Molecular generation method and device based on multi-modal optimal noise scheduling

Abstract

The invention provides a multi-modal optimal noise scheduling-based molecular generation method and device, and the method comprises the steps: representing any one joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time re-parameterization function, and building a unified parameterization frame of continuous modal noise and discrete modal noise; minimizing a pre-constructed generalized loss function through a pre-trained model, and training based on the generalized loss function to obtain a generation model for evaluating any joint noise scheduling function in the joint noise scheduling function space; based on an L2 norm square error evaluation result of each point of the discretized time grid by the generative model, searching and determining a minimum cumulative cost path from a starting point to an end point through a dynamic programming algorithm; and determining optimal joint noise scheduling according to the minimum cumulative cost path, and generating a drug molecular structure based on the optimal joint noise scheduling. Through the method provided by the invention, the geometric effectiveness and the calculation efficiency of molecule generation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence-assisted drug design, and in particular to a molecule generation method and device based on multimodal optimal noise scheduling. Background Art

[0002] Structure-Based Drug Design (SBDD) is a method that leverages the three-dimensional structural information of target proteins to design drug molecules with high affinity and selectivity. The core challenge lies in simultaneously modeling the continuous modalities (3D atomic coordinates) and discrete modalities (2D molecular topology) of a molecule to generate molecules that both conform to the geometric characteristics of the protein binding pocket and possess a reasonable chemical structure. In recent years, deep generative models (such as diffusion models and Bayesian flow networks) have made significant progress in SBDD, but the joint generation of multimodal data still faces a key challenge: the choice of noise scheduling strategy directly affects the geometric plausibility and bioactivity of the generated molecules. Noise scheduling determines the co-evolutionary path of the continuous and discrete modalities during the generation process. Its optimization objective is to maximize the variational lower bound (VLB), thereby improving the generation quality.

[0003] How to achieve drug molecule design based on multimodal optimal noise scheduling is a technical problem that needs to be solved at present. Summary of the Invention

[0004] The present invention provides a molecule generation method and device based on multimodal optimal noise scheduling, which are used to solve the defects in the prior art.

[0005] The present invention provides a molecule generation method based on multimodal optimal noise scheduling, comprising the following steps: A unified parameterization framework for continuous modal noise and discrete modal noise is established by expressing any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function; wherein the joint noise scheduling function space includes: a plurality of joint noise scheduling functions, each of the plurality of joint noise scheduling functions is composed of a continuous modal noise schedule and a discrete modal noise schedule; the time reparameterization function includes: a first reparameterization function and a second reparameterization function, corresponding to the continuous modal noise and the discrete modal noise, respectively; Minimizing a pre-constructed generalized loss function through a pre-trained model, and obtaining a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function; wherein the generalized loss function is: constructed based on the unified parameterized framework; Based on the L2 norm square error evaluation results of each point in the discretized time grid by the generative model, a dynamic programming algorithm is used to search and determine the minimum cumulative cost path from the starting point to the end point; An optimal joint noise schedule is determined according to the minimum cumulative cost path, and a drug molecular structure is generated based on the optimal joint noise schedule.

[0006] According to a molecule generation method based on multimodal optimal noise scheduling provided by the present invention, the first parameterized function is: , the second parameterized function is: ; The time reparameterization function is used to express any joint noise scheduling function in the predefined joint noise scheduling function space as a preset target form, and a unified parameterization framework for continuous modal noise and discrete modal noise is established, which is achieved by the following formula: in, represents the joint noise scheduling function in the form of a preset target, Represents a predefined monotonic function.

[0007] According to a molecule generation method based on multimodal optimal noise scheduling provided by the present invention, the pre-trained generative model minimizes a pre-constructed generalized loss function to obtain a target generalized loss function, which is implemented by the following formula: in, Represents the target generalized loss function on the input data The value on represents the latent variable, represents the continuous mode noise scheduling parameter, represents the discrete modal noise scheduling parameter, represents the probability distribution of the forward noise process, represents the output of the pre-trained generative model, represents the L2 norm squared error.

