A machine learning-guided biological sequence engineering modification method and device

Abstract

The present invention provides a method and apparatus for biosequence engineering guided by machine learning. Specifically, a Bayesian optimization-guided evolutionary algorithm (BO-EVO) is provided, which combines Bayesian optimization (BO) and evolutionary algorithm (EVO). EVO solves the problem of excessive computational complexity caused by BO locating the global optimum by brute force search of the entire design space. At the same time, the exploratory nature of BO is used to neutralize the greed and under-exploration shortcomings of EVO to achieve efficient iteration between machine learning models and robotic experiments, so as to economically obtain new protein variants of high value. The method of the present invention is expected to achieve efficient and scalable computing and exploration, and provide an efficient biosequence engineering solution.

Description

Technical Field

[0001] The present invention belongs to the field of bioengineering, and specifically relates to a method and device for biosequence engineering modification guided by machine learning. Background Art

[0002] Directed protein evolution techniques employ multiple rounds of gene mutation and phenotypic screening, often limited by the throughput of experimental techniques. Using model-based virtual screening techniques, machine learning has reduced the screening burden to a level acceptable for manual experimentation. Meanwhile, robotics allows for large-scale screening and rapid iteration in protein engineering; however, this technology has yet to be applied to machine learning-assisted directed evolution.

[0003] The protein fitness landscape describes a metaphorical, high-dimensional surface that links amino acid sequences to properties of interest, namely, fitness. Exploring this landscape in protein engineering is extremely challenging: 1. The search space grows exponentially with increasing amino acid sequence length; 2. Functional proteins are extremely rare; 3. High-value protein sequences decay exponentially with increasing fitness; 4. The fitness landscape is rugged due to epistatic effects; and 5. Experimental testing is expensive, inefficient, and labor-intensive. As an effective protein engineering strategy, directed evolution employs multiple rounds of mutation library construction and screening, but only the optimal mutations are fixed in each round. However, this greedy strategy can lead to local optima, especially on rugged fitness landscapes.

[0004] Machine learning algorithms are finding increasing application in both modeling protein fitness landscapes and guiding protein engineering. To learn sequence-structure-function relationships from labeled data, supervised learning is used to predict various properties, including three-dimensional structure, thermal stability, fluorescence intensity, ligand binding affinity, and catalytic performance. Meanwhile, gene and protein sequences are accumulating at an unprecedented rate in public databases (e.g., UniProt). Unsupervised learning models, such as UniRep, TAPE, ESM-1v, and ProtT5-XL-U50, learn protein representations by discovering hidden patterns from large amounts of unlabeled protein sequence data. Despite initial proof-of-concept results, existing fitness models are still limited by data scarcity and biases. For example, evolutionary data lack label diversity due to the absence of low-fitness and extremely high-fitness sequences. Furthermore, because functional characterization is expensive and laborious, sequence-function label data can only cover a limited region of the design space. Both of these factors lead to learned models being ill-conditioned in areas not covered by the data.

[0005] To guide protein engineering, both supervised and unsupervised models are being used to improve sample efficiency. For example, machine learning-assisted directed evolution (MLDE) first uses a physics-based ddG method to pre-screen all possible sequences in the design space, obtaining 384 candidate sequences for experimental characterization. This method then trains a fitness model, which is then used to prioritize the top 96 sequences. On the GB1 dataset, MLDE achieved an 81-fold success rate compared to directed evolution. Furthermore, ECNet, which leverages homologous sequence information as supervision to predict the effects of higher-order mutations, successfully engineered TEM-1 β-lactamase, achieving an 8-fold increase in ampicillin resistance after 2-6 mutations. Furthermore, by unsupervised learning of protein representations from over 20 million natural sequences, a fitness model for low-N protein engineering was obtained with only 24 sequence-function data points. This fitness model was optimized by simulated annealing, successfully identifying superior proteins on GFP and TEM-1 β-lactamase datasets. It is worth noting that, due to limited manpower, existing machine learning-guided protein engineering methods attempt to reduce experimental throughput and the number of iterations.

[0006] On the other hand, in bioengineering, biofoundries accelerate the "design-build-test-learn" closed-loop process through physical and information automation. When applied to protein engineering, automation allows for large-scale library construction and screening in a short period of time. Consequently, large quantities of high-quality new sequence-function data can be iteratively collected to improve model prediction and sequence design. However, this data capability is underutilized in algorithm-guided protein engineering.

[0007] Bayesian optimization, a principled approach for optimizing "black-box" functions, obtains candidate objects by optimizing a given sampling function, which is responsible for balancing exploration and exploitation. Bayesian optimization is widely used in protein engineering, pathway design, and fermentation strain design.

