Machine learning driven dna multiscale dynamic conformational reconstruction method of angularly correlated data

Abstract

The present application relates to the technical field of bioinformatics and computational molecular biology, and specifically discloses a DNA multi-scale dynamic conformation reconstruction method driven by machine learning and angle correlation data. The technical scheme of the present application acquires time-resolved angle correlation scattering data and extracts a time sequence feature matrix, uses a multi-scale coding network to extract global topology and local conformation time sequence features and generate an initial conformation trajectory, and then calculates the multiple differences in scattering intensity and experimental data on multiple scales by means of a scattering forward model, constructs a comprehensive loss function for iterative optimization, solves the problem that traditional inversion modeling relies on strong prior assumptions and is difficult to constrain multi-scale structures simultaneously under single low-dimensional observation, and solves the problem that molecular dynamics simulation and experimental data lack matching feedback, resulting in a lack of experimental constraints in reconstruction and high computational cost, and realizes the direct extraction of dynamic evolution information from angle correlation scattering data and the output of cross-scale consistent DNA dynamic conformation reconstruction results.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and computational molecular biology, and in particular to a method for multi-scale dynamic conformation reconstruction of DNA driven by machine learning-based angular correlation data. Background Technology

[0002] The three-dimensional conformation of DNA molecules in solution spans multiple orders of magnitude, from base pairs to chromatin fibers. Accurate reconstruction of their multi-scale dynamic structure is crucial for revealing gene expression regulation mechanisms. Angular correlation measurements, such as small-angle X-ray scattering curves, are essentially spherical averages of structural factors; phase and orientation information are completely lost during experimental acquisition. Traditional inversion modeling relies on strong prior assumptions or simplified physical models, making it difficult to simultaneously constrain local fine conformational parameters and global topological features from a single type of low-dimensional observation. This results in significant deviations in cross-scale consistency of reconstruction results. Furthermore, existing reconstruction frameworks are mostly designed for equilibrium ensemble averaging, failing to effectively extract and utilize the dynamic evolution information in time-resolved angular correlation data. While traditional molecular dynamics simulations can generate continuous conformational trajectories, computational costs increase dramatically with system scale, and the lack of direct matching and feedback correction mechanisms between simulation results and experimental angular correlation data hinders the dynamic reconstruction process from being experimentally constrained and difficult to complete within a reasonable timeframe, severely limiting the depth of research on the dynamic behavior of DNA in complex biological systems.

[0003] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a machine learning-driven method for multi-scale dynamic conformation reconstruction of DNA based on angular correlation data. The technical solution of this method is as follows: Time-resolved angular correlation scattering data of DNA molecules is obtained, wherein the time-resolved angular correlation scattering data includes scattering intensity curves of multiple time frames; The scattering intensity curves of each time frame are preprocessed, and the one-dimensional scattering feature vectors of each scattering intensity curve are extracted and arranged in chronological order to form a temporal feature matrix; The temporal feature matrix is ​​input into a multi-scale coding network to extract global topological temporal features and local conformational temporal features and map them to the latent space to obtain temporal latent variables. The temporal latent variables are input into the temporal generation network to generate three-dimensional all-atom conformations for each time frame, thus obtaining the initial conformation trajectory. The full atomic space coordinates of each time frame are input into the scattering forward model. The first difference between the reconstructed scattering intensity and the scattering intensity curve of the corresponding time frame, the second difference between the coordinate changes of adjacent time frames, the third difference between the global topological morphology parameters of the initial conformation trajectory and the global topological temporal features, and the fourth difference between the local fine conformation parameters and the local conformation temporal features are calculated. A comprehensive loss function is constructed based on the first difference, the second difference, the third difference, and the fourth difference. The initial conformational trajectory is iteratively optimized based on the comprehensive loss function until the comprehensive loss function converges, and the optimized conformational trajectory is output as the result of the multi-scale dynamic conformational reconstruction of the DNA molecule.

[0005] The technical solution of this invention acquires time-resolved angular correlation scattering data and extracts temporal feature matrices. It then uses a multi-scale coding network to extract global topological and local conformational temporal features and generate initial conformational trajectories. Subsequently, it uses a forward scattering model to calculate the multiple differences between the reconstructed scattering intensity and experimental data at multiple scales, constructs a comprehensive loss function for iterative optimization, and solves the problems of traditional inversion modeling relying on strong prior assumptions and making it difficult to simultaneously constrain multi-scale structures under a single low-dimensional observation, as well as the problems of lack of matching feedback between molecular dynamics simulations and experimental data, resulting in a lack of experimental constraints and excessively high computational costs for reconstruction. This invention enables the direct extraction of dynamic evolution information from angular correlation scattering data and outputs consistent DNA dynamic conformational reconstruction results across scales.

