Flexible manifold space generation molecular docking method and device based on interactive representation

Abstract

The invention provides a flexible manifold space generation molecular docking method and device based on interactive representation, and the method comprises the steps: obtaining a protein coordinate and a small molecule coordinate according to a sampling state; on the basis of a pre-trained interaction representation learning model, generating protein-molecule interaction representation according to the noisy protein pocket and the noisy ligand small molecule; based on a pre-trained generation model, generating a denoised state according to the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation; and obtaining a binding conformation of the protein pocket and the ligand micromolecule according to the denoised state. According to the method, interactive representation containing atomic-scale distance information between protein and small molecules is used as a generation condition of a translation and rotation dihedral angle of the ligand small molecules and a dihedral angle of a protein pocket side chain, and conformational changes of the ligand small molecules and the protein pocket side chain are generated in a flexible manifold space; the rapid and accurate butt joint of the protein pocket and the ligand small molecule is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of drug design, and in particular to a method and device for generating molecular docking in a flexible manifold space based on interactive representation. Background Art

[0002] Small molecule protein docking is a crucial step in drug discovery, aiming to predict the structure of small molecule drug binding to protein receptors. Traditional molecular dynamics-based docking methods, while highly accurate, are computationally expensive and slow, making them inadequate for rapid screening of drug candidates.

[0003] In recent years, deep learning-based small molecule protein docking methods have made rapid progress. However, most deep learning-based methods focus primarily on fixed protein structures and lack the ability to simulate changes in protein side chain flexibility, which limits their application in accurately predicting protein-drug binding structures. Some docking methods that do have this capability, such as AlphaFold3, have extremely high computational complexity and are mostly closed-source or semi-open-source (providing only the inference component), making further research and expansion difficult.

[0004] Therefore, there is an urgent need to develop a deep learning small molecule protein docking method that can not only efficiently simulate the flexible changes of protein side chains but also ensure high accuracy and scalability, so as to help accelerate and optimize the drug development process. Summary of the Invention

[0005] The present invention provides a flexible manifold space generation molecular docking method and device based on interactive representation, which is used to overcome the shortcomings of existing small molecule protein docking methods, such as the inability to efficiently simulate the flexible changes of protein side chains, slow speed, high cost and poor scalability, and realize the rapid and accurate docking of protein pockets with ligand small molecules.

[0006] On the one hand, the present invention provides a method for generating molecular docking in a flexible manifold space based on interaction representation, comprising: obtaining protein coordinates and small molecule coordinates according to a sampling state; wherein the sampling state includes the translational and rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies the manifold constraint; based on a pre-trained interaction representation learning model, generating a protein-molecule interaction representation according to a noisy protein pocket and a noisy ligand small molecule; the protein-molecule interaction representation contains atomic-level distance information between the protein pocket and the ligand small molecule; based on a pre-trained generation model, generating a denoised state according to the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation; wherein the denoised state includes the target translational and rotational dihedral angles of the ligand small molecule and the target dihedral angles of the protein pocket side chain, and the denoised state satisfies the manifold constraint; and obtaining the binding conformation of the protein pocket and the ligand small molecule according to the denoised state.

[0007] Furthermore, obtaining protein coordinates and small molecule coordinates based on the sampling state includes: projecting the sampling state onto a manifold through a projection function to obtain a sampling state of a flexible manifold space; and mapping the sampling state of the flexible manifold space to a Cartesian coordinate space to obtain the protein coordinates and small molecule coordinates.

[0008] Furthermore, the interaction representation learning model includes a protein encoder, a molecule encoder and an interaction model; accordingly, the interaction representation learning model based on pre-training generates a protein-molecule interaction representation according to a noisy protein pocket and a noisy ligand small molecule, including: encoding the noisy protein pocket into a protein representation based on the pre-trained protein encoder; encoding the noisy ligand small molecule into a small molecule representation based on the pre-trained molecule encoder; and generating a protein-molecule interaction representation according to the protein representation and the small molecule representation based on the pre-trained interaction model.

[0009] Furthermore, the pre-trained generative model generates a denoised state according to the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation, including: inputting the noisy protein pocket, protein coordinates, noisy ligand small molecule, small molecule coordinates, protein-molecule interaction representation and sampling time step into the pre-trained generative model to predict a scoring function; calculating the denoised state according to the scoring function and the sampling state of the flexible manifold space; wherein the scoring function includes the scoring functions corresponding to the translation vector, rotation matrix and dihedral angle transformation vector of the ligand small molecule, and the scoring function corresponding to the dihedral angle transformation vector of the protein pocket side chain.

