A method and device for generating molecular side chains based on a double control diffusion model

Abstract

The application relates to an artificial intelligence technology and discloses a molecule side chain generation method and device based on a double control diffusion model, which comprises the following steps: obtaining instances containing a protein binding pocket, a molecule core skeleton, target affinity and a connection anchor point, and performing forward noise addition on atomic features and coordinates of a molecule side chain; inputting the noisy side chain and a conditional environment into a double control isometric denoising network; the double control isometric denoising network calculates confidence based on node features and distances to anchor points by using a confidence interaction control module, constructs an edge interaction gate to gate, filter and update feature propagation, and uses an affinity context control module to map time steps and affinity conditions into residual gate modulation parameters and differentially inject intermediate node features; a node representation output is predicted based on affinity modulation, noise is estimated by gradually sampling through a reverse distribution, a clean molecule side chain is estimated, and the clean molecule side chain is connected to an anchor point to reconstruct a complete candidate molecule. The application can improve the accuracy of molecule side chain generation.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method and apparatus for generating molecular side chains based on a dual-control diffusion model. Background Technology

[0002] In recent years, the deep integration of artificial intelligence technology with structural biology, medicinal chemistry, and computational chemistry has catalyzed the development of intelligent drug design. In drug discovery, lead compound optimization is a crucial step connecting the discovery of seed compounds with the screening of clinical candidates. Its goal is to enhance the binding affinity, druggability, and synthetic accessibility of the molecule with the target protein by rationally modifying side chain groups while preserving the activity of the molecular core skeleton. Therefore, constructing molecular side chain generation models that can simultaneously perceive the three-dimensional structure of the protein binding pocket, the molecular core skeleton, and binding affinity constraints has become an important research direction in the field of structural-based drug design.

[0003] In existing technologies, mainstream molecular side chain generation and lead compound optimization schemes typically employ skeletal decoration or diffusion generation paradigms. These methods generally use a given protein binding pocket and molecular core skeleton as conditions, generating the atom types and spatial coordinates of the molecular side chain to be optimized through recurrent neural networks, isotropic graphical neural networks, or diffusion models. Some three-dimensional diffusion methods, to accommodate molecular side chains of varying lengths, often introduce pseudo-atom mechanisms to fill the space with real atoms to a preset maximum number, and jointly predict the atom types and three-dimensional coordinates during the denoising process. Meanwhile, existing affinity-guided methods typically use attention-conditional injection or classifier-guided sampling to incorporate affinity information to bias noise prediction results or sampling directions.

[0004] Although the above methods have achieved certain results in structural molecular generation tasks, they have the following drawbacks: (1) Contextual representation is easily contaminated: Pseudo-atoms or low-confidence atoms introduced in existing diffusion models will still participate in message passing as active nodes during the denoising process. Their unreliable features may spread to the atoms of protein binding pockets, molecular core skeletons and real molecular side chains, resulting in contamination of the denoising contextual representation and weakening the model's ability to perceive the real molecular environment; (2) Insufficient affinity control: Existing affinity guidance strategies mostly introduce conditional information at the network end or sampling direction, making it difficult to continuously regulate the intermediate representation between molecular side chains and skeleton-pocket environment throughout the denoising process. Molecular binding affinity essentially depends on the multi-level interaction between generated side chains and surrounding protein residues and molecular core skeletons, which can easily lead to insufficient response of the model to the target affinity; (3) Limited ability to balance generation quality and properties: Due to the lack of fine control over the propagation of unreliable messages and affinity context, the model may generate invalid molecules, overly simple side chain replacements, or sacrifice molecular properties such as synthetic accessibility and drug-likeness while improving binding affinity. When dealing with lead compound optimization tasks that heavily rely on three-dimensional pocket fitting and fine interaction modeling, the model is prone to generating results that fail to meet the needs of actual drug design due to insufficient context modeling or inadequate affinity constraints.

[0005] In summary, existing technologies result in low accuracy in the generation of molecular side chains. Summary of the Invention

[0006] This invention provides a method and apparatus for generating molecular side chains based on a dual-controlled diffusion model, which can improve the accuracy of molecular side chain generation.

[0007] To achieve the above objectives, this invention provides a method for generating molecular side chains based on a dual-control diffusion model, comprising: Obtain an input instance containing a protein binding pocket, molecular core backbone, molecular side chains, and target affinity conditions, and extract the connection anchors connecting the molecular side chains to the molecular core backbone. The molecular side chain is transformed into a set containing atom type vectors and spatial coordinates, and a forward diffusion process is performed to gradually add Gaussian noise to the atom type vectors and spatial coordinates of the molecular side chain to obtain the noisy molecular side chain state. The noisy molecule side chain state, molecular core backbone, protein binding pocket, current diffusion time step, and target affinity conditions are input into a pre-constructed dual-control isotropic denoising network, which is internally composed of an alternating confidence interaction control module and an affinity context control module. Using the confidence interaction control module, the node confidence is calculated based on the characteristics of the atomic nodes of the dual-control isotropic denoising network and the spatial distance from the spatial coordinates of the atomic nodes to the connection anchors. The edge-level interaction gate is constructed based on the node confidence and multiplication gating is performed on the converged message features during the feature propagation process of the isotropic graph neural network. The intermediate node features and atomic coordinates filtered by the node confidence are updated. Using the affinity context control module, the diffusion time step and the target affinity condition are embedded and mapped to generate time step modulation parameters and affinity modulation parameters respectively. By setting differentiated affinity modulation intensity, the fused residual gated modulation parameters are calculated and injected into the intermediate node features to obtain the node representation modulated by affinity condition. The node representation of the output prediction noise is used by the affinity condition modulated. The state of the clean molecular side chain is estimated stepwise by sampling based on the prediction noise using the inverse transfer distribution to obtain the final molecular side chain structure. The molecular side chain structure is then connected to the connection anchor to reconstruct the complete candidate molecule.

