Binding protein generation method based on electron density hidden space diffusion model

Abstract

The invention provides a combined protein generation method based on an electron density hidden space diffusion model. The method of the present disclosure utilizes a conditional diffusion model in a hidden space to generate a protein-bound electron density representation for structural related information of a given target spot. In the solution provided by the invention, related concepts of a reciprocal space are originally introduced to carry out encoding and decoding of a hidden space, and meanwhile, a mode of utilizing a guiding mask is provided to control the model to generate electron density with higher quality in a targeted manner. According to the method, unlabeled training is carried out by using a public data set (PDB), and a hidden space diffusion model capable of accurately generating protein electron density with binding characteristics in a specific region is obtained, so that a foundation is laid for subsequent realization of biological functions and design and development of related drugs.

Description

Technical Field

[0001] The present disclosure relates to protein production, and more particularly to a method for producing binding proteins based on an electron density latent space diffusion model. Background Art

[0002] In the field of traditional protein design, the more representative solutions currently include the Hallucination method proposed by Professor David Baker's team at the University of Washington and the SCUBA+ABACUS method proposed by Professor Liu Haiyan's team at the University of Science and Technology of China. Both solutions belong to the field of de novo protein design, which is to use computational methods to explore the huge sequence space and structural space in protein structure, so as to design proteins with new structures and functions.

[0003] The Hallucination scheme of David Baker's team is a protein computational design method based on the RoseTTA algorithm. It uses the energy function and sampling algorithm of RoseTTA to evaluate and optimize the generated sequence so that it can fold into the desired structure. This method can design proteins with novel topology, stability and function, such as new enzymes, nanoparticles, antibodies, etc. See the literature: Anishchenko, I., Pellock, SJ, Chidyausiku, TM et al. Denovo protein design by deep network hallucination. Nature 600, 547–552 (2021). https: / / doi.org / 10.1038 / s41586-021-04184-w.

[0004] SCUBA proposed by Liu Haiyan's team at the University of Science and Technology of China uses a new statistical learning strategy, using the kernel density estimation plus neural network fitting (NC-NN) method to fit the analytical energy function from the structural data, which can continuously and extensively search the main chain structure space and automatically generate a highly designable main chain. Subsequently, the developed ABACUS algorithm can be used to design artificial proteins with new structures and sequences on the generated main chain. See the literature: Huang, B., Xu, Y., Hu, X. et al. A backbone-centred energy function of neural networks for protein design. Nature 602, 523–528 (2022). https: / / doi.org / 10.1038 / s41586-021-04383-5.

[0005] In addition, scientists in David Baker's laboratory combined structure prediction networks and generative diffusion models in a new protein design method called RoseTTAFold Diffusion (RFDiffusion). This study revealed the powerful role of diffusion models in protein design, effectively improving the efficiency and accuracy of protein design. However, this scheme also has some defects: first, the design algorithm is only for the main chain, so further side chain design is required; second, this method requires the use of a series of algorithms owned by David Baker's team, which limits its application in the field of protein design. See the literature: Watson, JL, Juergens, D., Bennett, NR et al. Denovo design of protein structure and function with RFdiffusion. Nature (2023) (Published: 11 July 2023). https: / / doi.org / 10.1038 / s41586-023-06415-8.

[0006] Therefore, the protein design field hopes to propose a protein design method that can generate the corresponding three-dimensional protein electron density based on the three-dimensional electron density information of the input target protein, thereby quickly completing the required macromolecular design in an end-to-end sense. Summary of the invention

[0007] The present disclosure proposes a method for generating binding proteins based on an electron density diffusion model. The method particularly utilizes the electron density information of the target protein. The target protein, that is, the target area to be bound, can also be referred to as the target protein. In the present disclosure, the terms "target protein", "target protein", "target protein", "target point" and the like can be used interchangeably to refer to the same meaning. The method of the present disclosure specifically targets the three-dimensional protein electron density information of the input target protein, generates the corresponding three-dimensional protein electron density bound thereto, thereby quickly completing the required macromolecular design in an end-to-end sense.

[0008] According to a first aspect of the present disclosure, a method for generating a binding protein is provided. The method may include: performing a latent space transformation on the electron density of the target protein to obtain electron density latent space transformation data; and inputting the electron density latent space transformation data into an electron density latent space diffusion model to generate binding electron density data, wherein the binding protein represented by the binding electron density data can bind to the target protein.

[0009] In the method for generating a binding protein according to the first aspect of the present disclosure, the step of performing a latent space transformation on the electron density of the target protein to obtain electron density latent space transformation data may include: performing latent space encoding on the electron density of the target protein, transforming from real space to reciprocal latent space, and obtaining reciprocal latent space characteristics.

[0010] Preferably, the latent space encoding of the electron density of the target protein may include performing a data mode adaptation transformation on the electron density of the target protein. More preferably, the electron density of the target protein may be subjected to a Fourier transformation.

[0011] In the method for generating a binding protein according to the first aspect of the present disclosure, the inputting of electron density latent space transformation data into an electron density latent space diffusion model to generate binding electron density data may include: inputting the electron density latent space transformation data into an electron density latent space diffusion model; and performing latent space decoding on the output result of the electron density latent space diffusion model, transforming from reciprocal latent space to real space, and generating the electron density of the binding protein that can bind to the target protein.