[0006] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0008] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating an embodiment of a machine learning-driven DNA multi-scale dynamic conformation reconstruction method for angular correlation data according to the present invention. Detailed Implementation

[0009] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0010] Figure 1 This diagram illustrates a flowchart of an embodiment of a machine learning-driven DNA multi-scale dynamic conformation reconstruction method for angular correlation data provided by the present invention, executed by a control terminal. Figure 1 As shown, the machine learning-driven DNA multi-scale dynamic conformation reconstruction method for angular correlation data includes the following steps: S1. Obtain time-resolved angular correlation scattering data of DNA molecules, wherein the time-resolved angular correlation scattering data includes scattering intensity curves of multiple time frames.

[0011] DNA molecules refer to biological macromolecules composed of deoxyribonucleotides linked by phosphodiester bonds, exhibiting a multi-scale three-dimensional structure in solution, ranging from local base pair conformations to the global topology of chromatin fibers. For example, a nucleosome core DNA molecule containing 146 base pairs exhibits a curved, coiled conformation in solution. Time-resolved angular correlation scattering data refers to measurement data collected at multiple consecutive time points reflecting the angular distribution of the scattering signal of DNA molecules in solution. For example, small-angle X-ray scattering signals of a nucleosome core DNA molecule solution are collected at 100 time points (0 seconds, 1 second, 2 seconds up to 100 seconds) using a synchrotron radiation source, obtaining scattering data for 100 time frames.

[0012] The scattering intensity curve refers to the distribution curve of the scattered signal intensity as a function of the scattering vector amplitude within a single time frame. For example, within a 1-second time frame, the scattering intensity curve of the nucleosome core DNA molecule records the scattering vector amplitude q from 0.01 Å. -1 up to 0.5Å -1 Count the scattered photons at each sampling point within the range.

[0013] Specifically, time-resolved angle-correlated scattering data of DNA molecules is obtained. The time-resolved angle-correlated scattering data contains scattering intensity curves of multiple time frames. The scattering intensity curves of each time frame are acquired at multiple consecutive time points through small-angle X-ray scattering experiments. During the acquisition process, the distribution curve of the scattering signal intensity as a function of the scattering vector amplitude is recorded for each time frame. Each time frame has an equal time interval. The scattering intensity curves of all time frames together constitute the time-resolved angle-correlated scattering data.

[0014] S2. Preprocess the scattering intensity curves of each time frame, extract the one-dimensional scattering feature vectors of each scattering intensity curve, and arrange them in chronological order to form a time-series feature matrix.

[0015] Here, a one-dimensional scattering feature vector refers to a data representation extracted from a single scattering intensity curve, consisting of multiple values ​​arranged in a fixed order. For example, 200 intensity values ​​are uniformly sampled from the scattering intensity curve of a 1-second time frame, and these intensity values ​​are arranged into a 200-dimensional vector according to the scattering vector amplitude q from smallest to largest. A temporal feature matrix refers to a two-dimensional array formed by stacking the one-dimensional scattering feature vectors of multiple time frames in chronological order. For example, the 200-dimensional one-dimensional scattering feature vectors corresponding to each of the 100 time frames are arranged row by row in chronological order from 0 seconds to 100 seconds, forming a 100-row, 200-column temporal feature matrix.

[0016] Specifically, the scattering intensity curves of each time frame are preprocessed, including background subtraction, normalization, and noise filtering. One-dimensional scattering feature vectors are extracted from each scattering intensity curve after background subtraction, normalization, and noise filtering. The one-dimensional scattering feature vectors are composed of the intensity values ​​of each scattering intensity curve at multiple scattering vector amplitude sampling points arranged in ascending order of scattering vector amplitude. The one-dimensional scattering feature vectors of each time frame are stacked row by row in chronological order to form a temporal feature matrix in the form of a two-dimensional array.

[0017] S3. Input the temporal feature matrix into a multi-scale coding network, extract global topological temporal features and local conformational temporal features, and map them to the latent space to obtain temporal latent variables.

[0018] Among them, multi-scale coding network refers to a neural network model that can simultaneously extract features of different spatial scales from input data; for example, the temporal feature matrix is ​​input into a convolutional neural network containing a parallel dual-branch structure, and the network captures temporal features of global topological scale and local conformational scale respectively through convolutional kernels of different sizes.