[0008] To address the above problems, the present invention also provides a molecular side chain generation device based on a dual-controlled diffusion model, the device comprising: The forward noise-adding module is used to acquire an input instance containing protein binding pockets, molecular core backbone, molecular side chains, and target affinity conditions, and extract the connection anchors connecting the molecular side chains to the molecular core backbone; the molecular side chains are converted into a set containing atom type vectors and spatial coordinates, and a forward diffusion process is performed to gradually add Gaussian noise to the atom type vectors and spatial coordinates of the molecular side chains to obtain the noisy molecular side chain state. A dual-control denoising module is used to input the noisy molecule's side chain state, molecular core backbone, protein binding pocket, current diffusion time step, and target affinity condition into a pre-constructed dual-control equivariant denoising network. This network contains alternating series of confidence interaction control and affinity context control modules. The confidence interaction control module calculates node confidence based on the characteristics of the atomic nodes and the spatial distance from their spatial coordinates to the anchor points. It then constructs edge-level interaction gates based on these confidence levels and performs multiplicative gating on the converged message features during feature propagation in the equivariant graph neural network, updating the intermediate node features and atomic coordinates selected by node confidence. The affinity context control module performs embedding mapping on the diffusion time step and target affinity condition, generating time step modulation parameters and affinity modulation parameters. By setting differentiated affinity modulation intensities, it calculates the fused residual gating modulation parameters and injects them into the intermediate node features, obtaining node representations modulated by affinity conditions. The sampling and reconstruction module is used to output the prediction noise using the node representation modulated by affinity conditions. Based on the prediction noise, the state of the clean molecular side chain is estimated stepwise using the inverse transfer distribution to obtain the final generated molecular side chain structure. The molecular side chain structure is then connected to the connection anchor point to reconstruct the complete candidate molecule.

[0009] To address the above problems, the present invention also provides an electronic device, the electronic device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the molecular side chain generation method based on the dual-control diffusion model described above.

[0010] To address the aforementioned problems, the present invention also provides a computer-readable storage medium storing at least one computer program, which is executed by a processor in an electronic device to implement the molecular side chain generation method based on the dual-control diffusion model described above.

[0011] This invention extracts the connection anchor points between molecular side chains and the molecular core framework, ensuring that subsequently generated side chains can be seamlessly connected to specific positions in the framework. Furthermore, by transforming the molecular side chains into sets containing atomic type vectors and spatial coordinates, and performing a forward diffusion process to progressively add Gaussian noise to the atomic type vectors and spatial coordinates of the molecular side chains, a noisy molecular side chain state is obtained. Adding Gaussian noise only to the molecular side chains ensures that the molecular core framework and protein binding pocket remain in a pure state. Additionally, the noisy molecular side chain state, molecular core framework, protein binding pocket, current diffusion time step, and target affinity condition are input into a pre-constructed dual-control equivariant denoising network. This network alternately connects a confidence interaction control module and an affinity context control module. This alternating connection ensures that the model, in the same forward computation, can balance atomic-level information reliability with the integration of affinity conditions, achieving a unity of microscopic structural constraints and macroscopic property guidance. Moreover, edge-level interaction gates are constructed based on node confidence. Furthermore, during the feature propagation process of the isovariant graph neural network, edge interaction gates are used to multiply and gate the converged message features, updating the intermediate node features and atomic coordinates after node confidence screening. This further ensures that the molecular core backbone and protein binding pocket are in a pure state, allowing for effective characterization without contamination. Moreover, by setting differentiated affinity modulation intensities, the residual gating modulation parameters after fusion are calculated and injected into the intermediate node features, resulting in node representations modulated by affinity conditions. This not only adjusts the atomic representation of molecular side chains according to the target affinity but also synchronously adjusts the contextual representation of the molecular core backbone and protein binding pocket, thus achieving context-level affinity guidance. Finally, the node representations modulated by affinity conditions are used to output prediction noise. Based on the prediction noise, the state of clean molecular side chains is estimated stepwise using the inverse transition distribution, resulting in the final generated molecular side chain structure. Connecting the molecular side chain structure to the connection anchor point reconstructs the complete candidate molecule, improving the affinity matching degree between the generated molecular side chain and the protein binding pocket, thereby increasing the accuracy of molecular side chain generation. Attached Figure Description

[0012] Figure 1 This is a schematic flowchart of a molecular side chain generation method based on a dual-control diffusion model provided in an embodiment of the present invention. Figure 2 This is a schematic flowchart illustrating an example of a molecular side chain generation method based on a dual-control diffusion model provided in an embodiment of the present invention. Figure 3 A functional block diagram of a molecular side chain generation device based on a dual-control diffusion model provided in an embodiment of the present invention; Figure 4This is a schematic diagram of an electronic device for implementing a molecular side chain generation method based on a dual-control diffusion model, as provided in an embodiment of the present invention.

[0013] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0015] This application provides a molecular sidechain generation method based on a dual-control diffusion model. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the molecular sidechain generation method based on a dual-control diffusion model can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0016] Reference Figure 1 The diagram shown is a schematic flowchart of a molecular side chain generation method based on a dual-controlled diffusion model according to an embodiment of the present invention. In this embodiment, the molecular side chain generation method based on a dual-controlled diffusion model includes: S1. Obtain an input instance containing protein binding pockets, molecular core backbone, molecular side chains, and target affinity conditions, and extract the connection anchors connecting the molecular side chains to the molecular core backbone.

[0017] Understandably, a protein binding pocket refers to a three-dimensional spatial region on a target protein that specifically interacts with ligand molecules. The protein binding pocket serves as a fixed context environment that constrains the generation of side chains, and its structure is characterized by the types of heavy atoms it contains and their set of spatial coordinates.

[0018] Understandably, the molecular core framework refers to the basic compound structure in the drug lead compound whose atomic type and three-dimensional spatial conformation remain fixed throughout the entire side chain optimization and diffusion noise reduction process. It provides a basic attachment platform and steric constraint for the molecular side chains to be generated. Here, conformation refers to the different spatial arrangements in an organic compound molecule caused by the change in the relative positions of atoms (or groups of atoms) bonded to carbon atoms.

[0019] Understandably, a molecular side chain refers to a substituent group to be generated or optimized. It is generated based on the context of the protein binding pocket and the molecular core skeleton, as well as the guidance of the target affinity conditions. In the data space, it is characterized as a set of atomic type features and spatial coordinates containing real atoms and pseudo atoms.

[0020] Understandably, the target affinity condition refers to the constraint input on the binding strength between the desired complete candidate molecule and the protein binding pocket. It is continuously injected in a layer-by-layer feature modulation manner during the denoising process of the diffusion model to guide the generation direction of the molecular side chain to meet the expected drug-efficacy binding force.

[0021] Understandably, the connection anchor point refers to a specific atomic node or chemical bond breaking site located on the molecular core skeleton in advance. It serves as a binding hub between the molecular core skeleton and the molecular side chain, enabling the generated molecular side chain to be accurately connected to the molecular core skeleton to reconstruct a complete candidate molecular structure.