[0012] Preferably, the latent space decoding of the output result of the electron density latent space diffusion model may include performing a data modality adaptation transformation on the output result of the electron density latent space diffusion model. More preferably, the output result of the electron density latent space diffusion model may be subjected to an inverse Fourier transform.

[0013] The method for generating a binding protein according to the first aspect of the present disclosure may further include: obtaining a latent space mask by performing a latent space transformation on the guiding mask. Thus, the latent space mask is feature-fused with the electron density latent space transformation data, and input into the electron density latent space diffusion model to generate the electron density of the binding protein that can bind to the target protein.

[0014] Preferably, the guided mask is encoded in latent space and transformed from real space to reciprocal latent space to obtain a reciprocal latent space mask.

[0015] Preferably, the said latent space encoding of the guided mask may include performing a data modality adaptation transformation on the guided mask. More preferably, the guided mask may be subjected to a Fourier transformation.

[0016] In the method for generating a bound protein according to the first aspect of the present disclosure, the step of inputting the electron density latent space transformation data into the electron density latent space diffusion model to generate bound electron density data may include: using the electron density latent space diffusion model to calculate the diffusion result x at the current time step t. t Perform noise estimation and output the current step noise z t, and the denoising result x is calculated from this t-1 , as the diffusion result input for the next step, and the cycle is repeated again and again, stopping after T times.

[0017] Preferably, the initial input of the loop is a random sampling noise x in the latent space. T , the final output of the loop is the denoising result x0.

[0018] In the method for generating a bound protein according to the first aspect of the present disclosure, the step of inputting the electron density latent space transformation data into the electron density latent space diffusion model to generate bound electron density data may include: inputting the diffusion result x at the current time step t into the electron density latent space diffusion model to generate bound electron density data. t , the current time step t, the electron density latent space transformation data and the latent space mask are feature fused and input into the electron density latent space diffusion model.

[0019] Preferably, the denoising result x0 is latent space decoded to obtain the generated electron density of the bound protein.

[0020] Preferably, the method further comprises: using multiple sampling noise x T , input into the electron density latent space diffusion model to generate differentiated model output results for further analysis, identification and evaluation.

[0021] In the method for generating a binding protein according to the first aspect of the present disclosure, the electron density latent space diffusion model can be constructed by training with PDB data as training data.

[0022] Preferably, the PDB data used as training data includes high-resolution X-ray electron density data.

[0023] According to the second aspect of the present disclosure, a binding protein generation system is provided. The system may include: an acquisition module for acquiring the electron density of a target protein; a processing module for generating a binding protein capable of binding to the target protein according to the electron density of the target protein, wherein the processing module further includes the following submodules: a latent space transformation submodule for performing latent space transformation on the electron density of the target protein to obtain electron density latent space transformation data; and a generation submodule for inputting the electron density latent space transformation data into an electron density latent space diffusion model to generate binding electron density data, wherein the binding protein represented by the binding electron density data can bind to the target protein.

[0024] According to a third aspect of the present disclosure, a computer-implemented binding protein generation system is provided. The computer-implemented system can implement the following operations by executing a computer program: obtaining the electron density of the target protein; generating the electron density of the binding protein that can bind to the target protein according to the electron density of the target protein, and the operation further includes the following sub-operations: performing a latent space transformation on the electron density of the target protein to obtain electron density latent space transformation data; and inputting the electron density latent space transformation data into an electron density latent space diffusion model to generate binding electron density data, wherein the binding protein represented by the binding electron density data can bind to the target protein.

[0025] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium is provided for storing a computer program. The computer program includes instructions. When the instructions are executed by a processor of an electronic device, the electronic device implements the binding protein production method according to the first aspect of the present disclosure.

[0026] According to a fifth aspect of the present disclosure, a computer system is provided. The computer system comprises: a processor, a memory and a computer program. The computer program is stored in the memory and is configured to be executed by the processor. The computer program comprises instructions for implementing the binding protein generation method according to the first aspect of the present disclosure.

[0027] The present disclosure proposes a method for generating an electron density representation of a binding protein for structural information related to a given target using a conditional diffusion model on a latent space. In the solution proposed in the present disclosure, the relevant concepts of the reciprocal space are creatively introduced to encode and decode the latent space, and at the same time, the method of using guided masks is proposed to control the model to generate higher quality electron density in a targeted manner. Through unlabeled training using a public data set (PDB), a latent space diffusion model that can accurately generate protein electron density with binding characteristics in a specific area is obtained, thereby laying the foundation for the subsequent realization of biological functions and related drug design and development. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present disclosure will be more fully understood from the following detailed description taken in conjunction with the accompanying drawings, in which like elements are numbered in a similar manner, and in which:

[0029] Figure 1 is a flow chart of a method for producing a binding protein according to an embodiment of the present disclosure.

[0030] Figure 2 It is a schematic diagram of the association between the protein generation method and the diffusion model construction method according to an embodiment of the present disclosure.

[0031] Figure 3 Schematic diagram of the relationship between electron density, latent space transformation and diffusion model in the protein generation method according to an embodiment of the present disclosure.

[0032] Figure 4 It is a reference diagram for the method of selecting the three-dimensional reciprocal space coefficients when decoding the latent space.

[0033] Figure 5 The flowchart of the protein generation method according to the embodiment of the present disclosure is in the case of adding a spatial guidance mask.

[0034] Figure 6 is a schematic architecture of a diffusion model according to an embodiment of the present disclosure.

[0035] Figure 7 A loop structure of a protein production method according to an embodiment of the present disclosure is shown.

