusion, closed-loop feedback, and reinforcement learning. Compared with traditional methods, the prediction accuracy is greatly improved, the strain improvement cycle is shortened to 3 months, and the yield reaches 62 g / L, demonstrating significant technical advantages and economic value.

[0015] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0016] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 This is a schematic diagram of the AI-assisted microbial gene editing target prediction method proposed in this invention. Figure 2 This is a schematic diagram of the structure of the AI-assisted microbial gene editing target prediction system proposed in this invention. Detailed Implementation

[0017] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are for illustrative purposes only and are not intended to limit the scope of the invention. The invention is described more specifically in the following paragraphs by way of example with reference to the accompanying drawings. It should be noted that the drawings are in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0018] Please see Figures 1-2 This invention proposes an AI-assisted microbial gene editing target prediction system, which specifically includes: a multimodal data fusion and dynamic prediction model unit, a closed-loop feedback unit, and a reinforcement learning unit; The multimodal data fusion and dynamic prediction model unit is used to fuse gene sequence, protein structure, and metabolic flux data through a cross-attention mechanism, and to optimize prediction accuracy by combining a generative adversarial network. Specifically: (1) Data preprocessing Gene sequence data: The target gene sequence was aligned with the KEGG database using the BLAST alignment tool, and functional annotations (such as the pgsBCA operon's involvement in γ-PGA synthesis) were extracted.

[0019] The sequence is encoded into a One-Hot matrix (A / T / C / G correspond to 4-dimensional vectors), and the length is truncated to 1000 bp (padded with zeros if necessary).

[0020] Protein structure data: The three-dimensional structure of the target protein was predicted using AlphaFold2, and structural features (such as the proportion of secondary structures and the volume of binding pockets) were extracted.

[0021] The structural features are quantized into 128-dimensional vectors (by PCA dimensionality reduction).

[0022] Metabolic flux data: A metabolic network model was constructed using COBRApy, and metabolic fluxes (such as γ-PGA synthesis flux and glutamate consumption flux) were calculated using FBA simulation.

[0023] The flux values ​​are normalized to the [0,1] interval and used as dynamic feature inputs.

[0024] (2) Transformer model training Model architecture: Input layer: Gene sequence (1000×4), protein structure (128-dimensional), and metabolic flux (1-dimensional) are concatenated into a 1029-dimensional vector, and positional encoding is added.

[0025] Encoder layer: 6 Transformer Block layers, each containing multi-head self-attention (8 heads, attention dimension = 64) and feedforward network (2 fully connected layers, hidden layer dimension = 512).

[0026] Cross attention mechanism: Cross attention modules are introduced in layers 3 and 6 to allow gene sequences to interact with protein structures and metabolic fluxes.

[0027] Combat training: The generator (Transformer model) outputs the γ-PGA yield prediction, and the discriminator (3-layer fully connected network) judges whether the prediction is reasonable.

[0028] (3) Experimental verification Datasets: 1000 sets of wet experimental data (γ-PGA yield, gene expression, metabolite concentration) and 5000 sets of FBA simulation data.

[0029] Comparison results: Traditional LSTM model: prediction accuracy 78.3%, MAE=5.2 g / L.

[0030] This method (Transformer + adversarial training): prediction accuracy 89.1%, MAE = 3.1 g / L.

[0031] 2. The closed-loop feedback unit, through online Raman spectroscopy monitoring, CRISPR-Cas9 gene editing tools, and miniaturized fermentation chips, achieves full automation of the prediction, editing, and verification processes sequentially. Specifically: (1) Online monitoring module Raman spectroscopy detection: Using an Ocean Insight QE Pro spectrometer (wavelength range: 500-1800 cm⁻¹) -1 ), and the bacterial suspension spectrum was collected every 10 minutes.

[0032] Feature peak extraction: γ-PGA (1002 cm⁻¹) -1 ), glutamic acid (855 cm) -1 Concentration was predicted using a PLS regression model (R). 2 =0.95).

[0033] (2) Gene editing tools CRISPR-Cas9 system: gRNA design: PAM sequences (NGG) were screened using the CRISPRdirect tool, with an off-target rate of <0.1%.

[0034] Electroconversion parameters: Voltage 1.8 kV, Capacitor 25 μF, Resistance 200 Ω, Conversion efficiency >10 6 CFU / μg DNA.

[0035] (3) Miniaturized fermentation chip Chip design: Microfluidic channel: 100 μm wide, 50 μm deep, 1 μL single channel volume.

[0036] Colony culture area: a circular chamber with a diameter of 200 μm, each chip can hold 1024 colonies.

[0037] Verification process: Dilute the edited strain to 10. -6 CFU / mL, seeded onto the chip.

[0038] After 48 hours of fermentation, the yield was rapidly detected by fluorescent labeling (γ-PGA combined with SYBR Green).

[0039] High-yielding strains (yield > 60 g / L) were screened and transferred to shake flasks for further validation.

[0040] (4) Experimental results Traditional method: strain improvement cycle is 12 months, final yield is 45 g / L.