[0022] For example, an input instance containing a protein binding pocket, a molecular core backbone, molecular side chains, and a target affinity condition is obtained, where the input instance is (p, s, a, r), where (p) is the protein binding pocket, (s) is the molecular core backbone, (a) is the target affinity condition, and (r) is the molecular side chain to be generated or realistically labeled, and a dataset for model training is constructed. The protein binding pocket, molecular core backbone, and molecular side chains are all represented by heavy atom types and their spatial coordinates, and the generated molecular side chains are... .

[0023] For example, the extraction of molecular side chains and their connection anchors to the molecular core backbone is achieved by the following steps: based on a given protein-ligand complex, the protein binding pocket region, the molecular core backbone, and the molecular side chain connection anchors are determined, so that the generated molecular side chains can be connected to the predefined anchors while keeping the molecular core backbone unchanged.

[0024] S2. Transform the molecular side chain into a set containing atom type vectors and spatial coordinates, and perform a forward diffusion process to gradually add Gaussian noise to the atom type vectors and spatial coordinates of the molecular side chain to obtain the noisy molecular side chain state.

[0025] It is understood that an atom type vector refers to the feature code used to mathematically characterize the chemical element type and properties of each atom in the molecular side chain. In this embodiment of the invention, it is constructed in the form of a one-hot vector.

[0026] Understandably, the forward diffusion process refers to a Markov transfer process that injects random perturbations into the original, clear molecular side chain data in a discrete time step sequence. In this embodiment of the invention, the forward diffusion process is strictly limited to acting only on the atom type vectors and spatial coordinates of the molecular side chains, while keeping the structural state of the protein binding pocket and the molecular core backbone completely static, so as to avoid the context structure being destroyed during the forward diffusion process.

[0027] Understandably, Gaussian noise refers to a random disturbance signal that follows a preset mean and covariance distribution law. It follows a standard normal distribution and is superimposed on the atomic characteristics and spatial coordinates of the molecular side chain to mask the original true chemical properties and geometric structure information of the molecular side chain.

[0028] Understandably, the noisy molecular side chain state refers to the intermediate mixed state that the molecular side chain exhibits after a specific time step in the forward diffusion process, which combines the residual real structural signal with the injected Gaussian noise component.

[0029] Specifically, the step of converting the molecular side chain into a set containing atom type vectors and spatial coordinates includes: the atom type vectors are encoded using one-hot vectors, and the spatial coordinates are represented based on continuous three-dimensional coordinate system spatial vectors.

[0030] For example, the molecular side chain can be transformed into a set containing atom type vectors and spatial coordinates, and a forward diffusion process can be performed to progressively add Gaussian noise to the atom type vectors and spatial coordinates of the molecular side chain to obtain a noisy molecular side chain state. This can be achieved using the following implementation steps: Step 1: Molecular Side Chain Representation Construction: The molecular side chain is represented as a set of atom types and three-dimensional coordinates, where atom types are represented by one-hot vectors and coordinates are represented by three-dimensional Euclidean space coordinates. To support the generation of molecular side chains of different lengths, real atoms are filled to a preset maximum length, and pseudo-atoms are introduced for adaptive side chain length generation. Step 2, Molecular Side Chain Forward Diffusion Noise Addition: The diffusion process only acts on the molecular side chains, without adding noise to the protein binding pocket and the molecular core backbone. Let... Indicates the first The noisy molecular side chain states at each diffusion time step, among which Indicates the first Noisy atom type representation of each side chain atom, Indicates the first Noisy 3D coordinates of each side chain atom This indicates the number of atoms in the molecular side chains. The forward diffusion process is defined as: ; ; in, Indicates the time step from the previous diffusion step Transfer to the current diffusion time step The conditional probability distribution; This represents the state of the original, unnoisy molecular side chains. Directly obtain the first Noisy state at each diffusion time step ( The conditional probability distribution of ). The mean is Covariance is Gaussian distribution; This represents the initial state of the actual molecular side chain. , This represents the actual molecular side chain to be optimized or generated. Indicate diffusion time step, , This indicates the preset maximum number of diffusion steps.

[0031] ; ; in, Indicates the first The noise intensity coefficient for each diffusion time step is used to control the magnitude of noise added to the molecular side chain state at the current time step. This represents a preset noise scheduling sequence, whose function is to control the gradual increase of noise over time during the diffusion process; Indicates the first The single-step signal retention coefficient of each diffusion time step is used to represent the proportion of effective molecular side chain structure information retained from the previous time step state; Indicates from the first The time step to the The cumulative signal retention coefficient at each time step is used to measure the signal retention after [a certain number of time steps]. The overall proportion in which the original molecular side chain structure information is still preserved after noise addition; This represents the index of the intermediate time step in the cumulative multiplication operation; Represents the identity matrix, whose dimensions are 1 and 2. The feature dimensions are consistent, which is used to indicate that independent Gaussian noise is added to each feature dimension.

[0032] because During training, the noisy molecular side chain state at time step (t) can be directly sampled using reparameterization: ; in, This represents random noise sampled from a standard Gaussian distribution, used to simulate the perturbation of molecular side chain atom types and three-dimensional coordinates during diffusion. This represents the original molecular side chain structure information that is still retained at time step (t); This represents the noise information added at the (t)th time step. It is understood that, in this embodiment of the invention, as the time step... The increase, As the noise gradually decreases, the original molecular side chain information is gradually weakened, while the noise component gradually increases, thus achieving a gradual addition of noise to the molecular side chain state. The protein binding pocket and the molecular core backbone do not participate in the addition of noise in this process, but rather serve as structural conditions that constrain the generation direction of molecular side chains during the subsequent denoising process.

[0033] S3. Input the noisy molecule side chain state, molecular core backbone, protein binding pocket, current diffusion time step, and target affinity condition into a pre-constructed dual-control isotropic denoising network. The dual-control isotropic denoising network has a confidence interaction control module and an affinity context control module alternately connected in series.

[0034] Understandably, the pre-built dual-control equivariant denoising network refers to the core deep learning model used to predict and remove random noise contained in the noisy molecular side chain state during the reverse generation stage. The pre-built dual-control equivariant denoising network is built on the basis of graph neural networks that satisfy the equivariant property of three-dimensional space. Its internal architecture integrates a confidence gating mechanism for the message passing path of the graph structure and a conditional affine modulation mechanism for the node feature space, so that the model can achieve bidirectional constraints on the molecular side chain structure generation trajectory while maintaining the equivariance of spatial geometry.

