Drug design methods, devices, equipment, and media based on RNA basic model and thermodynamic guided discrete diffusion

Abstract

This application discloses a drug design method, apparatus, device, and medium based on an RNA basic model and thermodynamically guided discrete diffusion, relating to the field of drug design technology. The method includes: acquiring the original RNA drug sequence and inputting it into an RNA basic model to extract RNA feature information; performing inverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information to generate candidate chemically modified sequences; calculating the drug-likeness index of the candidate chemically modified sequences; determining the energy classifier gradient information at the current time step based on the drug-likeness index; using the energy classifier gradient information to correct the candidate chemically modified sequences to obtain the target chemically modified sequence at the current time step; using this target chemically modified sequence as the noise sequence for the next time step for iterative inverse denoising; and outputting the drug design result. This ensures that the corrected chemically modified sequence simultaneously satisfies stability, activity, and safety.

Description

Technical Field

[0001] This invention relates to the field of drug design technology, and in particular to drug design methods, apparatus, equipment and media based on RNA basic models and thermodynamically guided discrete diffusion. Background Technology

[0002] In the field of nucleic acid drugs, there are significant differences in the expression levels and in vivo half-life of unmodified RNA, with mRNA half-life differences reaching up to 40-fold in mammalian cells. For siRNA drugs, introducing artificial chemical modifications (such as 2'-F, 2'-O-Me, and thiomodification) is crucial to preventing nuclease degradation and increasing half-life. However, most existing sequence designs are limited to local optimization (similar to optimizing only the local codon fitness index, CAI), lacking a global perspective. Although recent deep generative models such as GEMORNA have emerged, capable of generating highly active UTRs and CDSs from scratch like "writing poetry," these models have encountered bottlenecks when dealing with the complex discrete chemical modification mapping of siRNA. Traditional Attention RNNs (GRUs) suffer catastrophic forgetting after long-term training (such as Epoch 960), with the test set recovery rate (RR) plummeting to 0.0973; even Transformer architectures (such as T5-small) only achieve a BLEU score of 0.6595 for extremely stringent antisense strand modifications. Furthermore, existing standard dynamic programming algorithms lack accurate free energy thermodynamic models after introducing chemical modifications (such as fluorination or m1Ψ), which may result in the generated modified sequences being thermodynamically unstable or containing undesirable motifs that can lead to immune off-target effects.

[0003] Overcoming the shortcomings of long short-term memory networks in extracting long-range dependencies due to the lack of attention mechanisms, and generating globally optimal drug design results in a discrete chemical modification space, is a technical problem that needs to be solved in this field. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a drug design method, apparatus, device, and medium based on an RNA-based model and thermodynamically guided discrete diffusion, which can overcome the deficiency of long short-term memory networks in lacking attention mechanisms and extracting long-range dependencies, and achieve the generation of globally optimal drug design results within a discrete chemical modification space. The specific scheme is as follows: In the first aspect, this application discloses a drug design method based on an RNA fundamental model and thermodynamically guided discrete diffusion, including: The original RNA drug sequence was obtained and input into the RNA basic model to extract RNA feature information containing sequence long-range dependence and structural motif information; A pre-trained discrete diffusion model is used to reverse-denoise the noise sequence at the current time step based on the RNA feature information to generate candidate chemical modification sequences. The druggability index of the candidate chemical modification sequences is calculated, and then the energy classifier gradient information at the current time step is determined based on the druggability index. The candidate chemical modification sequences are then corrected using the energy classifier gradient information to obtain the target chemical modification sequence at the current time step. This target chemical modification sequence is then used as the noise sequence for the next time step and iterative reverse-denoising is performed until the time step iteration termination condition is met. Finally, the corresponding target chemical modification sequence is output as the drug design result of the original RNA drug sequence.

[0005] Optionally, the step of inputting the original RNA drug sequence into an RNA basic model to extract RNA feature information containing long-range sequence dependence and structural motif information includes: The original RNA drug sequence is encoded nucleotide-by-nucleotide by RNA basic model to obtain an initial embedding vector, and the long-range dependencies between nucleotides at different positions in the initial embedding vector are calculated by self-attention mechanism. The contextual semantic information of nucleotides at each position is extracted using the aforementioned RNA basic model; The RNA basic model outputs the structural tendency code corresponding to each position as structural motif information. Based on the long-range dependency, the contextual semantic information, and the structural motif information, the RNA feature information of the original RNA drug sequence is determined.

[0006] Optionally, before performing reverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information, the method further includes: Obtain a training dataset, which contains multiple pairs of paired data, each pair of paired data including an unmodified RNA sequence and its corresponding chemically modified sequence; The training dataset is filtered using a strategy of retaining minimal residual values ​​to select target paired data that meet preset modification conditions to construct a training subset; A forward Markov transition matrix is ​​defined in the discrete chemical modification label space of the initial discrete diffusion model to degenerate the chemical modification sequence into a noise sequence with uniform random distribution step by step. The RNA base model is used to extract features from the unmodified RNA sequences in the training subset to obtain the RNA feature information of the unmodified RNA sequences; The initial discrete diffusion model is trained using the RNA feature information of the unmodified RNA sequence and the chemical modification sequence corresponding to the unmodified RNA sequence. The model learns the inverse process of reconstructing the uniformly randomly distributed noise sequence into the chemical modification sequence corresponding to the unmodified RNA sequence step by step, so as to obtain the discrete diffusion model under the current model parameters as a pre-trained discrete diffusion model.