[0010] Furthermore, the translational rotational dihedral angle of the ligand small molecule includes the translation vector, rotation matrix and dihedral angle transformation vector of the ligand small molecule, and the dihedral angle of the protein pocket side chain is the dihedral angle transformation vector of the protein pocket side chain; accordingly, the projection function is defined as follows: ; ; ; ; in, 、 、 Represent the translation vector, rotation matrix and dihedral angle transformation vector of the ligand small molecule respectively, 、 、 They represent the translation vector, rotation matrix and dihedral angle transformation vector of the projected ligand small molecule, respectively. represents the dihedral angle transformation vector of the protein pocket side chain, represents the dihedral angle transformation vector of the projected protein pocket side chain, 、 、 、 are all included in the sampling state of the flexible manifold space.

[0011] Furthermore, the generative model is trained and optimized, specifically including: initializing the translation and rotation dihedral angles corresponding to the small molecule samples and the dihedral angles corresponding to the protein pocket samples based on the small molecule samples and the protein pocket samples; adding random noise to the translation and rotation dihedral angles corresponding to the small molecule samples and the dihedral angles corresponding to the protein pocket samples, and projecting them onto the manifold to obtain a projection state; calculating the score functions for the translation component, the rotation component and the dihedral angle component based on the projection state; mapping the projection state to a Cartesian coordinate space to obtain the small molecule sample coordinates and the protein pocket sample coordinates; based on the generative model, predicting the conditional score function according to the initial molecular conformation of the small molecule sample, the small molecule sample coordinates, the protein pocket sample coordinates, the initial pocket conformation of the protein pocket sample, the noisy time step and the interaction representation; iteratively optimizing the generative model through the target loss function based on the translation component score function, the rotation component score function and the dihedral angle component score function, as well as the predicted conditional score function.

[0012] Furthermore, the objective loss function is defined as follows: ; ; ; ; ; in, 、 、 Represents the translation vectors of the small molecule samples predicted during model training , rotation matrix and the dihedral transformation vector The score function of Represents the dihedral angle transformation vector of the protein pocket sample predicted during generative model training The score function of 、 、 They represent the translation component score, rotation component score, and dihedral component score corresponding to the small molecule sample, respectively. represents the dihedral angle component score corresponding to the protein pocket sample, 、 、 They represent the initial values ​​of the translation vector, rotation matrix, and dihedral angle conversion vector corresponding to the small molecule sample. represents the initial value of the dihedral angle transformation vector corresponding to the protein pocket sample, 、 、 Respectively represent the translation vector, rotation matrix and dihedral angle transformation vector of the projected small molecule sample, represents the dihedral angle transformation vector of the projected protein pocket sample, represents the mean of random noise, represents the variance of random noise, Indicates the correction factor.

[0013] In a second aspect, the present invention also provides a flexible manifold space-based molecular docking device based on interactive representation, comprising: a coordinate acquisition module for acquiring protein coordinates and small molecule coordinates according to a sampling state; wherein the sampling state includes the translational and rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies the manifold constraint; an interactive representation generation module for generating a protein-molecule interaction representation based on a pre-trained interactive representation learning model according to a noisy protein pocket and a noisy ligand small molecule; the protein-molecule interaction representation contains atomic-level distance information between the protein pocket and the ligand small molecule; a denoised state generation module for generating a denoised state based on the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation based on a pre-trained generation model; wherein the denoised state includes the target translational and rotational dihedral angles of the ligand small molecule and the target dihedral angles of the protein pocket side chain, and the denoised state satisfies the manifold constraint; a small molecule docking module for obtaining the binding conformation of the protein pocket and the ligand small molecule according to the denoised state.

[0014] In a third aspect, 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 flexible manifold space based on interactive representation and molecular docking method as described above is implemented.

[0015] In a fourth 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, implements any of the above-described methods for generating molecular docking in a flexible manifold space based on interactive representation.

[0016] The present invention provides a method for generating molecular docking in a flexible manifold space based on interactive representation, which obtains protein coordinates and small molecule coordinates according to a sampling state; wherein the sampling state includes the translational and rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies the manifold constraint; based on a pre-trained interactive representation learning model, a protein-molecule interaction representation is generated according to a noisy protein pocket and a noisy ligand small molecule; the protein-molecule interaction representation contains atomic-level distance information between the protein pocket and the ligand small molecule; based on a pre-trained generation model, a denoised state is generated according to the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation; wherein the denoised state includes the target translational and rotational dihedral angles of the ligand small molecule and the target dihedral angles of the protein pocket side chain, and the denoised state satisfies the manifold constraint; based on the denoised state, the binding conformation of the protein pocket and the ligand small molecule is obtained. This method uses the interaction representation containing atomic-level distance information between protein and small molecule as the generation condition for the translational and rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, generates conformational changes of the ligand small molecule and the protein pocket side chain on the flexible manifold space, and realizes the rapid and precise docking of the protein pocket and the ligand small molecule. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces 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 a method for generating molecular docking in a flexible manifold space based on interactive representation provided by an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the pre-training of the interactive representation learning model provided by an embodiment of the present invention.