[0035] Understandably, the confidence interaction control module refers to the computational unit within the denoising network responsible for monitoring the reliability of feature propagation between nodes. Specifically, it calculates the true atom confidence score by evaluating the features of the current atom node and its spatial distance relative to the connection anchor point, and constructs a multiplication interaction gate tensor acting on the directed edges accordingly. This allows for dynamic truncation or allowance of the transmitted message flow during the local neighbor feature aggregation phase of the network, thereby preventing pseudo-atoms or low-confidence atoms from outputting invalid features to reliable context nodes.

[0036] Understandably, the affinity context control module refers to the feature modulation unit within the denoising network responsible for fusing macroscopic attribute targets into the feature representations of microscopic graph nodes. It utilizes diffusion time-step embedding and affinity conditional embedding to generate scaling and translation parameters for the affine transformation. Simultaneously, it assigns differentiated modulation intensities to molecular side-chain atoms and context atoms using a mask matrix. Subsequently, through a residual gating mechanism, the fused feature modulation parameters are successively injected into the node representations of each hidden layer of the network, ensuring that the denoising network receives a unified affinity conditional constraint signal throughout the entire iteration cycle.

[0037] For example, the construction process of the pre-built dual-control equivariant denoising network is as follows: In the state of noisy molecular side chains Molecular core backbone (s), protein binding pocket (p), diffusion time step (t), and affinity conditions. As input, an equivariant graph neural network is used to predict the noise introduced during the diffusion process. The formula is as follows: in, It is a dual-control, equal-variable denoising network.

[0038] S4. Using the confidence interaction control module, the node confidence is calculated based on the characteristics of the atomic nodes of the dual-control equivariant denoising network and the spatial distance from the spatial coordinates of the atomic nodes to the connecting anchor points. The edge-level interaction gate is constructed based on the node confidence. During the feature propagation process of the equivariant graph neural network, the edge interaction gate is used to perform multiplication gating on the converged message features, and the intermediate node features and atomic coordinates filtered by the node confidence are updated.

[0039] Understandably, edge-level interaction gates refer to dynamically adjusted weight matrices or filtering coefficients defined on directed edges between nodes in the message passing graph structure of an equivariant graph neural network. They are calculated based on the combined node confidence of both the sending and receiving nodes, acting as a multiplicative switch or attenuation filter for message flow at the algorithmic level. Their core physical function is to prevent the propagation of invalid, contaminating features—including pseudo-atoms and low-confidence atoms—to the reliable context environment during local feature aggregation and spatial coordinate aggregation at nodes, thereby ensuring the purity and reliability of the overall denoised context space representation.

[0040] Specifically, node confidence is calculated based on the characteristics of atomic nodes in the dual-control isotropic denoising network and the spatial distance from the spatial coordinates of the atomic nodes to the connecting anchor points. Then, edge-level interaction gates are constructed based on the node confidence, including: The distance from the atomic node to the connecting anchor point is encoded by the radial basis function and concatenated with the feature of the atomic node. This concatenation is then input into a learnable gating network to obtain the true atomic confidence score of the atomic node. By combining the mask matrix of molecular side chains and the mask matrix of node validity, the true atom confidence score is mapped to the final node confidence score. Extract the final confidence scores of the information receiving node and the information sending node, and take the intersection of the complements of the final confidence scores of the information receiving node and the information sending node to generate the edge-level interaction gate of the corresponding directed edge.

[0041] Understandably, radial basis function encoding refers to using radial basis function encoding operators to map the calculated continuous scalar distance into a high-dimensional discrete space embedded feature vector, enabling neural networks to capture minute differences in physical distance.

[0042] Furthermore, during the feature propagation process of the isovariant graph neural network, edge interaction gates are used to perform multiplication gating on the converged message features, updating the intermediate node features and atomic coordinates that have been filtered by node confidence, including: The gated message features are obtained by multiplying the original message features between the edge-level interaction gate and the node bitwise, and the updated node features are obtained by summing the current node features with the gated message features gathered from the neighboring nodes. The equivalent relative coordinate offset is obtained by multiplying the same side-level interaction gate with the relative coordinate difference and the coordinate update weight. The equivalent relative coordinate offset is then filtered according to the mask matrix of the molecular side chain. The filtered equivalent relative coordinate offset is then added to the spatial coordinates of the current node, and the updated atomic coordinates are output.

[0043] For example, using the confidence interaction control module, the node confidence is calculated based on the features of the atomic nodes of the dual-control isotropic denoising network and the spatial distance from the spatial coordinates of the atomic nodes to the connection anchors. An edge-level interaction gate is then constructed based on the node confidence. During the feature propagation process of the isotropic graph neural network, the edge interaction gate is used to perform multiplication gating on the converged message features, updating the intermediate node features and atomic coordinates filtered by node confidence. This is achieved through the following implementation steps: Step 1, Confidence Interaction Control: For each atomic node in the l-th layer of the equivariant graph neural network, based on the current node features... The node confidence is estimated by its distance from the anchor point. Assuming the molecular system is centered at the anchor point, then the node... relative distance of link anchor points for: ;in, The three-dimensional coordinates of the current node. It is the second norm of Euclidean space.

[0044] The distance is encoded using a radial basis function and concatenated with node features, then the confidence score is obtained through a learnable gated network. ; in, Represents a node The base confidence score for real atoms. This represents the Sigmoid function. Indicates the first Layered learnable gating networks, For radial basis function encoding operators, These are the node characteristics of the information receiving node.

[0045] Step 2, Node Confidence and Edge Interaction Gate Construction: Let... For the molecular side chain mask, If the node's validity mask is used, then the node's final confidence level is... Defined as: ; For directed edges ,in As an information receiving node, Construct an edge interaction gate for the information sending node. : ;in, This represents the final confidence level of the information sending node.

[0046] This edge interaction gate is used to suppress the propagation of unreliable information from low-confidence atoms or pseudo-atoms to high-confidence nodes, while preserving information exchange between reliable atoms.

[0047] Step 3, Gated Message Propagation and Coordinate Update: During feature propagation, the message strength is adjusted using an edge interaction gate to obtain the gated message. : ;in, Representing edge attribute features, representing the node receiving connection information. With information sending node Vectorized representation of the physicochemical connections between them. For message generation networks, The node characteristics of the information sending node; And update node characteristics: ; in, Representing edge attribute characteristics, Indicates the information receiving node The neighborhood group, To update the network for features, For information receiving nodes Updated node characteristics.

[0048] Furthermore, the same interaction gate is used for isotropic coordinate updates: ;in, For information sending nodes Updated node features These are spatial relative direction vectors. For isotropic coordinate step size mapping, ;in, For the displacement increment of the isotropic coordinates, These are the updated atomic coordinates.