[0036] Figure 8 The decoding and screening process of the protein production method according to an embodiment of the present disclosure is shown.

[0037] Fig. 9 Target sites for protein 6EXX are shown.

[0038] Fig.10 Comparison of the model generation results with the true residues is shown.

[0039] Fig.11 is a schematic block diagram of a binding protein production system according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0040] Unless otherwise defined, technical and scientific terms used in the present disclosure have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs.

[0041] The technical solution of the present disclosure is further described in detail below through examples and in conjunction with the accompanying drawings. Unless otherwise specified, the methods, tools and materials of the embodiments described below are conventional methods, tools and products that can be obtained through public channels or purchased on the market. Those skilled in the art of the present disclosure will understand that the methods, tools and materials described below are only exemplary and should not be regarded as limiting the scope of the present disclosure.

[0042] From the overall perspective of protein production methods, Figure 1 A flow chart of a method for producing a binding protein according to an embodiment of the present disclosure is shown.

[0043] like Figure 1As shown in , the binding protein generation method 100 according to the present disclosure starts at step S110 , in which the electron density of the target protein is subjected to latent space transformation to obtain electron density latent space transformation data.

[0044] Then, in step S120, the electron density latent space transformation data is input into the electron density latent space diffusion model to generate binding electron density data, wherein the binding protein represented by the binding electron density data can bind to the target protein.

[0045] Those skilled in the art should understand that the "diffusion model" is used in the present disclosure to generate binding proteins. From the context of the present disclosure, it can be seen that the input and output data of the diffusion model are both electron density data. In other words, although the protein generation method of the present disclosure is used to design and generate binding proteins, specifically with respect to the diffusion model of the present disclosure, its output is the electron density data of the generated binding protein that can bind to the target protein.

[0046] It should be noted here that, in the context of the present disclosure, the so-called binding protein capable of binding to the target protein refers to a protein that can form a complementary binding relationship with the target protein in an all-round manner. Although the present disclosure abstracts the design of binding proteins as a problem of geometric completion, the scope is not limited to this. In fact, the diffusion model of the present disclosure can generalize the properties that match the target protein in shape to the matching relationship in physical, chemical and biological properties through learning from the real protein binding electron density data, and realize the required biological function.

[0047] As mentioned above, the difference between the method disclosed in the present invention and the prior art is that the three-dimensional electron density information of the input target protein can be used to generate the corresponding three-dimensional protein electron density associated with it, thereby quickly completing the required macromolecular design in an end-to-end sense. The difference between the method disclosed in the present invention and the prior art, that is, the technical means particularly emphasized in the present invention, will be described later in conjunction with specific embodiments.

[0048] Figure 2 A schematic diagram of the association between the protein generation method and the diffusion model construction method according to an embodiment of the present disclosure is given.

[0049] exist Figure 2 In the association diagram shown, the left path shows the method of constructing a diffusion model, while the right path shows the method of using the diffusion model to produce proteins.

[0050] Figure 2 The process on the right side is actually Figure 1 The method for producing a binding protein according to an embodiment of the present disclosure is shown. Figure 2The process on the left is described in detail below (as part of the method for generating conjugated proteins). Figure 2 It can be seen that both the construction of the diffusion model and the actual generation process of the protein require the use of latent space transformation, which will be introduced in detail below. In the process of building the diffusion model, the protein electron density data in the training sample is used to perform latent space transformation and train the diffusion model. The fully trained diffusion model can be used in combination with the actual generation process of the protein, that is, Figure 2 The process on the right side of the diagram.

[0051] If Figure 2 As a whole method, the whole process is divided into 5 steps (such as Figure 2 Numbers 1-5):

[0052] 1. Perform latent space transformation on the public PDB experimental electron density to obtain unlabeled training data;

[0053] 2. Use the obtained training data to train the generative diffusion model;

[0054] 3. Perform the same latent space transformation on the electron density of the target protein to be designed to form test data;

[0055] 4. Input the test data into the trained diffusion model and let the model output the test results;

[0056] 5. The test results are sampled multiple times and comprehensively evaluated to obtain the final generated electron density of the bound protein.

[0057] A more detailed process of the method for producing the binding protein according to an embodiment of the present disclosure is introduced below.

[0058] Diffusion Model

[0059] The algorithm disclosed in the present invention is based on the diffusion model, which is a new generative model for deep learning that has been proven to be very effective. In general, generative models in deep learning can generate new data samples by learning the probability distribution of training data. Specifically, these models have the ability to map the input random vector to a completely new and reasonable sampled sample, and these generated data sampled samples have similar characteristics and distribution to the original data. Compared with existing generative models, such as variational autoencoders (VAE) and generative adversarial networks (GANs), the diffusion model is a new type of model with better stability and higher generation quality.