[0008] In order to use Bayesian optimization to guide protein engineering on combinatorial fitness landscapes, Bayesian optimization needs to be improved. First, although BO only obtains one sample at a time, batches of samples can be constructed and tested in parallel by robots. Batch BO has been used to design biological sequences. Batch methods are divided into two categories: iterative generation of batch data or one-time acquisition of batch data. However, how to balance the exploration and utilization of single batch data remains an open problem. Second, BO locates the global optimum by brute force searching the entire design space, targeting the combinatorial mutation library of proteins (N sites ~ 20 N design space), the computational complexity of BO increases exponentially with the problem dimension. Summary of the Invention

[0009] Faced with how to balance the exploration of single batch data in the batch processing process, and how to reduce the excessive computational load of the mutant protein library BO, as well as the pathological problems of machine learning models, insufficient exploration capabilities and scalability problems in biological sequence design tasks. The present invention proposes an evolutionary algorithm guided by Bayesian optimization (BO-EVO) to achieve efficient iteration between machine learning models and robotic experiments, so as to economically obtain new protein variants of high value. The method of the present invention combines BO and EVO, and solves the problem of excessive computational load caused by BO when locating the global optimum through a brute force search of the entire design space through EVO. At the same time, the exploratory nature of BO is used to neutralize the greed and under-exploration shortcomings of the evolutionary algorithm. Therefore, in combination with EVO and BO, it is expected to achieve efficient and scalable computing and exploration, and provide an efficient biological sequence engineering transformation solution.

[0010] In one aspect, the present invention provides a method for biological sequence engineering guided by machine learning, comprising the following steps:

[0011] S1) Data acquisition: obtaining biological sequences and their corresponding fitness data;

[0012] S2) Model training: Machine learning is performed on the obtained biological sequences and their corresponding fitness data to obtain a fitness prediction model and model uncertainty for the biological sequences;

[0013] S3) Obtaining a first seed: Obtaining a first seed sequence from the biological sequence obtained in step S1) and its corresponding fitness data through fitness-based sampling;

[0014] S4) generating a mutation subspace: performing random mutations on the sites to be mutated in the seed sequence obtained in step S3) or S6), and forming a mutation subspace with a set of mutated biological sequences, wherein the single-point mutation rate of the random mutation is the inverse of the number of sites to be mutated;

[0015] S5) Mutation Subspace Bayesian Optimization: Bayesian optimization is used on the mutation subspace to select a single candidate sequence for experimental query, where the sampling function is the upper confidence bound (UCB) and the surrogate model is the fitness prediction model obtained in S2);

[0016] S6) New seed generation: First, the uncertainty of the candidate sequence obtained in S5) is predicted using the prediction model obtained in S2). If the uncertainty is no greater than twice the uncertainty of the first seed sequence in this round, the candidate sequence is used as the next seed sequence in this round and is used to form the next mutation subspace. If the uncertainty is greater than twice the uncertainty of the first seed sequence in this round, the next seed sequence in this round is obtained according to step S3) and is used to form the next mutation subspace.

[0017] S7) repeating the cycle from S4) to S6) until a preset throughput cycle is met, thereby obtaining a batch of candidate sequences to be experimentally queried;

[0018] S8) performing experiments on the batch of candidate sequences to be queried in the experiment obtained in S7) to obtain candidate sequences and detecting their corresponding fitness values;

[0019] S9) adding the candidate sequence and its corresponding fitness value to the biological sequence and its corresponding fitness data in S1) and using them for the next round of model training to obtain the fitness prediction model and model uncertainty of the biological sequence in the next round;

[0020] S10) Using the latest seed sequence obtained in S6) as the first seed sequence for the next round, and repeating S4) to S9) until the preset rounds are met or the expected mutants are screened out.

[0021] Furthermore, in step S1), the biological sequence and its corresponding fitness data are obtained from a known fitness landscape, from a known sequence and a corresponding fitness dataset, or from an experimentally measured sequence and corresponding fitness data.

[0022] Furthermore, in step S1), the sequence and its corresponding fitness data obtained from the experimentally measured sequence and the corresponding fitness data can be the above-mentioned candidate sequence and the fitness result obtained after testing it, or can be other known sequences and corresponding fitness data test results.

[0023] Furthermore, the sequence is a protein sequence, a polypeptide sequence, a ribonucleic acid sequence, or a deoxyribonucleic acid sequence.

[0024] Furthermore, in step S2), the prediction model is selected from a Gaussian process model, a Bayesian neural network prediction model, an integrated model, an evidence deep learning model or other prediction uncertainty models.

[0025] Furthermore, in step S2), before the prediction model learning process, the sequence is first expressed in a digital form.

[0026] Furthermore, in step S3), the first seed sequence obtained may be a natural sequence or a mutant sequence.

[0027] Furthermore, in step S4), the number of mutated biological sequences in the mutation subspace is 1.0-3.0 times the number of sites to be mutated.

[0028] Furthermore, in step S5), the value of β in UCB is 0.05-0.25.

[0029] Further, in step S7), the preset flux is limited to between 200-500.