[0020] Figure 3 It is a structural schematic diagram of a flexible manifold space generation molecular docking device based on interactive representation provided by an embodiment of the present invention.

[0021] Figure 4 It is a schematic diagram of the physical structure of an electronic device provided by an embodiment of 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 embodiments of the present invention, not all embodiments. 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] It should be noted that small molecule-protein docking is a crucial task in drug development. Some traditional molecular dynamics-based methods are slow and consume significant computational resources. Most current deep learning-based small molecule docking methods lack the ability to model flexible docking with changes in protein side chains. Even those that do have this capability suffer from high computational complexity, are mostly closed-source or semi-open-source (only the inference portion), lack scalability, and are difficult to use. Therefore, there is an urgent need to develop a fast and accurate small molecule docking method to fill this gap.

[0024] In view of this, the present invention proposes a flexible manifold space generation molecular docking method based on interactive representation, specifically, Figure 1 A schematic flow chart of a method for generating molecular docking in a flexible manifold space based on interactive representation provided by an embodiment of the present invention is shown.

[0025] like Figure 1 As shown, the method includes steps S110-S140, and steps S110-S140 and related steps will be described in detail below.

[0026] S110, obtaining protein coordinates and small molecule coordinates according to a sampling state; wherein the sampling state includes the translational rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies the manifold constraint.

[0027] It is easy to understand that the initial state can be sampled from the standard normal distribution, which is recorded as the sampling state. The sampling state includes the translation and rotation dihedral angles of the ligand small molecule (i.e., the translation vector , rotation matrix and the dihedral transformation vector ) and the dihedral angles of the protein pocket side chains (i.e., the dihedral angle transformation vector ).

[0028] Then, the sampling state is projected onto the corresponding manifold (i.e., flexible manifold space) through an innovatively designed projection function, so that the sampling state can satisfy the manifold constraints (dihedral angles satisfy the constraints on the torus, and the rotation quaternion corresponding to the rotation matrix satisfies the constraints on the sphere). Then, the projected sampling state is mapped back to the Cartesian coordinate space to obtain the protein coordinates of the current sampling time step. and small molecule coordinates .

[0029] Optionally, the innovatively designed projection function is defined as follows (1)-(4).

[0030] (1).

[0031] (2).

[0032] (3).

[0033] (4).

[0034] In formulas (1)-(4), 、 、 Represent the translation vector, rotation matrix and dihedral angle transformation vector of the ligand small molecule respectively, 、 、 They represent the translation vector, rotation matrix and dihedral angle transformation vector of the projected ligand small molecule, respectively. represents the dihedral angle transformation vector of the protein pocket side chain, represents the dihedral angle transformation vector of the projected protein pocket side chain, 、 、 、 are all included in the sampling state of the flexible manifold space.

[0035] Based on the projection functions (1)-(4), the sampling state of the flexible manifold space can be obtained, namely 、 、 、 .

[0036] Furthermore, optionally, the sampling state of the flexible manifold space (also called product space) can be mapped to the Cartesian coordinate space through an existing mapping function to obtain the protein coordinates and small molecule coordinates .

[0037] S120, based on the pre-trained interaction representation learning model, generates a protein-molecule interaction representation according to the noisy protein pocket and the noisy ligand small molecule; the protein-molecule interaction representation contains atomic-level distance information between the protein pocket and the ligand small molecule.

[0038] It is easy to understand that in this embodiment, an interaction representation learning model is pre-trained, which can generate corresponding protein-molecule interaction representations based on given noisy protein pockets and ligand small molecules.

[0039] Optional, Figure 2 FIG. 1 shows a schematic diagram of the pre-training of the interactive representation learning model provided by an embodiment of the present invention. Figure 2 As shown in the figure, the interactive representation learning model includes a protein encoder (Pocket Encoder), a molecule encoder (MolEncoder), and an interaction model (Interaction Module).

[0040] Accordingly, based on the pre-trained interaction representation learning model, a protein-molecule interaction representation is generated according to the noisy protein pocket and the noisy ligand small molecule, including: based on the pre-trained protein encoder, encoding the noisy protein pocket into a protein representation; based on the pre-trained molecule encoder, encoding the noisy ligand small molecule into a small molecule representation; based on the pre-trained interaction model, generating a protein-molecule interaction representation according to the protein representation and the small molecule representation.

[0041] Specifically, a noisy protein pocket is input into a pre-trained protein encoder to produce an output protein representation. A noisy protein pocket refers to a protein pocket with noise added to its side chain dihedral angles. A noisy ligand small molecule is input into a pre-trained molecular encoder to produce an output small molecule representation. A noisy ligand small molecule refers to a ligand small molecule with noise added to its translational and rotational dihedral angles. The protein representation and small molecule representation are then input into a pre-trained interaction model to produce an output protein-molecule interaction representation. The protein-molecule interaction representation contains atomic-level distance information between the protein pocket and the ligand small molecule, which serves as a condition to guide the generation of the generative model.