[0008] According to a molecule generation method based on multimodal optimal noise scheduling provided by the present invention, the method determines the minimum cumulative cost path from the starting point to the end point by searching through a dynamic programming algorithm based on the L2 norm square error evaluation result of each point of the discretized time grid based on the generation model, including: Based on the L2 norm square error evaluation results of each point in the discretized time grid by the generative model and the estimated first-order derivative of the noise scheduling function with respect to time, the minimum cumulative cost path from the starting point [0,0] to the end point [1,1] is determined by searching through a dynamic programming algorithm.

[0009] According to a molecule generation method based on multimodal optimal noise scheduling provided by the present invention, the minimum cumulative cost path satisfies the following recursive relationship: in, represents the minimum cumulative cost, represents the continuous mode noise scheduling parameter, represents the discrete modal noise scheduling parameter, represents the step vector, represents the discretized noise level parameter, which is used to estimate the first-order derivative of the noise scheduling function with respect to time, represents the joint noise scheduling function, Indicates the instantaneous cost of the current grid point.

[0010] According to the present invention, a molecule generation method based on multimodal optimal noise scheduling is provided, wherein the optimal joint noise scheduling is determined according to the minimum cumulative cost path, and the drug molecular structure is generated based on the optimal joint noise scheduling, including: Mapping the minimum cumulative cost path to the joint noise scheduling function space to determine the optimal joint noise scheduling; Drug molecular structures are generated based on the optimal joint noise scheduling.

[0011] The present invention also provides a molecule generation device based on multimodal optimal noise scheduling, comprising the following modules: A module is established for expressing any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function, thereby establishing a unified parameterization framework for continuous mode noise and discrete mode noise; wherein the joint noise scheduling function space includes: a plurality of joint noise scheduling functions, each of the plurality of joint noise scheduling functions is composed of a continuous mode noise schedule and a discrete mode noise schedule; and the time reparameterization function includes: a first reparameterization function and a second reparameterization function, corresponding to the continuous mode noise and the discrete mode noise, respectively; A construction module is configured to minimize a pre-constructed generalized loss function through a pre-trained model, and train a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function; wherein the generalized loss function is constructed based on the unified parameterized framework; A determination module, configured to search and determine a minimum cumulative cost path from a starting point to an end point through a dynamic programming algorithm based on an L2 norm square error evaluation result of each point in the discretized time grid according to the generative model; A generation module is used to determine an optimal joint noise schedule according to the minimum cumulative cost path, and generate a drug molecular structure based on the optimal joint noise schedule.

[0012] According to a molecule generation device based on multimodal optimal noise scheduling provided by the present invention, the first parameterized function is: , the second parameterized function is: ; The time reparameterization function is used to express any joint noise scheduling function in the predefined joint noise scheduling function space as a preset target form, and a unified parameterization framework for continuous modal noise and discrete modal noise is established, which is achieved by the following formula: in, represents the joint noise scheduling function in the form of a preset target, Represents a predefined monotonic function.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the molecule generation method based on multimodal optimal noise scheduling as described above is implemented.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for generating molecules based on multimodal optimal noise scheduling.

[0015] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-described methods for generating molecules based on multimodal optimal noise scheduling.

[0016] The present invention provides a molecule generation method and device based on multimodal optimal noise scheduling, which expresses any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function, and establishes a unified parameterization framework for continuous modal noise and discrete modal noise; wherein, the joint noise scheduling function space includes: multiple joint noise scheduling functions, each of the multiple joint noise scheduling functions consists of continuous modal noise scheduling and discrete modal noise scheduling; the time reparameterization function includes: a first reparameterization function and a second reparameterization function, which respectively correspond to the continuous modal noise and the discrete modal noise. state noise and discrete modal noise; minimizing a pre-constructed generalized loss function through a pre-trained model, and obtaining a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function training; wherein, the generalized loss function is: constructed based on the unified parameterized framework; based on the L2 norm square error evaluation results of each point of the discretized time grid by the generative model, the minimum cumulative cost path from the starting point to the end point is searched and determined by a dynamic programming algorithm; the optimal joint noise scheduling is determined according to the minimum cumulative cost path, and the drug molecular structure is generated based on the optimal joint noise scheduling. It can be seen that the present invention realizes the dynamic coordination of the two modes in the generation process by establishing a unified parameterized framework for continuous mode and discrete mode, and designing a generalized loss function based on double integration; using a dynamic programming algorithm to search for the minimum cumulative cost path in the two-dimensional noise space defined by the generalized loss function, the VLB optimization is converted into a path integral problem; the optimal joint noise scheduling obtained by time reparameterization improves the geometric validity and computational efficiency of molecule generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 It is a flow chart of the molecule generation method based on multimodal optimal noise scheduling provided by the present invention.