[0007] Optionally, the step of calculating the druggability index of the candidate chemically modified sequence and then determining the energy classifier gradient information at the current time step based on the druggability index includes: The minimum free energy, RISC assembly affinity, and off-target effect penalty term of the candidate chemical modification sequence are calculated, and the minimum free energy, the RISC assembly affinity, and the off-target effect penalty term are weighted to form a drug-likeness index. The gradient information of the candidate chemically modified sequence is calculated using the drug-likeness index to obtain the energy classifier gradient information at the current time step.

[0008] Optionally, the step of using the energy classifier gradient information to refine the candidate chemical modification sequence to obtain the target chemical modification sequence at the current time step includes: The perturbation vector is obtained based on the gradient information of the energy classifier and the preset guiding step size coefficient. The perturbation vector is superimposed on the candidate chemical modification sequence to obtain the target chemical modification sequence at the current time step.

[0009] Optionally, the original RNA drug sequence is a siRNA drug sequence; Accordingly, the step of performing reverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information includes: The positive and negative strands of the noise sequence at the current time step are reversed by using a dual-track decoding network based on a pre-trained discrete diffusion model and the RNA feature information.

[0010] Optionally, the process of performing reverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information further includes: Conditional and unconditional diffusion models trained based on RNA feature information are used to perform conditional and unconditional predictions on the noise sequence at the current time step, respectively, to obtain conditional and unconditional prediction output information. The conditional prediction output information and the unconditional prediction output information are linearly combined according to a preset combination weight to obtain denoised output information for correcting the candidate chemical modification sequence at the current time step.

[0011] Secondly, this application discloses a drug design device based on an RNA basic model and thermodynamically guided discrete diffusion, comprising: The information extraction module is used to obtain the original RNA drug sequence and input the original RNA drug sequence into the RNA basic model to extract RNA feature information containing sequence long-range dependence and structural motif information; The drug design module is used to perform reverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information to generate candidate chemical modification sequences. It then calculates the druggability index of the candidate chemical modification sequences, determines the energy classifier gradient information at the current time step based on the druggability index, and uses the energy classifier gradient information to correct the candidate chemical modification sequences to obtain the target chemical modification sequence for the current time step. This target chemical modification sequence is then used as the noise sequence for the next time step for iterative reverse denoising until the time step iteration termination condition is met. Finally, the corresponding target chemical modification sequence is output as the drug design result of the original RNA drug sequence.

[0012] Thirdly, this application discloses an electronic device, comprising: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the aforementioned disclosed drug design method based on an RNA fundamental model and thermodynamically guided discrete diffusion.

[0013] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed drug design method based on an RNA fundamental model and thermodynamically guided discrete diffusion.

[0014] As can be seen, this application discloses obtaining the original RNA drug sequence and inputting the original RNA drug sequence into an RNA basic model to extract RNA feature information containing long-range sequence dependence and structural motif information; using a pre-trained discrete diffusion model and based on the RNA feature information to perform reverse denoising on the noise sequence at the current time step to generate candidate chemical modification sequences, and calculating the druggability index of the candidate chemical modification sequences; then determining the energy classifier gradient information at the current time step based on the druggability index, and using the energy classifier gradient information to correct the candidate chemical modification sequences to obtain the target chemical modification sequence at the current time step, which is used as the noise sequence for the next time step for iterative reverse denoising until the time step iteration end condition is met, and outputting the corresponding target chemical modification sequence as the drug design result of the original RNA drug sequence. Therefore, it can be seen that training the encoder using the RNA-based model to extract deep features can capture the far-end interactions of sequences, that is, obtain global long-range dependencies, which can be used to coordinate the modification scheme globally. By adopting the discrete diffusion model, the Markov process that defines the discrete label space can directly generate effective discrete modification combinations, avoiding the invalid output caused by continuous space mapping. In each step of denoising, gradient information is determined and the output is perturbed, so that the chemically modified sequence output after correction can simultaneously satisfy stability, activity and safety, rather than just statistical fitting. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a flowchart of a drug design method based on RNA fundamental model and thermodynamically guided discrete diffusion disclosed in this application; Figure 2 This is a schematic diagram of a drug design device based on an RNA fundamental model and thermodynamically guided discrete diffusion disclosed in this application. Figure 3 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] In the field of nucleic acid drugs, there are significant differences in the expression levels and in vivo half-life of unmodified RNA, with mRNA half-life differences reaching up to 40-fold in mammalian cells. For siRNA drugs, introducing artificial chemical modifications (such as 2'-F, 2'-O-Me, and thiomodification) is crucial to preventing nuclease degradation and increasing half-life. However, most existing sequence designs are limited to local optimization (similar to optimizing only the local codon fitness index, CAI), lacking a global perspective. Although recent deep generative models such as GEMORNA have emerged, capable of generating highly active UTRs and CDSs from scratch like "writing poetry," these models have encountered bottlenecks when dealing with the complex discrete chemical modification mapping of siRNA. Traditional Attention RNNs (GRUs) suffer catastrophic forgetting after long-term training (such as Epoch 960), with the test set recovery rate (RR) plummeting to 0.0973; even Transformer architectures (such as T5-small) only achieve a BLEU score of 0.6595 for extremely stringent antisense strand modifications. Furthermore, existing standard dynamic programming algorithms lack accurate free energy thermodynamic models after introducing chemical modifications (such as fluorination or m1Ψ), which may result in the generated modified sequences being thermodynamically unstable or containing undesirable motifs that can lead to immune off-target effects.