[0019] Global topological temporal features refer to the feature representations reflecting the changes in the overall morphology of DNA molecules, such as bending, twisting, and end-to-end distance, over time. For example, the first branch of the convolutional neural network extracts from the temporal feature matrix the overall morphological temporal change pattern of the nucleosome core DNA molecule, showing the bending angle changing from 65 degrees to 82 degrees and the end-to-end distance shrinking from 55 Å to 48 Å over 100 time frames. Local conformational temporal features refer to the feature representations reflecting the changes in the small-scale fine structure of DNA molecules, such as the relative orientation of base pairs, helical parameters, and dihedral angles of the backbone, over time. For example, the second branch of the convolutional neural network extracts from the temporal feature matrix the fluctuating temporal patterns of the twisting angle, helical ascent angle, and roll angle of each base step of the nucleosome core DNA molecule over 100 time frames.

[0020] In this context, the latent space refers to the low-dimensional continuous vector space to which the encoding network maps the input data, where features at different scales are coupled and encoded into a unified latent representation. For example, global topological temporal features and local conformational temporal features are jointly mapped into a 128-dimensional latent space vector. Temporal latent variables refer to the latent variable representations obtained in the latent space that contain cross-scale coupling information and evolve over time. For example, the 128-dimensional latent space vectors corresponding to each of the 100 time frames are arranged in chronological order to form a 100-row, 128-column temporal latent variable matrix.

[0021] Specifically, the temporal feature matrix is ​​input into a multi-scale coding network, which includes a first coding branch and a second coding branch set in parallel. The receptive field of the convolutional kernel of the first coding branch is larger than that of the convolutional kernel of the second coding branch. Global topological temporal features are extracted through the first coding branch, and local conformational temporal features are extracted through the second coding branch. The global topological temporal features and local conformational temporal features are concatenated and fused in the latent space to obtain temporal latent variables.

[0022] S4. Input the temporal latent variables into the temporal generation network to generate three-dimensional all-atom conformations for each time frame, and obtain the initial conformation trajectory.

[0023] The temporal generative network refers to a recurrent neural network model that generates a sequence of three-dimensional spatial coordinates frame by frame, using temporal latent variables as conditional inputs. For example, a 100-row, 128-column temporal latent variable matrix is ​​input row by row into a recurrent generative network based on gated recurrent units, and the network outputs the three-dimensional atomic coordinates of each time frame sequentially. The three-dimensional full-atom conformation refers to the complete set of coordinates of all non-hydrogen atoms in a DNA molecule in three-dimensional space. For example, the nucleosome core DNA molecule contains approximately 4000 non-hydrogen atoms, and the three-dimensional full-atom conformation of each time frame is the set of three-dimensional coordinates corresponding to these atoms in 100 time frames.

[0024] The initial conformation trajectory refers to the sequence of all three-dimensional full-atom conformations of all time frames directly generated by the temporal generation network in chronological order. For example, the recursive generation network outputs the three-dimensional full-atom coordinate set of each of the 100 time frames in sequence, forming an initial conformation trajectory from 0 seconds to 100 seconds.

[0025] Specifically, the temporal latent variables are input into the temporal generation network, which adopts a recursive generation network structure based on an autoregressive strategy. The network generates the three-dimensional full-atom conformation of each time frame in chronological order, starting from the first time frame. When generating the three-dimensional full-atom conformation of the current time frame, the recursive generation network uses the spatial coordinates of the terminal bases generated in the previous time frame as the starting conformation anchor point of the current time frame generation process. The initial conformation trajectory is obtained after the three-dimensional full-atom conformations of each time frame are arranged in chronological order.

[0026] S5. Input the full atomic space coordinates of each time frame into the scattering forward model, calculate the first difference between the reconstructed scattering intensity and the scattering intensity curve of the corresponding time frame, the second difference between the coordinate changes of adjacent time frames, the third difference between the global topological morphology parameters of the initial conformation trajectory and the global topological temporal features, and the fourth difference between the local fine conformation parameters and the local conformation temporal features, and construct a comprehensive loss function based on the first difference, the second difference, the third difference and the fourth difference.

[0027] Here, the all-atom spatial coordinates refer to the numerical positions of each atom in the three-dimensional all-atom conformation within the three-dimensional Cartesian coordinate system. For example, the all-atom spatial coordinates of the phosphorus atom of the first nucleotide in the nucleosome core DNA molecule at a 1-second time frame are (12.3, 45.8, -2.1) Å. The forward scattering model refers to a mathematical model that calculates the theoretical scattering intensity curve based on the three-dimensional all-atom conformation. For example, the all-atom spatial coordinates of the nucleosome core DNA molecule at a 1-second time frame are input into the forward scattering model. The model calculates the distances between each atom pair based on the Debye formula and superimposes the scattering amplitudes to output a theoretical scattering intensity curve.