[0060] The diffusion model is based on the physical diffusion process, which can be roughly represented as starting from the initial sample x0 and continuously adding noise until the standard Gaussian noise x TThis process is called the forward process:

[0061]

[0062] This process is a Markov process, and the transition probability of each step is determined by the normal distribution. From the application point of view, what the disclosed solution really needs is the reverse process of the sampling process, that is, gradually filtering out the noise from the Gaussian random distribution and finally obtaining the sample generation process:

[0063] p:x T →x T-1 →…→x1→x0~P data

[0064] This is exactly the transformation that the deep generative model needs to learn, and it is also the part that the diffusion model needs to train and learn. It can be found that what the diffusion model needs to learn at each step is a model that predicts how much noise there is in that step. Its specific loss function can be derived using the classic variational lower bound method (VLB, Variational Lower Bound) in probabilistic learning, such as the most commonly used DDPM framework (see Ho, Jonathan, Ajay Jain, and Pieter Abbeel. "Denoising diffusion probabilistic models." Advances in Neural Information Processing Systems 33 (2020): 6840-6851, hereby incorporated by reference into the present disclosure in its entirety, making it part of the content of the present disclosure), a simplified loss function is:

[0065]

[0066] Among them, z θ is the diffusion model used in the present disclosure, t is the time step, and the integers [1, T] are uniformly valued. is a parameter related to the variance at each time step, z t is a random number at time t, and x0 is the training data that is not contaminated by noise. By optimizing the parameter θ, the diffusion model can be trained to simulate the generation of data using the denoising process. When sampling a new data sample, a Gaussian distributed noise x is generated first. T , and then iterates time step by time, gradually removing the noise, and finally getting the required sampling points. This is the principle of sampling in the diffusion model.

[0067] Electron density

[0068] By analyzing the diffraction data of crystals to X-rays, the structure of crystalline materials at the atomic level can be reversely inferred to obtain detailed structural data. In addition to obtaining the parameters of the unit cell, the X-ray diffraction method can also fit the phase angle of the diffraction points based on the overall information of the collected diffraction points through methods such as the direct method. In this way, the electron density function in the unit cell can be calculated through Fourier transform, and the atomic arrangement in various complex molecules can be clearly displayed through the electron density map. From this perspective, the electron density is the direct result of the experiment and contains the most original and richest structure-related information. The three-dimensional molecular structure generally used in chemistry and biology is the derived result obtained by fitting the electron density function through structure building software, so there may be a certain amount of information loss.

[0069] Since the 1930s, X-ray diffraction has been used to study the structure of biological macromolecules. Scientists have determined the structures of a variety of biological macromolecules (such as proteins, nucleic acids, and complex compositions), and have compiled the structural data obtained by diffraction and analysis into the PDB (Protein Data Bank), which is accessible free of charge to scientists around the world. Currently, the PDB contains about 200,000 structures, and each structure is assigned a unique identifier, called a PDB ID. Currently, the structures obtained by X-ray diffraction in the PDB will have an accompanying MTZ diffraction file. Through the corresponding calculation programs (such as cctbx, Coot, etc.), the diffraction data can be converted from reciprocal space to real space and an electron density representation can be obtained.

[0070] The algorithm disclosed in the present invention directly generates the electron density of the binding protein from the electron density of the target, and is directly based on the electron density representation, which is close to the experimental results and avoids the loss of information conversion. In addition, the algorithm disclosed in the present invention uses the electron density information of the PDB database for training, because the overall algorithm framework of the present invention is a generative diffusion model, so no additional annotation is required on the training data.

[0071] Latent Space Transformation

[0072] The algorithm disclosed in the present invention is based on a stable diffusion model, or a latent diffusion model. For example, see Rombach, Robin, et al. "High-resolution image synthesis with latent diffusion models." Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2022. By reference here, the entire content of the above document is incorporated into the present disclosure to make it a part of the content of the present disclosure. It is a variant of the diffusion model. The original model can generate images corresponding to text according to the content of the text. The present disclosure also uses the idea of ​​latent space to encode and decode the original electron density space. Here, borrowing the principle of X-ray diffraction in crystallography, the electron density of a given area is transformed into a reciprocal space by three-dimensional Fourier transform, and the reciprocal frequency space is truncated to discard the high-frequency parts that are not particularly important. Generally speaking, the area of ​​concern in the present disclosure is a cube with a side length of approximately The entire sampling point is about 1M to 10M. At this time, taking 50K to 300K independent coefficient components in the reciprocal space can achieve a relatively good reconstruction effect. Figure 3 A schematic diagram is used to show the relationship between electron density, latent space transformation and model.

[0073] The fast Fourier transform (FFT) used in the present disclosure is a three-dimensional FFT (3DFFT), which is specifically expressed by the following formula:

[0074]

[0075] For FFT, assuming that the array f is an N×N×N real input, the output coefficients g have the following conjugation symmetry:

[0076] g[i,j,k]=g * [i mod N,j mod N,k mod N]

[0077] Where x mod N∈[0,N). Due to the existence of symmetry, only half of the coefficients need to be taken, or to be more precise, those coefficients where 2i≤N. For a given reciprocal space length L, first generate all index sets {(i,j,k)|i,j,k∈[0,N)}, and then calculate the folded index i',j',k' according to the following formula:

[0078] i′=min{i,Ni},j′=min{j,Nj},k′=min{k,Nk}

[0079] And follow i' 2 +j' 2 +k' 2 Stable sorting, take the coefficients corresponding to the first L indexes. When decoding, the coefficients at the (Ni, Nj, Nk) position must also be filled in according to the complex conjugate symmetry. The method of taking coefficients can refer to Figure 4 . Figure 4 The four corners on the left side of the cube space are the coefficients that need to be taken, and the four corners on the right side are the coefficients that can be filled in according to symmetry.