[0030] Furthermore, in step S10), the preset rounds are at least 2 rounds.

[0031] Furthermore, in step S8), the method for conducting experiments on a batch of candidate sequences to be queried is to construct biological sequences and test fitness values using an engineering method.

[0032] Furthermore, the method for conducting the experiment in step S8) is to first obtain the candidate sequences to be queried by the experiment using a chemical or biological method, and then actually test the fitness value of each candidate sequence.

[0033] Another aspect of the present invention provides a machine learning-guided biological sequence engineering device, which is a device capable of implementing the above steps S1)-S10).

[0034] The biological sequence engineering modification device includes the following modules: a biological sequence and fitness data acquisition module, a machine learning model module, and a biological sequence recommendation module;

[0035] The biological sequence and fitness data acquisition module is used to store and call sequences and their fitness data, convert the called sequences into digital codes; and return the corresponding fitness values based on the batch of candidate sequences generated by the biological sequence recommendation module;

[0036] The machine learning model module is used to call the biological sequence from the biological sequence and fitness data acquisition module, perform machine learning based on the called data, and form a prediction model and model uncertainty between the biological sequence and its fitness;

[0037] The biological sequence recommendation module is used to obtain a batch of candidate sequences to be experimentally queried under the guidance of the machine learning model module. The biological sequence recommendation module can call data in the biological sequence and fitness data acquisition module, and based on the first seed sequence obtained by fitness sampling, perform random mutations on the first seed sequence, and form a mutation subspace with the random mutation sequence. Bayesian optimization is used in the mutation subspace to select a single candidate sequence, wherein the sampling function is UCB and the proxy model is the prediction model formed by the machine learning model module;

[0038] Confirm the uncertainty of a single candidate sequence. If the uncertainty is no more than 2 times the uncertainty of the first seed sequence in this round, the candidate sequence will be used as the next seed sequence in this round and used to form the mutation subspace of this round. If the uncertainty is more than 2 times the uncertainty of the first seed sequence in this round, the next seed sequence in this round will be obtained from the called sequence based on fitness sampling. The next cycle will be started until the preset throughput and batch of candidate sequences are reached.

[0039] Furthermore, the biological sequence recommendation module also includes a unit for starting the next round, which transmits the generated batch of candidate sequences to the biological sequence and its fitness data acquisition module, and guides the batch of candidate sequences generated by the biological sequence and its fitness data acquisition module to return corresponding fitness values; the batch of candidate sequences generated by the biological sequence and its fitness data acquisition module and their corresponding fitness values are incorporated into the existing biological sequence and its fitness data; and is used to guide the machine learning model module to further generate a new prediction model, and the new prediction model in the machine learning model module is used to guide the next round of recommendations of the biological sequence recommendation module.

[0040] Furthermore, the biological sequence and fitness data acquisition module includes a recording unit for recording information, an encoding unit for converting the called sequence into a digital code, and a batch experiment unit for returning corresponding fitness values based on the batch of candidate sequences generated in the biological sequence recommendation module.

[0041] Beneficial effects

[0042] 1) The present invention's biological sequence engineering method utilizes batch "dry-wet" iterations for optimization, which not only reduces the experimental cost and volume of wet experiments, but also improves algorithm performance. "Dry" or "dry experiment" refers to computer-generated data collection, training, simulation, or prediction, while "wet" or "wet experiment" refers to actual experiments. The present invention utilizes an algorithm and wet experiment iteration optimization approach—the algorithm guides wet experiments in exploring the fitness landscape, thereby obtaining new sequence-fitness relationship data. The data obtained from the wet experiments is then used to update the machine learning model, thereby improving algorithm performance and further guiding wet experiment exploration.

[0043] 2) Usually, the optimization of the sampling function in the Bayesian optimization method usually uses "brute force" to evaluate the entire search space. For protein sequences, the types of its constituent units - amino acids are at least 20 (natural amino acids). As the protein sequence is extended, the amount of data is very large. The conventional Bayesian optimization method is used to perform brute force exploration of the entire space, which requires too much computation, resulting in a significant increase in the exploration cost. The present invention generates a subspace through an evolutionary method and uses Bayesian optimization in the subspace to effectively alleviate the scalability of Bayesian optimization. By obtaining the subspace in batches and screening out candidate sequences from the subspace, the scalability and data scalability are guaranteed at the same time. Therefore, the method of the present invention is fast, requires less computing resources, and can be used for engineering tasks of longer sequences.

[0044] 3) The present invention not only considers the relationship between fitness and sequence during the modeling process, but also the uncertainty of the predicted values, establishing an uncertainty model for this purpose. Because machine learning models are difficult to accurately predict outside the data support through extrapolation, estimating and utilizing the uncertainty of the predicted values can effectively prevent the optimization algorithm from "using" the model to make inaccurate predictions outside the data support.