[0041] This method shortens the strain improvement cycle to 3 months and increases the final yield to 62 g / L.

[0042] The reinforcement learning unit uses a deep Q-network (DQN) combined with a metabolic network model (CBM) to dynamically recommend the optimal editing strategy and avoid toxicity risks. Specifically: (1) State space Gene expression (1000 dimensions), metabolite concentration (5 dimensions), editing history (5000 dimensions).

[0043] (2) Action space Twelve atomic operations (such as gene knockout, overexpression, and promoter replacement) can be combined into 144 strategies.

[0044] (3) Reward function ; in, =1, =0.5, =2, toxicity risk is determined by the concentration threshold of toxic metabolites predicted by the metabolic network model; (4) Training and verification Hardware configuration: NVIDIA A100 GPU, training time <72 hours.

[0045] Comparison results: Manual strategy: Output increased by 28.7%, 12 edits required.

[0046] This method (DQN) resulted in a 42.3% increase in yield and required 7 edits.

[0047] Example 1: Rapid Improvement of High-Yielding γ-PGA Strains Initial strain: wild-type Bacillus subtilis (γ-PGA yield: 15 g / L).

[0048] Closed-loop iterative process: Round 1: Predicted knockout of the ywtD gene, resulting in an increased yield of 32 g / L after editing.

[0049] Round 3: Predicted overexpression of the pgsC gene, yield increased to 58 g / L.

[0050] Round 5: The predicted replacement promoter is P43, with a final yield of 62 g / L.

[0051] Cycle and cost: Total duration: 3 months (vs. 12 months using traditional methods).

[0052] Total cost: 150,000 yuan (vs. 800,000 yuan for traditional methods).

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the accompanying drawings and the above description. However, any modifications, alterations, or variations made by those skilled in the art without departing from the scope of the present invention, utilizing the disclosed technical content, are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. An AI-assisted method for predicting microbial gene editing targets, characterized in that, The specific steps include the following: The first step involves fusing gene sequence, protein structure, and metabolic flux data through a cross-attention mechanism, and then combining this with a generative adversarial network to optimize prediction accuracy. The second step involves automating the entire process of prediction, editing, and verification sequentially by using online Raman spectroscopy monitoring, CRISPR-Cas9 gene editing tools, and miniaturized fermentation chips. The third step involves using a deep Q-network combined with a metabolic network model to dynamically recommend the optimal editing strategy and mitigate toxicity risks.

2. The AI-assisted microbial gene editing target prediction method according to claim 1, characterized in that, The cross-attention mechanism is introduced in layers 3 and 6 of the Transformer encoder, allowing dynamic interaction between gene sequence features and protein structure features and metabolic flux features; the number of interaction layers is 2-6, the number of attention heads is 8, and the attention dimension is 64.

3. An AI-assisted microbial gene editing target prediction system, characterized in that, Includes a multimodal data fusion and dynamic prediction model unit, a closed-loop feedback unit, and a reinforcement learning unit; The multimodal data fusion and dynamic prediction model unit is used to fuse gene sequence, protein structure, and metabolic flux data through a cross-attention mechanism, and combine it with a generative adversarial network to optimize prediction accuracy. The closed-loop feedback unit achieves full automation of the prediction, editing, and verification process by using online Raman spectroscopy monitoring, CRISPR-Cas9 gene editing tools, and miniaturized fermentation chips. The reinforcement learning unit uses a deep Q-network combined with a metabolic network model to dynamically recommend the optimal editing strategy and avoid toxicity risks.

4. The AI-assisted microbial gene editing target prediction system according to claim 3, characterized in that, The miniaturized fermentation chip includes a microfluidic channel and a colony culture area. The microfluidic channel has a width of 100 μm, a depth of 50 μm, and a single channel volume of 1 μL. The colony culture area is a circular chamber with a diameter of 200 μm, and each chip can hold 1024 colonies.

5. The AI-assisted microbial gene editing target prediction system according to claim 4, characterized in that, The gRNA of the CRISPR-Cas9 gene editing tool was designed using the CRISPRdirect tool, with the PAM sequence being NGG, and the off-target rate was less than 0.1%.

6. The AI-assisted microbial gene editing target prediction system according to claim 5, characterized in that, The state space of the deep Q-network includes gene expression data, metabolite concentration data, and editing history data.

7. The AI-assisted microbial gene editing target prediction system according to claim 6, characterized in that, The reward function of the deep Q-network is defined as: ; in, =1, =0.5, =2, toxicity risk is determined by the concentration threshold of toxic metabolites predicted by the metabolic network model.

8. The AI-assisted microbial gene editing target prediction system according to claim 7, characterized in that, The metabolic network model calculates metabolic flux through constraint optimization, providing metabolic feasibility constraints for editing strategies; The constraints of the metabolic network model include substrate consumption rate, product formation rate, and toxic metabolite accumulation threshold.

9. The AI-assisted microbial gene editing target prediction system according to claim 8, characterized in that, This system is suitable for improving the production strains of high-value-added biomaterials such as γ-polyglutamic acid and polyhydroxy fatty acid esters.

Images