[0080] Specifically, in Figure 1 In the step S110 shown, the electron density of the target protein is subjected to latent space transformation to obtain the electron density latent space transformation data, which may further include: the electron density of the target protein is subjected to latent space encoding, and the transformation is performed from real space to reciprocal latent space to obtain reciprocal latent space features. The latent space encoding here may be some appropriate data modality adaptation transformation. Specifically in the example disclosed in the present invention, it may be a Fourier transform, such as FFT, and more specifically 3DFFT. Of course, in practice, in order to adapt the model, some other appropriate data modality adaptation transformations may also be used.

[0081] Guided Mask

[0082] In general generative models, all generated elements are generally equal, or in other words, their generated results are equally weighted. For example, when using a stable diffusion model to generate a painting, the algorithm will generate it as a whole, and generally will not pay specific attention to certain parts of the painting. In the model of the present disclosure, in order to improve the generation quality of certain specific areas, a guiding mask is used to allow the model to focus on areas that are more difficult to generate with high quality. The guiding mask is only used for the input of the model, guiding the area where the model focuses, and there is no masked part in the output. In order to adapt to the denoising operation of the diffusion model in the reciprocal space of the entire model, the present disclosure will use the latent space encoding module to convert the mask specified in the spatial domain to the reciprocal space, and organically integrate it with other inputs. The specific steps can be obtained by Figure 5 Flowchart representation of .

[0083] Specifically, the spatial guidance mask can be represented by a Boolean value at each spatial point. For example, a value of 1 indicates that the model needs to pay special attention to the generation of this position, otherwise it is 0. Other values ​​can also be used to represent specific meanings. The mask can be determined according to the needs of the actual task. For example, if the quality of generation of a certain residue position is not particularly good, a guiding mask can be added to the range of one or several specific residue spaces. The mask will also be encoded into the reciprocal space through the latent space:

[0084] F reciprocal =3DFFT(Mask)

[0085] Then it is fused with other features and finally input into the diffusion model for generation.

[0086] That is to say, Figure 1 The method 100 further includes: similar to step S110, obtaining a latent space mask by performing a latent space transformation on the guiding mask; then, in step S120, performing feature fusion on the latent space mask and the electron density latent space transformation data, inputting the feature into the electron density latent space diffusion model, and generating the electron density of the binding protein that can bind to the target protein.

[0087] Similarly, the step of transforming the guided mask into a latent space to obtain a latent space mask includes: encoding the guided mask into a latent space, transforming the real space into the reciprocal latent space, and obtaining the reciprocal latent space mask. The latent space encoding here can be some appropriate data modality adaptation transformation. Specifically in the example disclosed in the present invention, it can be a Fourier transform, such as FFT, more specifically 3DFFT. Of course, in practice, in order to adapt the model, some other appropriate data modality adaptation transformations can also be used.

[0088] Model framework

[0089] The overall algorithm framework of the present disclosure is a diffusion model, and its specific architecture is as follows: Figure 6 shown.

[0090] The structure of the algorithm framework is described in detail below.

[0091] First, from the previous description, we can see that the diffusion model is essentially a diffusion result x at the current time step t. t The model for noise estimation, whose output is the current step noise z t , and use this to calculate the denoising result x t-1 , and at the same time, it is used as the input for the next step and the cycle repeats T times before stopping.

[0092] That is to say, Figure 1Step S120 includes: using the electron density latent space diffusion model to calculate the diffusion result x of the current time step t t Perform noise estimation and output the current step noise z t , and the denoising result x is calculated from this t-1 , as the diffusion result input for the next step, and the cycle is repeated again and again, stopping after T times.

[0093] The initial input of the loop is the random sampling noise x in the latent space T , and finally undergoes a series of denoising cycles to obtain the generated result x0. For a specific time step t, the algorithm framework of the present disclosure includes the following parts:

[0094] Input part: In addition to the current step diffusion result x just mentioned t In addition to the time step t, because the algorithm of the present disclosure generates the protein electron density that can bind to the target based on the density of the area to be bound near the target, the present disclosure also includes the electron density c around the binding target as a condition for generating the model. At the same time, as mentioned above, in order to make the model more focused, the present disclosure introduces a spatial guiding mask m.

[0095] Latent space encoding part: The combined target electron density c and the guiding mask m are both features in the real space. In order to be able to input the model, the present disclosure proposes to use a latent space encoder to encode the combined target electron density c and the guiding mask m:

[0096] c′=3DFFT(c),m′=3DFFT(m)

[0097] Thus, the reciprocal latent space feature c' combined with the target and the mask m' on the reciprocal latent space are obtained.

[0098] Electron density latent space diffusion model: First, take the inputs mentioned above, including the diffusion result x t , time step t, combine the target reciprocal feature c' and the reciprocal mask m' to perform a feature fusion. The fusion here can vary according to different model settings. There are two typical fusion methods. One is vector concatenation, including direct concatenation (R d ×R d →R 2d ), or splicing of extended dimensions (R d ×R d →R d×2) and so on. Another way is to use the attention mechanism (see Vaswani, Ashish, et al. "Attention is all you need." Advances in neural information processing systems 30 (2017), which is incorporated into the present disclosure by reference and becomes part of the present disclosure) to extract the association of all features. That is, the diffusion result x at the current time step t needs to be t , the current time step t, the electron density latent space transformation data and the latent space mask are feature fused and input into the electron density latent space diffusion model.