[0019] Figure 2 This is a complete flow chart of the molecule generation method based on multimodal optimal noise scheduling provided by the present invention.

[0020] Figure 3 It is a structural schematic diagram of a molecule generation device based on multimodal optimal noise scheduling provided by the present invention.

[0021] Figure 4 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0023] The following combination Figures 1-4 The present invention describes a molecule generation method and device based on multimodal optimal noise scheduling.

[0024] It should be noted that this paper proposes VLB-Optimal Scheduling (VOS), a principled approach for achieving optimal probabilistic paths in structure-based drug design (SBDD) by analyzing and determining the optimal noise schedule of discrete two-dimensional topology and continuous three-dimensional atomic positions. Furthermore, by combining VOS with an advanced equivariant network architecture, MolPilot is proposed, achieving the best generation results. This is specifically explained from the following three aspects: 1. The path-dependent VLB corresponding to the distorted probability path induced by a specific noise schedule in the function space is theoretically analyzed.

[0025] 2. We formally describe the complete function space of all possible noise schedules and propose to navigate this space via temporal reparameterization.

[0026] 3. We propose a path-invariant generalized loss objective that can effectively estimate the VLB potential energy surface over the entire space of noisy scheduling functions. We then describe a method for searching for the optimal schedule that maximizes VLB on this landscape.

[0027] Figure 1 is a flow chart of the molecule generation method based on multimodal optimal noise scheduling provided by the present invention, such as Figure 1 As shown, the method includes the following: Step 100: Express any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function, and establish a unified parameterization framework for continuous modal noise and discrete modal noise; wherein the joint noise scheduling function space includes: multiple joint noise scheduling functions, each of the multiple joint noise scheduling functions consists of continuous modal noise scheduling and discrete modal noise scheduling; the time reparameterization function includes: a first reparameterization function and a second reparameterization function, corresponding to continuous modal noise and discrete modal noise, respectively.

[0028] It should be noted that this embodiment establishes a key conclusion for the distorted probability path: joint noise scheduling Will induce path-dependent VLB.

[0029] A key foundation in unimodality is the continuous time loss (VLB) versus noise scheduling function. The shape of This makes it possible to design different scheduling strategies to improve training efficiency. As a natural extension, this embodiment extends this invariance to Multimodal scenarios: (1) in, represents any continuous mode noise scheduling function, represents any discrete modal noise scheduling function, both of which are time The monotonic function on and , Indicates that after variable substitution, Expressed as The probability distribution of the forward noise process after the function, represents the output of the pre-trained generative model after variable substitution, Represents the L2 norm square error after variable substitution.

[0030] This equation shows that the generalized loss is invariant to decoupled scheduling, which is expressed as The surface integral on .

[0031] However, this generalized loss This no longer corresponds to the goal of generative modeling, since VLB should be the path integral along a curve in the plane: (2) This curve is equivalent to a specific coupled noise schedule Therefore, it is concluded that even if the endpoint and The VLB under different coupling modes will also change, and VLB is no longer The intermediate trajectory of remains unchanged, which is essentially different from unimodal generative modeling.

[0032] Therefore, the core challenge is to find a joint schedule from the design space Z that can get the optimal VLB along the path integration. : (3) To formalize the design space of noise scheduling, this embodiment defines a monotonically increasing function space , then the coupled function space for: (4) in, and are the monotonic scheduling functions for continuous mode and discrete mode respectively.