[0019] To this end, the present invention provides a drug design scheme based on RNA basic model and thermodynamically guided discrete diffusion, which can overcome the deficiency of long short-term memory network in lack of attention mechanism extraction long-range dependence, and achieve the generation of globally optimal drug design results in discrete chemical modification space.

[0020] like Figure 1 As shown, this invention provides a drug design method based on an RNA fundamental model and thermodynamically guided discrete diffusion, comprising: Step S11: Obtain the original RNA drug sequence and input the original RNA drug sequence into the RNA basic model to extract RNA feature information containing sequence long-range dependence and structural motif information.

[0021] In this embodiment, the original RNA drug sequence is processed nucleotide-by-nucleotide embedding encoding using an RNA basic model to obtain an initial embedding vector. Long-range dependencies between nucleotides at different positions in the initial embedding vector are calculated using a self-attention mechanism. The RNA basic model extracts the contextual semantic information of each nucleotide position. The RNA basic model outputs the structural tendency encoding corresponding to each position as structural motif information. Based on the long-range dependencies, the contextual semantic information, and the structural motif information, the RNA feature information of the original RNA drug sequence is determined. It is understood that, firstly, the original RNA drug sequence is obtained, for example, a 20-nucleotide siRNA positive strand sequence: 5'-AUGCAUGCAUGCAUGCAAUG-3'. This sequence is input into a pre-trained RNA basic model, which can be the Nuwa RNA basic model or the RNA-FM model. These models, through self-supervised pre-training with large-scale RNA sequence data, possess the ability to perform multimodal unified representation of RNA sequences, structures, and functions. The model first performs embedding encoding on each nucleotide position of the input sequence. Specifically, a learnable embedding vector is assigned to each nucleotide type (A, U, C, G), and positional encoding is superimposed to preserve sequence order information, thus obtaining an initial embedding vector for each position. For example, for a sequence of length L, the model generates an L×d initial embedding matrix, where d is the embedding dimension (e.g., 768 dimensions). The initial embedding vector is then input into a multi-layer Transformer encoder of the RNA-based model. Each Transformer layer contains a multi-head self-attention sublayer. In self-attention computation, each nucleotide position can focus on nucleotides at all other positions in the sequence, capturing long-range dependencies between different positions by calculating attention weights. For example, a nucleotide at the 5' end of the sequence can establish a direct association with the nucleotide at the 3' end through a self-attention mechanism, thereby understanding potential constraints on distal base pairing or structural folding. Through layer-by-layer transformation of the multi-layer Transformer, the output hidden state of each position has aggregated contextual information from its surroundings and even the global sequence. The model aggregates the output hidden states of each layer (e.g., taking the output of the last layer or weighted summation of all layers) to obtain the context-dependent embedding of each nucleotide position. This embedding not only contains the base type of the position itself, but also encodes semantic information about its upstream and downstream sequences, such as local sequence preferences and possible secondary structural units. During pre-training, the RNA basic model implicitly learns the association patterns between sequences and structures through self-supervised tasks such as masked language modeling. During inference, the model can output the structural tendency encoding for each position, such as the probability distribution or continuous vector representation of the structural motifs that the position is likely to form, such as stem-loops, convex loops, inner loops, or free ends.This structural bias encoding directly reflects the natural structural preferences of the original sequence without chemical modifications, providing prior information for subsequent thermodynamic guidance. Finally, the model fuses the calculated long-range dependencies (aggregated features of the self-attention weight matrix), contextual semantic information (contextual embedding vectors at each position), and structural motif information (structural bias encoding) to form a set of high-dimensional latent vectors. These high-dimensional latent vectors represent RNA feature information, typically with a dimension of L×d' (e.g., 20×1024), containing rich representations of each position in the sequence and its interactions.

[0022] In this way, the basic RNA model transforms raw, discrete nucleotide sequences into continuous, biologically meaningful feature representations.

[0023] Step S12: By pre-training a discrete diffusion model and performing reverse denoising on the noise sequence at the current time step based on the RNA feature information, candidate chemical modification sequences are generated. The druggability index of the candidate chemical modification sequences is calculated. Then, the energy classifier gradient information at the current time step is determined based on the druggability index. The candidate chemical modification sequences are corrected using the energy classifier gradient information to obtain the target chemical modification sequence at the current time step. This target chemical modification sequence is then used as the noise sequence for the next time step and subjected to iterative reverse denoising until the time step iteration termination condition is met. The corresponding target chemical modification sequence is then output as the drug design result of the original RNA drug sequence.