[0045] 4) The method of the present invention improves the efficiency of biological sample transformation and evolution. By sampling less than 1% of all possible mutations, the method of the present invention can achieve a performance improvement of more than 7 times.

[0046] 5) The Bayesian optimization-guided evolutionary algorithm used in the present invention does not require target object related structures or homologous sequences. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Flowchart of the method of the present invention.

[0048] Figure 2 Schematic diagram of the BO-EVO solution.

[0049] a is the conceptual diagram of BO-EVO. Iteratively, candidate offspring sequences are generated by random mutation of the parent sequence, and BO selects a sequence from the candidate sequences.

[0050] BO-EVO consists of modules (b), including the biological sequence and fitness data acquisition module, the machine learning model module, and the biological sequence recommendation module. The biological sequence recommendation module interacts with the other two modules to propose batches of sequences. First, sequences and their corresponding fitness data are obtained from a known database or experimental results. Machine learning is then applied to the data obtained from the data acquisition module to develop a prediction model. Sequences are predicted based on this prediction model and guided to generate new query sequences in the recommendation module. New sequences and fitness data are synthesized and tested on the sequences in the recommendation module and then used to learn the prediction model.

[0051] Figure 3 This is the result graph of the fitness exploration algorithm.

[0052] a is the success rate achieved at the end of each round.

[0053] b is the maximum (top) and average (bottom) fitness of all sequences obtained. DETAILED DESCRIPTION

[0054] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below, but it should not be understood as limiting the scope of implementation of the present invention.

[0055] Combine Figure 1 and Figure 2 For illustration, some specific embodiments of the present invention provide a method for biological sequence engineering guided by machine learning, comprising the following steps:

[0056] S1) Data acquisition: obtaining biological sequences and their corresponding fitness data;

[0057] The sequence and its corresponding fitness data can be obtained from a known fitness landscape, from a known sequence and its corresponding fitness data set, or from an experimentally measured sequence and its corresponding fitness data. The sequence and its corresponding fitness data obtained from experimental data can be the query sequence obtained according to the method of the present invention and the fitness result obtained after testing the query sequence, or can be other known sequences and corresponding fitness data test results.

[0058] The biological sequence is a protein sequence, a polypeptide sequence, a ribonucleic acid sequence, or a deoxyribonucleic acid sequence.

[0059] In a specific embodiment, the sequence and its corresponding fitness data are derived from a known fitness landscape, which can be derived from an empirical fitness landscape or a statistical fitness landscape. For example, the sequence and its corresponding fitness data are derived from an empirical 4-site combination of (20 4=160,000) GB1 fitness landscape, empirical fitness landscape PhoQ and statistical NK landscape, and combined with wet experiments to transform the RhlA enzyme. By using different fitness landscapes, it is proved that the method of the present invention has universality.

[0060] S2) Model training: Machine learning is performed on the obtained biological sequences and their corresponding fitness data to obtain a fitness prediction model and model uncertainty for the predicted sequence;

[0061] The prediction model can be a Gaussian process model, a Bayesian neural network prediction model, an ensemble model, an evidence-based deep learning model, or other models with prediction uncertainty.

[0062] In a specific embodiment, a Gaussian process model (GPR) is used as a prediction model. The Gaussian process model is completely described by a mean function and a covariance function. The mean corresponds to the fitness and the covariance is used as an estimate of uncertainty. The kernel used is RBF, The RBF kernel parameters k and γ are predicted using maximum likelihood, using GPyTorch with CUDA acceleration as a more efficient Gaussian process implementation than scikit-learn, and employing gradient descent and Adam optimization for maximum marginal likelihood estimation.

[0063] In the process of predictive model learning, the sequence needs to be digitized, for example, the amino acid sequence is converted into a digital expression using the ESM protein language model.

[0064] S3) Obtaining a first seed: Obtaining a first seed sequence from the biological sequence obtained in step S1) and its corresponding fitness data based on fitness;

[0065] The obtained seed sequence can be a natural sequence or a mutant sequence.

[0066] S4) generating a mutation subspace: performing random mutations on the sites to be mutated in the seed sequence obtained in step S3) or S6), and forming a mutation subspace with a mutation rate that is the inverse of the number of sites to be mutated;

[0067] For example, if the sequence to be screened is 50 amino acids long, with four amino acid sites to be mutated, the point mutation rate is 0.25, which means that for each site to be mutated, the mutated biological sequence will be a single-point mutation sequence. By limiting the mutation rate of random mutations, the locality of the subspace is guaranteed.

