A gene editing tool protein directed evolution optimization method based on a multi-scale mutation perception graph network and adaptive time sequence evolution modeling

Abstract

The application discloses a gene editing tool protein directed evolution optimization method and system, and belongs to the cross field of bioinformatics and artificial intelligence. The method comprises the following steps: converting protein variant sequences into a multi-layer graph structure containing short-range, medium-range and long-range contact relationships through multi-scale Seq2Graph; extracting mutation site perception features by using a multi-scale graph attention network, and enhancing mutation signals by using a mutation site mask biasing mechanism; simultaneously predicting editing efficiency and specificity scores by using a double-branch regression network; dynamically determining an optimal historical round segmentation point based on sequence distribution drift measurement and mutation cumulative density distribution, and realizing adaptive time sequence evolution modeling; constructing a prediction Pareto front and dynamically adjusting double-target weights according to sparsity to uniformly cover a non-dominated solution set; and correcting systematic prediction bias by using a confidence calibration sub-network, and realizing AI prediction and wet experiment closed-loop iteration. The method is used for improving the editing efficiency of base editors deaminase and Cas protein variants and reducing off-target rates.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bioinformatics, artificial intelligence and genetic engineering. Specifically, it relates to a method and system for the directed evolution optimization of gene editing tool proteins based on multi-scale mutation-aware graph networks and adaptive temporal evolution modeling. In particular, it relates to a method for guiding the directed evolution of base editor deaminase components and CRISPR-Cas proteins using multi-scale graph attention networks, adaptive temporal segmentation strategies and Pareto dynamic trade-off mechanisms. Background Technology

[0002] Technical bottlenecks in the directed evolution of gene-editing tool proteins

[0003] Gene editing technology is a revolutionary tool in the biomedical field. Adenine base editors (ABEs) and cytosine base editors (CBEs) can achieve site-directed single-base conversions without introducing double-strand breaks in DNA, showing significant promise for the treatment of point mutation-related genetic diseases such as sickle cell anemia and phenylketonuria. The core component of ABE is TadA deaminase, the core component of CBE is APOBEC1 deaminase, and the core component of the CRISPR system is the Cas protein. Their editing efficiency, specificity, and fidelity directly determine the safety and efficacy of gene editing therapy.

[0004] However, naturally derived gene-editing tools often fail to meet clinical treatment needs in terms of editing efficiency and specificity. Currently, these proteins are mainly modified and optimized through directed evolution. Traditional directed evolution involves introducing random mutations, constructing mutant libraries, and high-throughput screening to select mutants with superior performance, but it suffers from drawbacks such as low screening throughput, long cycle time, and high cost.

[0005] Existing AI-assisted directed evolution technology and its shortcomings

[0006] In recent years, AI-assisted directed evolution methods have gradually emerged. Yang et al. (Nature Methods, 2019) proposed the Machine Learning Guided Directed Evolution (MLDE) method, which uses a Gaussian process model to predict the fitness of protein mutants. However, it is mainly based on the one-hot encoding of the sequence or physicochemical features, without utilizing the three-dimensional structural information of the protein, and is insufficient in modeling the propagation effect of mutations in structural networks.

[0007] Jha et al. (Scientific Reports, 2022) published a protein graph representation method that uses amino acid residues as nodes, constructs edges based on PDB structural atomic distances, extracts node features using a pre-trained protein language model, and then uses a graph attention network for downstream prediction. However, this method is aimed at static protein function prediction tasks and does not address the temporal dependencies of multi-round experimental data in directed evolution, nor does it model the multi-scale propagation paths of mutation effects in protein structure networks.

[0008] Chinese patent CN202211231553.9 discloses a method for predicting protein interaction sites based on graph neural networks, which represents proteins as graph structures and uses graph neural networks to extract features. However, this method is aimed at static protein interaction prediction tasks and does not involve the dynamic mapping relationship between sequence variants and functional data during directed evolution, nor does it consider the temporal dependence of multi-round experimental data.

[0009] TadABench-1M (NeurIPS 2024 / bioRxiv 2024) employed temporal data splitting in its TadA deaminase benchmark, using early-cycle data as the training set and later-cycle data as the test set to simulate the iterative characteristics of protein engineering. However, this work only uses temporal splitting as a data partitioning method for benchmark evaluation, without deeply integrating it with the model training process to guide subsequent optimization, nor designing adaptive model training strategies to address the distribution drift problem in temporal data. Its Seq2Graph method constructs a variant-enrichment relationship graph, rather than a residue-structure contact graph.

[0010] SPIRED-Fitness (Nature Communications, 2024) disclosed a GAT geometric encoder combining ESM-2 protein language model sequence embedding and structure prediction for mutational fitness prediction. However, this method uses a single-scale structure representation, fails to distinguish the differential contributions of short-range, mid-range, and long-range structural interactions to mutational effects, and does not design a training framework for the temporal evolutionary characteristics in directed evolution scenarios.

[0011] ALDE (Nature Communications, 2025) employs active learning to achieve alternating iterations of multiple rounds of wet experiments and machine learning models. However, the core of ALDE lies in the optimization of the sampling strategy (uncertainty sampling). Its model training itself still uses static data partitioning, does not design an adaptive training framework for the temporal characteristics of directed evolutionary data, and does not introduce a prediction confidence calibration mechanism.