[0024] In this embodiment, before performing reverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information, the method further includes: acquiring a training dataset containing multiple pairs of paired data, each pair including an unmodified RNA sequence and its corresponding chemical modification sequence; filtering the training dataset using a minimum residual value retention strategy to select target paired data that meet preset modification conditions to construct a training subset; defining a forward Markov transition matrix in the discrete chemical modification label space of the initial discrete diffusion model to degenerate the chemical modification sequence to a uniformly randomly distributed noise sequence step by step; extracting features from the unmodified RNA sequences in the training subset using the RNA base model to obtain the RNA feature information of the unmodified RNA sequences; training the initial discrete diffusion model using the RNA feature information of the unmodified RNA sequences and the chemical modification sequences corresponding to the unmodified RNA sequences to learn the inverse process of reverting the uniformly randomly distributed noise sequence to the chemical modification sequence corresponding to the unmodified RNA sequence step by step, so as to obtain the discrete diffusion model under the current model parameters as the pre-trained discrete diffusion model. Understandably, the model training process includes steps such as data preparation, data filtering, defining the forward diffusion process, feature extraction, and inverse denoising training. The specific execution flow of each step is as follows: Obtain a training dataset for supervised learning. This dataset contains multiple pairs of paired data, each consisting of an unmodified RNA sequence and its corresponding experimentally validated or expert-designed real chemical modification sequence. For example, the dataset may contain thousands to tens of thousands of siRNA sequences, with each unmodified sequence such as 5'-AUGCAUGCAU-3' and its corresponding chemical modification sequence such as 5'-AmUmGfCfAmUfGfCfAmU-3' (where m represents 2'-O-Me modification and f represents 2'-F modification). This data can be derived from patent literature, published papers, or publicly available experimental data.

[0025] Furthermore, to improve the model's convergence efficiency and generation quality, the training dataset is preprocessed. A strategy of retaining minimal residual values ​​is used for data purification. Specifically, the residual value is calculated for each chemical modification sequence. This value quantifies the regularity and complexity of the modification pattern. The lower the residual value, the stronger the regularity and the lower the complexity of the modification sequence (e.g., uniform distribution of modification sites and no abnormal combinations). Target pairs with residual values ​​below a preset threshold are selected to form the training subset. By eliminating redundant interference and noisy data, the modification syntax learned by the model is made purer, avoiding overfitting to low-quality samples.

[0026] Then, an initial discrete diffusion model (e.g., D3PM) is constructed, defining a discrete chemical modification label space containing all possible modification types and their corresponding unmodified bases. For example, the label space could include: A, U, C, G (unmodified), Af, Uf, Cf, Gf (2'-F modification), Am, Um, Cm, Gm (2'-O-Me modification), sus (thiomodification), etc. Each label represents a discrete state. In this discrete space, a forward Markov transition matrix Qt (t=1,...,T) is defined to describe the state transition probability from time step t-1 to t. For each time step, the transition matrix is ​​designed so that as these transition matrices are applied step-by-step, the true chemically modified sequence (e.g., Af, Um, Cf...) is gradually noisier, eventually degenerating into a uniformly randomly distributed noise sequence at time step T (each position is independently and uniformly sampled from the label space). This forward process is a fixed, learnable Markov chain.

[0027] For each unmodified RNA sequence in the training subset, its RNA feature information is extracted using a pre-trained RNA base model (e.g., the Nuwa RNA base model or RNA-FM). Specifically, the unmodified RNA sequence is input into the RNA base model, and through operations such as nucleotide-by-nucleotide embedding, multi-layer Transformer encoding, and self-attention computation, a high-dimensional latent vector containing long-range dependencies, sequence context semantics, and structural motif information is obtained. This latent vector serves as the conditional input for the subsequent diffusion model; its specific extraction method is described in the aforementioned RNA feature information extraction process and will not be repeated here.

[0028] The training process for inverse denoising is as follows: The initial discrete diffusion model includes an inverse denoising network (usually using the same number of time steps T as the forward process, e.g., T=500). The training objective of this network is to learn to gradually reconstruct the true chemical modification sequence corresponding to the unmodified RNA sequence from a uniformly randomly distributed noise sequence, guided by RNA feature information. Specifically, for each training sample: starting with the true chemical modification sequence x0 corresponding to the unmodified sequence, a noisy sequence xt is obtained at one time step t (t is randomly selected from {1,...,T}) according to the forward transition matrix Qt. The current time step t, the noisy sequence xt, and RNA feature information (as conditions) are input into the inverse denoising network. This network predicts a denoised distribution or directly predicts an estimate of x0. The loss between the predicted result and the true x0 (e.g., cross-entropy loss) is calculated, and the parameters of the inverse denoising network are updated through gradient backpropagation. This process is repeated until the model converges, meaning the inverse denoising network can accurately reconstruct the original chemically modified sequence from the noisy sequence at any time step. After training, the discrete diffusion model with the current model parameters is the pre-trained discrete diffusion model. Thus, in actual drug design, for a new, unseen sequence, simply load the pre-trained model parameters, start with a pure noisy sequence (t=T), and iterate step-by-step using the inverse denoising network (t decreasing from T to 1), introducing thermodynamic guidance at each step, to output an optimized chemically modified sequence without retraining the model.