[0033] To explore any scheduling configuration, this embodiment introduces a time reparameterization function and , so that any Through predefined And implicit function and The various forms of expression are: (5) There is the following theorem to ensure that this is based on of The general form is sufficient to cover the entire function space .

[0034] Theorem 1: Assume that there exists a monotonic function ,make is any such monotonic function. Then there exists a time-reparameterized function , making In fact, and have the same monotonicity.

[0035] That is, in formula (5) The general form is sufficient to cover the entire function space .

[0036] Note 2: From the monotonicity of the scheduling function, it can be seen that only the independent setting Any combination of noise levels can be achieved.

[0037] Step 200: Minimize a pre-constructed generalized loss function through a pre-trained model, and obtain a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function training; wherein the generalized loss function is: constructed based on the unified parameterized framework.

[0038] Specifically, find the optimal The first obstacle is to obtain a Generative Models Evaluated on This embodiment replaces the variable and ,train To minimize the generalized loss described in formula (1) , so that the objective function is converted to: (6) in, Represents the target generalized loss function on the input data The value on represents the latent variable, represents the continuous mode noise scheduling parameter, represents the discrete modal noise scheduling parameter, represents the probability distribution of the forward noise process, represents the output of the pre-trained generative model, Denotes the L2 norm squared error. The present invention theoretically proves that the generalized loss function is invariant with respect to the variational lower bound of any joint noise scheduling function in the joint noise scheduling function space. The following proposition holds true for this: Proposition 3: Assume that there exists a generative model trained by equation (6) ,and for Any monotonically increasing function in. Then the line integral in equation (2) corresponds to Negative VLB.

[0039] Step 300: Based on the L2 norm square error evaluation results of each point in the discretized time grid according to the generative model, a minimum cumulative cost path from the starting point to the end point is searched and determined by a dynamic programming algorithm.

[0040] Furthermore, this embodiment proposes that Methods for navigating in function space. According to Remark 2, it is possible to discretize Therefore, we need to find the optimal VLB. can be converted into a discretized Step trajectory Search up from [0, 0] to [1, 1] for the solution with the minimum cumulative cost.

[0041] Figure 2 This is a complete flow chart of the molecule generation method based on multimodal optimal noise scheduling provided by the present invention, such as Figure 2 As shown, through the evaluation Trained generative model The KL divergence on a batch of samples may be estimated and The cost matrix is placed on the value grid and a smooth loss surface is fitted using B-spline interpolation.

[0042] Given noise level As The implicit function of , whose cost is defined as: (7) This embodiment uses a dynamic programming method to solve the search problem for the minimum cumulative cost. The minimum cumulative cost path satisfies the following recursive relationship: (8) in, represents the minimum cumulative cost, represents the continuous mode noise scheduling parameter, represents the discrete modal noise scheduling parameter, represents the step vector, The effective range of is approximated by the gradient of the smooth surface. This method can ensure that a scheduling strategy with the maximum VLB is obtained. represents the discretized noise level parameter, which is used to estimate the first-order derivative of the noise scheduling function with respect to time, represents the joint noise scheduling function, Indicates the instantaneous cost of the current grid point.

[0043] Step 400: Determine an optimal joint noise schedule according to the minimum cumulative cost path, and generate a drug molecular structure based on the optimal joint noise schedule.

[0044] Specifically, Corresponding to all The optimal unbiased Monte Carlo estimate of VLB in the generative model trained is obtained by Back to Path You can get the optimal schedule .

[0045] Continue to see Figure 2, the derived optimal scheduling strategy is visualized in Figure 2 B. Intuitively, this temporal reparameterization corresponds to a two-stage probabilistic path: (1) Shape-driven sketch generation stage: First, the generation process focuses on the generation of continuous atomic positions and basically ignores the discrete topological structure. When , the generative model begins to fit the possible two-dimensional molecular graph to the rough shape of the fixed three-dimensional conformation. (2) Topology-driven docking phase: In the final phase , the generation process enters the docking phase, adjusting the conformation according to the discrete molecular topology. By promoting this principle-based probabilistic path, the optimal noise scheduling strategy effectively integrates information from different modalities, thereby achieving the best variational lower bound (VLB) and sample generation quality.