[0012] DeepDE (iScience, 2025) proposed an iterative deep learning-guided evolutionary algorithm that implements multiple rounds of iterative evolution, with each round using a trained model to predict the next batch of candidate variants. However, when DeepDE retrains the model in each iteration, it adopts a random partitioning training strategy after merging historical data, which does not strictly follow the temporal boundaries of the directed evolution data. This results in the model being unable to learn the cumulative effect of mutations and the drift pattern of sequence distribution during directed evolution.

[0013] The aforementioned existing technologies share the following unresolved technical problems:

[0014] First, the structural characterization accuracy is insufficient. Existing methods often use a single distance threshold (e.g., 8 Å) to construct protein contact maps, conflating strong interactions (<5 Å, such as hydrogen bonds and salt bridges) with weak interactions (5–8 Å, such as hydrophobic interactions), failing to accurately model the differentiated propagation paths of mutational effects at different spatial scales. Furthermore, existing methods do not perform attention enhancement on the mutation site itself, resulting in the mutation signal being diluted by a large number of non-mutated neighbor nodes during message passing.

[0015] Second, the temporal modeling is rigid. Existing technologies often use fixed round ratios for temporal partitioning (such as a simple first 50% / last 50%), without considering the dynamic changes in sequence distribution drift speed and mutation accumulation density during directed evolution. When mutation exploration is insufficient in early rounds or drastic distribution changes occur in later rounds, fixed partitioning can lead to insufficient model training or distorted validation.

[0016] Third, the dual-objective optimization is unbalanced. When optimizing editing efficiency and specificity simultaneously, existing methods often use fixed-weight comprehensive score ranking, which cannot dynamically adjust the optimization direction according to the sparsity of the current Pareto front. This causes the recommended candidate variants to cluster around the extreme point of a single objective, missing the intermediate solution that is optimal in the dual-objective tradeoff.

[0017] Fourth, there is a lack of prediction confidence calibration. Existing closed-loop iterative methods directly feed wet experimental data into the training set without distinguishing between high-confidence prediction errors (systematic bias) and low-confidence prediction errors (random errors). This results in erroneous samples continuously contaminating the training data, leading to the accumulation of bias in the model during the iteration process.

[0018] The technical problem to be solved by the present invention

[0019] This invention aims to overcome the aforementioned deficiencies of the prior art and provide a method for the directed evolution of gene editing tools that can: (a) accurately model the short-range, medium-range, and long-range propagation paths of mutation effects in the three-dimensional structure network of proteins through multi-scale structure diagrams; (b) dynamically optimize the training-validation boundary through an adaptive temporal segmentation framework using temporal evolution information from directed evolution experimental data; (c) achieve synergistic optimization of editing efficiency and specificity through the Pareto dynamic trade-off mechanism; (d) distinguish and correct systematic prediction biases through a confidence-calibrated subnetwork; and (e) realize a closed-loop iteration of AI prediction and wet experimental validation. Summary of the Invention

[0020] Technical Solution Overview

[0021] This invention provides a method for directed evolutionary optimization of gene-editing tool proteins based on multi-scale mutation-aware graph networks and adaptive temporal evolution modeling, comprising the following steps:

[0022] (1) Multi-scale graph structure transformation: The multi-scale Seq2Graph method was used to transform protein variant sequences into multi-layer graph structure data containing short-range (<5 Å), mid-range (5–8 Å), and long-range (8–12 Å) contact relationships. The initial features of each layer node were the concatenation features of the 1280-dimensional residue embedding vector extracted by the pre-trained protein language model (ESM-1b or ESM-2) and the one-hot encoding vector of the mutation type. The three-dimensional structure was derived from the experimentally determined structure or the AlphaFold2 predicted structure.

[0023] (2) Mutation Site Awareness Feature Extraction: Multi-Scale Graph Attention Network (MSMA-GAT) is used to extract features from multi-layer graph structure data. MSMA-GAT includes a near-range attention head group (4 heads), a mid-range attention head group (4 heads), and a long-range attention head group (4 heads). Each head group independently processes the graph layers at its corresponding scale and then fuses them using learnable scale-gated weights. A mutation site masking bias mechanism is introduced, applying an initial gain bias of 0.2–0.5 to known mutation sites and their first-order structural neighbor nodes during attention calculation to prevent the mutation signal from being diluted during message passing. After three layers of message passing, the node features of each layer are weighted and aggregated and then global average pooled to obtain a 1024-dimensional overall representation vector that integrates the multi-scale structural context and the mutation effect propagation path.

[0024] (3) Dual-target sequence-activity prediction: A 1024-dimensional overall representation vector is input into a two-branch fully connected regression network. The first two layers of the network are shared hidden layers (1024→512→256), and from the third layer onwards, it is divided into two independent sub-branches, which output the predicted values ​​of editing efficiency and specificity score, respectively. The shared layer captures the common representation of sequence-activity, and the independent branches capture the specific representation of each target. The loss function is the weighted sum of the mean squared error of editing efficiency and the mean squared error of specificity score, and the optimizer is Adam.

[0025] (4) Adaptive temporal evolution modeling: Historical directed evolution experimental data with a total of M rounds (M≥10) are used as the training benchmark. The drift metric (Kernel MMD or KL divergence) of variant sequence distribution between adjacent rounds and the cumulative spatial kernel density distribution of mutation sites in each round are calculated. The optimal historical round split point N is dynamically determined based on the drift metric and the cumulative density distribution: traversing N=3 to N=M-2, the value of N that makes the weighted sum of the Spearman correlation coefficient of the validation set and the drift metric of the training-validation distribution optimal is selected. The model is trained with data from round 1 to round N, and validated with data from round N+1 to round M. The joint evaluation metrics include the Spearman correlation coefficient (measuring ordination consistency) and the ordination stability variance (repeated K times of training-validation, K≥5), ensuring that the model not only predicts accurately but also that the prediction results are stable.