[0042] It is worth mentioning that the pre-training process of the protein encoder in this embodiment is innovatively designed. Specifically, the side chains of the protein pocket samples are perturbed during training, and the protein pocket samples with added side chain dihedral angle noise and atomic coordinate Gaussian noise are used as input, and the predicted noise is used as output, so that the restored protein pocket samples can be obtained.

[0043] according to Figure 2As can be seen, training the interaction model involves two main tasks: distance prediction (Distance Map) and mask ligand reconstruction (Mask Ligand). In the distance prediction task, dihedral-perturbed protein and small molecule representations are input, and the distance matrix between the protein and small molecule representations is output. In the mask ligand reconstruction task, partially masked small molecule and protein representations are input, and the predicted masked portion of the small molecule is output.

[0044] Among them, the protein representation samples and the small molecule representation samples can be generated based on the above-mentioned pre-trained protein encoder and molecule encoder according to the protein pocket samples and small molecule samples.

[0045] S130, based on a pre-trained generative model, generating a denoised state according to the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation; wherein the denoised state satisfies a manifold constraint.

[0046] Optionally, the noisy protein pocket, protein coordinates, noisy ligand small molecule, small molecule coordinates, protein-molecule interaction representation, and sampling time step are input into the pre-trained generative model to predict the score function. The score function here includes the score function corresponding to the translation vector of the ligand small molecule, the score function corresponding to the rotation matrix, the score function corresponding to the dihedral angle transformation vector, and the score function corresponding to the dihedral angle transformation vector of the protein pocket side chain. Subsequently, the predicted score function is compared with the sampling state of the flexible manifold space (i.e. 、 、 、 ), the denoised state can be calculated.

[0047] Among them, the denoised state includes the target translation and rotation dihedral angles of the ligand small molecule (as well as the target translation vector, target rotation matrix and target dihedral angle transformation vector of the ligand small molecule) and the target dihedral angles of the protein pocket side chain (that is, the target dihedral angle transformation vector of the protein pocket side chain), and the denoised state satisfies the manifold constraint.

[0048] It should be noted that since both the initial sampling state and the denoised state need to satisfy the corresponding manifold constraints, this process is a constrained generation process. This embodiment develops a constrained diffusion generation framework, which will be elaborated in detail in the following embodiments.

[0049] It should also be noted that the specific process of training and optimizing the generation model will be expanded in detail in the embodiments below and will not be described in detail here.

[0050] S140, obtaining the binding conformation of the protein pocket and the ligand small molecule according to the denoised state.

[0051] It is easy to understand that, once the target translational and rotational dihedral angles of the ligand small molecule and the target dihedral angles of the protein pocket side chains are determined, the target translational and rotational dihedral angles of the ligand small molecule are converted into target small molecule coordinates in Cartesian coordinate space, and the target dihedral angles of the protein pocket side chains are converted into target protein coordinates in Cartesian coordinate space. Thereafter, the protein pocket and the ligand small molecule are docked based on the target small molecule coordinates and the target protein coordinates to obtain the binding conformation of the protein pocket and the ligand small molecule, i.e., a protein-small molecule complex.

[0052] In this embodiment, protein coordinates and small molecule coordinates are obtained according to the sampling state; wherein the sampling state includes the translational rotation dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies the manifold constraint; based on the pre-trained interaction representation learning model, a protein-molecule interaction representation is generated according to the noisy protein pocket and the noisy ligand small molecule; the protein-molecule interaction representation contains the atomic-level distance information between the protein pocket and the ligand small molecule; based on the pre-trained generative model, a denoised state is generated according to the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation; wherein the denoised state includes the target translational rotation dihedral angles of the ligand small molecule and the target dihedral angles of the protein pocket side chain, and the denoised state satisfies the manifold constraint; based on the denoised state, the binding conformation of the protein pocket and the ligand small molecule is obtained. This method uses the interaction representation containing atomic-level distance information between protein and small molecule as the generation condition for the translational and rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, generates conformational changes of the ligand small molecule and the protein pocket side chain on the flexible manifold space, and realizes the rapid and precise docking of the protein pocket and the ligand small molecule.

[0053] On the basis of the above embodiments, the training optimization process of the generation model will be further described in detail below.

[0054] First, based on the small molecule samples and protein pocket samples, the translation and rotation dihedral angles corresponding to the small molecule samples and the dihedral angles corresponding to the protein pocket samples are initialized.

[0055] Specifically, for a given small molecule sample, the tool RDKit is used to generate the initial molecular conformation of the small molecule sample. , and perform arbitrary translation, rotation, and dihedral angle transformations on the initial molecular conformation to obtain the molecular conformation space , this space can also be called flexible manifold space.