[0049] Step 3 above allows the atomic coordinates of the molecular side chains to be gradually denoised based on reliable context while maintaining isovariability.

[0050] S5. Using the affinity context control module, embedding mapping is performed on the diffusion time step and the target affinity condition to generate time step modulation parameters and affinity modulation parameters respectively. By setting differentiated affinity modulation intensity, the fused residual gated modulation parameters are calculated and injected into the intermediate node features to obtain the node representation modulated by affinity conditions.

[0051] Understandably, the time-step modulation parameter refers to the dynamic feature adjustment coefficient generated by calculating the global time scale representing the current denoising or denoising progress through specific embedding network mapping and affine transformation. Its core function is to enable each layer of the isovariant denoising graph neural network to adaptively adjust the activation state of node features and denoising intensity according to the severity of the disturbance and damage to the current molecular side chain structure.

[0052] Understandably, affinity modulation parameters refer to the affine bias coefficients in the feature space generated by mapping the target value of the macroscopic biochemical binding strength between the desired candidate molecule side chain and the target protein binding pocket through a dedicated conditional coding network. Their core function is to apply mathematical fine-tuning—multiplicative scaling and additive translation—to the hidden features of microscopic atoms, thereby progressively and continuously embedding external affinity constraint signals into the representation space of the molecule's side chain and context. This forces the denoising network to strictly follow an optimized trajectory to enhance the target binding force when reconstructing the chemical properties and three-dimensional geometric conformation of atoms.

[0053] Specifically, by setting differentiated affinity modulation intensities, the fused residual gated modulation parameters are calculated, including: The diffusion time step and target affinity are encoded using time step embedding function and affinity embedding function, respectively; The encoded embedded features are input into independent time step generators and affinity generators, which output diffusion stage modulation parameters and target affinity modulation parameters, respectively.

[0054] Furthermore, the residual gate modulation parameters are injected into the intermediate node features to obtain the node representation after affinity condition modulation, including: Molecular side chain masks are used to distinguish molecular side chain atomic nodes from framework pocket context atomic nodes, and independent learnable affinity modulation coefficients are assigned to the two types of nodes. Based on the affinity modulation coefficient of the corresponding node, the time step modulation parameter and the affinity modulation parameter are linearly superimposed, and the feature linear affine transformation is performed on the intermediate node features after layer normalization. The intermediate node features, time step embeddings, and affinity embeddings before the affine transformation are concatenated. An activation function is then used to construct a node-level feature gating. The node-level feature gating is then combined with the affine transformation result through a Hadamard product. Finally, the residual connection method is used to superimpose the node representation onto the initial intermediate node features to obtain the node representation modulated by affinity conditions.

[0055] For example, using the affinity context control module, embedding mapping is performed on the diffusion time step and the target affinity condition to generate time step modulation parameters and affinity modulation parameters respectively; by setting differentiated affinity modulation intensities, the fused residual gated modulation parameters are calculated, and the residual gated modulation parameters are injected into the intermediate node features to obtain the node representation modulated by the affinity condition. This is achieved using the following implementation steps: Step 1, Affinity Context Control: To continuously inject target affinity conditions during the denoising process, the diffusion time step is controlled. Affinity with the target By embedding, we get: ; in, For time step embedding functions, For affinity embedding function, Embedded for time steps, Embedded for affinity.

[0056] Step 2, FiLM modulation parameter generation: Two independent FiLM generators are used to generate modulation parameters based on time step embedding and affinity embedding, respectively. : ;in, This is a time-step FiLM generator.

[0057] ;in, For affinity FiLM generator.

[0058] in, Used to generate diffusion time step modulation parameters Used to generate target affinity modulation parameters.

[0059] Step 3: Calculation of Differential Affinity Modulation Intensity: Since the atoms of the molecular side chain and the atoms of the framework-pocket context contribute differently to the affinity, a learnable modulation intensity parameter is set to give the atoms of the molecular side chain and the context atoms different affinity modulation intensities. : ; in, For molecular side chain masking, This indicates the affinity modulation intensity of atoms in the molecular side chain. This indicates the affinity modulation intensity of the skeleton-pocket context atoms.

[0060] Step 4, Residual Gated FiLM Modulation: The time-step modulation parameters and affinity modulation parameters are fused to obtain: ;in, This is for comprehensive scaling parameters.

[0061] ;in, For comprehensive translation parameters.

[0062] After performing layer normalization on the node representation, the FiLM update amount is calculated. : ;in, The node characteristics of the information receiving node. This refers to the normalized node characteristics of the information receiving node. Further construct node-level feature gating : ; Finally, the node representation after affinity conditional modulation is obtained. : ; The modulated node representations described above serve as input to the subsequent equivariant message passing module, enabling the model to continuously perceive the target affinity condition throughout the denoising process.

[0063] S6. The node representation of the output prediction noise is modulated by affinity conditions. Based on the prediction noise, the state of the clean molecular side chain is estimated by stepwise sampling using the inverse transfer distribution to obtain the final generated molecular side chain structure. The molecular side chain structure is then connected to the connection anchor point to reconstruct the complete candidate molecule.

[0064] Understandably, reverse transition distribution stepwise sampling refers to the iterative deduction process of extracting state samples from the conditional probability model in reverse time order, based on the Markov chain principle, during the reverse decoding stage of the generative diffusion model.

[0065] Specifically, the state of the clean molecular side chains is estimated stepwise by sampling based on the predicted noise using the inverse transition distribution, including: During the model training phase, a classifier free guidance strategy is adopted, and the network parameters are updated by forcibly replacing part of the real target affinity conditions in the input with empty conditional embedding vectors according to the preset discard probability. At any time step in the denoising generation stage, forward inference is performed by inputting the target affinity condition and the empty condition embedding vector respectively to obtain the conditional noise prediction component and the unconditional noise prediction component. The difference extension term between the conditional noise prediction component and the unconditional noise prediction component is calculated based on the preset guidance intensity hyperparameter. The difference extension term is then compensated into the unconditional noise prediction component to generate the final guidance noise. The state of the clean molecular side chain is estimated using the preset sampling formula.

[0066] Understandably, the state of a clean molecular side chain refers to the zero-noise molecular side chain estimated structure obtained by directly stripping away the noise component within any reverse diffusion time step in the denoising generation stage.