[0028] Here, the reconstructed scattering intensity refers to the theoretical scattering intensity value calculated by the forward scattering model based on the three-dimensional all-atom conformation; for example, the forward scattering model is used when the scattering vector amplitude q = 0.5 Å. -1 The reconstructed scattering intensity of the nucleosome core DNA molecule at a 1-second time frame was calculated to be 1523 arbitrary units. The first difference refers to the measure of the difference between the reconstructed scattering intensity and the experimentally measured scattering intensity curve; for example, the reconstructed scattering intensity curve of the nucleosome core DNA molecule at a 1-second time frame is compared with the experimental scattering intensity curve at the same time frame in logarithmic space, and the sum of the squares of the intensity differences at each scattering vector amplitude point q is calculated as the first difference.

[0029] The second difference refers to the measure of the difference in the spatial coordinates of all atoms between adjacent time frames in the initial conformation trajectory. For example, the root mean square deviation of all atomic spatial displacements of the nucleosome core DNA molecule between the 1-second time frame and the 0-second time frame is calculated, and the root mean square deviation is the second difference between adjacent frame pairs.

[0030] The global topological morphology parameter refers to a numerical indicator extracted from the three-dimensional all-atom conformation to describe the overall geometric morphology of the DNA molecule. For example, the bending angle of the nucleosome core DNA molecule in a 1-second timeframe is calculated to be 72 degrees, the twist angle to be 34 degrees, and the end-to-end distance to be 51 Å. These three parameters together constitute the global topological morphology parameter. The third difference refers to a measure of the difference between the global topological morphology parameter of the initial conformational trajectory and the global topological temporal features extracted by the multi-scale coding network. For example, the global topological morphology parameter of the nucleosome core DNA molecule in a 1-second timeframe is compared with the corresponding global topological temporal features in the same timeframe, and the weighted sum of the squares of the differences in bending angle, twist angle, and end-to-end distance is calculated as the third difference.

[0031] The local fine conformation parameters refer to numerical indicators extracted from the three-dimensional full-atom conformation to describe the local fine structure of base pairs and base steps in a DNA molecule. For example, the torsion angle, helical angle, roll angle, and glycosidic bond dihedral angle of each base step are calculated from the three-dimensional full-atom conformation of the nucleosome core DNA molecule in a 1-second timeframe; these parameters collectively constitute the local fine conformation parameters. The fourth difference refers to the measure of the difference between the local fine conformation parameters of the initial conformation trajectory and the temporal features of the local conformation extracted by the multi-scale coding network. For example, the local fine conformation parameters of the nucleosome core DNA molecule in a 1-second timeframe are compared with the temporal features of the corresponding local conformation in the same timeframe, and the weighted sum of the squares of the differences of each parameter at all base steps is calculated as the fourth difference.

[0032] The comprehensive loss function refers to the overall objective function for optimizing conformational trajectories, which is constructed by combining the first difference, the second difference, the third difference, and the fourth difference according to their respective weight coefficients. For example, the weight of the first difference of the nucleosome core DNA molecule is set to 0.3, the weight of the second difference is set to 0.2, the weight of the third difference is set to 0.3, and the weight of the fourth difference is set to 0.2. The comprehensive loss function value is obtained by weighting and summing the four terms.

[0033] Specifically, the full atomic space coordinates of each time frame are input into the scattering forward model. The scattering forward model calculates the reconstructed scattering intensity corresponding to each time frame based on the full atomic space coordinates of each time frame. The first difference between the reconstructed scattering intensity and the scattering intensity curve of the corresponding time frame is calculated. The change between the full atomic space coordinates of adjacent time frames is calculated as the second difference. Global topological morphology parameters are extracted from the full atomic space coordinates of each time frame, and the third difference between the global topological morphology parameters and the global topological temporal features is calculated. Local fine conformation parameters are extracted from the full atomic space coordinates of each time frame, and the fourth difference between the local fine conformation parameters and the local conformation temporal features is calculated. The first, second, third, and fourth differences are weighted linearly combined according to their respective weight coefficients to obtain the comprehensive loss function.

[0034] S6. Iteratively optimize the initial conformational trajectory based on the comprehensive loss function until the comprehensive loss function converges, and output the optimized conformational trajectory as the multi-scale dynamic conformational reconstruction result of the DNA molecule.

[0035] The optimized conformational trajectory refers to the final conformational sequence obtained by iteratively correcting the initial conformational trajectory with the goal of minimizing the comprehensive loss function. For example, after 200 iterations of optimization, the comprehensive loss function value of the initial conformational trajectory of the nucleosome core DNA molecule decreased from the initial 0.52 to 0.08, and the three-dimensional full-atom coordinates of each time frame were gradually corrected to form the optimized conformational trajectory.

[0036] Among them, the multi-scale dynamic conformation reconstruction result refers to the final output of a three-dimensional conformation sequence that includes all time frames and is consistent with the experimental data at all spatial scales. For example, the optimized conformation trajectory of the nucleosome core DNA molecule includes the full atomic coordinates of 100 time frames. The trajectory matches the experimental scattering data at both local scales such as the base step torsion angle and global scales such as the bending angle, and serves as the multi-scale dynamic conformation reconstruction result of the molecule within a time window of 0 to 100 seconds.

