Method for mRNA sequence optimization based on generative models

Abstract

A kind of mRNA sequence optimization method based on generation model, in offline stage, the generated model constructed is trained using the database of random generation or natural mRNA;In the online stage, the mRNA sequence to be optimized is input into the trained generation model, and the corresponding weight is generated for the codon sequence probability distribution and the codon probability distribution based on the fitness of the codon at each position of the mRNA sequence to be optimized by greedy or probabilistic random sampling method, to obtain the optimized mRNA sequence.The present application optimizes mRNA coding region codon by generation model, and realizes the joint optimization of multiple indicators by means of the weighted sum of probability, significantly improves the optimization efficiency, and the probability weighting framework brings better scalability and flexibility to the algorithm, providing a more powerful optimization tool for mRNA design.

Description

Technical Field

[0001] This invention relates to a technology in the field of bioengineering, specifically a method for optimizing messenger RNA (mRNA) sequences based on a generative model. Background Technology

[0002] Existing messenger RNA (mRNA) vaccines still face some key limitations in terms of storage, distribution, and efficacy: because mRNA molecules are chemically unstable and easily affected by the external environment, they have a relatively short lifespan in the body, which leads to insufficient protein expression and thus affects the immunogenicity of the vaccine. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies, such as the inability to freely select the weights of CAI and MFE during the optimization process and the slow optimization speed. It proposes a generative model-based mRNA sequence optimization method, which optimizes the codons in the mRNA coding region through a generative model and achieves joint optimization of multiple indicators by using probabilistic weighted summation. This significantly improves the optimization efficiency, while the probabilistic weighted framework brings better scalability and flexibility to the algorithm, providing a more powerful optimization tool for mRNA design.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a generative model-based method for optimizing mRNA sequences. In the offline stage, a generative model is trained using a database of randomly generated or natural mRNAs. In the online stage, the mRNA sequence to be optimized is input into the trained generative model, and corresponding weights are generated for the codon sequence probability distribution and the codon probability distribution based on the fitness of the codon at each position in the mRNA sequence to be optimized using a greedy or probabilistic random sampling method, thereby obtaining the optimized mRNA sequence.

[0006] The generative model includes an Actor unit and a Critic unit, wherein the Actor unit learns a parameterized behavioral strategy, that is, for different mRNA sequence inputs, it generates a corresponding codon sequence probability distribution; the Critic unit evaluates the generation effect of the generative unit to obtain a reference value of the MFE of the generated sequence.

[0007] The reference value is preferably divided by the length of the mRNA sequence to facilitate learning by the neural network.

[0008] The training process involves taking randomly generated mRNA sequences or mRNA sequences from a database as input, performing forward inference to obtain the probability distribution of codon sequences, sampling the distribution to obtain codon sequences, calculating their minimum free energy value, combining this value with their generation probability to calculate the loss function, and updating the module parameters by performing gradient backpropagation on the loss function to achieve neural network training.

[0009] The loss function L = αL actor +(1-α)L critic ,in: M * The actual MFE value of the optimized mRNA sequence generated by the Actor unit, where N is the length of the RNA sequence, b is the output of the Critic unit, and P is the value of the Critic unit. π (c i ) represents the i-th codon c generated by the Actor under the current behavior policy π after optimization. i The probability of L; actor and L critic The loss functions for the Actor and Critic units are respectively, and the two are weighted by a manually set weight α and then summed to obtain the total loss function L.

[0010] Technical effect

[0011] This invention utilizes a generative model to calculate the generation probability of all synonymous codons at each codon position in mRNA, integrating the optimization information corresponding to the MFE index into the probability. This method, which generates probabilities through a generative model and then samples to obtain the optimized sequence, is significantly faster than traditional optimization methods. Using probability as a medium, the optimization requirements of different indicators can be easily combined according to the designer's desired weights, achieving flexible joint optimization effects. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the system of the present invention;

[0013] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0014] like Figure 1As shown, this embodiment relates to a generative model-based mRNA sequence optimization system, which includes: a minimum free energy (MFE) module, a codon fitness index (CAI) module, and a joint optimization module. The MFE module generates a codon sequence probability distribution optimized based on the MFE index from the input mRNA sequence; the CAI module calculates the fitness of the codon at each position and obtains a codon probability distribution optimized for the CAI index; and the joint optimization module generates a codon sequence that simultaneously optimizes both indices based on the two probability distributions.

[0015] The MFE module includes an Actor unit and a Critic unit, wherein the Actor unit learns a parameterized behavioral strategy, that is, for different mRNA sequence inputs, it generates a corresponding codon sequence probability distribution; the Critic unit evaluates the generation effect of the Actor unit to obtain a reference value for the MFE of the generated sequence.

[0016] The learning objective of the Actor unit is to minimize the expected value of the MFE of the probability distribution of the generated codon sequence.

[0017] The CAI module includes a relative codon usage frequency calculation unit (RSCU) and a codon sequence probability distribution generation unit. The RSCU calculates the usage frequency of different synonymous codons at each position in the selected optimization environment based on the input RNA sequence to obtain the relative usage frequency. The codon sequence probability distribution generation unit obtains the generation probability of different synonymous codons at each position by normalizing the relative usage frequency.

[0018] The joint optimization module includes a probability-weighted summation unit and a sampling unit, wherein: the probability-weighted summation unit obtains the total codon sequence probability distribution by weighted summation based on the codon sequence probability distributions output by the MFE module and the CAI module. The sampling unit obtains the optimized codon sequence based on the total codon sequence probability distribution through random sampling or greedy sampling.