[0067] For example, the predicted noise is represented by nodes modulated by affinity conditions. The state of the clean molecular side chains is estimated stepwise using the back-transfer distribution based on the predicted noise to obtain the final molecular side chain structure. The molecular side chain structure is then connected to the connection anchors to reconstruct the complete candidate molecule. This can be achieved using the following implementation steps: Step 1: Training the Dual-Controlled Isovariant Denoising Network: During training, the dual-controlled isovariant denoising network is trained with real noise. To supervise the signal, the loss function is defined as minimizing the mean square error loss between the predicted noise and the actual noise, and is further defined as follows: : ; in, For mathematical expectation operators, To predict noise, The term represents the trainable parameters in a dual-control isovariant denoising network. By minimizing this loss function, the model learns to recover the atom types and three-dimensional spatial coordinates of the true molecular side chains from a noisy molecular side chain representation, given the protein binding pocket, molecular core backbone, and affinity. This is the characteristic error sum-of-squares operator.

[0068] Step 2, Classifier-guided sampling: During the generation phase, conditional noise predictions are calculated separately. And unconditional noise prediction : ,in, Embedded with real conditions; ,in, Null conditional embedding; And according to the guiding strength Affinity-guided noise prediction : ; in, Used to control the guiding strength of target affinity conditions on the generation process.

[0069] Step 4: Reconstruction of complete candidate molecule side chains: Estimate the state of clean molecule side chains based on affinity-guided noise prediction and sample step by step until complete candidate molecule side chains are obtained. : ,in, The signal recovery amplification factor is a global amplification compensation coefficient that is forcibly applied to compensate for the loss of characteristic scales due to irreversible signal attenuation during the initial noise diffusion process.

[0070] This invention extracts the connection anchor points between molecular side chains and the molecular core framework, ensuring that subsequently generated side chains can be seamlessly connected to specific positions in the framework. Furthermore, by transforming the molecular side chains into sets containing atomic type vectors and spatial coordinates, and performing a forward diffusion process to progressively add Gaussian noise to the atomic type vectors and spatial coordinates of the molecular side chains, a noisy molecular side chain state is obtained. Adding Gaussian noise only to the molecular side chains ensures that the molecular core framework and protein binding pocket remain in a pure state. Additionally, the noisy molecular side chain state, molecular core framework, protein binding pocket, current diffusion time step, and target affinity condition are input into a pre-constructed dual-control equivariant denoising network. This network alternately connects a confidence interaction control module and an affinity context control module. This alternating connection ensures that the model, in the same forward computation, can balance atomic-level information reliability with the integration of affinity conditions, achieving a unity of microscopic structural constraints and macroscopic property guidance. Moreover, edge-level interaction gates are constructed based on node confidence. Furthermore, during the feature propagation process of the isovariant graph neural network, edge interaction gates are used to multiply and gate the converged message features, updating the intermediate node features and atomic coordinates after node confidence screening. This further ensures that the molecular core backbone and protein binding pocket are in a pure state, allowing for effective characterization without contamination. Moreover, by setting differentiated affinity modulation intensities, the residual gating modulation parameters after fusion are calculated and injected into the intermediate node features, resulting in node representations modulated by affinity conditions. This not only adjusts the atomic representation of molecular side chains according to the target affinity but also synchronously adjusts the contextual representation of the molecular core backbone and protein binding pocket, thus achieving context-level affinity guidance. Finally, the node representations modulated by affinity conditions are used to output prediction noise. Based on the prediction noise, the state of clean molecular side chains is estimated stepwise using the inverse transition distribution, resulting in the final generated molecular side chain structure. Connecting the molecular side chain structure to the connection anchor point reconstructs the complete candidate molecule, improving the affinity matching degree between the generated molecular side chain and the protein binding pocket, thereby increasing the accuracy of molecular side chain generation.

[0071] Reference Figure 2 The diagram shown is a schematic flowchart of an example of a molecular side chain generation method based on a dual-control diffusion model provided in an embodiment of the present invention.

[0072] like Figure 3 The diagram shown is a functional block diagram of a molecular side chain generation device based on a dual-control diffusion model provided in an embodiment of the present invention.

[0073] The molecular side chain generation device 100 based on a dual-control diffusion model described in this invention can be installed in an electronic device. Depending on the functions implemented, the molecular side chain generation device 100 based on a dual-control diffusion model may include a forward noise addition module 101, a dual-control noise reduction module 102, and a sampling and reconstruction module 103.

[0074] The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0075] In this embodiment, the functions of each module / unit are as follows: The forward noise module 101 is used to acquire an input instance containing a protein binding pocket, a molecular core backbone, molecular side chains, and target affinity conditions, and extract the connection anchor points connecting the molecular side chains to the molecular core backbone; convert the molecular side chains into a set containing atom type vectors and spatial coordinates, and perform a forward diffusion process to gradually add Gaussian noise to the atom type vectors and spatial coordinates of the molecular side chains to obtain a noisy molecular side chain state.

[0076] The dual-control denoising module 102 is used to input the noisy molecule side chain state, molecular core backbone, protein binding pocket, current diffusion time step, and target affinity condition into a pre-constructed dual-control equivariant denoising network. The dual-control equivariant denoising network contains an alternating series of a confidence interaction control module and an affinity context control module. Using the confidence interaction control module, node confidence is calculated based on the characteristics of the atomic nodes in the dual-control equivariant denoising network and the spatial distance from the spatial coordinates of the atomic nodes to the connecting anchor points. An edge-level interaction gate is constructed based on the node confidence, and during the feature propagation process of the equivariant graph neural network, the edge interaction gate is used to perform multiplication gating on the converged message features, updating the intermediate node features and atomic coordinates filtered by node confidence. Using the affinity context control module, embedding mapping is performed on the diffusion time step and target affinity condition respectively, generating time step modulation parameters and affinity modulation parameters. By setting differentiated affinity modulation intensities, the fused residual gating modulation parameters are calculated and injected into the intermediate node features to obtain the node representation modulated by the affinity condition.

[0077] The sampling and reconstruction module 103 is used to output prediction noise using node representations modulated by affinity conditions, estimate the state of clean molecular side chains by gradually sampling based on the prediction noise using the inverse transfer distribution, obtain the final generated molecular side chain structure, and connect the molecular side chain structure to the connection anchor point to reconstruct the complete candidate molecule.

[0078] like Figure 4The diagram shown is a schematic representation of an electronic device that implements a molecular side chain generation method based on a dual-control diffusion model, according to an embodiment of the present invention.

[0079] The electronic device may include a processor 10, a memory 11, a communication bus 12, and a communication interface 13. It may also include a computer program stored in the memory 11 and capable of running on the processor 10, such as a molecular side chain generation method program based on a dual-control diffusion model.

[0080] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 (e.g., executing a molecular side-chain generation method program based on a dual-control diffusion model), and calls data stored in the memory 11 to perform various functions of the electronic device and process data.