[0037] Specifically, the initial conformational trajectory is iteratively optimized based on the comprehensive loss function. In each iteration, a gradient descent-type optimizer is used to update the full atomic space coordinates of each time frame in the initial conformational trajectory. After each iteration, a rigid constraint correction is applied to the full atomic space coordinates to keep the bond lengths and bond angles between adjacent atoms in the DNA molecule within a preset physical range. The iteration is repeated until the comprehensive loss function converges. After the comprehensive loss function converges, the conformational trajectory corresponding to the termination of the iterative optimization is output as the optimized conformational trajectory. The optimized conformational trajectory is used as the result of the multi-scale dynamic conformational reconstruction of the DNA molecule.

[0038] The technical solution of this embodiment acquires time-resolved angular correlation scattering data and extracts temporal feature matrices. It then uses a multi-scale coding network to extract global topological and local conformational temporal features and generate initial conformational trajectories. Subsequently, it uses a forward scattering model to calculate the multiple differences between the reconstructed scattering intensity and experimental data at multiple scales, constructs a comprehensive loss function for iterative optimization, and solves the problems of traditional inversion modeling relying on strong prior assumptions and making it difficult to simultaneously constrain multi-scale structures under a single low-dimensional observation, as well as the problems of lack of matching feedback between molecular dynamics simulations and experimental data, resulting in a lack of experimental constraints and excessively high computational costs for reconstruction. This solution enables the direct extraction of dynamic evolution information from angular correlation scattering data and outputs consistent DNA dynamic conformational reconstruction results across scales.

[0039] In one alternative approach, the multi-scale coding network includes a first coding branch and a second coding branch configured in parallel. The receptive field of the convolutional kernel of the first coding branch is larger than that of the convolutional kernel of the second coding branch. The first coding branch outputs the global topological temporal features, and the second coding branch outputs the local conformational temporal features. The global topological temporal features and the local conformational temporal features are spliced ​​and fused in the latent space to obtain the temporal latent variables.

[0040] The first encoding branch refers to a branch structure in a multi-scale encoding network that uses convolutional operations with a larger receptive field to extract global topological temporal features. For example, in a convolutional neural network, the first encoding branch uses an 11×11 convolutional kernel to convolve the temporal feature matrix, capturing the overall morphological change pattern across multiple time frames. The second encoding branch refers to a branch structure in a multi-scale encoding network that uses convolutional operations with a smaller receptive field to extract local conformational temporal features. For example, in a convolutional neural network, the second encoding branch uses a 3×3 convolutional kernel to convolve the temporal feature matrix, capturing the local change details between adjacent scattering intensity points within each time frame.

[0041] The receptive field of a convolutional kernel refers to the size of the local region covered by a single convolution operation on the input data in a convolutional neural network. For example, the receptive field of the convolutional kernel of the first encoding branch covers 11 consecutive time frames and all feature dimensions in the temporal feature matrix, while the receptive field of the convolutional kernel of the second encoding branch only covers the local region within the current time frame and its adjacent frames.

[0042] In the above-mentioned optional methods, a parallel dual-branch structure is further used to extract global topological temporal features and local conformational temporal features with different receptive fields, and then spliced ​​and fused in the latent space. This enables the network to simultaneously capture large-scale morphological changes and local fine structural information of DNA, resulting in a more complete and coordinated multi-scale feature expression.

[0043] In one alternative approach, the time-series generation network is a recursive generation network based on an autoregressive strategy. When the recursive generation network generates the three-dimensional all-atom conformation sequentially in each of the time frames, it uses the spatial coordinates of the last base generated in the previous time frame as the anchor point of the starting conformation of the current time frame generation process.

[0044] The autoregressive strategy refers to a recursive generation method in which the temporal generation network relies on the information of the previously generated conformation in the previous frame when generating the conformation of the current time frame. For example, when generating the 3D all-atom conformation of the 2-second time frame, the recursive generation network first uses the terminal base coordinates generated in the 1-second time frame as known conditions, and then generates the remaining atomic coordinates of the current frame. The recursive generation network refers to a generation network that uses a recurrent neural network structure and generates the conformations of each frame sequentially in chronological order. For example, a recursive generation network based on gated recurrent units generates the 3D all-atom conformations of the 1-second, 2-second, and up to 100-second time frames sequentially, starting from the 0-second time frame.