[0099] Next, it is the core module of the diffusion model. Generally speaking, this module includes an autoencoder (AE, AutoEncoder) structure, which includes continuous down-sampling feature extraction and subsequent up-sampling generation modules. A more common implementation method is to use the U-Net module commonly used in common segmentation models, because it can more effectively capture feature associations at different resolutions, and can flexibly configure different numbers of layers and resolutions. For U-Net, for example, see Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation." Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015. Springer International Publishing, 2015. The entire contents of the above-mentioned documents are incorporated into the present disclosure by reference, making them part of the contents of the present disclosure.

[0100] Output part: The output of the model is the predicted current step noise z in the latent space t . Get the noise z t After that, combine the current step diffusion result x with the input tAccording to the principle of the diffusion model (for example, refer to the previously mentioned literature Ho, Jonathan, Ajay Jain, and Pieter Abbeel. "Denoising diffusion probabilistic models." Advances in Neural Information Processing Systems 33(2020): 6840-6851), the denoising result x can be obtained by sampling the normal distribution. t-1 .

[0101] Model Training

[0102] According to an embodiment of the present disclosure, public PDB data is used for training. That is, the electron density latent space diffusion model is constructed by training with PDB data as training data. In a preferred embodiment of the present disclosure, a structure with high-resolution X-ray electron density is selected. That is, the PDB data used as training data includes high-resolution X-ray electron density data. At the same time, the generated result is generally a short secondary structure fragment, so for a PDB, multiple sets of input data may be divided as needed. Overall, the training data of the embodiments of the present disclosure is around 150,000 to 200,000. Each set of data contains the main side chains to be generated, the corresponding peripheral electron density information (used as target points) and the guiding mask (mainly the side chain part).

[0103] More specifically, in model training, an example of constructing training data based on PDB data can be as follows: the required secondary structure is cut out, and the cut-out part is the binding protein, which can distinguish the main chain and the side chain; the remaining part is used as the target structure, in which the division of the main chain and the side chain is not concerned, but as the overall background. The input of the model is the electron density data of the target, that is, the background with the secondary structure cut out; the small segment of the secondary structure cut out is used as the target that the model needs to generate, that is, the binding protein to be generated. As mentioned above, multiple groups of input data can be split as needed, that is, different secondary structures are cut out of each group to form multiple groups of training data.

[0104] In addition, the main reason for adding guiding masks to side chains is that the quality of side chain generation is often lower than that of the main chain.

[0105] The model parameters of the present disclosure are about 200M. In the specific training environment, 8 A10080G GPUs are used for training. According to an embodiment of the present disclosure, the Adam optimizer is used, and the learning rate is kept at 10 -5The batch size is 32. In the embodiment of the present disclosure, about 80 full sample training processes (epochs) are trained, and the overall training time is greater than 100 hours (>100h). Due to the use of the diffusion model, the training process is more stable than adversarial training such as adversarial generative network (GAN), and there is basically no model crash caused by gradient divergence.

[0106] Algorithm operation and preliminary verification results

[0107] When using the trained model for inference, first use Figure 7 The T-step cycle shown above starts with a random sample of inverse latent spatial noise x T Gradually remove the noise and get the final inverse hidden space generation result x0. Of course, you can use the noise x sampled multiple times T , input into the electron density latent space diffusion model to generate model output results with differences, ensuring the diversity of generated results.

[0108] Then, if Figure 8 As shown, the latent space denoising result x0 is decoded in the latent space, and the generated binding protein electron density is obtained. The multiple generated results can be further analyzed, identified and evaluated to further screen the binding proteins that meet the biological function requirements.

[0109] Specifically, in Figure 1 In step S120, the step of inputting the electron density latent space transformation data into the electron density latent space diffusion model to generate the bound electron density data may include: inputting the electron density latent space transformation data into the electron density latent space diffusion model; then, performing latent space decoding on the output result of the electron density latent space diffusion model, transforming from the reciprocal latent space to the real space, and generating the electron density of the bound protein that can bind to the target protein.

[0110] like Figure 8 As shown in , the latent space decoding of the output result of the electron density latent space diffusion model includes performing appropriate data mode adaptation transformation on the output result of the electron density latent space diffusion model. Specifically in this example of the present disclosure, an inverse Fourier transform, such as IFFT, more specifically 3DIFFT, can be performed. Of course, in practice, some other appropriate data mode adaptation transformations can also be performed.

[0111] For protein 6EXX (such as Fig. 9 The helical part of the sample was tested 1000 times and the recognition statistics were calculated. The results show that compared with the ground truth (GT) residues, Fig.10As shown, the model of the present disclosure can achieve a recall rate of 92.3%.

[0112] Compared with the prior art, the present disclosure proposes the following technical means:

[0113] 1. The present disclosure uses a generative model based on electron density representation, and emphasizes the goal of generating protein electron density that is combined with a given target electron density.

[0114] 2. The core algorithm of the present invention uses a diffusion model framework based on U-Net, which has much better training stability than traditional generation models such as GAN and is easier to use.

[0115] 3. The data processing disclosed in the present invention uses latent space encoding and decoding, the idea of ​​which is borrowed from the reciprocal space theory of X-ray crystallography, which can effectively reduce the amount of data and improve the model throughput.

[0116] 4. In order to achieve practical generation quality, the present disclosure creatively proposes the concept of guiding mask.

[0117] 5. Although the concept of the structural submodules of the core model of the present invention (such as diffusion model, U-Net, etc.) is not disclosed for the first time in the present invention, combining them and applying them to the method for generating a combined protein is disclosed for the first time in the present invention.