[0029] In this embodiment, if the original RNA drug sequence is an siRNA drug sequence, the corresponding step of performing reverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information includes: performing reverse denoising on the sense and antisense strands parsed from the noise sequence at the current time step using a dual-track decoding network of the pre-trained discrete diffusion model and based on the RNA feature information. It is understood that siRNA drugs include a sense strand and an antisense strand, which form a double-stranded structure in a complementary pairing. In this embodiment, the complete siRNA double-stranded sequence is first obtained, for example: sense strand (20nt): 5'-AUGCAUGCAUGCAUGCAAUG-3', antisense strand (20nt): 5'-CAUUGCAUGCAUGCAUGCAU-3' (complementary to the sense strand). These two strands are then treated as independent RNA drug sequences, and their corresponding RNA feature information is extracted using an RNA basic model to obtain the sense strand feature vector h_sense and the antisense strand feature vector h_antisense. Simultaneously, the randomly initialized noise sequences of the two strands are used as the starting points for their respective inverse denoising. The inverse denoising network of the pre-trained discrete diffusion model is then a dual-track decoding network, comprising two parallel decoding tracks. One track handles the inverse denoising process of the sense strand, with a higher modification tolerance threshold. This is because the sense strand is typically degraded after siRNA takes effect, requiring relatively lower precision in modification and allowing for more flexible modification patterns to enhance stability. The other track handles the inverse denoising process of the antisense strand, with a lower modification tolerance threshold and higher RISC affinity weights. This is because the antisense strand needs to accurately identify target mRNA and efficiently load it into the RISC complex, requiring more stringent requirements on modification sites and types. The two decoding tracks can share some underlying parameters (e.g., attention mechanisms, embedding layers) to reduce overfitting and improve training efficiency, but each has its own independent output layer or independent guiding weight adjustment module, thus achieving asymmetric generation capabilities. During the inverse denoising process, the processing of the sense and antisense strands is independent and parallel. For each time step t (decreasing from T to 1), perform the following operations: For the inverse denoising of the positive chain, the positive chain noise sequence xt1, the time step index t, and the positive chain feature vector h_sense at the current time step are input into the positive chain track of the dual-track decoding network. The positive chain track outputs the predicted clean sequence or the predicted noise component, generating a positive chain candidate chemical modification sequence. The druggability index (MFE, RISC affinity, off-target penalty) of the candidate sequence is calculated, and the energy classifier gradient is generated according to the preset high tolerance threshold to correct the output, thus obtaining the positive chain target chemical modification sequence at the current time step.

[0030] For inverse denoising of the antisense chain, the antisense chain noise sequence xt2, time step index t, and antisense chain feature vector h_antisense at the current time step are input into the antisense chain track of the dual-track decoding network. The antisense chain track outputs the predicted clean sequence, generating candidate chemical modification sequences for the antisense chain. The druggability index of these candidate sequences is calculated, but a lower modification tolerance threshold and a higher RISC affinity weight are used to generate the corresponding energy classifier gradient, which corrects the output to obtain the target chemical modification sequence for the antisense chain at the current time step. The inverse denoising processes of the two tracks are performed simultaneously within the same pre-trained discrete diffusion model without interference. After each iteration, the corrected target sequences for the sense chain and antisense chain are used as the noise sequences for the next time step and input into the corresponding tracks.

[0031] The different requirements for chemical modifications of the sense and antisense strands stem from their different biological functions. The sense strand primarily functions to form a double-stranded structure to protect the antisense strand, and its sequence specificity requirements are not high. Therefore, a higher proportion of 2'-O-Me or thiomodification can be used to enhance resistance to nuclease degradation, and even a dense distribution of modification sites will not significantly affect RNA interference activity. The antisense strand needs to bind to the Argonaute protein of the RISC complex, and the nucleotide type and modification state at its 5' end directly affect the silencing efficiency. Over-modification or inappropriate modification can hinder RISC loading, leading to off-target effects or loss of activity. Therefore, the modification design of the antisense strand must be more conservative, typically avoiding modification in the seed region (positions 2-8) and using modifications such as 2'-F only at the ends or specific positions. Dual-track decoding networks achieve automated adaptation to these biological differences by setting different guiding parameters (modification tolerance threshold, RISC affinity weight) for the two strands, eliminating the need for manual rule design for each strand.

[0032] After the inverse denoising iteration is completed (time step t=0), the final chemical modification sequences of the sense and antisense strands are output. For example, the chemical modification result of the sense strand modification scheme after inverse denoising is 5'-AmUmGfCfAmUfGfCmAfUfGfCmAfAmUfG-3', and the chemical modification result of the antisense strand modification scheme is 5'-mCfAfUmUfGfCmAfUfGfCmAfUfGfCmAfU-3'. The modification patterns of the two strands exhibit asymmetric characteristics, with a higher modification density in the sense strand and a lower modification density in the antisense strand, where key sites (such as the 5' end) remain unmodified or undergo specific modification types.