[0046] The molecular generation method based on multimodal optimal noise scheduling provided by the present invention establishes a unified parameterization framework for continuous modes (3D atomic coordinates) and discrete modes (2D molecular topology) and designs a generalized loss function based on double integrals to achieve dynamic coordination of the two modes in the generation process; a dynamic programming algorithm is used to search for the minimum cumulative cost path in the two-dimensional noise space defined by the generalized loss function, and the VLB optimization is converted into a path integral problem. The optimal scheduling obtained by time reparameterization effectively improves the training convergence speed and VLB of the generation model; the optimal scheduling automatically presents a two-stage feature of shape drive (t∈[0.3,0.8]) → topology drive (t>0.8), which solves the modal imbalance problem in the prior art; through the strategy of pre-training model + offline dynamic programming, only a single training is required to support the optimal scheduling search of any protein pocket. Compared with traditional methods, the computational time of the inference stage is significantly reduced while maintaining the same generation quality.

[0047] The above is a description of the steps of the molecule generation method based on multimodal optimal noise scheduling provided by the present invention. From the description of the above steps, it can be seen that according to the molecule generation method based on multimodal optimal noise scheduling provided by the present invention, any joint noise scheduling function in the predefined joint noise scheduling function space is expressed as a preset target form through a time reparameterization function, and a unified parameterization framework for continuous modal noise and discrete modal noise is established; wherein, the joint noise scheduling function space includes: multiple joint noise scheduling functions, each of the multiple joint noise scheduling functions consists of continuous modal noise scheduling and discrete modal noise scheduling; the time reparameterization function includes: a first reparameterization function and a second reparameterization function, which are respectively for Applicable to continuous modal noise and discrete modal noise; minimize a pre-constructed generalized loss function through a pre-trained model, and obtain a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function training; wherein the generalized loss function is: constructed based on the unified parameterized framework; based on the L2 norm square error evaluation results of each point of the discretized time grid by the generative model, determine the minimum cumulative cost path from the starting point to the end point through a dynamic programming algorithm search; determine the optimal joint noise scheduling based on the minimum cumulative cost path, and generate the drug molecular structure based on the optimal joint noise scheduling. It can be seen that the present invention realizes the dynamic coordination of the two modes in the generation process by establishing a unified parameterized framework for continuous and discrete modes and designing a generalized loss function based on double integration; adopts a dynamic programming algorithm to search for the minimum cumulative cost path in the two-dimensional noise space defined by the generalized loss function, and transforms the VLB optimization into a path integral problem; the optimal joint noise scheduling obtained by time reparameterization improves the geometric validity and computational efficiency of molecule generation.

[0048] The following describes a molecule generation device based on multimodal optimal noise scheduling provided by the present invention. The molecule generation device based on multimodal optimal noise scheduling described below and the molecule generation method based on multimodal optimal noise scheduling described above can refer to each other.

[0049] Figure 3 Schematic diagram of the structure of the molecular generation device based on multimodal optimal noise scheduling provided by the present invention, such as Figure 3 As shown, the molecule generation device based on multimodal optimal noise scheduling provided by the present invention includes: Establishing module 301, for expressing any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function, and establishing a unified parameterization framework for continuous mode noise and discrete mode noise; wherein the joint noise scheduling function space includes: a plurality of joint noise scheduling functions, each of the plurality of joint noise scheduling functions is composed of a continuous mode noise scheduling and a discrete mode noise scheduling; the time reparameterization function includes: a first reparameterization function and a second reparameterization function, corresponding to the continuous mode noise and the discrete mode noise, respectively; A construction module 302 is configured to minimize a pre-constructed generalized loss function using a pre-trained model, and train a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function; wherein the generalized loss function is constructed based on the unified parameterized framework; A determination module 303 is configured to determine a minimum cumulative cost path from a starting point to an end point by searching using a dynamic programming algorithm based on an L2 norm square error evaluation result of each point in the discretized time grid according to the generative model; The generating module 304 is configured to determine an optimal joint noise schedule according to the minimum cumulative cost path, and generate a drug molecular structure based on the optimal joint noise schedule.