[0081] The memory 11 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device, such as a plug-in portable hard drive, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc. Furthermore, the memory 11 can include both internal and external storage units of the electronic device. The memory 11 can be used not only to store application software and various types of data installed on the electronic device, such as the code of a molecular side chain generation method program based on a dual-control diffusion model, but also to temporarily store data that has been output or will be output.

[0082] The communication bus 12 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. The bus is configured to enable communication between the memory 11 and at least one processor 10, etc.

[0083] The communication interface 13 is used for communication between the aforementioned electronic device and other devices, including a network interface and a user interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, Bluetooth interface, etc.), typically used to establish communication connections between the electronic device and other electronic devices. The user interface may be a display, an input unit (such as a keyboard), or optionally, a standard wired or wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device and to display a visual user interface.

[0084] Figure 4 Only electronic devices with components are shown; those skilled in the art will understand that... Figure 4 The structure shown does not constitute a limitation on the electronic device and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0085] For example, although not shown, the electronic device may also include a power supply (such as a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 10 through a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0086] It should be understood that the embodiments described are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0087] The program for a molecular side chain generation method based on a dual-control diffusion model, stored in the memory 11 of the electronic device, is a combination of multiple instructions. When run in the processor 10, it can achieve the following: Obtain an input instance containing a protein binding pocket, molecular core backbone, molecular side chains, and target affinity conditions, and extract the connection anchors connecting the molecular side chains to the molecular core backbone. The molecular side chain is transformed into a set containing atom type vectors and spatial coordinates, and a forward diffusion process is performed to gradually add Gaussian noise to the atom type vectors and spatial coordinates of the molecular side chain to obtain the noisy molecular side chain state. The noisy molecule side chain state, molecular core backbone, protein binding pocket, current diffusion time step, and target affinity conditions are input into a pre-constructed dual-control isotropic denoising network, which is internally composed of an alternating confidence interaction control module and an affinity context control module. Using the confidence interaction control module, the node confidence is calculated based on the characteristics of the atomic nodes of the dual-control isotropic denoising network and the spatial distance from the spatial coordinates of the atomic nodes to the connection anchors. The edge-level interaction gate is constructed based on the node confidence and multiplication gating is performed on the converged message features during the feature propagation process of the isotropic graph neural network. The intermediate node features and atomic coordinates filtered by the node confidence are updated. Using the affinity context control module, the diffusion time step and the target affinity condition are embedded and mapped to generate time step modulation parameters and affinity modulation parameters respectively. By setting differentiated affinity modulation intensity, the fused residual gated modulation parameters are calculated and injected into the intermediate node features to obtain the node representation modulated by affinity condition. The node representation of the output prediction noise is used by the affinity condition modulated. The state of the clean molecular side chain is estimated stepwise by sampling based on the prediction noise using the inverse transfer distribution to obtain the final molecular side chain structure. The molecular side chain structure is then connected to the connection anchor to reconstruct the complete candidate molecule.

[0088] Specifically, the specific implementation method of the processor 10 for the above instructions can be referred to the description of the relevant steps in the corresponding embodiment of the accompanying drawings, and will not be repeated here.

[0089] Furthermore, if the modules / units integrated in the electronic device 1 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0090] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor of an electronic device, can perform the following: Obtain an input instance containing a protein binding pocket, molecular core backbone, molecular side chains, and target affinity conditions, and extract the connection anchors connecting the molecular side chains to the molecular core backbone. The molecular side chain is transformed into a set containing atom type vectors and spatial coordinates, and a forward diffusion process is performed to gradually add Gaussian noise to the atom type vectors and spatial coordinates of the molecular side chain to obtain the noisy molecular side chain state. The noisy molecule side chain state, molecular core backbone, protein binding pocket, current diffusion time step, and target affinity conditions are input into a pre-constructed dual-control isotropic denoising network, which is internally composed of an alternating confidence interaction control module and an affinity context control module. Using the confidence interaction control module, the node confidence is calculated based on the characteristics of the atomic nodes of the dual-control isotropic denoising network and the spatial distance from the spatial coordinates of the atomic nodes to the connection anchors. The edge-level interaction gate is constructed based on the node confidence and multiplication gating is performed on the converged message features during the feature propagation process of the isotropic graph neural network. The intermediate node features and atomic coordinates filtered by the node confidence are updated. Using the affinity context control module, the diffusion time step and the target affinity condition are embedded and mapped to generate time step modulation parameters and affinity modulation parameters respectively. By setting differentiated affinity modulation intensity, the fused residual gated modulation parameters are calculated and injected into the intermediate node features to obtain the node representation modulated by affinity condition. The node representation of the output prediction noise is used by the affinity condition modulated. The state of the clean molecular side chain is estimated stepwise by sampling based on the prediction noise using the inverse transfer distribution to obtain the final molecular side chain structure. The molecular side chain structure is then connected to the connection anchor to reconstruct the complete candidate molecule.

[0091] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0092] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0093] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0094] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0095] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0096] The blockchain referred to in this invention is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0097] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0098] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a system claim may also be implemented by a single unit or device through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for generating molecular side chains based on a dual-control diffusion model, characterized in that, The method includes: Obtain an input instance containing a protein binding pocket, molecular core backbone, molecular side chains, and target affinity conditions, and extract the connection anchors connecting the molecular side chains to the molecular core backbone. The molecular side chain is transformed into a set containing atom type vectors and spatial coordinates, and a forward diffusion process is performed to gradually add Gaussian noise to the atom type vectors and spatial coordinates of the molecular side chain to obtain the noisy molecular side chain state. The noisy molecule side chain state, molecular core backbone, protein binding pocket, current diffusion time step, and target affinity conditions are input into a pre-constructed dual-control isotropic denoising network, which is internally composed of an alternating confidence interaction control module and an affinity context control module. Using the confidence interaction control module, the node confidence is calculated based on the characteristics of the atomic nodes of the dual-control isotropic denoising network and the spatial distance from the spatial coordinates of the atomic nodes to the connection anchors. The edge-level interaction gate is constructed based on the node confidence and multiplication gating is performed on the converged message features during the feature propagation process of the isotropic graph neural network. The intermediate node features and atomic coordinates filtered by the node confidence are updated. Using the affinity context control module, the diffusion time step and the target affinity condition are embedded and mapped to generate time step modulation parameters and affinity modulation parameters respectively. By setting differentiated affinity modulation intensity, the fused residual gated modulation parameters are calculated and injected into the intermediate node features to obtain the node representation modulated by affinity condition. The node representation of the output prediction noise is used by the affinity condition modulated. The state of the clean molecular side chain is estimated stepwise by sampling based on the prediction noise using the inverse transfer distribution to obtain the final molecular side chain structure. The molecular side chain structure is then connected to the connection anchor to reconstruct the complete candidate molecule.