[0056] Then, the actual optimal molecular conformation Based on the molecular conformation space Find a molecular conformation that is closest to the optimal molecular conformation and record it as the target molecular conformation, that is, .in, is the root mean square error distance, used for the optimal molecular conformation and molecular conformation in molecular conformation space similarity.

[0057] Next, through the mapping function The target molecular conformation is converted into the translational rotation dihedral angle in the product space, that is, .in, represents the translation vector of the small molecule sample, represents the rotation matrix of the small molecule sample, Represents the dihedral transformation vector of the small molecule sample.

[0058] At the same time, through the mapping function The initial conformation of the protein pocket sample Convert to dihedral angle in product space, that is .

[0059] Thus, the translational rotational dihedral angle corresponding to the small molecule sample can be obtained Dihedral angles corresponding to protein pocket samples .

[0060] Then, random noise is added to the translation and rotation dihedral angles corresponding to the small molecule samples and the dihedral angles corresponding to the protein pocket samples, and they are projected onto the manifold to obtain the projected state.

[0061] Specifically, you first need to use the function The rotation matrix Convert to unit quaternion, that is The unit quaternion is a The representation on is used to describe the rotation.

[0062] Then, randomly sample time steps , according to the noise scheduling, the sampling noise is added to the translation and rotation dihedral angles corresponding to the small molecule sample and the dihedral angles corresponding to the protein pocket sample, that is, .in, That is, the translation and rotation dihedral angles of the small molecule sample and the dihedral angles of the protein pocket sample after the introduction of noise.

[0063] Next, through the projection function Will Projected onto the corresponding manifold, Through this projection, it can be guaranteed is a three-dimensional vector, is the unit quaternion, The angle is limited to , The angle is limited to The specific content of the projection function can be found in equations (1)-(4).

[0064] Then, according to the projection state , the score functions are calculated for the translation component, rotation component and dihedral angle component respectively.

[0065] See the following equations (5)-(8).

[0066] (5).

[0067] (6).

[0068] (7).

[0069] (8).

[0070] In equations (5)-(8), represents the mean of random noise, represents the variance of random noise. Equation (5) is used to calculate the conditional score of the translation component, Equation (6) is used to calculate the conditional score of the rotation component, Equation (7) is used to calculate the conditional score of the molecular dihedral angle component, and Equation (8) is used to calculate the conditional score of the pocket dihedral angle component.

[0071] Expanding, the score function 、 、 、 By expanding, we can obtain the following equations (9)-(12), which can be used to solve the conditional scores of the translation component, rotation component and dihedral angle component.

[0072] (9).

[0073] (10).

[0074] (11).

[0075] (12).

[0076] In formulas (11) and (12), represents the correction coefficient, which is defined as follows (13)-(15).

[0077] (13).

[0078] (14).

[0079] (15).

[0080] Next, the projection state Mapped to the Cartesian coordinate space, we obtain the small molecule sample coordinates and protein pocket sample coordinates.

[0081] Optionally, use a mapping function The molecular conformation of the current product space Mapping back to Cartesian coordinate space, that is .

[0082] At the same time, using the mapping function The pocket configuration of the current product space Mapping back to Cartesian coordinate space, that is .

[0083] From this, we can get the coordinates of the small molecule sample and protein pocket sample coordinates .

[0084] Furthermore, based on the generative model, the initial conformation of the small molecule sample , small molecule sample coordinates , protein pocket sample coordinates , initial conformation of the protein pocket sample , noise time step and interactive representation , prediction conditional score function .

[0085] Among them, the interactive representation is generated based on the pre-trained interactive representation learning model according to the noisy small molecule samples and the noisy protein pocket samples.

[0086] Finally, based on the translation component score function, rotation component score function, dihedral component score function, and the predicted conditional score function, the generative model is iteratively optimized through the target loss function.

[0087] The objective loss function is defined as follows (16).

[0088] (16).

[0089] In formula (16), 、 、 Represents the translation vectors of the small molecule samples predicted during model training , rotation matrix and the dihedral transformation vector The score function of Represents the dihedral angle transformation vector of the protein pocket sample predicted during generative model training The score function of 、 、 They represent the translation component score, rotation component score, and dihedral component score corresponding to the small molecule sample, respectively. represents the dihedral angle component score corresponding to the protein pocket sample, 、 、 They represent the initial values ​​of the translation vector, rotation matrix, and dihedral angle conversion vector corresponding to the small molecule sample. represents the initial value of the dihedral angle transformation vector corresponding to the protein pocket sample, 、 、 Respectively represent the translation vector, rotation matrix and dihedral angle transformation vector of the projected small molecule sample, represents the dihedral angle transformation vector of the projected protein pocket sample, represents the mean of random noise, represents the variance of random noise.