[0026] (5) Pareto Dynamic Trade-off Recommendation: A pre-trained AI optimization model is used to predict the candidate variant library, constructing a two-dimensional Pareto frontier based on editing efficiency and specificity scores. The weight of the comprehensive score is dynamically adjusted according to the sparsity of the frontier in each dimension: α is increased (up to 0.8) when the sparsity of the efficiency dimension is higher, and β is increased (up to 0.5) when the sparsity of the specificity dimension is higher. When making recommendations, candidate variants that uniformly cover the non-dominated solution set of the Pareto frontier are preferred, rather than simply taking the extreme value of the comprehensive score, to ensure the diversity of the recommendation list in the dual-objective space.

[0027] (6) Confidence Calibration Closed-Loop Validation: Recommended candidate variants are synthesized and functionally validated through wet experiments to obtain measured values. Prediction bias is calculated, and variants with biases exceeding the preset confidence interval are marked as anomalies. The anomaly-marked samples are input into a confidence calibration subnetwork composed of a two-layer fully connected residual network for residual learning, and the predicted value correction is output. The main network is updated using the calibrated model parameters to achieve closed-loop iteration. A model update is performed after each round of wet experimental validation, and iteration stops when the Spearman correlation coefficient of the validation set increases by less than 1% for three consecutive rounds.

[0028] Core technology principles

[0029] 1. Multi-scale Seq2Graph: Hierarchical deconstruction of structural interactions

[0030] Protein function depends not only on its amino acid sequence but also on the interactions of amino acid residues at different scales in three-dimensional space. This invention deconstructs the traditional single-threshold (8 Å) contact map into three scales: a near-field contact layer (<5 Å): primarily capturing strong interactions such as hydrogen bonds, salt bridges, and coordination bonds, corresponding to the direct impact of mutations on the local conformation of the protein's active pocket.

[0031] Mid-range contact layer (5–8 Å): mainly captures moderate-intensity interactions such as hydrophobic interactions and van der Waals forces, and corresponding mutations indirectly affect active sites through the structural microenvironment.

[0032] Remote contact layer (8–12 Å): mainly captures long-distance regulation such as allosteric effects and long-range electrostatic interactions, corresponding to the impact of mutations on the overall stability and dynamic conformation of proteins.

[0033] By inputting layers at three scales into independent attention heads, the model can adaptively learn the differentiated contributions of different spatial scales to editing efficiency and specificity. Scale-gated weights are automatically optimized through backpropagation, enabling the model to dynamically adjust its dependence on structures at each scale when predicting mutations at different sites.

[0034] 2. Mutation site masking bias mechanism: to prevent the dilution of mutation signals.

[0035] In standard graph attention networks (GATs), each node uniformly aggregates information from all its neighbors. When a mutation site is located on the protein surface and has many neighbors, the mutation signal is diluted by a large number of non-critical neighbors. The mutation site masking bias mechanism of this invention applies a learnable gain bias (initial value 0.2–0.5) to the mutation site and its first-order structural neighbors before the attention coefficients are calculated. This allows GATs to focus on the mutation region and its immediate structural environment at the beginning of message passing, and then gradually diffuse the mutation effect to the remote regulatory region through multi-layer propagation. This mechanism significantly improves the model's ability to capture epistasis.

[0036] 3. Adaptive temporal evolution modeling: dynamic optimal segmentation

[0037] Directed evolution is a non-stationary stochastic process: in the early rounds, mutation sites are sparse and the sequence distribution is close to that of the wild type; in the middle rounds, beneficial mutations begin to accumulate and the sequence distribution drifts rapidly; in the later rounds, mutations become saturated and the sequence distribution tends to converge. A fixed proportion of the first N rounds of training cannot adapt to the evolutionary dynamics of different proteins.

[0038] The adaptive temporal segmentation framework of this invention dynamically determines the optimal N through two quantization metrics:

[0039] Distribution Shift Metric (D): The Kernel MMD of adjacent rounds is calculated based on the ESM embedding of the variant sequences. The larger the D value, the greater the difference in sequence distribution between rounds, and the more historical rounds the model needs to cover the distribution changes.

[0040] Cumulative mutation density (C): Estimated based on the spatial kernel density of newly added mutation sites in the three-dimensional structure in each round. A slowdown in the growth of the C value indicates that mutation exploration is approaching saturation, and the number of training rounds can be reduced to avoid overfitting.

[0041] The criterion for determining the optimal N is to maximize the Spearman correlation coefficient ρ of the validation set and minimize the drift of the training-validation distribution D: N* = argmax_N [w1·ρ(N) — w2·D(N)].

[0042] Furthermore, ranking stability variance is introduced as a model reliability indicator: the variance σ² of the K independent training-validation cycles is calculated. The model is considered stable only when σ² is below a preset threshold (e.g., 0.005), thus avoiding spurious high correlation caused by random initialization.

[0043] 4. Pareto Dynamic Trade-off: Optimal Coverage in Bi-Objective Space

[0044] There is often a trade-off between editing efficiency and specificity scores: increasing efficiency may reduce specificity, and vice versa. A composite score with fixed weights (such as 0.6 / 0.4) may oversample in a certain region of the Pareto front, leading to homogenization of recommendation variants.