[0068] In some specific embodiments, the number of randomly mutated biological sequences generated in the mutation subspace is 1.0-3.0 times the number of sites to be mutated. The number of biological sequences in the mutation subspace can be adjusted as needed, and it is usually not necessary to include all biological sequences that meet the point mutation rate within the mutation subspace. Although the mutation subspace can complete the present invention regardless of its size, its size may affect the efficiency of the method of the present invention. The size of the mutation subspace can be selected based on the degree of expansion of the target sequence, or it can be screened and confirmed based on the possible mutation sites given in the prior art. For example, if the extended sequence range is such that 4 amino acid mutant proteins may be produced (i.e., the sites to be mutated are 4 amino acids), 4-12 mutant sequences can be randomly mutated and included in the mutation subspace for Bayesian optimization. The number of sites to be mutated is the number of the smallest units of sites to be mutated in the biological sequence. For example, the smallest unit in a polypeptide or protein sequence is an amino acid, while the smallest unit in a DNA sequence is a nucleotide.

[0069] S5) Mutation Subspace Bayesian Optimization: Bayesian optimization is used on the mutation subspace to select a single candidate sequence for experimental query, where the sampling function is UCB and the proxy model is the fitness prediction model obtained in S2);

[0070] In some specific embodiments, the value of β in UCB is 0.05-0.25, for example, 0.1, 0.2;

[0071] S6) New seed generation: First, the uncertainty of the candidate sequence obtained in S5) is predicted using the prediction model obtained in S2). If the uncertainty is no greater than twice the uncertainty of the first seed sequence in this round, the candidate sequence is used as the next seed sequence in this round and is used to form the next mutation subspace. If the uncertainty is greater than twice the uncertainty of the first seed sequence in this round, the next seed sequence in this round is obtained according to step S3) and is used to form the next mutation subspace.

[0072] S7) repeating the cycle from S4) to S6) until a preset throughput cycle is met, thereby obtaining a batch of candidate sequences to be experimentally queried;

[0073] The preset throughput is limited to between 200 and 500 cycles, that is, 200 to 500 candidate sequences to be experimentally queried are obtained.

[0074] Obtaining a candidate sequence to be queried is a cycle. After completing a cycle, the process further includes evaluating and determining the first seed for the next cycle using step S6).

[0075] S8) Experimenting on the batch of candidate sequences to be queried in the experiment obtained in S7) to obtain candidate sequences and detecting their corresponding fitness values.

[0076] The method for conducting the experiment in step S8) is to first obtain the candidate sequences to be queried by the experiment using a chemical or biological method, and then actually test the fitness value of each candidate sequence.

[0077] In step S8), the method for conducting experiments on a batch of candidate sequences to be queried is to synthesize biological sequences using an engineering method and test fitness values.

[0078] S9) adding the candidate sequence and its corresponding fitness value to the biological sequence and its corresponding fitness data in S1) and using them for the next round of model training to obtain the fitness prediction model and model uncertainty of the biological sequence in the next round;

[0079] S10) Using the latest seed sequence obtained in S6) as the first seed sequence for the next round, and repeating S4) to S9) until the preset rounds are met or the expected mutants are screened out.

[0080] Obtaining batches of candidate sequence experimental results constitutes a round. The resulting sequences and their corresponding fitness data are then added to the biological sequences and their corresponding fitness data from step S1), marking the beginning of the next round. In the next round, the supplemented biological sequences and their corresponding fitness data are used to train the prediction model for the next round. Starting from the second round, the seed sequence obtained in step S6 of the first round is used as the first seed sequence for this round, and this cycle repeats.

[0081] The preset rounds are at least 2 rounds.

[0082] Combine Figure 2 b. Description Some specific embodiments of the present invention further provide a machine learning-guided biological sequence engineering device, which is a device capable of implementing the above steps S1)-S10).

[0083] The biological sequence engineering modification device includes the following modules: a biological sequence and its fitness data acquisition module, a machine learning model module and a biological sequence recommendation module;

[0084] The biological sequence and fitness data acquisition module is used to store and call sequences and their fitness data, convert the called sequences into digital codes; and return the corresponding fitness values based on the batch of candidate sequences generated by the biological sequence recommendation module;

[0085] Specifically, the biological sequence and fitness data acquisition module includes a recording unit for recording information, an encoding unit for converting the called sequence into a digital code, and a batch experiment unit for returning the corresponding fitness value based on the batch candidate sequences generated in the biological sequence recommendation module.

[0086] The machine learning model module is used to call the biological sequence from the biological sequence and fitness data acquisition module, perform machine learning and training based on the called data, and form a prediction model and model uncertainty between the biological sequence and its fitness;

[0087] The biological sequence recommendation module is used to obtain a batch of candidate sequences to be experimentally queried under the guidance of the machine learning model module. The biological sequence recommendation module can call data in the biological sequence and fitness data acquisition module, and based on the first seed sequence obtained by fitness sampling, perform random mutations on the first seed sequence, and form a mutation subspace with the random mutation sequence. Bayesian optimization is used in the mutation subspace to select a single candidate sequence, wherein the sampling function is UCB and the proxy model is the prediction model formed by the machine learning model module;

[0088] Confirm the uncertainty of a single candidate sequence. If the uncertainty is no more than 2 times the uncertainty of the first seed sequence in this round, the candidate sequence will be used as the next seed sequence in this round and used to form the mutation subspace of this round. If the uncertainty is more than 2 times the uncertainty of the first seed sequence in this round, the next seed sequence in this round will be obtained from the called sequence based on fitness sampling. The next cycle will be started until the preset throughput and batch of candidate sequences are reached.