[0050] The molecular generation device based on multimodal optimal noise scheduling provided by the present invention expresses any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function, and establishes a unified parameterization framework for continuous modal noise and discrete modal noise; wherein, the joint noise scheduling function space includes: a plurality of joint noise scheduling functions, each of the plurality of joint noise scheduling functions is composed of a continuous modal noise scheduling and a discrete modal noise scheduling; the time reparameterization function includes: a first reparameterization function and a second reparameterization function, which respectively correspond to the continuous modal noise Acoustic and discrete modal noise; minimizing a pre-constructed generalized loss function through a pre-trained model, and obtaining a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function training; wherein the generalized loss function is: constructed based on the unified parameterized framework; based on the L2 norm square error evaluation results of each point of the discretized time grid by the generative model, the minimum cumulative cost path from the starting point to the end point is searched and determined by a dynamic programming algorithm; the optimal joint noise scheduling is determined according to the minimum cumulative cost path, and the drug molecular structure is generated based on the optimal joint noise scheduling. It can be seen that the present invention realizes the dynamic coordination of the two modes in the generation process by establishing a unified parameterized framework for continuous and discrete modes and designing a generalized loss function based on double integration; the dynamic programming algorithm is used to search for the minimum cumulative cost path in the two-dimensional noise space defined by the generalized loss function, and the VLB optimization is converted into a path integral problem; the optimal joint noise scheduling obtained by time reparameterization improves the geometric validity and computational efficiency of molecule generation.

[0051] Based on the above embodiment, in this embodiment, the first parameterized function is: , the second parameterized function is: ; The time reparameterization function is used to express any joint noise scheduling function in the predefined joint noise scheduling function space as a preset target form, and a unified parameterization framework for continuous modal noise and discrete modal noise is established, which is achieved by the following formula: in, represents the joint noise scheduling function in the form of a preset target, Represents a predefined monotonic function.

[0052] Based on the above embodiment, in this embodiment, the generalized loss function is implemented by the following formula: in, Represents the target generalized loss function on the input data The value on represents the latent variable, represents the continuous mode noise scheduling parameter, represents the discrete modal noise scheduling parameter, represents the probability distribution of the forward noise process, represents the output of the pre-trained generative model, represents the L2 norm squared error.

[0053] Based on the above embodiment, in this embodiment, the determining module 303 is specifically configured to: Based on the L2 norm square error evaluation results of each point in the discretized time grid by the generative model and the estimated first-order derivative of the noise scheduling function with respect to time, the minimum cumulative cost path from the starting point [0,0] to the end point [1,1] is determined by searching through a dynamic programming algorithm.

[0054] Based on the above embodiment, in this embodiment, the minimum cumulative cost path satisfies the following recursive relationship: in, represents the minimum cumulative cost, represents the continuous mode noise scheduling parameter, represents the discrete modal noise scheduling parameter, represents the step vector, represents the discretized noise level parameter, which is used to estimate the first-order derivative of the noise scheduling function with respect to time, represents the joint noise scheduling function, Indicates the instantaneous cost of the current grid point.

[0055] Based on the above embodiment, in this embodiment, the generating module 304 is specifically configured to: Mapping the minimum cumulative cost path to the joint noise scheduling function space to determine the optimal joint noise scheduling; Drug molecular structures are generated based on the optimal joint noise scheduling.