[0045] Before each recommendation, this invention first constructs a predicted Pareto front for the current candidate variant library and calculates the sparsity of the front in both the efficiency and specificity dimensions (the mean Euclidean distance between adjacent non-dominated solutions). If the sparsity in the efficiency dimension is higher, it indicates that there is still room for exploration in that direction, and α is increased to guide the recommendation to expand towards the high-efficiency region; if the sparsity in the specificity dimension is higher, β is increased. This mechanism ensures that the candidate variants recommended in each round uniformly cover the optimal solution set of the dual-objective tradeoff, avoiding resource waste at a single objective extremum point.

[0046] 5. Confidence-based calibration closed loop: distinguishing between systematic bias and random error.

[0047] Traditional closed-loop iteration treats all wet experimental data equally as feedback, resulting in a lack of ability to correct for systematic biases (such as persistent overestimation of specific structural domain abrupt changes). The confidence calibration subnetwork of this invention specifically learns the residual patterns of the main network's prediction bias:

[0048] Samples with prediction bias within the 95% confidence interval are considered normal prediction errors and are directly added to the training set.

[0049] For anomalous samples where the bias exceeds the confidence interval, input the calibration subnetwork (the input is a concatenation of the overall representation vector and the predicted bias, a 2-layer residual structure) to learn the mapping relationship between the bias and the protein representation.

[0050] In the next round of prediction, the main network output is corrected by the calibration subnetwork to gradually eliminate systematic biases.

[0051] Beneficial effects

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. Improved Structural Representation Accuracy: By employing multi-scale contact maps and mutation site masking bias, the short-range direct impact, mid-range indirect transmission, and long-range allosteric regulation of mutation effects are modeled hierarchically, significantly improving the prediction accuracy of epistatic effects. On the TadA deaminase dataset, the prediction accuracy (Spearman correlation coefficient) of this invention reaches 0.91, which is superior to single-scale GAT (0.89), GCN (0.82), and Graph Transformer (0.87).

[0054] 2. Adaptive Temporal Modeling: Compared to fixed partitioning (first 5 training rounds / last 5 validation rounds), the adaptive temporal segmentation framework improves the Spearman correlation coefficient by 0.03 and reduces the ranking stability variance by 40% on the validation set. The model can dynamically and automatically adjust the training boundary according to the actual evolution, avoiding underfitting or overfitting.

[0055] 3. Dual-objective optimization balance: Pareto's dynamic trade-off mechanism improves the coverage uniformity of recommended candidate variants in the editing efficiency-specificity space by 35%, and the proportion of variants that outperform wild-type in both objectives among the top 20 recommendations increases from 60% to 85%.

[0056] 4. Accelerated convergence of closed-loop iteration: The confidence calibration subnetwork improves the convergence speed of the model in closed-loop iteration by 50%. After three rounds of iteration, the Spearman correlation coefficient of the validation set increased from the initial 0.91 to 0.94, while the traditional direct feedback strategy only increased to 0.92.

[0057] 5. Significantly reduced wet experiment costs: By replacing large-scale random screening with AI prediction, the number of variants constructed is reduced by 99.4%, the number of wet experiment verifications is reduced by 99.55%, and the cycle is shortened by 66.7%.

[0058] 6. High versatility: Applicable to the directed evolution optimization of various gene editing tool proteins such as ABE deaminase components, CBE deaminase components, and Cas protein variants. Attached Figure Description

[0059] Figure 1This is a schematic diagram of the overall process of the gene editing tool protein directed evolution optimization method based on multi-scale mutation-aware graph network and adaptive temporal evolution modeling of the present invention.

[0060] Figure 2 This is a schematic diagram of how the multi-scale Seq2Graph method in this invention converts protein sequences into multi-layer graph structure data. Solid lines represent near-range contact edges (<5 Å), dashed lines represent mid-range contact edges (5–8 Å), and dotted lines represent far-range contact edges (8–12 Å).

[0061] Figure 3 This is a schematic diagram of the architecture of the Multi-Scale Graph Attention Network (MSMA-GAT) in this invention, including a near-range attention head group, a mid-range attention head group, a long-range attention head group, a scale-gated fusion layer, a mutation site masking bias mechanism, and a global pooling layer.

[0062] Figure 4 This is a schematic diagram of the adaptive temporal evolution modeling framework in this invention, illustrating the process of distribution drift measurement, mutation cumulative density calculation, and determination of the dynamic optimal split point N.

[0063] Figure 5 This is a schematic diagram of the Pareto dynamic trade-off recommendation mechanism in this invention, illustrating the process of predicting the Pareto frontier, evaluating sparsity, and recommending non-dominated solution sets.

[0064] Figure 6 This is a schematic diagram of the confidence calibration closed-loop verification mechanism in this invention, illustrating the process of anomaly labeling, calibration subnetwork residual learning, and model update.

[0065] Figure 7 This is a schematic diagram of the closed-loop iteration structure of the system of the present invention, the wet experimental verification platform, and the automated screening platform.

[0066] Figure 8 This is a Pareto front comparison of TadA deaminase editing efficiency and specificity scores before and after optimization of the method of this invention. Detailed Implementation

[0067] The present invention will be further described in detail below with reference to specific embodiments and accompanying drawings. These embodiments are only for explaining the present invention and do not constitute any limitation on its scope.