[0033] In this embodiment, conditional and unconditional diffusion models trained based on RNA feature information are used to perform conditional and unconditional predictions on the noise sequence at the current time step, respectively, to obtain conditional and unconditional prediction output information. The conditional and unconditional prediction output information are then linearly combined according to preset combination weights to obtain denoised output information for correcting candidate chemical modification sequences at the current time step. It is understood that, without the gradient calculation and correction steps of the energy classifier, the guided optimization of candidate chemical modification sequences can be achieved using a classifier-free guided method. Specifically, a unified diffusion model (containing both conditional and unconditional diffusion models) that can simultaneously support conditional and unconditional predictions is pre-trained. The structure of this model is the same as that of a conventional discrete diffusion inverse denoising network, but a conditional random discard strategy is adopted during the training phase. Specifically, for each training sample (unmodified RNA sequence and its corresponding real chemical modification sequence), conditional information is randomly discarded according to probability p (e.g., p=0.1). In other words, RNA feature information is input into the inverse denoising network with probability 1-p (conditional mode); with probability p, the conditional input is replaced with an empty vector or a zero vector (unconditional mode). In unconditional mode, the model does not receive any RNA feature information and relies solely on its prior knowledge of the distribution of modification markers for prediction. In this way, the same model simultaneously learns both conditional prediction (generating matching modification sequences based on input RNA feature information) and unconditional prediction (generating general modification sequences based solely on the statistical regularity of modification markers). After training, a unified diffusion model is obtained, which has two functional branches: a conditional branch (with RNA feature information as input) and an unconditional branch (with empty conditions as input). The two branches share most of the network parameters, differing only in the conditional inputs during the forward computation. Therefore, during the inference phase, they are linearly combined to generate the guided denoised output. Specifically, in the inverse denoising process (time step t decreases from T to 1), for the noise sequence xt at the current time step, two forward computations are performed simultaneously. Conditional prediction: The current noise sequence xt, time step index t, and RNA feature information h (extracted from the original RNA drug sequence) are input into the model to obtain the conditional prediction output. Unconditional prediction: The current noise sequence xt, time step index t, and an empty conditional vector (e.g., a zero vector) are input into the model to obtain the unconditional prediction output. Then, according to the preset combination weight w, the two prediction outputs are linearly combined to generate the guided denoised output. When w=0, it degenerates into unconditional generation; when w=1, it is equivalent to standard conditional generation; when w>1, the difference between conditional prediction and unconditional prediction is amplified, thereby enhancing the guidance strength.The denoised output after guidance is directly used to calculate the denoised result of the current time step (i.e., to obtain x{t-1} from xt), which is equivalent to the result after correcting the candidate chemical modification sequence using the gradient of the energy classifier.

[0034] In this embodiment, the minimum free energy, RISC assembly affinity, and off-target effect penalty term of the candidate chemically modified sequence are calculated, and the minimum free energy, RISC assembly affinity, and off-target effect penalty term are weighted into a druggability index. The gradient information of the candidate chemically modified sequence is calculated using the druggability index to obtain the energy classifier gradient information at the current time step. It can be understood that the guided optimization process for the candidate chemically modified sequence under the energy classifier gradient calculation and correction step is as follows: First, the energy classifier gradient information is determined. Specifically, for the candidate chemically modified sequence generated at the current time step (e.g., a 20nt siRNA positive strand modification sequence: 5'-AmUmGfCfAmUfGfCmAfUfGfCmAfAmUfG-3'), the minimum free energy is calculated using a nearest-neighbor thermodynamic parameter model. Since the sequence contains non-natural chemical modifications (2'-F, 2'-O-Me), an extended thermodynamic parameter set (correcting standard RNA parameters according to the modification type) or a fast prediction model based on machine learning potential is required. The lower the MFE value (the larger the negative value), the more stable the secondary structure formed by the sequence and the stronger its resistance to degradation by nucleases in vivo. For specific calculations, dynamic programming algorithms or deep learning prediction models (such as modified adaptations of RNAfold and LinearFold) can be used to output the MFE value of the sequence, in kcal / mol. For example, the MFE calculation result of the above candidate sequence is -32.4 kcal / mol, which is the minimum free energy. Predicting RISC assembly affinity using a pre-trained regression model or empirical scoring function involves inputting the modification type distribution of the candidate sequence (especially the modification status of the 1st position at the 5' end, the 2nd to 8th positions in the seed region, and the 3' end), and outputting a normalized affinity score (range 0-1, higher values ​​indicate higher loading efficiency). The RISC assembly affinity score reflects the efficiency of the antisense strand being effectively loaded into the RNA-induced silencing complex after the siRNA double helix unwinds. This affinity is highly dependent on the chemical modification patterns at the 5' end and key sites. For example, when the 5' end of the antisense strand retains an unmodified phosphate group and the second position is modified with 2'-F, the affinity score can reach 0.85; if the 5' end is modified with 2'-O-Me, the affinity may drop below 0.3. The off-target penalty term is used to quantify the risk of non-specific binding between the siRNA antisense strand and non-target mRNA. Specifically, the modified sequence of the seed region (positions 2-8) of the antisense strand is extracted and compared with the complementary regions of all non-target mRNAs in the transcriptome to calculate the number of potential off-target sites and their binding strength. The off-target penalty term can be designed as a monotonically increasing function of the off-target score, for example: The penalty term is calculated as α × N + β × |E|, where N represents the number of potential off-target sites, E represents the free energy of the strongest off-target binding, and α and β are the corresponding preset weighting coefficients.