2. The molecular side chain generation method based on a dual-control diffusion model as described in claim 1, characterized in that, The step of converting the molecular side chain into a set containing atomic type vectors and spatial coordinates includes: the atomic type vectors are encoded using one-hot vectors, and the spatial coordinates are represented based on spatial vectors in a continuous three-dimensional coordinate system.

3. The molecular side chain generation method based on the dual-control diffusion model as described in claim 1, characterized in that, The step of calculating node confidence based on the characteristics of atomic nodes in the dual-control isotropic denoising network and the spatial distance from the spatial coordinates of the atomic nodes to the connecting anchor points, and constructing edge-level interaction gates based on the node confidence, includes: The distance from the atomic node to the connecting anchor point is encoded by the radial basis function and concatenated with the feature of the atomic node. This concatenation is then input into a learnable gating network to obtain the true atomic confidence score of the atomic node. By combining the mask matrix of molecular side chains and the mask matrix of node validity, the true atom confidence score is mapped to the final node confidence score. Extract the final confidence scores of the information receiving node and the information sending node, and take the intersection of the complements of the final confidence scores of the information receiving node and the information sending node to generate the edge-level interaction gate of the corresponding directed edge.

4. The molecular side chain generation method based on the dual-control diffusion model as described in claim 1, characterized in that, The process of feature propagation in the isotropic graph neural network utilizes edge interaction gates to perform multiplication gating on the converged message features, updating the intermediate node features and atomic coordinates after node confidence screening, including: The gated message features are obtained by multiplying the original message features between the edge-level interaction gate and the node bitwise, and the updated node features are obtained by summing the current node features with the gated message features gathered from the neighboring nodes. The equivalent relative coordinate offset is obtained by multiplying the same side-level interaction gate with the relative coordinate difference and the coordinate update weight. The equivalent relative coordinate offset is then filtered according to the mask matrix of the molecular side chain. The filtered equivalent relative coordinate offset is then added to the spatial coordinates of the current node, and the updated atomic coordinates are output.

5. The molecular side chain generation method based on the dual-control diffusion model as described in claim 1, characterized in that, The step of calculating the fused residual gated modulation parameters by setting differentiated affinity modulation intensities includes: The diffusion time step and target affinity are encoded using time step embedding function and affinity embedding function, respectively; The encoded embedded features are input into independent time step generators and affinity generators, which output diffusion stage modulation parameters and target affinity modulation parameters, respectively.

6. The molecular side chain generation method based on the dual-control diffusion model as described in claim 1, characterized in that, The step of injecting residual gated modulation parameters into intermediate node features to obtain node representations modulated by affinity conditions includes: Molecular side chain masks are used to distinguish molecular side chain atomic nodes from framework pocket context atomic nodes, and independent learnable affinity modulation coefficients are assigned to the two types of nodes. Based on the affinity modulation coefficient of the corresponding node, the time step modulation parameter and the affinity modulation parameter are linearly superimposed, and the feature linear affine transformation is performed on the intermediate node features after layer normalization. The intermediate node features, time step embeddings, and affinity embeddings before the affine transformation are concatenated. An activation function is then used to construct a node-level feature gating. The node-level feature gating is then combined with the affine transformation result through a Hadamard product. Finally, the residual connection method is used to superimpose the node representation onto the initial intermediate node features to obtain the node representation modulated by affinity conditions.

7. The molecular side chain generation method based on a dual-control diffusion model as described in claim 1, characterized in that, The stepwise sampling estimation of the state of the clean molecular side chain based on the predicted noise using the inverse transfer distribution includes: During the model training phase, a classifier free guidance strategy is adopted, and the network parameters are updated by forcibly replacing part of the real target affinity conditions in the input with empty conditional embedding vectors according to the preset discard probability. At any time step in the denoising generation stage, forward inference is performed by inputting the target affinity condition and the empty condition embedding vector respectively to obtain the conditional noise prediction component and the unconditional noise prediction component. The difference extension term between the conditional noise prediction component and the unconditional noise prediction component is calculated based on the preset guidance intensity hyperparameter. The difference extension term is then compensated into the unconditional noise prediction component to generate the final guidance noise. The state of the clean molecular side chain is estimated using the preset sampling formula.

8. A molecular side chain generation device based on a dual-controlled diffusion model, characterized in that, The apparatus can realize the molecular side chain generation method based on the dual-controlled diffusion model as described in any one of claims 1 to 7, and the apparatus comprises: The forward noise-adding module is used to acquire an input instance containing protein binding pockets, molecular core backbone, molecular side chains, and target affinity conditions, and extract the connection anchors connecting the molecular side chains to the molecular core backbone; the molecular side chains are converted into a set containing atom type vectors and spatial coordinates, and a forward diffusion process is performed to gradually add Gaussian noise to the atom type vectors and spatial coordinates of the molecular side chains to obtain the noisy molecular side chain state. A dual-control denoising module is used to input the noisy molecule's side chain state, molecular core backbone, protein binding pocket, current diffusion time step, and target affinity condition into a pre-constructed dual-control equivariant denoising network. This network contains alternating series of confidence interaction control and affinity context control modules. The confidence interaction control module calculates node confidence based on the characteristics of the atomic nodes and the spatial distance from their spatial coordinates to the anchor points. It then constructs edge-level interaction gates based on these confidence levels and performs multiplicative gating on the converged message features during feature propagation in the equivariant graph neural network, updating the intermediate node features and atomic coordinates selected by node confidence. The affinity context control module performs embedding mapping on the diffusion time step and target affinity condition, generating time step modulation parameters and affinity modulation parameters. By setting differentiated affinity modulation intensities, it calculates the fused residual gating modulation parameters and injects them into the intermediate node features, obtaining node representations modulated by affinity conditions. The sampling and reconstruction module is used to output the prediction noise using the node representation modulated by affinity conditions. Based on the prediction noise, the state of the clean molecular side chain is estimated stepwise using the inverse transfer distribution to obtain the final generated molecular side chain structure. The molecular side chain structure is then connected to the connection anchor point to reconstruct the complete candidate molecule.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the molecular side chain generation method based on the dual-control diffusion model as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the molecular side chain generation method based on the dual-control diffusion model as described in any one of claims 1 to 7.

Images