[0068] Example 1: Directed Evolutionary Optimization of TadA Deaminase Based on Multi-Scale Mutation-Aware Graph Network and Adaptive Temporal Evolution Modeling

[0069] 1. Data Acquisition and Dataset Construction

[0070] The target protein was TadA deaminase, a core component of the adenine base editor (ABE). The wild-type amino acid sequence of TadA was obtained from the UniProt database (UniProt ID: P0A816). Amino acid sequences of TadA variants and their corresponding editing efficiency (%) and editing specificity scores (0–100) were obtained from previous directed evolution wet experiments at our institution. A total of 5000 sequence-function data pairs of TadA variants were obtained, distributed across 10 rounds of directed evolution experiments (500 variants per round).

[0071] Each data record includes: variant number, amino acid sequence (including mutation site and mutation type), editing efficiency (%), editing specificity score (0–100), and evolutionary cycle. This data is used to construct a protein variant sequence-function mapping dataset, stored in CSV format.

[0072] 2. Multi-scale Seq2Graph graph structure transformation

[0073] The protein sequences in the dataset were converted into multi-level graph structure data using the multi-scale Seq2Graph method:

[0074] (1) Node feature construction: Based on the protein sequence, a 1280-dimensional embedding vector for each residue was extracted using the pre-trained protein language model ESM-1b (version 1.0). Simultaneously, each residue was encoded with its mutation type: wild-type residues were encoded as [0,0,0], substitution mutations as [1,0,0], insertion mutations as [0,1,0], and deletion mutations as [0,0,1]. The 1280-dimensional embedding was concatenated with the 3-dimensional mutation type encoding to obtain the initial 1283-dimensional node features.

[0075] (2) Construction of multi-layer edge relationships: Based on the three-dimensional structure of TadA predicted by AlphaFold2 (version 2.3), the Euclidean distance between Cα atoms of each residue was calculated. - Near-range contact edge: distance < 5 Å; - Mid-range contact edge: 5 Å ≤ distance < 8 Å; - Long-range contact edge: 8 Å ≤ distance < 12 Å.

[0076] (3) Multi-layer graph construction: Construct a three-layer undirected graph G={G_near, G_mid, G_far}, where the node set of each layer is the same (|V|=167), and the edge set corresponds to the three distance thresholds mentioned above.

[0077] 3. Feature extraction using multi-scale graph attention networks

[0078] Multiscale Graph Attention Network (MSMA-GAT) is used to extract features from multi-layer graph structure data:

[0079] (1) Network architecture: 3-layer GAT, with 12 attention heads in each layer, divided into short-range head group (4 heads), mid-range head group (4 heads), and long-range head group (4 heads). The first layer has an input dimension of 1283 and an output dimension of 1024; the second layer has an input dimension of 1024 and an output dimension of 1024; the third layer has an input dimension of 1024 and an output dimension of 1024.

[0080] (2) Mutation site mask bias: For known mutation sites and their first-order structural neighbors (directly connected by short-range contact edges), an initial gain bias of 0.3 is applied before the attention coefficient calculation. This bias is used as a learnable parameter and is fine-tuned during training through backpropagation.

[0081] (3) Scale-gated fusion: After each head group independently calculates its attention, it is weighted and fused using learnable scale-gated weights (three scalars, normalized by Softmax).

[0082] (4) Global pooling: After three layers of message passing, the features of each layer node are weighted and aggregated and global average pooled to obtain a 1024-dimensional protein sequence overall representation vector.

[0083] 4. Establishment of a dual-target sequence-activity prediction model

[0084] Input the 1024-dimensional global representation vector into the two-branch fully connected regression network:

[0085] (1) Shared layers: Layer 1 has 1024-dimensional input and 512-dimensional output; Layer 2 has 512-dimensional input and 256-dimensional output. The activation function is ReLU, and the dropout rate is 0.1.

[0086] (2) Independent Branch Layer: Starting from the 3rd layer, it is divided into two sub-branches. Editing efficiency branch: 256-dimensional input, 1-dimensional output; Specificity scoring sub-branch: 256-dimensional input, 1-dimensional output. The output layer has no activation function.

[0087] (3) Training configuration: loss function is MSE (efficiency) + MSE (specificity), optimizer is Adam, learning rate is 0.001, batch size is 64, training rounds are 200, and the early stopping strategy is to stop when the validation set loss does not decrease for 20 consecutive rounds.

[0088] 5. Adaptive temporal evolution modeling

[0089] The training baseline was 10 rounds of historical directed evolution experiments.

[0090] (1) Distribution drift metric calculation: Based on the ESM-1b embedding of the variant sequences in each round, the Kernel MMD between adjacent rounds was calculated. The MMD value between rounds 5 and 6 was 0.12, and that between rounds 6 and 7 was 0.08, indicating that the distribution drift slowed down after round 5.

[0091] (2) Cumulative mutation density calculation: Three-dimensional spatial nuclear density estimation (bandwidth 3Å) was performed on the newly added mutation sites in each round. The growth of the peak nuclear density slowed down after the 5th round, indicating that mutation exploration tended to saturate.

[0092] (3) Determining the dynamic split point: Traverse N=3 to N=8 and calculate the Spearman correlation coefficient ρ and training-validation MMD value for each N value. The weighted sum is optimal when N=5 (ρ=0.91, MMD=0.15); when N=6, ρ increases slightly to 0.912 but MMD increases to 0.18. Therefore, the optimal split point is determined to be N*=5.