[0045] The terminal base spatial coordinates refer to the three-dimensional spatial coordinates of all atoms in the last base pair at the end of the double helix of a DNA molecule. For example, the set of three-dimensional coordinates of all atoms in the 146th base pair of the nucleosome core DNA molecule constitutes the terminal base spatial coordinates, which are located in the region (48.2, 23.7, 31.5) Å at a 1-second timeframe. The starting conformation anchor point refers to the terminal base spatial coordinates of the previous frame, which serve as a fixed reference point when generating the three-dimensional full-atom conformation of the current timeframe. For example, when the recursive generation network generates the full-atom conformation of a 2-second timeframe, it uses the terminal base spatial coordinates of the 1-second timeframe as the starting conformation anchor point, and the position of the anchor point remains unchanged during the generation process.

[0046] In the above-mentioned optional methods, a self-regressive recursive strategy is further adopted to generate conformations. The spatial coordinates of the bases at the end of the previous time frame are used as the anchor point of the starting conformation of the current time frame to ensure the continuity of the conformations in physical space between adjacent time frames, so that the initial conformation trajectory is closer to the temporal evolution characteristics of real molecular motion.

[0047] In one alternative approach, the kernel size of the first encoding branch is larger than that of the second encoding branch. Both the first and second encoding branches include a batch normalization layer and a non-linear activation layer. The concatenation and fusion operation involves connecting the global topological temporal features and the local conformational temporal features along the feature dimension. It should be noted that the stride of the first encoding branch is larger than the stride of the second encoding branch.

[0048] Here, the kernel size refers to the spatial dimension of the filter in the convolution operation; for example, the kernel size of the first coding branch is 11×11, and the kernel size of the second coding branch is 3×3.

[0049] In this context, a batch normalization layer refers to a network layer structure that normalizes the output data of convolutional layers to accelerate network training and improve stability. For example, a batch normalization layer is connected after the 11×11 convolutional layer in the first encoding branch, and this layer performs a normalization transformation on each channel of the convolutional output, setting the mean to zero and the variance to 1. A nonlinear activation layer refers to a network layer structure that applies a nonlinear transformation to the output of convolutional or fully connected layers to enhance the network's expressive power. For example, a ReLU nonlinear activation layer is connected after the batch normalization layer, and this layer sets negative values ​​in the convolutional output to zero while keeping positive values ​​unchanged.

[0050] In the above-mentioned optional methods, further differential convolution kernel size combined with batch normalization and nonlinear activation is used to achieve stable hierarchical extraction of global and local features. The method of splicing and fusing along the feature dimension preserves the original expression of information at each scale, which makes it easier for the generator network to complete the conformation reconstruction based on multi-scale latent variables in subsequent steps.

[0051] In one optional approach, before extracting the global topological temporal features and the local conformational temporal features, the multi-scale coding network first encodes the temporal positional information of each time frame in the temporal feature matrix through a positional coding layer, and then superimposes the encoded temporal positional information into the temporal feature matrix. It should be noted that the encoding function used in the positional coding layer is a relative positional coding function containing learnable parameters, and the output dimension of the relative positional coding function is consistent with the feature dimension of the temporal feature matrix.

[0052] The encoded temporal position information refers to the position representation in which the absolute position or relative order of each time frame in the temporal sequence is encoded into a vector form and superimposed on the original features. For example, the position encoding layer generates a 200-dimensional position encoding vector for each of the 100 time frames from 0 seconds to 100 seconds. The dimensions of the vector change periodically with the frame number according to the sine and cosine functions. The encoded temporal position information is superimposed frame by frame on the corresponding row of the temporal feature matrix.

[0053] In the above-mentioned optional approach, the temporal position information of each time frame is further encoded by a position encoding layer before feature extraction and superimposed onto the temporal feature matrix, so that the network can perceive the order of time frames, avoid the dilution of temporal information in convolution operation, and enhance the network's ability to characterize the time dependence in the dynamic evolution of DNA.

[0054] In one alternative approach, the comprehensive loss function is a weighted linear combination of the first difference, the second difference, the third difference, and the fourth difference, wherein the weight coefficients corresponding to the first difference, the second difference, the third difference, and the fourth difference change according to a predetermined scheduling strategy during the iterative optimization process.

[0055] The predetermined scheduling strategy refers to the scheme for adjusting the weight coefficients of each difference item according to pre-defined rules during the iterative optimization process. For example, in the iterative optimization process of the nucleosome core DNA molecule, the weight coefficient of the second difference decreases from the initial value of 0.3 to the final value of 0.1 in a manner that decreases to 0.9 times the current value every 10 iterations.

[0056] In the above-mentioned optional methods, the four differences are further combined into a weighted linear combination to form a comprehensive loss function, and the weight coefficients are dynamically adjusted according to a predetermined scheduling strategy during the iterative optimization process, so that different constraints play a leading role in different stages of optimization and balance the emphasis between experimental data fitting and physical rationality.