[0089] The biological sequence recommendation module also includes a unit for starting the next round, which transmits the generated batch of candidate sequences to the biological sequence and its fitness data acquisition module, and guides the batch of candidate sequences generated by the biological sequence and its fitness data acquisition module to return corresponding fitness values; the batch of candidate sequences generated by the biological sequence and its fitness data acquisition module and their corresponding fitness values are incorporated into the existing biological sequence and its fitness data; and is used to guide the machine learning model module to further generate a new prediction model, and the new prediction model in the machine learning model module is used to guide the next round of recommendations of the biological sequence recommendation module.

[0090] Example 1 Developing BO-EVO Algorithm with GB1 Adaptability Terrain

[0091] S1) Data acquisition: The biological sequences and their corresponding fitness data were obtained from the GB1 fitness landscape; the GB1 landscape is a combined empirical fitness landscape of four sites (V39, D40, G41, and V54) in the G protein B1 domain, and its wild-type (WT) sequence is from the Protein Structure Database (PDB ID: 2GI9). The landscape consists of 20 4= 160,000 sequences. A total of 149,361 sequences were measured experimentally, and the remaining sequences were estimated using the measured data. Sequence fitness is determined by stability (i.e., the fraction of folded protein) and function (i.e., binding affinity to IgG Fc). Fitness is normalized by subtracting the global minimum from the global maximum fitness, resulting in a value between 0 and 1, with a WT fitness of approximately 0.1.

[0092] S2) Model training: Machine learning is performed on the obtained biological sequences and their corresponding fitness data using Gaussian process regression (GPR) to obtain a fitness prediction model and model uncertainty for the biological sequences;

[0093] S3) Obtaining a first seed: Obtaining a first seed sequence from the biological sequence obtained in step S1) and its corresponding fitness data through fitness-based sampling;

[0094] S4) Mutation subspace generation: Random mutations are performed on the seed sequence obtained in step S3) or S6), and the resulting set of biological sequences forms a mutation subspace, where the single-point mutation rate of random mutations is the inverse of the number of sites to be mutated, which is 4; the number of biological sequences in the subspace is 1-3 times the number of sites to be mutated, which is 4, and 8 sequences are selected within this range for the experiment;

[0095] S5) Mutation Subspace Bayesian Optimization: Bayesian optimization is used on the mutation subspace to select a single candidate sequence for experimental query, where the sampling function is UCB and the proxy model is the fitness prediction model obtained in S2); β = 0.2.

[0096] S6) New seed generation: First, the uncertainty of the candidate sequence obtained in S5) is predicted using the prediction model obtained in S2). If the uncertainty is no greater than twice the uncertainty of the first seed sequence in this round, the candidate sequence is used as the first seed sequence in the next round and is used to form the mutation subspace in the next round. If the uncertainty is greater than twice the uncertainty of the first seed sequence in this round, the first seed sequence in the next round is obtained according to step S3).

[0097] S7) Repeat S4) to S6) until the preset throughput is met, and 384 candidate sequences to be experimentally queried are obtained;

[0098] S8) querying the fitness values of the batch of candidate sequences to be queried in the experiment obtained in S7).

[0099] S9) adding the candidate sequence and its corresponding fitness value to the biological sequence and its corresponding fitness data in S1) and using them for the next round of model training to obtain the fitness prediction model and its uncertainty for the next round of biological sequence;

[0100] S10) Using the latest seed sequence obtained in S6) as the first seed sequence for the next round, and repeating S4) to S9) until the preset rounds are met or the expected mutants are screened out.

[0101] The BO-EVO algorithm suitable for biological sequence modification was developed through the GB1 fitness landscape. This algorithm can greatly improve the computing speed and reduce the computing requirements.

[0102] Example 2 Validation of the BO-EVO Algorithm Using the Empirical Fitness Landscape PhoQ and the NK Model as a Protein Fitness Landscape

[0103] The BO-EVO algorithm obtained in Example 1 was further verified using different fitness terrains.

[0104] Similar to the GB1 topography, the PhoQ topography is also an empirical four-site topography. PhoQ uses enrichment rate as fitness. Using a fitness normalization strategy, the WT fitness is approximately 0.02.

[0105] NK terrain is transformed from the original NK terrain.

[0106] The BO-EVO algorithm of the present invention was further validated using PhoQ terrain and NK terrain using a method similar to Example 1. The experimental results show that the BO-EVO algorithm of the present invention is applicable to various types of terrains.