[0118] In addition, for some of the technical means mentioned above, those skilled in the art should also note:

[0119] 1. Diffusion models, including latent diffusion models (see Rombach, Robin, et al. "High-resolution image synthesis with latent diffusion models." Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2022), are not disclosed for the first time in the present disclosure, but the embodiments of the present disclosure make a number of improvements thereto. Including:

[0120] a. Application to the generation of three-dimensional electron density: The original model input is text and the output is an image; this is obviously very different from the target electron density input and protein electron density binding disclosed in the present invention. In order to achieve this goal, the embodiments of the present invention have made a very large transformation at the code level. In general, the model of the present invention only draws on the ideas of existing models, including the use of the existing U-Net structure.

[0121] b. Introducing reciprocal space as latent space: The latent diffusion model in the above literature has its latent space obtained by the image pre-training model. However, the latent space disclosed in this paper is aimed at the application purpose of electron density generation, and the reciprocal space in crystallography is creatively used. On the one hand, it is more theoretically based, and on the other hand, it is convenient to integrate the data processing flow of electron density.

[0122] c. Use of guided masks: This is an idea independently proposed in this disclosure, which is mainly to focus on the capabilities of the model, emphasize the generation of the side chain part, and guide the model to generate higher quality models.

[0123] 2. The idea of ​​generating a binding protein electron density based on the target electron density is disclosed for the first time in this disclosure. No predecessors have used similar ideas before. The whole idea and the process of getting results from the diffusion model are a process pipeline that is very different from traditional protein design. Therefore, the successful development of related algorithms and technologies based on this idea is itself a very big innovation.

[0124] In summary, the present disclosure proposes a process for generating binding proteins using a generative model. The process uses public data and does not require manual annotation to generate results, and is very practical.

[0125] Those skilled in the art will appreciate that, based on the binding protein production method disclosed herein, a binding protein production system can be developed. Fig.11 11 is a schematic block diagram of a binding protein production system according to an embodiment of the present disclosure. Specifically, the binding protein production system 1100 includes an acquisition module 1110 and a processing module 1120 .

[0126] The acquisition module 1110 is used to acquire the electron density of the target protein.

[0127] The processing module 1120 is used to generate a binding protein that can bind to the target protein according to the electron density of the target protein. Fig.11 As shown in , the processing module 1120 may further include: a latent space transformation submodule 1121, used to perform latent space transformation on the electron density of the target protein to obtain electron density latent space transformation data; and a generation submodule 1122, used to input the electron density latent space transformation data into the electron density latent space diffusion model to generate bound electron density data, wherein the bound protein represented by the bound electron density data is capable of binding to the target protein.

[0128] Those skilled in the art should understand that the above-mentioned binding protein generation system can be implemented by a computer. For example, the following operations can be implemented by executing a computer program: obtaining the electron density of the target protein; generating a binding protein capable of binding to the target protein according to the electron density of the target protein, and the operation further includes the following sub-operations: performing a latent space transformation on the electron density of the target protein to obtain electron density latent space transformation data; and inputting the electron density latent space transformation data into an electron density latent space diffusion model to generate binding electron density data, wherein the binding protein represented by the binding electron density data can bind to the target protein.

[0129] That is to say, although in the binding protein production system according to the present disclosure, the various operating functions corresponding to the production method are described as modules or sub-modules, those skilled in the art should understand that such modules or sub-modules may be circuit components or other physical components, or may not be composed of circuit components or other physical components, but may be functional modules constructed by computer programs.

[0130] In addition, it should be recognized by those skilled in the art that the method of the present disclosure can be implemented as a computer program. As described above in conjunction with the accompanying drawings, the method of the above embodiment is executed by one or more programs, and the instructions in the program cause the computer or processor to execute the algorithm described in conjunction with the accompanying drawings. These programs can be stored and provided to the computer or processor using various types of non-transient computer-readable media. Non-transient computer-readable media include various types of tangible storage media. Examples of non-transient computer-readable media include magnetic recording media (such as floppy disks, tapes, and hard disk drives), magneto-optical recording media (such as magneto-optical disks), CD-ROMs (compact disk read-only memories), CD-Rs, CD-R / Ws, and semiconductor memories (such as ROMs, PROMs (programmable ROMs), EPROMs (erasable PROMs), flash ROMs, and RAMs (random access memories)). Further, these programs can be provided to the computer using various types of transient computer-readable media. Examples of transient computer-readable media include electrical signals, optical signals, and electromagnetic waves. Transient computer-readable media can be used to provide programs to the computer through wired communication paths or wireless communication paths such as wires and optical fibers.

[0131] For example, according to one embodiment of the present disclosure, a non-transitory computer-readable storage medium may be provided for storing a computer program, wherein the computer program includes instructions that, when executed by a processor of an electronic device, enable the electronic device to implement the binding protein generation method as described above.

[0132] In addition, according to the content disclosed in the present disclosure, a computer system can also be provided, the computer system comprising: a processor; a memory; and a computer program. The computer program is stored in the memory and is configured to be executed by the processor. The computer program includes instructions for implementing the above-described binding protein generation method.

[0133] The implementation methods of the present disclosure are not limited to the above-mentioned embodiments. Without departing from the spirit and scope of the present disclosure, ordinary technicians in this field can make various changes and improvements to the present disclosure in form and details, and these are considered to fall within the protection scope of the present disclosure.