[0093] (4) Ranking stability verification: K=5 independent training-verifications were performed with N=5. The Spearman correlation coefficients were 0.905, 0.912, 0.908, 0.914 and 0.910, respectively, and the variance σ²=0.0011, which is lower than the threshold of 0.005, so the model is considered stable.

[0094] 6. Pareto Dynamic Trade-off Recommendation

[0095] The trained AI optimization model was used to predict the TadA full-site saturation mutation library (167×19=3173 single-point mutation variants):

[0096] (1) Construct a two-dimensional prediction scatter plot of editing efficiency-specificity score and extract the Pareto front non-dominated solution set (47 variants in total).

[0097] (2) Calculate the sparsity of the frontier: the average spacing of the efficiency dimension is 0.18, and the average spacing of the specificity dimension is 0.12. The sparsity of the efficiency dimension is higher, so the weights are adjusted to α=0.7 and β=0.3.

[0098] (3) Sort by comprehensive score, recommend the top 20 candidate variants from the non-dominated solution set, and output the amino acid mutation scheme.

[0099] 7. Confidence calibration closed-loop verification

[0100] The 20 recommended candidate variants were synthesized using gene technology (Sangon Biotech (Shanghai) Co., Ltd.), and expression vectors were constructed and transfected into the HEK293T cell line (ATCC CRL-1573).

[0101] (1) Wet experiment determination: For the three endogenous target sites, the editing efficiency was calculated by Sanger sequencing and TIDE analysis (version 2.0); off-target sites were detected by whole genome sequencing (sequencing depth 30×), and the specificity score was calculated.

[0102] (2) Anomaly labeling and calibration: Calculate prediction bias. The prediction efficiency of the candidate variant TadA-M12 is 78.5%, while the measured efficiency is 52.3%. The bias exceeds the 95% confidence interval and is therefore labeled as an anomaly. Its representation vector and bias value are input into the confidence calibration subnetwork (2-layer residual network, 128-dimensional hidden layer) to learn the bias correction mode.

[0103] (3) Model update: Update the main network using the calibrated parameters and start the next iteration.

[0104] 8. Iterative optimization results

[0105] After 3 rounds of closed-loop iteration:

[0106] Round 1: The optimal variant TadA-V3 (P48V, R51H, whose complete amino acid sequence is shown in SEQ ID NO:2) had an editing efficiency of 61.3%, a specificity score of 82, and an off-target rate of 0.21%.

[0107] At the same time, the second and third rounds will also be standardized:

[0108] Round 2: The optimal variant TadA-V6 (P48V, R51H, L145F, whose complete amino acid sequence is shown in SEQ ID NO:3) had an editing efficiency of 74.8%, a specificity score of 89, and an off-target rate of 0.15%.

[0109] Round 3: The optimal variant TadA-V9 (P48V, R51H, L145F, D147T, I156R, the complete amino acid sequence of which is shown in SEQ ID NO:4) had an editing efficiency of 89.2%, a specificity score of 96, and an off-target rate of 0.08%.

[0110] Compared to wild-type TadA (editing efficiency 26.3%, specificity score 65, off-target rate 0.38%), TadA-V9 exhibits approximately 3.39-fold improved editing efficiency and a 78.9% reduced off-target rate. This variant contains five co-mutation sites: P48V and R51H are located within the 4.2 Å active pocket, L145F and D147T are located in the surface flexible region, and I156R is located in the mid-range regulatory region between the two.

[0111] Example 2: Applicability verification of different gene editing tool proteins

[0112] Following the method in Example 1, APOBEC1 (CBE deaminase, UniProt ID: P41238) and SpCas9 (UniProt ID: Q99ZW2) were optimized respectively.

[0113] APOBEC1 optimization results: Wild-type editing efficiency 32.1%, specificity score 58, off-target rate 0.42%; after 3 rounds of iterations, the optimal variant editing efficiency was 83.2%, specificity score 93, off-target rate 0.14%. Efficiency improved by 2.59 times, and off-target rate decreased by 66.7%.

[0114] SpCas9 optimization results: Wild-type editing efficiency 45.3%, specificity score 62, off-target rate 0.56%; after 3 iterations, the optimal variant editing efficiency was 90.1%, specificity score 95, off-target rate 0.19%. Efficiency improved by 1.99 times, and off-target rate decreased by 66.1%.

[0115] The above results demonstrate that the method of the present invention is applicable to a variety of gene editing tool proteins and has good versatility.

[0116] Example 3: Comparative Experiment of Multi-Scale Graph Attention Networks and Single-Scale Networks

[0117] Based on the data and process of Example 1, the following four architectures are compared:

[0118] Single-scale GAT (8 Å) no 0.89 6 / 10 Multi-scale GAT (This invention, without masking) no 0.90 7 / 10 Multi-scale GAT (This invention, with masking) yes 0.91 9 / 10 Graph Transformer no 0.87 7 / 10

[0119] Experimental results show that multi-scale structural modeling improves prediction accuracy by 0.02 compared to single-scale modeling, and further improves it by 0.01 after adding mutation site masking bias, significantly increasing the recommendation success rate. The reason for this is that multi-scale modeling distinguishes the differentiated contributions of strong / weak / long-range interactions, and mutation masking bias prevents the key mutation signals in the active pocket from being diluted by residues exposed to a large amount of solvent on the surface.