[0090] After sufficient training until the target loss function converges, the trained generative model is output .

[0091] In this embodiment, a generative model is iteratively optimized using a target loss function. Based on the pre-trained generative model, a denoised state is generated based on the protein coordinates, small molecule coordinates, and protein-molecule interaction representation. The denoised state includes the target translational and rotational dihedral angles of the ligand small molecule and the target dihedral angles of the protein pocket side chains, and the denoised state satisfies the manifold constraint. Based on the denoised state, the binding conformation of the protein pocket and the ligand small molecule is obtained. This method generates conformational changes of the ligand small molecule and the protein pocket side chains on a flexible manifold space by using the interaction representation containing atomic-level distance information between the protein and small molecule as the generation conditions for the translational and rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chains, thereby achieving rapid and accurate docking of the protein pocket and the ligand small molecule.

[0092] On the basis of the above embodiments, the overall process of the method for generating molecular docking in a flexible manifold space based on interactive representation will be described in detail below.

[0093] First, the parameters are initialized.

[0094] Specifically, set the initial time , and sample the initial state from a standard normal distribution Project the sampled initial state onto the manifold, i.e. , to ensure that each parameter satisfies the manifold constraint. Use the mapping function to map each parameter of the manifold space back to the Cartesian coordinate space, that is , .

[0095] Then, loop the sampling.

[0096] when When: Based on the pre-trained interaction representation learning model, the interaction representation is generated according to the noisy small molecule samples and protein pocket samples, that is, ; Use the generative model to predict the translation vector, rotation matrix and score function corresponding to the dihedral angle transformation vector of the ligand small molecule, as well as the score function corresponding to the dihedral angle transformation vector of the protein pocket side chain, that is, ; Calculate and update model parameters , indicating that from arrive The noise attenuation coefficient of ; calculate the state after denoising, that is .

[0097] like , then Gaussian noise is added to ensure the randomness of sampling, that is, , and project the state after adding noise onto the manifold, that is .

[0098] The time step decreases, At this time, the projected state is mapped back to the Cartesian coordinate space to obtain the protein coordinates and small molecule coordinates of the current time step, that is, , .

[0099] when When the final protein coordinates and small molecule coordinates are returned, that is, the target protein coordinates and target small molecule coordinates By docking according to the target protein coordinates and the target small molecule coordinates, the binding conformation of the protein pocket and the ligand small molecule can be obtained.

[0100] As can be seen from the above, in the flexible manifold space generation molecular docking method based on interactive representation provided in this embodiment, a diffusion generation framework is introduced, which gradually adds noise to the initial conformations of the protein pocket and the ligand small molecule, and then learns to reverse this process to remove the noise, thereby generating new conformations of the protein pocket and the ligand small molecule.

[0101] Corresponding to the molecular docking method based on interactive representation and flexible manifold space generation described in the above embodiments, the present invention also proposes a molecular docking device based on interactive representation and flexible manifold space generation.

[0102] Specifically, Figure 3 A schematic structural diagram of a flexible manifold space-based molecular docking device based on interactive representation provided by an embodiment of the present invention is shown.

[0103] like Figure 3 As shown, the device includes: a coordinate acquisition module 310, which is used to acquire protein coordinates and small molecule coordinates according to a sampling state; wherein the sampling state includes the translational rotation dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies the manifold constraint; an interaction representation generation module 320, which is used to generate a protein-molecule interaction representation based on a pre-trained interaction representation learning model according to the noisy protein pocket and the noisy ligand small molecule; the protein-molecule interaction representation contains atomic-level distance information between the protein pocket and the ligand small molecule; a denoised state generation module 330, which is used to generate a denoised state based on the pre-trained generation model according to the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation; wherein the denoised state includes the target translational rotation dihedral angles of the ligand small molecule and the target dihedral angles of the protein pocket side chain, and the denoised state satisfies the manifold constraint; a small molecule docking module 340, which is used to obtain the binding conformation of the protein pocket and the ligand small molecule according to the denoised state.

[0104] In this embodiment, the coordinate acquisition module 310 acquires the protein coordinates and the small molecule coordinates according to the sampling state; wherein, the sampling state includes the translational rotation dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies the manifold constraint; the interaction representation generation module 320 generates a protein-molecule interaction representation based on the noisy protein pocket and the noisy ligand small molecule based on the pre-trained interaction representation learning model; the protein-molecule interaction representation contains the atomic-level distance information between the protein pocket and the ligand small molecule; the denoised state generation module 330 generates a denoised state based on the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation based on the pre-trained generation model; wherein, the denoised state includes the target translational rotation dihedral angles of the ligand small molecule and the target dihedral angles of the protein pocket side chain, and the denoised state satisfies the manifold constraint; the small molecule docking module 340 obtains the binding conformation of the protein pocket and the ligand small molecule based on the denoised state. The device uses the interaction representation containing atomic-level distance information between proteins and small molecules as the generation condition for the translational and rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, generates conformational changes of the ligand small molecule and the protein pocket side chain on the flexible manifold space, and realizes the rapid and precise docking of the protein pocket and the ligand small molecule.