[0035] Finally, the three indicators mentioned above are weighted and combined into a unified comprehensive druggability score S. The weighting coefficients can be adjusted according to the drug design objectives, and there are no specific limitations on this. Then, the above druggability index S is regarded as an energy function of the candidate chemical modification sequence x_cand (discrete label sequence). In order to guide the generation direction in the inverse denoising process of the diffusion model, it is necessary to calculate the gradient of S with respect to x_cand. Since x_cand is discrete (each position is a classification label), continuous relaxation techniques or direct gradient approximation in the label embedding space are usually used. Specifically, each modification label in the candidate sequence is mapped to a continuous embedding vector ei, and the druggability index S can be regarded as a function of the embedding vector S(e1, e2,..., eL). The partial derivative of S with respect to each embedding vector is calculated to obtain the gradient vector. This gradient vector reflects which direction fine-tuning the label distribution under the current candidate sequence embedding representation can maximize the comprehensive druggability score. This gradient vector is used as the gradient information output of the energy classifier for subsequent perturbation of the output direction of the denoising network.

[0036] In this embodiment, a perturbation vector is obtained based on the gradient information of the energy classifier and a preset guiding step size coefficient. The perturbation vector is then superimposed on the candidate chemical modification sequence to obtain the target chemical modification sequence at the current time step. It is understood that in practical implementation, the energy classifier gradient information can be multiplied by a step size coefficient and then superimposed on the prediction of the denoising network, thereby guiding the generation process towards a region with better drug-like properties.

[0037] As can be seen, this application discloses obtaining the original RNA drug sequence and inputting the original RNA drug sequence into an RNA basic model to extract RNA feature information containing long-range sequence dependence and structural motif information; using a pre-trained discrete diffusion model and based on the RNA feature information to perform reverse denoising on the noise sequence at the current time step to generate candidate chemical modification sequences, and calculating the druggability index of the candidate chemical modification sequences; then determining the energy classifier gradient information at the current time step based on the druggability index, and using the energy classifier gradient information to correct the candidate chemical modification sequences to obtain the target chemical modification sequence at the current time step, which is used as the noise sequence for the next time step for iterative reverse denoising until the time step iteration end condition is met, and outputting the corresponding target chemical modification sequence as the drug design result of the original RNA drug sequence. Therefore, it can be seen that training the encoder using the RNA-based model to extract deep features can capture the far-end interactions of sequences, that is, obtain global long-range dependencies, which can be used to coordinate the modification scheme globally. By adopting the discrete diffusion model, the Markov process that defines the discrete label space can directly generate effective discrete modification combinations, avoiding the invalid output caused by continuous space mapping. In each step of denoising, gradient information is determined and the output is perturbed, so that the chemically modified sequence output after correction can simultaneously satisfy stability, activity and safety, rather than just statistical fitting.

[0038] like Figure 2 As shown, the present invention also discloses a drug design device based on an RNA basic model and thermodynamically guided discrete diffusion, comprising: The information extraction module 11 is used to obtain the original RNA drug sequence and input the original RNA drug sequence into the RNA basic model to extract RNA feature information containing sequence long-range dependence and structural motif information; The drug design module 12 is used to perform reverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information to generate candidate chemical modification sequences, calculate the druggability index of the candidate chemical modification sequences, determine the energy classifier gradient information at the current time step based on the druggability index, and use the energy classifier gradient information to correct the candidate chemical modification sequences to obtain the target chemical modification sequence at the current time step. This target chemical modification sequence is then used as the noise sequence for the next time step for iterative reverse denoising until the time step iteration termination condition is met. Finally, the corresponding target chemical modification sequence is output as the drug design result of the original RNA drug sequence.

[0039] Thus, by injecting structural and functional constraints into the three dimensions of prior coding of the RNA basic model, discrete state transition, and energy classifier guidance, a unified chemical modification generation model adapted to the heterogeneity of siRNA positive and negative strands was constructed, which can be used to achieve global optimization design of highly active, long half-life nucleic acid drug sequences.

[0040] Furthermore, embodiments of this application also disclose an electronic device, Figure 3 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0041] Figure 3 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the drug design method based on the RNA fundamental model and thermodynamically guided discrete diffusion disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0042] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0043] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0044] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0045] The operating system 221 manages and controls the various hardware devices and computer programs 222 on the electronic device 20 to enable the processor 21 to perform calculations and processing on the massive amounts of data 223 in the memory 22. It can be Windows Server, Netware, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the drug design method based on the RNA fundamental model and thermodynamically guided discrete diffusion disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the electronic device from external devices, as well as data collected by its own input / output interface 25.

[0046] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned drug design method based on an RNA fundamental model and thermodynamically guided discrete diffusion. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0047] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0048] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly in hardware, software modules executed by a processor, or a combination of both. The software module may be located in random access memory (RAM), memory, read-only memory (ROM), electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, removable disks, CD-ROMs (Compact Disc-Read Only Memory), or any other form of storage medium known in the art.

[0049] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0050] The solution provided by the present invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A drug design method based on an RNA fundamental model and thermodynamically guided discrete diffusion, characterized in that, include: The original RNA drug sequence was obtained and input into the RNA basic model to extract RNA feature information containing sequence long-range dependence and structural motif information; A pre-trained discrete diffusion model is used to reverse-denoise the noise sequence at the current time step based on the RNA feature information to generate candidate chemical modification sequences. The druggability index of the candidate chemical modification sequences is calculated, and then the energy classifier gradient information at the current time step is determined based on the druggability index. The candidate chemical modification sequences are then corrected using the energy classifier gradient information to obtain the target chemical modification sequence at the current time step. This target chemical modification sequence is then used as the noise sequence for the next time step and iterative reverse-denoising is performed until the time step iteration termination condition is met. Finally, the corresponding target chemical modification sequence is output as the drug design result of the original RNA drug sequence.

