CN-122024817-A - RNA sequence design method, device, equipment and medium based on reinforcement learning and latent space diffusion model

CN122024817ACN 122024817 ACN122024817 ACN 122024817ACN-122024817-A

Abstract

The application discloses an RNA sequence design method, device, equipment and medium based on reinforcement learning and a latent space diffusion model, which relate to the technical field of artificial intelligence, and are characterized in that a mixed attention mechanism in the latent space diffusion model is utilized to process an initial RNA sequence and an RNA tertiary structure, rotation position coding, activation function optimization and mask processing are carried out on processed data to obtain a first RNA sequence, a reinforcement learning optimization engine is utilized to calculate a reward value of a second RNA sequence based on a preset reward mechanism, so as to obtain each reward value, fusion and structural analysis are carried out on each reward value, a target reward value is obtained, a third RNA sequence corresponding to the target reward value is determined from the second RNA sequence, multi-target optimization is carried out on the third RNA sequence by utilizing a Pareto front edge method, so that target RNA is constructed, the model training time and the resource consumption are reduced, and RNA sequence design based on reinforcement learning and the latent space diffusion model is realized.

Inventors

GUO XIN
HAN LIMEI
CHENG YUAN
SI QI
LIU XUYANG
JIANG CHEN
QI YUAN

Assignees

复旦大学
上海科学智能研究院

Dates

Publication Date: 20260512
Application Date: 20260122

Claims (10)

1. An RNA sequence design method based on reinforcement learning and a latent space diffusion model, which is characterized by being applied to a preset calculation frame, wherein the preset calculation frame comprises a latent space diffusion model and a reinforcement learning optimization engine, and the method comprises the following steps: Acquiring an initial RNA sequence and an RNA tertiary structure, processing the initial RNA sequence and the RNA tertiary structure by utilizing a mixed attention mechanism in a latent space diffusion model to obtain processed data, and performing rotary position coding, activation function optimization and mask processing on the processed data to obtain a first RNA sequence; Screening a second RNA sequence from the first RNA sequence, calculating the reward value of the second RNA sequence by using a reinforcement learning optimization engine based on a preset reward mechanism to obtain each reward value in reinforcement learning, fusing and structurally analyzing each reward value to obtain a target reward value, and determining a third RNA sequence corresponding to the target reward value from the second RNA sequence; And performing multi-objective optimization on the third RNA sequence by utilizing a pareto front edge method so as to construct target RNA.
2. The method of claim 1, wherein the processing the initial RNA sequence and the RNA tertiary structure using a mixed attention mechanism in the latent spatial diffusion model comprises: combining the sliding window attention with the global attention to build a mixed attention mechanism; the mixed-attention mechanism is introduced into an encoder layer in a latent spatial diffusion model so as to process an initial RNA sequence and an RNA tertiary structure by using the mixed-attention mechanism, wherein the RNA tertiary structure comprises an RNA backbone structure and corresponding spatial structure coordinates.
3. The method for designing an RNA sequence based on reinforcement learning and latent space diffusion model according to claim 1, wherein the performing rotational position coding, activation function optimization, masking processing on the processed data to obtain a first RNA sequence comprises: encoding the processed data using a rotational position encoding technique; Optimizing the encoded data by using a gating Gaussian error linear unit activation function; And masking the optimized data to obtain a first RNA sequence.
4. The method for designing an RNA sequence based on reinforcement learning and latent space diffusion model according to claim 3, wherein masking the optimized data to obtain a first RNA sequence comprises: Randomly selecting a mask position from the optimized data and determining a first nucleotide at the mask position; judging whether the first nucleotide has a base pairing relationship with a second nucleotide in the initial RNA sequence; If the base pairing relationship exists, storing the first nucleotide and the second nucleotide in a masking range, and masking the nucleotides in the masking range to obtain a first RNA sequence.
5. The method of claim 1, wherein the calculating the reward value for the second RNA sequence using the reinforcement learning optimization engine and based on a predetermined reward mechanism comprises: Calculating a first reward value of a second RNA sequence meeting a preset high-frequency condition by using a reinforcement learning optimization engine and based on a preset first reward mechanism; And performing second prize value calculation on the second RNA sequence meeting the preset low-frequency condition by using the reinforcement learning optimization engine based on a preset second prize mechanism.
6. The method for designing an RNA sequence based on reinforcement learning and latent space diffusion model according to claim 1, wherein the fusing and structural analysis of the prize values to obtain the target prize value comprises: Determining a target weight according to service requirements, and fusing all the reward values by using the target weight to obtain fused reward values; And carrying out structural analysis on the fused reward value by using a preset RNA structural analysis tool to obtain a target reward value, wherein the preset RNA structural analysis tool comprises VIENNARNA.
7. The method of RNA sequence design based on reinforcement learning and latent spatial diffusion model according to any one of claims 1 to 6, wherein the multi-objective optimization of the third RNA sequence using pareto front edge method comprises: Determining weight to be adjusted according to product requirements; And prioritizing the third RNA sequence by utilizing a pareto front edge method, and performing multi-objective optimization on the third RNA sequence based on the prioritized priority and the weight to be adjusted.
8. An RNA sequence design device based on reinforcement learning and a latent space diffusion model, characterized in that it is applied to a preset calculation frame, the preset calculation frame comprising a latent space diffusion model and a reinforcement learning optimization engine, the device comprising: The data processing module is used for acquiring an initial RNA sequence and an RNA tertiary structure, processing the initial RNA sequence and the RNA tertiary structure by utilizing a mixed attention mechanism in a potential space diffusion model to obtain processed data, and performing rotary position coding, activation function optimization and mask processing on the processed data to obtain a first RNA sequence; The reward value calculation module is used for screening a second RNA sequence from the first RNA sequence, calculating the reward value of the second RNA sequence by using a reinforcement learning optimization engine based on a preset reward mechanism to obtain each reward value in reinforcement learning, fusing and structurally analyzing each reward value to obtain a target reward value, and determining a third RNA sequence corresponding to the target reward value from the second RNA sequence; and the multi-objective optimization module is used for carrying out multi-objective optimization on the third RNA sequence by utilizing the pareto front edge method so as to construct the target RNA.
9. An electronic device, comprising: A memory for storing a computer program; a processor for executing the computer program to implement the reinforcement learning and latent space diffusion model-based RNA sequence design method according to any one of claims 1 to 7.
10. A computer-readable storage medium for storing a computer program, wherein the computer program when executed by a processor implements the reinforcement learning and latent space diffusion model-based RNA sequence design method according to any one of claims 1 to 7.