[0120] Example 4: Comparative Experiment of Adaptive Temporal Segmentation and Fixed Temporal Segmentation

[0121] Based on the data from Example 1, the following three training strategies are compared:

[0122] Strategy A (Adaptive in this invention) N*=5 0.91 0.0011 0.88 Strategy B (fixed first 5 rounds) N=5 0.89 0.0032 0.84 Strategy C (fixed first 7 rounds) N=7 0.87 0.0045 0.82

[0123] Experimental results show that adaptive temporal segmentation not only improves prediction accuracy but also significantly reduces the ranking stability variance, demonstrating that the model is more robust to random initialization. The external validation set also outperforms the fixed segmentation strategy.

[0124] Example 5: Comparative Experiment of Pareto Dynamic Trade-offs and Fixed Weight Recommendations

[0125] Fixed weight 12 / 20 62% 0.72 Pareto dynamic trade-offs 17 / 20 89% 0.75

[0126] Based on the data from the third iteration of Example 1, the recommendation performance of fixed weights (α=0.6, β=0.4) and Pareto dynamic tradeoffs is compared:

[0127] Pareto's dynamic tradeoffs increased the proportion of variants that outperformed wild-type in both objectives from 60% to 85%, and improved Pareto frontier coverage by 27%, significantly improving recommendation quality.

[0128] Example 6: Comparative Experiment of Confidence Calibration Closed Loop and Traditional Direct Feedback

[0129] Three rounds of closed-loop iteration were performed according to Example 1, and the confidence calibration subnetwork was compared with the traditional direct data feedback:

[0130] Traditional direct feedback 0.89 0.91 0.92 18% Confidence calibration closed loop 0.89 0.92 0.94 8%

[0131] The confidence calibration closed loop improved the accuracy of the third round of verification by 0.02 and reduced the proportion of outliers by 10 percentage points, proving that the calibration subnetwork effectively identified and corrected systematic biases.

[0132] Example 7: Comparison of AI Optimization Model and Wet Experiment Costs

[0133] Number of variants 10,000 60 (3 rounds x 20) 99.4% wet experiment verification number 10,000 45 (3 rounds x 15) 99.55% wet test cycle 12 weeks 4 weeks 66.7% Reagent and labor costs 500,000 yuan 120,000 yuan 76%

Claims

1. A method for directed evolutionary optimization of gene-editing tool proteins based on multi-scale mutation-aware graph networks and adaptive temporal evolution modeling, characterized in that, Includes the following steps: (1) Multi-scale graph structure conversion steps: A multi-scale Seq2Graph method was used to convert protein variant sequences into multi-layer graph structure data. This multi-layer graph structure data includes a short-range contact layer, a mid-range contact layer, and a long-range contact layer. Specifically: the edges of the short-range contact layer represent strong contact relationships where the spatial distance between the Cα atoms of amino acid residues is less than 5 Å; the edges of the mid-range contact layer represent weak contact relationships where the spatial distance between the Cα atoms of amino acid residues is greater than or equal to 5 Å and less than 8 Å; and the edges of the long-range contact layer represent long-distance regulatory relationships where the spatial distance between the Cα atoms of amino acid residues is greater than or equal to 8 Å and less than 12 Å. The initial features of each layer node are the concatenation vector of the residue embedding vector extracted from the pre-trained protein language model and the mutation type encoding vector. (2) Steps for extracting features from mutation sites: A multi-scale graph attention network is used to extract features from the multi-layer graph structure data. The multi-scale graph attention network includes: a short-range attention head group, a mid-range attention head group, and a long-range attention head group. Each head group independently processes the graph layer at its corresponding scale. A mutation site masking bias mechanism is used to apply an initial gain bias to known mutation sites and their first-order structural neighbor nodes during attention calculation. After message passing at each scale, the features of each layer node are weighted, aggregated, and globally averaged to obtain an overall representation vector that integrates the multi-scale structural context and the mutation effect propagation path. (3) Dual-target sequence-activity prediction steps: The overall representation vector is input into a two-branch fully connected regression network, which simultaneously outputs the predicted values ​​of editing efficiency and specificity score. The two branches of the two-branch fully connected regression network share the first two hidden layers, and from the third layer onwards, they are divided into independent sub-branches corresponding to editing efficiency and specificity score, respectively. (4) Adaptive temporal evolution modeling steps: Using multi-round historical directed evolution experimental data as the training benchmark, an adaptive temporal segmentation framework is employed to train an AI optimization model. This adaptive temporal segmentation framework includes: calculating the drift metric of variant sequence distribution between adjacent rounds; calculating the cumulative density distribution of mutation sites in each round; dynamically determining the optimal historical round segmentation point N based on the drift metric and cumulative density distribution, ensuring that the round data included in the training set satisfies the balance between sequence distribution continuity and mutation information saturation; training the sequence-activity prediction model using data from rounds 1 to N, and using data from round N+1 and subsequent rounds as the validation set; using the Spearman correlation coefficient and ordination stability variance of the predicted and measured values ​​on the validation set as joint evaluation indicators, and determining the optimal model when the ordination stability variance is below a preset threshold and the Spearman correlation coefficient reaches its peak value. (5) Pareto dynamic trade-off recommendation steps: A trained AI optimization model is used to predict the editing efficiency and specificity score of each variant in the candidate variant library, and a Pareto front is constructed. The combined score weight of editing efficiency and specificity score is dynamically adjusted according to the sparsity distribution of the predicted Pareto front, so that the recommended candidate variants uniformly cover the non-dominated solution set of the Pareto front. The top N candidate variants are recommended according to the combined score, and the corresponding amino acid mutation schemes are output. (6) Confidence calibration closed-loop verification steps: Recommended candidate variants are synthesized and functionally verified through wet experiments to obtain measured editing efficiency and measured specificity scores. The prediction deviation between the predicted and measured values ​​of each candidate variant is calculated, and variants with prediction deviations exceeding a preset confidence interval are marked as anomalies. The anomaly-marked samples are input into the confidence calibration subnetwork for residual learning, and the sequence-activity prediction model is updated using the calibrated model parameters to achieve a closed-loop iteration of AI prediction and wet experimental verification.