[0107] Example 3: Confirmation of the BO-EVO algorithm using RhlA enzyme modification

[0108] RhlA is a key enzyme in the synthesis of the lipid portion of rhamnolipids (RLs), an important biosurfactant. The enzyme specificity of RhlA determines the chemical structure of the lipid portion, which in turn influences the physicochemical and biological activities of the corresponding RL molecule. However, it is difficult to modify the enzyme specificity of RhlA through (semi-)rational design or directed evolution. This example applies BO-EVO to modify the enzyme specificity of RhlA through iterative feedback between a machine learning model and robotic experiments.

[0109] To evaluate the specificity of the RhlA enzyme, matrix-assisted laser desorption / ionization time-of-flight (MALDI-ToF) mass spectrometry (MS) was used to quantify the four RL products, Rha-C18, Rha-C20, Rha-C22, and Rha-C24, in liquid cultures of recombinant Escherichia coli as the production host.

[0110] In microplate culture, Rha-C20 is the primary product of E. coli cells harboring wild-type RhlA (UniProt ID: Q51559). This experiment aimed to increase the yield and proportion (fitness) of the product Rha-C18 by modifying RhlA. The fitness was normalized to 1, using the yield and proportion of Rha-C18 corresponding to wild-type RhlA as a reference.

[0111] To apply BO-EVO, R74, A101, L148, and S173 were selected as four target residues for combinatorial mutations, as existing knowledge suggests that many mutations on these residues significantly enhance the production of Rha-C18. For the BO-EVO iterations, we observed that the cumulative maximum fitness increased with each round, reaching 7.35 in round 4, while the number of mutants that required robotic experimental quantification was less than 1% of the entire design space. Experimental results Figure 3 The results show that the method of the present invention can be combined with robotic experiments to effectively improve the specificity of RhlA enzyme.

[0112] Example 4 Comparative Experiment of Fitness Terrain Exploration Algorithms

[0113] AdaLead and BO are used as the benchmarks for BO-EVO to examine the necessity of combining these two exploration strategies. The performance of random mutation (Random) is also evaluated as the benchmark, which is expected to be the worst among the four algorithms ( Figure 2 ). The success rate of the algorithm in searching for the global optimal sequence after five rounds ( Figure 2 a) and the maximum and average fitness of all sequences recommended by the algorithm ( Figure 2 b), AdaLead outperforms random mutation, but this pure evolutionary algorithm is significantly inferior to BO-EVO or BO, both of which use the UCB sampling function to balance exploration and exploitation. These results show the importance of considering both sequence fitness and model uncertainty when exploring rugged adaptive environments. On the other hand, by performing a brute force search of the entire design space (160,000 sequences) in each iteration, BO alone achieves better performance than BO-EVO, but the latter only evaluates 3072 sequences per round (only 1.92% of the entire design space) (see Table 1 for a detailed comparison of the settings of the four exploration algorithms. Although BO-EVO does not perform as well as pure BO, the computational time of BO-EVO is almost constant when exploring the combinatorial mutation landscape, while the computational time of pure BO grows exponentially with the number of target residues, making it not scalable. The algorithm of the present invention is scalable and can be used for engineering tasks of longer sequences.

[0114] Comprehensive experimental conditions and results show that this invention, by combining BO and EVO, significantly reduces the number of evaluation sequences in the proxy model, achieving high speed and low computational resources. Furthermore, obtaining the top sequences through UCB better balances exploration and exploitation.

[0115] Table 1

[0116]

Claims

1. A method for biological sequence engineering guided by machine learning, characterized in that: It includes the following steps: S1) Data acquisition: obtaining biological sequences and their corresponding fitness data; S2) Model training: Machine learning is performed on the obtained biological sequences and their corresponding fitness data to obtain a fitness prediction model and model uncertainty for the biological sequences; S3) Obtaining a first seed: Obtaining a first seed sequence from the biological sequence obtained in step S1) and its corresponding fitness data through fitness-based sampling; S4) generating a mutation subspace: performing random mutations on the sites to be mutated in the seed sequence obtained in step S3) or S6), and forming a mutation subspace with a set of mutated biological sequences, wherein the single-point mutation rate of the random mutation is the inverse of the site to be mutated; S5) Mutation Subspace Bayesian Optimization: Bayesian optimization is used on the mutation subspace to select a single candidate sequence for experimental query, where the sampling function is UCB and the proxy model is the fitness prediction model obtained in S2); S6) New seed generation: First, the uncertainty of the candidate sequence obtained in S5) is predicted using the prediction model obtained in S2). If the uncertainty is no greater than twice the uncertainty of the first seed sequence in this round, the candidate sequence is used as the next seed sequence in this round and is used to form the next mutation subspace. If the uncertainty is greater than twice the uncertainty of the first seed sequence in this round, the next seed sequence in this round is obtained according to step S3) and is used to form the next mutation subspace. S7) repeating the cycle from S4) to S6) until a preset throughput cycle is met, thereby obtaining a batch of candidate sequences to be experimentally queried; S8) performing experiments on the batch of candidate sequences to be queried in the experiment obtained in S7) to obtain candidate sequences and detecting their corresponding fitness values; S9) adding the candidate sequence and its corresponding fitness value to the biological sequence and its corresponding fitness data in S1) and using them for the next round of model training to obtain the fitness prediction model and model uncertainty of the biological sequence in the next round; S10) Using the latest seed sequence obtained in S6) as the first seed sequence for the next round, and repeating S4) to S9) until the preset rounds are met or the expected mutants are screened out.