[0057] In one alternative approach, the predetermined scheduling strategy is such that the weight coefficient corresponding to the second difference decreases as the number of iterations increases, while the weight coefficient corresponding to the first difference increases as the number of iterations increases, and the scheduling strategy keeps each weight coefficient constant after the number of iterations reaches a predetermined step threshold.

[0058] The predetermined step threshold refers to the boundary value of the number of iterations used to trigger the state switching of the scheduling strategy during the iterative optimization process. For example, in the iterative optimization process of the nucleosome core DNA molecule, the predetermined step threshold is set to 100 times. Before reaching 100 iterations, the weight coefficients change according to the decreasing rule. After reaching 100 iterations, the weight coefficients remain constant and no longer change.

[0059] Among the above-mentioned optional methods, a scheduling strategy of decreasing weight for temporal smoothing constraint and increasing weight for experimental data fitting is further adopted. In the early stage, the focus is on ensuring the continuity of the conformation trajectory, and in the later stage, the matching accuracy with the experimental scattering data is enhanced. After reaching a predetermined number of steps, the weight is fixed to achieve stable convergence.

[0060] In one alternative approach, the step of iteratively optimizing the initial conformation trajectory based on the comprehensive loss function includes: A gradient descent-type optimizer is used to update the all-atomic space coordinates of each time frame in the initial conformation trajectory. After each iteration, a rigid constraint correction is applied to the all-atomic space coordinates. This rigid constraint correction is used to keep the bond lengths and bond angles between adjacent atoms in the DNA molecule within a preset physical range. It should be noted that the rigid constraint correction uses a projection gradient method. In each iteration of the projection gradient method, the all-atomic space coordinates are first updated along the negative gradient direction of the comprehensive loss function. Then, the updated coordinates are projected onto a manifold that satisfies the bond length and bond angle constraints. The projection operation is achieved by solving a constrained least squares problem.

[0061] Among them, gradient descent optimizers refer to optimization algorithms that update variable values ​​along the gradient descent direction based on the gradient information of the optimization variables using a loss function. For example, an adaptive optimizer using first-order and second-order moment estimation updates the all-atom spatial coordinates of the nucleosome core DNA molecule at each time frame. In each iteration, the optimizer calculates the partial derivative of the comprehensive loss function with respect to each atom coordinate and adjusts the coordinate values ​​along the negative gradient direction. Rigid constraint correction refers to the adjustment of atom positions after coordinate updates to maintain chemical bond lengths and bond angles within a reasonable range. For example, after each iteration of coordinate updates, if the bond length between adjacent atoms deviates from the standard value of 1.0 Å by more than 0.05 Å or the bond angle deviates from the standard value of 109.5° by more than 5°, the constraint correction algorithm projects the atom positions back onto the manifold that satisfies the bond length and bond angle constraints.

[0062] The preset physical range refers to the reasonable range of bond length and bond angle determined in advance based on the physicochemical properties of chemical bonds. For example, the preset physical range of phosphorus-oxygen bonds in the nucleosome core DNA molecule is 1.45 Å to 1.55 Å, the preset physical range of carbon-nitrogen bonds is 1.30 Å to 1.40 Å, and the preset physical range of bond angle of tetrahedral carbon is 105° to 115°.

[0063] In the above-mentioned optional methods, after the gradient descent optimizer updates the all-atom space coordinates, a rigid constraint correction is applied to the coordinates to limit the bond lengths and bond angles of adjacent atoms in the DNA molecule to a preset physical range, preventing chemical structure distortions during the optimization process and ensuring that the final reconstructed DNA conformation conforms to the basic constraints of molecular mechanics.

[0064] In an alternative approach, the first difference between the reconstructed scattering intensity and the scattering intensity curve of the corresponding time frame is calculated according to a first difference formula that includes logarithmic space residuals and solvent plate scattering correction, the first difference formula being: ; Where L1 represents the first difference, T is the total number of frames of the plurality of time frames, t is the time frame index, q is the scattering vector amplitude index, Q is the total number of sampling points of the scattering vector amplitude, and I rec (q,t) represents the reconstructed scattering intensity calculated using a forward scattering model from the full atomic space coordinates of time frame t in the initial conformation trajectory. exp (q,t) represents the intensity value of the scattering intensity curve at time frame t in the time-resolved angularly correlated scattering data at the scattering vector amplitude index q. β is a numerical stability constant, and β is a solvent plate scattering correction coefficient. The value of the solvent plate scattering correction coefficient β is related to the solvent electron density of the DNA molecule in the solution environment.