Claims

1. A method for producing a binding protein, characterized in that: The method comprises: Performing latent space transformation on the electron density of the target protein to obtain electron density latent space transformation data; and The electron density latent space transformation data is input into the electron density latent space diffusion model to generate binding electron density data, wherein the binding protein represented by the binding electron density data can bind to the target protein.

2. The method according to claim 1, characterized in that The step of performing latent space transformation on the electron density of the target protein to obtain electron density latent space transformation data comprises: The electron density of the target protein is encoded in latent space and transformed from real space to reciprocal latent space to obtain the reciprocal latent space features.

3. The method according to claim 2, characterized in that The step of encoding the electron density of the target protein in hidden space includes: Perform data modality adaptation transformation on the electron density of the target protein.

4. The method according to claim 3, characterized in that The step of encoding the electron density of the target protein in hidden space includes: Perform a Fourier transform on the electron density of the target protein.

5. The method according to claim 1, characterized in that The step of inputting the electron density latent space transformation data into the electron density latent space diffusion model to generate combined electron density data includes: inputting the electron density latent space transformation data into the electron density latent space diffusion model; and The output results of the electron density latent space diffusion model are decoded in latent space, transformed from reciprocal latent space to real space, and the electron density of the binding protein that can bind to the target protein is generated.

6. The method according to claim 5, characterized in that The latent space decoding of the output result of the electron density latent space diffusion model includes: The output results of the electron density latent space diffusion model are transformed into data mode adaptation.

7. The method according to claim 6, characterized in that The latent space decoding of the output result of the electron density latent space diffusion model includes: Perform an inverse Fourier transform on the output of the electron density latent space diffusion model.

8. The method according to claim 1, characterized in that The method further comprises: Obtaining a latent space mask by transforming the guided mask into a latent space; and The above-mentioned inputting of the electron density latent space transformation data into the latent space diffusion model to generate the combined electron density data further includes: feature fusion of the latent space mask and the electron density latent space transformation data, inputting the latent space transformation data into the electron density latent space diffusion model, and generating the electron density of the binding protein that can bind to the target protein.

9. The method according to claim 8, characterized in that The step of obtaining a latent space mask by performing a latent space transformation on the guided mask includes: performing latent space encoding on the guided mask, transforming from real space to reciprocal latent space, and obtaining a reciprocal latent space mask.

10. The method according to claim 9, characterized in that The step of encoding the guiding mask into a latent space comprises: Perform data modality adaptation transformation on the guided mask.

11. The method according to claim 10, characterized in that The step of encoding the guiding mask into a latent space comprises: Take the Fourier transform of the guided mask.

12. The method according to claim 1, characterized in that The step of inputting the electron density latent space transformation data into the electron density latent space diffusion model to generate combined electron density data includes: The diffusion result x of the current time step t is calculated using the electron density latent space diffusion model. t Perform noise estimation and output the current step noise z t , and the denoising result x is calculated from this t-1 , as the diffusion result input for the next step, and the cycle is repeated again and again, stopping after T times.

13. The method according to claim 12, characterized in that The initial input of the loop is the random sampling noise x in the latent space T , the final output of the loop is the denoising result x0.

14. The method according to claim 8, characterized in that The step of inputting the electron density latent space transformation data into the electron density latent space diffusion model to generate combined electron density data includes: The diffusion result x of the current time step t t , the current time step t, the electron density latent space transformation data and the latent space mask are feature fused and input into the electron density latent space diffusion model.

15. The method according to claim 13, characterized in that The generating of the binding electron density data comprises: performing latent space decoding on the denoising result x0 to obtain the electron density of the binding protein.

16. The method according to claim 13, characterized in that The method further comprises: using multiple sampling noise x T , input into the electron density latent space diffusion model to generate differentiated model output results for further analysis, identification and evaluation.

17. The method according to claim 1, characterized in that The electron density latent space diffusion model is constructed by training with PDB data as training data.

18. The method according to claim 17, characterized in that The PDB data used as training data includes high-resolution X-ray electron density data.

19. A binding protein production system, characterized in that: The system comprises: An acquisition module is used to obtain the electron density of the target protein; A processing module is used to generate a binding protein capable of binding to the target protein according to the electron density of the target protein, and the processing module further includes the following submodules: A latent space transformation submodule, used to perform latent space transformation on the electron density of the target protein to obtain electron density latent space transformation data; and The generation submodule is used to input the electron density latent space transformation data into the electron density latent space diffusion model to generate binding electron density data, wherein the binding protein represented by the binding electron density data can bind to the target protein.

20. A computer-implemented system for producing a binding protein, wherein the computer-implemented system implements the following operations by executing a computer program: Obtain electron density of target protein; According to the electron density of the target protein, a binding protein capable of binding to the target protein is generated, and the operation further includes the following sub-operations: Performing latent space transformation on the electron density of the target protein to obtain electron density latent space transformation data; and The electron density latent space transformation data is input into the electron density latent space diffusion model to generate combined electron density data, where: The binding protein represented by the binding electron density data is capable of binding to the target protein.

21. A non-transitory computer-readable storage medium for storing a computer program, the computer program comprising instructions which, when executed by a processor of an electronic device, cause the electronic device to implement the binding protein production method according to any one of claims 1 to 18.

22. A computer system, comprising: processor; Memory; and A computer program, wherein the computer program is stored in the memory and is configured to be executed by the processor, the computer program comprising instructions for implementing the binding protein production method according to any one of claims 1 to 18.

Images