2. The biological sequence engineering method according to claim 1, characterized in that: In step S1), the biological sequence and its corresponding fitness data are obtained from a known fitness landscape, from a known sequence and a corresponding fitness dataset, or from an experimentally measured biological sequence and corresponding fitness data.

3. The biological sequence engineering method according to claim 2, characterized in that: In step S1), the sequences obtained from the experimentally measured biological sequences and the corresponding fitness data and their corresponding fitness data are candidate sequences and the fitness results obtained after testing them, or are other known biological sequences and the corresponding fitness data test results.

4. The biological sequence engineering method according to claim 1, characterized in that: The biological sequence is a protein sequence, a polypeptide sequence, a ribonucleic acid sequence, or a deoxyribonucleic acid sequence.

5. The biological sequence engineering method according to claim 1, characterized in that: In step S2), the prediction model is selected from a Gaussian process model, a Bayesian neural network prediction model, an integrated model, an evidence deep learning model or other models of prediction uncertainty.

6. The biological sequence engineering method according to claim 1, characterized in that: In step S2), the sequence is first expressed in digital form before the prediction model learning process.

7. The biological sequence engineering method according to claim 1, characterized in that: In step S4), the number of mutated biological sequences in the mutation subspace is 1.0-3.0 times the number of sites to be mutated.

8. The biological sequence engineering method according to claim 1, characterized in that: In step S5), the value of β in UCB is 0.05-0.

25.

9. The biological sequence engineering method according to claim 1, characterized in that: In step S7), the preset flux is limited to between 200-500.

10. A machine learning-guided biological sequence engineering device, characterized in that: The biological sequence engineering device is a device that can implement steps S1) to S10) of the biological sequence engineering method according to any one of claims 1 to 9.

11. The biological sequence engineering device according to claim 10, characterized in that: The biological sequence engineering modification device includes the following modules: a biological sequence and fitness data acquisition module, a machine learning model module, and a biological sequence recommendation module; The biological sequence and fitness data acquisition module is used to store and call sequences and their fitness data, convert the called sequences into digital codes; and return the corresponding fitness values based on the batch of candidate sequences generated by the biological sequence recommendation module; The machine learning model module is used to call the biological sequence from the biological sequence and fitness data acquisition module, perform machine learning and training based on the called data, and form a prediction model and model uncertainty between the biological sequence and its fitness; The biological sequence recommendation module is used to obtain a batch of candidate sequences to be experimentally queried under the guidance of the machine learning model module. The biological sequence recommendation module can call data in the biological sequence and fitness data acquisition module, and based on the first seed sequence obtained by fitness sampling, perform random mutations on the first seed sequence, and form a mutation subspace with the random mutation sequence. Bayesian optimization is used in the mutation subspace to select a single candidate sequence, wherein the sampling function is UCB and the proxy model is the prediction model formed by the machine learning model module; Confirm the uncertainty of a single candidate sequence. If the uncertainty is no more than 2 times the uncertainty of the first seed sequence in this round, the candidate sequence will be used as the first seed sequence in the next round and used to form the mutation subspace in the next round. If the uncertainty is more than 2 times the uncertainty of the first seed sequence in this round, the first seed sequence in the next round will be obtained from the called sequence based on fitness sampling. The next round will be started until the preset number of rounds and batches of candidate sequences are reached.

12. The biological sequence engineering device according to claim 11, characterized in that: The biological sequence recommendation module also includes a unit for starting the next round, which transmits the generated batch of candidate sequences to the biological sequence and its fitness data acquisition module, and guides the batch of candidate sequences generated by the biological sequence and its fitness data acquisition module to return corresponding fitness values; the batch of candidate sequences generated by the biological sequence and its fitness data acquisition module and their corresponding fitness values are incorporated into the existing biological sequence and its fitness data; and is used to guide the machine learning model module to further generate a new prediction model, and the new prediction model in the machine learning model module is used to guide the next round of recommendations of the biological sequence recommendation module.

13. The biological sequence engineering device according to any one of claims 10 to 12, characterized in that: The biological sequence and fitness data acquisition module includes a recording unit for recording information, an encoding unit for converting the called sequence into a digital code, and a batch experiment unit for returning the corresponding fitness value based on the batch candidate sequences generated in the biological sequence recommendation module.

Images