[0056] Figure 4 An example of a physical structure diagram of an electronic device is shown below. Figure 4As shown, the electronic device may be a robot or other electronic device, and may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 may call the logic instructions in the memory 430 to execute the molecule generation method based on multimodal optimal noise scheduling, including: A unified parameterization framework for continuous modal noise and discrete modal noise is established by expressing any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function; wherein the joint noise scheduling function space includes: a plurality of joint noise scheduling functions, each of the plurality of joint noise scheduling functions is composed of a continuous modal noise schedule and a discrete modal noise schedule; the time reparameterization function includes: a first reparameterization function and a second reparameterization function, corresponding to the continuous modal noise and the discrete modal noise, respectively; Minimizing a pre-constructed generalized loss function through a pre-trained model, and obtaining a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function; wherein the generalized loss function is: constructed based on the unified parameterized framework; Based on the L2 norm square error evaluation results of each point in the discretized time grid by the generative model, a dynamic programming algorithm is used to search and determine the minimum cumulative cost path from the starting point to the end point; An optimal joint noise schedule is determined according to the minimum cumulative cost path, and a drug molecular structure is generated based on the optimal joint noise schedule.

[0057] Furthermore, the logic instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0058] On the other hand, the present invention further provides a computer program product, comprising a computer program, which may be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is capable of executing the molecule generation method based on multimodal optimal noise scheduling provided by the above methods, including: A unified parameterization framework for continuous modal noise and discrete modal noise is established by expressing any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function; wherein the joint noise scheduling function space includes: a plurality of joint noise scheduling functions, each of the plurality of joint noise scheduling functions is composed of a continuous modal noise schedule and a discrete modal noise schedule; the time reparameterization function includes: a first reparameterization function and a second reparameterization function, corresponding to the continuous modal noise and the discrete modal noise, respectively; Minimizing a pre-constructed generalized loss function through a pre-trained model, and obtaining a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function; wherein the generalized loss function is: constructed based on the unified parameterized framework; Based on the L2 norm square error evaluation results of each point in the discretized time grid by the generative model, a dynamic programming algorithm is used to search and determine the minimum cumulative cost path from the starting point to the end point; An optimal joint noise schedule is determined according to the minimum cumulative cost path, and a drug molecular structure is generated based on the optimal joint noise schedule.

[0059] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the molecule generation method based on multimodal optimal noise scheduling provided by the above methods, including: A unified parameterization framework for continuous modal noise and discrete modal noise is established by expressing any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function; wherein the joint noise scheduling function space includes: a plurality of joint noise scheduling functions, each of the plurality of joint noise scheduling functions is composed of a continuous modal noise schedule and a discrete modal noise schedule; the time reparameterization function includes: a first reparameterization function and a second reparameterization function, corresponding to the continuous modal noise and the discrete modal noise, respectively; Minimizing a pre-constructed generalized loss function through a pre-trained model, and obtaining a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function; wherein the generalized loss function is: constructed based on the unified parameterized framework; Based on the L2 norm square error evaluation results of each point in the discretized time grid by the generative model, a dynamic programming algorithm is used to search and determine the minimum cumulative cost path from the starting point to the end point; An optimal joint noise schedule is determined according to the minimum cumulative cost path, and a drug molecular structure is generated based on the optimal joint noise schedule.

[0060] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0061] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A molecule generation method based on multimodal optimal noise scheduling, characterized in that: include: A unified parameterization framework for continuous modal noise and discrete modal noise is established by expressing any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function; wherein the joint noise scheduling function space includes: a plurality of joint noise scheduling functions, each of the plurality of joint noise scheduling functions is composed of a continuous modal noise schedule and a discrete modal noise schedule; the time reparameterization function includes: a first reparameterization function and a second reparameterization function, corresponding to the continuous modal noise and the discrete modal noise, respectively; Minimizing a pre-constructed generalized loss function through a pre-trained model, and obtaining a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function; wherein the generalized loss function is: constructed based on the unified parameterized framework; Based on the L2 norm square error evaluation results of each point in the discretized time grid by the generative model, a dynamic programming algorithm is used to search and determine the minimum cumulative cost path from the starting point to the end point; An optimal joint noise schedule is determined according to the minimum cumulative cost path, and a drug molecular structure is generated based on the optimal joint noise schedule.