[0105] It should be noted that the molecular docking device based on interactive representation of flexible manifold space provided in the embodiment of the present invention can correspond to the molecular docking method based on interactive representation of flexible manifold space described in the above embodiments, and will not be repeated here.

[0106] Figure 4 An example of a physical structure diagram of an electronic device is shown below. Figure 4As shown, the electronic device may include: a processor (processor) 410 , a communication interface (Communications Interface) 420 , a memory (memory) 430 and a communication bus 440 , wherein the processor 410 , the communication interface 420 , and the memory 430 communicate with each other via the communication bus 440 . The processor 410 can call the logic instructions in the memory 430 to execute a flexible manifold space generation molecular docking method based on interaction representation, the method comprising: obtaining protein coordinates and small molecule coordinates according to a sampling state; wherein the sampling state includes the translational rotation dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies the manifold constraint; based on a pre-trained interaction representation learning model, generating a protein-molecule interaction representation according to the noisy protein pocket and the noisy ligand small molecule; the protein-molecule interaction representation contains atomic-level distance information between the protein pocket and the ligand small molecule; based on a pre-trained generation model, generating a denoised state according to the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation; wherein the denoised state includes the target translational rotation dihedral angles of the ligand small molecule and the target dihedral angles of the protein pocket side chain, and the denoised state satisfies the manifold constraint; and obtaining the binding conformation of the protein pocket and the ligand small molecule according to the denoised state.

[0107] 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.

[0108] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the flexible manifold space generation molecular docking method based on interaction representation provided by the above methods, the method comprising: obtaining protein coordinates and small molecule coordinates according to a sampling state; wherein the sampling state includes the translational and rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies the manifold constraint; based on a pre-trained interaction representation learning model, generating a protein-molecule interaction representation according to the noisy protein pocket and the noisy ligand small molecule; the protein-molecule interaction representation contains atomic-level distance information between the protein pocket and the ligand small molecule; based on a pre-trained generation model, generating a denoised state according to the protein coordinates, the small molecule coordinates and the protein-molecule interaction representation; wherein the denoised state includes the target translational and rotational dihedral angles of the ligand small molecule and the target dihedral angles of the protein pocket side chain, and the denoised state satisfies the manifold constraint; and obtaining the binding conformation of the protein pocket and the ligand small molecule according to the denoised state.

[0109] 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.

[0110] 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.

[0111] 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 flexible manifold space generation molecular docking method based on interactive representation, characterized in that: include: Obtaining protein coordinates and small molecule coordinates according to a sampling state; wherein the sampling state includes the translational and rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies a manifold constraint; Based on a pre-trained interaction representation learning model, a protein-molecule interaction representation is generated from a noisy protein pocket and a noisy ligand small molecule. The protein-molecule interaction representation contains atomic-level distance information between the protein pocket and the ligand small molecule. Generate a denoised state based on the protein coordinates, the small molecule coordinates, and the protein-molecule interaction representation based on a pre-trained generative model; wherein the denoised state includes target translational and rotational dihedral angles of the ligand small molecule and target dihedral angles of the protein pocket side chain, and the denoised state satisfies a manifold constraint; According to the denoised state, the binding conformation of the protein pocket and the ligand small molecule is obtained.

2. The method for generating molecular docking based on flexible manifold space based on interactive representation according to claim 1, characterized in that: The step of obtaining protein coordinates and small molecule coordinates according to the sampling state includes: Projecting the sampling state onto the manifold by a projection function to obtain a sampling state of the flexible manifold space; The sampling state of the flexible manifold space is mapped to a Cartesian coordinate space to obtain the protein coordinates and the small molecule coordinates.

3. The method for generating molecular docking in a flexible manifold space based on interactive representation according to claim 1, characterized in that: The interactive representation learning model includes a protein encoder, a molecular encoder and an interactive model; Accordingly, the pre-trained interaction representation learning model generates protein-molecule interaction representations based on noisy protein pockets and noisy ligand small molecules, including: encoding the noisy protein pocket into a protein representation based on a pre-trained protein encoder; Based on a pre-trained molecular encoder, encoding the noisy ligand small molecule into a small molecule representation; Based on the pre-trained interaction model, a protein-molecule interaction representation is generated according to the protein representation and the small molecule representation.