2. The drug design method based on RNA fundamental model and thermodynamically guided discrete diffusion according to claim 1, characterized in that, The step of inputting the original RNA drug sequence into an RNA basic model to extract RNA feature information containing long-range sequence dependence and structural motif information includes: The original RNA drug sequence is encoded nucleotide-by-nucleotide by RNA basic model to obtain an initial embedding vector, and the long-range dependencies between nucleotides at different positions in the initial embedding vector are calculated by self-attention mechanism. The contextual semantic information of nucleotides at each position is extracted using the aforementioned RNA basic model; The RNA basic model outputs the structural tendency code corresponding to each position as structural motif information. Based on the long-range dependency, the contextual semantic information, and the structural motif information, the RNA feature information of the original RNA drug sequence is determined.

3. The drug design method based on RNA fundamental model and thermodynamically guided discrete diffusion according to claim 1, characterized in that, Before performing reverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information, the process further includes: Obtain a training dataset, which contains multiple pairs of paired data, each pair of paired data including an unmodified RNA sequence and its corresponding chemically modified sequence; The training dataset is filtered using a strategy of retaining minimal residual values ​​to select target paired data that meet preset modification conditions to construct a training subset; A forward Markov transition matrix is ​​defined in the discrete chemical modification label space of the initial discrete diffusion model to degenerate the chemical modification sequence into a noise sequence with uniform random distribution step by step. The RNA base model is used to extract features from the unmodified RNA sequences in the training subset to obtain the RNA feature information of the unmodified RNA sequences; The initial discrete diffusion model is trained using the RNA feature information of the unmodified RNA sequence and the chemical modification sequence corresponding to the unmodified RNA sequence. The model learns the inverse process of reconstructing the uniformly randomly distributed noise sequence into the chemical modification sequence corresponding to the unmodified RNA sequence step by step, so as to obtain the discrete diffusion model under the current model parameters as a pre-trained discrete diffusion model.

4. The drug design method based on RNA fundamental model and thermodynamically guided discrete diffusion according to claim 1, characterized in that, The calculation of the druggability index of the candidate chemically modified sequence, and then the determination of the energy classifier gradient information at the current time step based on the druggability index, includes: The minimum free energy, RISC assembly affinity, and off-target effect penalty term of the candidate chemical modification sequence are calculated, and the minimum free energy, the RISC assembly affinity, and the off-target effect penalty term are weighted to form a drug-likeness index. The gradient information of the candidate chemical modification sequence is calculated using the drug-likeness index to obtain the energy classifier gradient information at the current time step.

5. The drug design method based on RNA fundamental model and thermodynamically guided discrete diffusion according to claim 1, characterized in that, The step of using the gradient information of the energy classifier to correct the candidate chemical modification sequence to obtain the target chemical modification sequence at the current time step includes: The perturbation vector is obtained based on the gradient information of the energy classifier and the preset guiding step size coefficient. The perturbation vector is superimposed on the candidate chemical modification sequence to obtain the target chemical modification sequence at the current time step.

6. The drug design method based on RNA fundamental model and thermodynamically guided discrete diffusion according to claim 1, characterized in that, The original RNA drug sequence is a siRNA drug sequence; Accordingly, the step of performing reverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information includes: The positive and negative strands of the noise sequence at the current time step are reversed by using a dual-track decoding network based on a pre-trained discrete diffusion model and the RNA feature information.

7. The drug design method based on RNA fundamental model and thermodynamically guided discrete diffusion according to claim 1, characterized in that, The process of reverse denoising the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information also includes: Conditional and unconditional diffusion models trained based on RNA feature information are used to perform conditional and unconditional predictions on the noise sequence at the current time step, respectively, to obtain conditional and unconditional prediction output information. The conditional prediction output information and the unconditional prediction output information are linearly combined according to a preset combination weight to obtain denoised output information for correcting the candidate chemical modification sequence at the current time step.

8. A drug design device based on an RNA fundamental model and thermodynamically guided discrete diffusion, characterized in that, include: The information extraction module is used to obtain the original RNA drug sequence and input the original RNA drug sequence into the RNA basic model to extract RNA feature information containing sequence long-range dependence and structural motif information; The drug design module is used to perform reverse denoising on the noise sequence at the current time step using a pre-trained discrete diffusion model and based on the RNA feature information to generate candidate chemical modification sequences. It then calculates the druggability index of the candidate chemical modification sequences, determines the energy classifier gradient information at the current time step based on the druggability index, and uses the energy classifier gradient information to correct the candidate chemical modification sequences to obtain the target chemical modification sequence for the current time step. This target chemical modification sequence is then used as the noise sequence for the next time step for iterative reverse denoising until the time step iteration termination condition is met. Finally, the corresponding target chemical modification sequence is output as the drug design result of the original RNA drug sequence.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the drug design method based on RNA fundamental model and thermodynamically guided discrete diffusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when executed by a processor, the computer programs implement the steps of the drug design method based on the RNA basic model and thermodynamically guided discrete diffusion as described in any one of claims 1 to 7.

Images