2. The molecule generation method based on multimodal optimal noise scheduling according to claim 1, characterized in that: The first parameterized function is: , the second parameterized function is: ; The time reparameterization function is used to express any joint noise scheduling function in the predefined joint noise scheduling function space as a preset target form, and a unified parameterization framework for continuous modal noise and discrete modal noise is established, which is achieved by the following formula: in, represents the joint noise scheduling function in the form of a preset target, Represents a predefined monotonic function.

3. The molecule generation method based on multimodal optimal noise scheduling according to claim 1, characterized in that: The generalized loss function is implemented by the following formula: in, Represents the target generalized loss function on the input data The value on represents the latent variable, represents the continuous mode noise scheduling parameter, represents the discrete modal noise scheduling parameter, represents the probability distribution of the forward noise process, represents the output of the pre-trained generative model, represents the L2 norm squared error.

4. The molecule generation method based on multimodal optimal noise scheduling according to claim 1, characterized in that: The method of determining the minimum cumulative cost path from the starting point to the end point by searching for the L2 norm square error evaluation result of each point of the discretized time grid based on the generative model includes: Based on the L2 norm square error evaluation results of each point in the discretized time grid by the generative model and the estimated first-order derivative of the noise scheduling function with respect to time, the minimum cumulative cost path from the starting point [0,0] to the end point [1,1] is determined by searching through a dynamic programming algorithm.

5. The molecule generation method based on multimodal optimal noise scheduling according to claim 4, characterized in that: The minimum cumulative cost path satisfies the following recursive relationship: in, represents the minimum cumulative cost, represents the continuous mode noise scheduling parameter, represents the discrete modal noise scheduling parameter, represents the step vector, represents the discretized noise level parameter, which is used to estimate the first-order derivative of the noise scheduling function with respect to time, represents the joint noise scheduling function, Indicates the instantaneous cost of the current grid point.

6. The molecule generation method based on multimodal optimal noise scheduling according to claim 1, characterized in that: The determining of an optimal joint noise schedule according to the minimum cumulative cost path, and generating a drug molecular structure based on the optimal joint noise schedule, includes: Mapping the minimum cumulative cost path to the joint noise scheduling function space to determine the optimal joint noise scheduling; Drug molecular structures are generated based on the optimal joint noise scheduling.

7. A molecular generation device based on multimodal optimal noise scheduling, characterized in that: include: A module is established for expressing any joint noise scheduling function in a predefined joint noise scheduling function space as a preset target form through a time reparameterization function, thereby establishing a unified parameterization framework for continuous mode noise and discrete mode noise; wherein the joint noise scheduling function space includes: a plurality of joint noise scheduling functions, each of the plurality of joint noise scheduling functions is composed of a continuous mode noise schedule and a discrete mode noise schedule; and the time reparameterization function includes: a first reparameterization function and a second reparameterization function, corresponding to the continuous mode noise and the discrete mode noise, respectively; A construction module is configured to minimize a pre-constructed generalized loss function through a pre-trained model, and train a generative model for evaluating any joint noise scheduling function in the joint noise scheduling function space based on the generalized loss function; wherein the generalized loss function is constructed based on the unified parameterized framework; A determination module, configured to search and determine a minimum cumulative cost path from a starting point to an end point through a dynamic programming algorithm based on an L2 norm square error evaluation result of each point in the discretized time grid according to the generative model; A generation module is used to determine an optimal joint noise schedule according to the minimum cumulative cost path, and generate a drug molecular structure based on the optimal joint noise schedule.

8. The molecule generation device based on multimodal optimal noise scheduling according to claim 7, characterized in that: The first parameterized function is: , the second parameterized function is: ; The time reparameterization function is used to express any joint noise scheduling function in the predefined joint noise scheduling function space as a preset target form, and a unified parameterization framework for continuous modal noise and discrete modal noise is established, which is achieved by the following formula: in, represents the joint noise scheduling function in the form of a preset target, Represents a predefined monotonic function.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the molecule generation method based on multimodal optimal noise scheduling according to any one of claims 1 to 6 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the molecule generation method based on multimodal optimal noise scheduling according to any one of claims 1 to 6 is implemented.

Images