Description

RNA sequence design method, device, equipment and medium based on reinforcement learning and latent space diffusion model Technical Field The invention relates to the technical field of artificial intelligence, in particular to an RNA sequence design method, device, equipment and medium based on reinforcement learning and a latent space diffusion model. Background RNA (ribonucleic acid) reverse folding is the basis for rationally designing functional RNA molecules, and the method is mainly divided into two types, namely a physical model-based model and a deep learning-based model, wherein the models are mainly applied to the field of image or text generation and are mainly applied to the field of image or text generation by combining reinforcement learning, such as RhoDesign, RDesign, the models are used for generating sequences through learning sequence-to-structure mapping relations, but are difficult to process complex data distribution and long-range dependency relations and are also not capable of directly optimizing non-microscopic structural indexes, the models are used for generating sequences through iterative denoising, the models are usually directly operated in sequence space or high-dimensional feature space, co-evolution information is not fully utilized, and the training targets are still sequence recovery (such as cross entropy loss), and the methods are mainly applied to the field of image or text generation by combining reinforcement learning, and are used for calculating rewards and updating models through sampling complete and multi-step denoising tracks, and are too high in calculation cost for molecular generation tasks. From the above, how to solve the problem of low efficiency when the diffusion model is optimized by the traditional reinforcement learning, reduce the training time and the resource consumption of the model, and realize the RNA sequence design based on the reinforcement learning and the latent space diffusion model is a problem to be solved in the field. Disclosure of Invention In view of the above, the present invention aims to provide a method, a device, an apparatus and a medium for designing an RNA sequence based on reinforcement learning and a latent space diffusion model, which can solve the problem of low efficiency when optimizing the diffusion model in the conventional reinforcement learning, reduce the training time and the resource consumption of the model, and realize the RNA sequence design based on reinforcement learning and the latent space diffusion model. The specific scheme is as follows: in a first aspect, the application discloses an RNA sequence design method based on reinforcement learning and a latent space diffusion model, which is applied to a preset calculation frame, wherein the preset calculation frame comprises the latent space diffusion model and a reinforcement learning optimization engine, and the method comprises the following steps: Acquiring an initial RNA sequence and an RNA tertiary structure, processing the initial RNA sequence and the RNA tertiary structure by utilizing a mixed attention mechanism in a latent space diffusion model to obtain processed data, and performing rotary position coding, activation function optimization and mask processing on the processed data to obtain a first RNA sequence; Screening a second RNA sequence from the first RNA sequence, calculating the reward value of the second RNA sequence by using a reinforcement learning optimization engine based on a preset reward mechanism to obtain each reward value in reinforcement learning, fusing and structurally analyzing each reward value to obtain a target reward value, and determining a third RNA sequence corresponding to the target reward value from the second RNA sequence; And performing multi-objective optimization on the third RNA sequence by utilizing a pareto front edge method so as to construct target RNA. Optionally, the processing of the initial RNA sequence and the RNA tertiary structure using a mixed-attention mechanism in a latent spatial diffusion model comprises: combining the sliding window attention with the global attention to build a mixed attention mechanism; the mixed-attention mechanism is introduced into an encoder layer in a latent spatial diffusion model so as to process an initial RNA sequence and an RNA tertiary structure by using the mixed-attention mechanism, wherein the RNA tertiary structure comprises an RNA backbone structure and corresponding spatial structure coordinates. Optionally, the performing rotation position coding, activation function optimization, and masking processing on the processed data to obtain a first RNA sequence includes: encoding the processed data using a rotational position encoding technique; Optimizing the encoded data by using a gating Gaussian error linear unit activation function; And masking the optimized data to obtain a first RNA sequence. Optionally, the masking processing is performed on the opti