2. The method according to claim 1, characterized in that, The gene editing tool proteins include the deaminase component of the adenine base editor (ABE), the deaminase component of the cytosine base editor (CBE), or a Cas protein variant of the CRISPR-Cas system.

3. The method according to claim 1, characterized in that, The pre-trained protein language model in step (1) is ESM-1b or ESM-2, and the extracted residue embedding dimension is 1280. The mutation type encoding vector includes one-hot encoding of substitution mutation, insertion mutation and deletion mutation. The three-dimensional structure is derived from the experimentally determined structure or the AlphaFold2 predicted structure.

4. The method according to claim 1, characterized in that, The multi-scale graph attention network in step (2) has 3 layers, each containing 12 attention heads, including 4 near-range attention heads, 4 mid-range attention heads, and 4 long-range attention heads. Each head group independently calculates its attention coefficients and then performs weighted fusion using learnable scale-gated weights. The mutation site mask bias mechanism applies an initial gain bias of 0.2 to 0.5 to the mutation site and its first-order neighbors.

5. The method according to claim 1, characterized in that, The drift metric in step (4) is the kernel maximum mean difference (Kernel MMD) or KL divergence based on variant sequence embedding; the cumulative density distribution is the spatial kernel density estimate of the newly added mutation sites in the three-dimensional structure of the protein in each round; the method for dynamically determining the optimal historical round split point N is: traversing N=3 to N=M-2, and selecting the N value that makes the weighted sum of the Spearman correlation coefficient of the validation set and the drift of the training-validation distribution optimal; the ranking stability variance is the variance of the Spearman correlation coefficient after K repetitions of training-validation, where K≥5.

6. The method according to claim 1, characterized in that, The specific method of Pareto dynamic trade-off in step (5) is as follows: when the sparsity of the predicted Pareto front in the editing efficiency dimension is higher than that in the specificity score dimension, the editing efficiency weight α is increased and the specificity score weight β is decreased; when the sparsity of the specificity score dimension is even higher, β is increased and α is decreased; where α+β=1, α∈[0.5,0.8], β∈[0.2,0.5]; the non-dominated solution set is the solution set that is not simultaneously dominated by other candidate variants in both the editing efficiency and specificity score objectives.

7. The method according to claim 1, characterized in that, The confidence calibration subnetwork in step (6) is a 2-layer fully connected residual network. The input is the concatenation feature of the overall representation vector and the prediction bias, and the output is the prediction correction amount. The closed-loop iteration performs a model update once after each round of wet experimental verification. The iteration stops when the Spearman correlation coefficient of the validation set increases by less than 1% for three consecutive rounds of iteration.

8. A directed evolutionary optimization system for gene-editing tool proteins based on multi-scale mutation-aware graph networks and adaptive temporal evolution modeling, characterized in that, include: The module includes a multi-scale graph structure conversion module, which uses the multi-scale Seq2Graph method to convert protein variant sequences into multi-layer graph structure data containing short-range, mid-range, and long-range contact relationships; a mutation site-aware feature extraction module, which uses a multi-scale graph attention network to extract features from the multi-layer graph structure data, including a short-range attention head group, a mid-range attention head group, and a long-range attention head group, and includes a mutation site masking bias mechanism; and a dual-target sequence-activity prediction module, which inputs the extracted overall representation vector into a two-branch fully connected regression network and simultaneously outputs edit efficiency prediction values ​​and specificity score prediction values. The adaptive temporal evolution modeling module is used to dynamically determine the optimal historical round split point N based on sequence distribution drift metric and mutation cumulative density distribution using multi-round historical directed evolution experimental data as the training benchmark, and to train the model using data from round 1 to round N, and validate the model using data from round N+1 and subsequent rounds. The Pareto dynamic trade-off recommendation module is used to construct a predicted Pareto frontier and dynamically adjust the comprehensive score weights according to the frontier sparsity, so that the recommended candidate variants uniformly cover the non-dominated solution set. The confidence calibration closed-loop verification module is used to receive wet experiment verification results, mark variants whose prediction deviation exceeds the confidence interval as anomalies, and update model parameters through the confidence calibration sub-network. A memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method according to any one of claims 1 to 7.

9. The system according to claim 8, characterized in that, The system communicates with the directed evolution automated screening platform via an API interface to realize the transmission of instructions from prediction results to automated screening and the data feedback of screening results; the confidence calibration closed-loop verification module also includes a wet experimental data quality assessment submodule, which is used to filter the feedback data based on the Sanger sequencing peak quality value and the whole genome sequencing coverage depth, and only use the data that meets the quality standards for model updates.

10. A gene-editing tool protein optimized by the method according to any one of claims 1 to 7, characterized in that, The gene editing tool protein contains at least three co-mutation sites, at least one of which is located within 5 angstroms of the protein's active pocket, and at least one is located in a flexible region on the protein surface; the editing efficiency of the gene editing tool protein is significantly improved compared to the wild-type protein, and its off-target rate is significantly reduced compared to the wild-type protein.

Images