[0019] like Figure 2 As shown, this embodiment uses an mRNA sequence optimization method based on the above system. In the offline stage, the constructed generative model is trained using randomly generated or natural mRNA databases. In the online stage, the mRNA sequence to be optimized is input into the trained generative model. Corresponding weights are generated for the codon sequence probability distribution and the codon probability distribution based on the fitness of the codon at each position of the mRNA sequence to be optimized through greedy or probabilistic random sampling methods, so as to obtain the optimized mRNA sequence.

[0020] Through specific experiments, under the condition that the mRNA sequence length is fixed at 1749 (corresponding to 583 amino acids), the MFE module was trained for more than 6000 rounds using randomly generated mRNA sequences. The batch size for each training session was 1, and the Adam optimizer was used with a learning rate of 0.001. The trained model was then used to optimize 9 random mRNA sequences, and the optimized MFE results were compared with the traditional LinearDesign optimization method (Chinese Patent Document No. EP4030438(A1) Publication Date 20220720). The comparison results are shown in Table 1 below. All experiments in this specification were conducted on a server equipped with an NVIDIA GeForce RTX3090 graphics card. The server's CPU was an AMD Ryzen Threadripper 3970X 32-Core Processor. This invention uses the PyTorch framework for training and testing the neural network model.

[0021] Table 1

[0022]

[0023] As shown in the table, the MFE module of this invention achieves 72.97% of the optimization effect in 0.21% of the time required by the traditional method. This indicates that the generative method used in this invention not only significantly improves computational efficiency in MFE index optimization but also achieves good index optimization results. Furthermore, comparison with natural mRNA reveals that the MFE index optimized by this method is closer to the MFE value of natural mRNA of approximately the same length.

[0024] The effectiveness of this invention in practical applications was tested in the design task of the Victoria HA protein. Protein expression levels were measured using flow cytometry for the original mRNA sequence, the optimized sequence of this invention, and the sequence optimized using the traditional LinearDesign method. The mRNA sequence designed by this method showed better expression.

[0025] Compared with existing technologies, the total computation time of this method is only 0.21% of that of existing technologies, and the optimized MFE index is closer to the MFE value of natural mRNA of approximate length (approximately -600 kcal / mol to -800 kcal / mol).

[0026] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for optimizing mRNA sequences based on a generative model, characterized in that, The constructed generative model is trained using a database of randomly generated or natural mRNAs during the offline phase; In the online phase, the generative model trained with the mRNA sequence to be optimized is input. The generative model includes an Actor unit and a Critic unit. The Actor unit learns a parameterized behavior strategy, that is, for different mRNA sequence inputs, it generates corresponding codon sequence probability distributions. The Critic unit evaluates the generation effect of the generative unit to obtain a reference value of the MFE of the generated sequence. Then, the total codon sequence probability distribution is obtained by weighted summation based on the codon sequence probability distribution. The optimized codon sequence is obtained by random sampling or greedy sampling based on the total codon sequence probability distribution.

2. The mRNA sequence optimization method based on a generative model according to claim 1, characterized in that, The reference value is divided by the length of the mRNA sequence to facilitate learning by the neural network.

3. The mRNA sequence optimization method based on a generative model according to claim 1, characterized in that, The training process involves taking randomly generated mRNA sequences or mRNA sequences from a database as input, performing forward inference to obtain the probability distribution of codon sequences, sampling the distribution to obtain codon sequences, calculating their minimum free energy value, combining this value with their generation probability to calculate the loss function, and updating the module parameters by performing gradient backpropagation on the loss function to achieve neural network training.

4. The mRNA sequence optimization method based on a generative model according to claim 3, characterized in that, The loss function ,in: , , The actual MFE value of the optimized mRNA sequence generated by the Actor unit, where N is the length of the RNA sequence and b is the output of the Critic unit. Represents the Actor's current behavioral strategy Generate the i-th codon of the optimized sequence. The probability of; and These are the loss functions for the Actor unit and the Critic unit, respectively, with their weights set manually. The weighted sums are used to obtain the total loss function L.

5. A generative model-based mRNA sequence optimization system for implementing the method of any one of claims 1-4, characterized in that, include: The system comprises a minimum free energy (MFE) module, a codon fitness index (CAI) module, and a joint optimization module. Specifically, the MFE module generates a codon sequence probability distribution optimized based on the MFE index from the input mRNA sequence; the CAI module calculates the fitness of the codon at each position and obtains a codon probability distribution optimized for the CAI index; and the joint optimization module generates a codon sequence that simultaneously optimizes both indices based on the two probability distributions.

6. The mRNA sequence optimization system according to claim 5, characterized in that, The MFE module includes an Actor unit and a Critic unit, wherein: the Actor unit learns a parameterized behavioral strategy, that is, for different mRNA sequence inputs, it generates a corresponding codon sequence probability distribution; the Critic unit evaluates the generation effect of the Actor unit to obtain a reference value of the MFE of the generated sequence; the learning objective of the Actor unit is to minimize the expected value of the MFE of the generated codon sequence probability distribution.

7. The mRNA sequence optimization system according to claim 5, characterized in that, The CAI module includes a relative codon usage frequency calculation unit (RSCU) and a codon sequence probability distribution generation unit. The RSCU calculates the usage frequency of different synonymous codons at each position in the selected optimization environment based on the input RNA sequence to obtain the relative usage frequency. The codon sequence probability distribution generation unit obtains the generation probability of different synonymous codons at each position by normalizing the relative usage frequency.

8. The mRNA sequence optimization system according to claim 5, characterized in that, The joint optimization module includes a probability weighted summation unit and a sampling unit, wherein: the probability weighted summation unit obtains the total codon sequence probability distribution by weighted summation based on the codon sequence probability distributions output by the MFE module and the CAI module; the sampling unit obtains the optimized codon sequence by random sampling or greedy sampling based on the total codon sequence probability distribution.

Images