4. The method for generating molecular docking in a flexible manifold space based on interactive representation according to claim 2, characterized in that: The pre-trained generative model generates a denoised state according to the protein coordinates, the small molecule coordinates, and the protein-molecule interaction representation, including: The noisy protein pocket, protein coordinates, noisy ligand small molecule, small molecule coordinates, protein-molecule interaction representation, and sampling time step are input into a pre-trained generative model to predict a scoring function; Calculating a denoised state according to the score function and the sampling state of the flexible manifold space; The scoring function includes scoring functions corresponding to the translation vector, rotation matrix and dihedral angle transformation vector of the ligand small molecule, and scoring function corresponding to the dihedral angle transformation vector of the protein pocket side chain.

5. The method for generating molecular docking in a flexible manifold space based on interactive representation according to claim 2, characterized in that: The translational rotational dihedral angle of the ligand small molecule includes the translation vector, rotation matrix and dihedral angle transformation vector of the ligand small molecule, and the dihedral angle of the protein pocket side chain is the dihedral angle transformation vector of the protein pocket side chain; Accordingly, the projection function is defined as follows: ； ； ； ； in, 、 、 Represent the translation vector, rotation matrix and dihedral angle transformation vector of the ligand small molecule respectively, 、 、 They represent the translation vector, rotation matrix and dihedral angle transformation vector of the projected ligand small molecule, respectively. represents the dihedral angle transformation vector of the protein pocket side chain, represents the dihedral angle transformation vector of the projected protein pocket side chain, 、 、 、 are all included in the sampling state of the flexible manifold space.

6. The method for generating molecular docking in a flexible manifold space based on interactive representation according to any one of claims 1 to 5, characterized in that: Training and optimizing the generative model specifically includes: According to the small molecule samples and protein pocket samples, the translation and rotation dihedral angles corresponding to the small molecule samples and the dihedral angles corresponding to the protein pocket samples are initialized; adding random noise to the translational and rotational dihedral angles corresponding to the small molecule sample and the dihedral angles corresponding to the protein pocket sample, and projecting them onto the manifold to obtain a projected state; According to the projection state, score functions are calculated for the translation component, the rotation component and the dihedral angle component respectively; Mapping the projected state to a Cartesian coordinate space to obtain small molecule sample coordinates and protein pocket sample coordinates; Based on the generative model, the conditional score function is predicted according to the initial molecular conformation of the small molecule sample, the small molecule sample coordinates, the protein pocket sample coordinates, the initial pocket conformation of the protein pocket sample, the noisy time step, and the interaction representation; Based on the translation component score function, the rotation component score function, the dihedral component score function, and the predicted conditional score function, the generation model is iteratively optimized through the target loss function.

7. The method for generating molecular docking based on flexible manifold space based on interactive representation according to claim 6, characterized in that: The objective loss function is defined as follows: ； ； ； ； ； in, 、 、 Represents the translation vectors of the small molecule samples predicted during model training , rotation matrix and the dihedral transformation vector The score function, Represents the dihedral angle transformation vector of the protein pocket sample predicted during generative model training The score function, 、 、 They represent the translation component score, rotation component score, and dihedral component score corresponding to the small molecule sample, respectively. represents the dihedral angle component score corresponding to the protein pocket sample, 、 、 They represent the initial values ​​of the translation vector, rotation matrix, and dihedral angle conversion vector corresponding to the small molecule sample. represents the initial value of the dihedral angle transformation vector corresponding to the protein pocket sample, 、 、 Respectively represent the translation vector, rotation matrix and dihedral angle transformation vector of the projected small molecule sample, represents the dihedral angle transformation vector of the projected protein pocket sample, represents the mean of random noise, represents the variance of random noise, Indicates the correction factor.

8. A flexible manifold space generation molecular docking device based on interactive representation, characterized in that: include: A coordinate acquisition module, configured to acquire protein coordinates and small molecule coordinates based on a sampling state, wherein the sampling state includes the translational and rotational dihedral angles of the ligand small molecule and the dihedral angles of the protein pocket side chain, and the sampling state satisfies a manifold constraint; An interaction representation generation module is used to generate a protein-molecule interaction representation based on a pre-trained interaction representation learning model, based on a noisy protein pocket and a noisy ligand small molecule; the protein-molecule interaction representation contains atomic-level distance information between the protein pocket and the ligand small molecule; a denoised state generation module, configured to generate a denoised state based on the protein coordinates, the small molecule coordinates, and the protein-molecule interaction representation based on a pre-trained generative model; wherein the denoised state includes target translational and rotational dihedral angles of the ligand small molecule and target dihedral angles of the protein pocket side chain, and the denoised state satisfies a manifold constraint; The small molecule docking module is used to obtain the binding conformation of the protein pocket and the ligand small molecule according to the denoised state.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the flexible manifold space generation molecular docking method based on interactive representation is implemented according to any one of claims 1 to 6.

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 flexible manifold space generation molecular docking method based on interactive representation is implemented according to any one of claims 1 to 7.

Images