[0065] It should be noted that the principle behind the construction of the first difference formula is to transform the scattering intensity to logarithmic space to eliminate the interference of differences in the magnitude of intensity across different scattering vector amplitude ranges on the fitting. Simultaneously, a weighting factor inversely proportional to the sum of the square of the scattering vector amplitude and the solvent plate scattering correction coefficient is introduced to enhance the contribution of structural information in the low scattering angle region and correct for the plate scattering effect of the solvent background. The function of the first difference formula is to quantitatively measure the overall deviation between the reconstructed scattering intensity and the experimentally measured scattering intensity curve within the entire scattering vector amplitude range, providing experimental data constraints for subsequent iterative optimization. The dimension of the first difference on the left-hand side of the first difference formula depends on the weighting factor. The sum of the squared term of the scattering vector amplitude and the solvent plate scattering correction coefficient is the reciprocal of the sum of the two terms. The dimension of the scattering vector amplitude is the reciprocal of the length, and the dimension of the solvent plate scattering correction coefficient is the same as that of the squared scattering vector amplitude, which is the negative square of the length. The dimension of the reciprocal of the sum of the two terms is the square of the length. The squared term of the logarithmic space residual is dimensionless. Therefore, the overall dimension of the right side of the equation is the square of the length. Since the first difference is a relative difference measure, its value depends on the definition of the scattering intensity unit and the scattering vector amplitude sampling density. Under fixed experimental conditions and data processing protocols, this dimension does not affect the relative comparison and optimization direction of the difference measure. It belongs to the permissible dimensional loss function form.

[0066] In the above-mentioned optional methods, the first difference is further calculated by combining logarithmic space residual with solvent plate scattering correction, which reduces the dominance of strong scattering signal on the loss function, while compensating for the background contribution of the solution environment to the scattering curve, thereby improving the matching degree between the reconstructed scattering intensity and the experimental scattering intensity curve.

[0067] In one optional approach, the second difference in the coordinate changes of adjacent time frames and the third difference in the global topological morphological parameters of the initial conformation trajectory and the global topological temporal features are calculated uniformly according to a joint constraint formula. The calculation result of the joint constraint formula is the joint constraint value, which is used as a unified substitute for the second difference and the third difference, respectively. The joint constraint formula is as follows: ; Among them, L 2,3 N represents the joint constraint value of the second difference and the third difference. t The total number of time frames is t, where t is the time frame index, N is the total number of atoms in the DNA molecule, i is the atom index, and r is the atom index. i (t) is the full atomic space coordinate vector of atom i at time frame t, r i (t-1) is the all-atom spatial coordinate vector of atom i at time frame t-1, σ is the scale normalization factor, and P topo(t) is the global topological morphological parameter vector of time frame t in the initial conformational trajectory. The global topological morphological parameter vector is composed of the bending angle, twisting angle, and end-to-end distance length of the DNA molecule. latent,t Let C be the global topological temporal feature vector corresponding to time frame t in the temporal latent variables, and let C be the covariance matrix of the global topological temporal feature vector in the latent space. The covariance matrix C is a positive definite matrix. -1 is the inverse of the covariance matrix.

[0068] It should be noted that the joint constraint formula for the second and third differences is constructed by using the exponential decay function of the displacement of the full atomic spatial coordinates in adjacent time frames as the weighting factor of the Mahalanobis distance of the global topological difference. This ensures that the global topological difference is fully accounted for in time periods with small atomic motion amplitudes, while the constraint contribution of the global topological difference is adaptively suppressed in time periods with large atomic motion amplitudes to avoid excessive influence of transient fluctuations on the overall optimization. The function of the joint constraint formula is to unify the physical requirement of temporal continuity with the requirement of consistency of global topological shape in the same constraint term. By using exponential weights, it achieves nonlinear modulation of the contribution of displacement magnitude to topological difference, thereby achieving synergistic constraint. The smoothness of the conformational trajectory in the time dimension and the accuracy of the global geometric shape; the joint constraint value on the left side of the joint constraint formula is dimensionless, and the numerator of the exponential term on the right side of the equation is the sum of the squares of the atomic displacements of adjacent frames, whose dimension is the square of the length, and the denominator is the square of the scale normalization factor, whose dimension is consistent with the square of the length. After dividing the two, the exponential term is dimensionless. In the Mahalanobis distance term, the dimensions of each component of the difference between the global topological morphological parameter vector and the global topological temporal feature vector depend on the definition of the specific topological parameters. The difference vector is multiplied on the left by the inverse of the covariance matrix and then multiplied on the right by the difference vector to obtain the Mahalanobis distance scalar value. This scalar value is dimensionless. Therefore, the right side of the equation is dimensionless as a whole, which corresponds completely with the dimension of the joint constraint value on the left side of the equation.

[0069] In the above optional approach, the temporal smoothing constraint and the global topological constraint are further calculated in a unified manner through a joint constraint formula. The influence range of the coordinate changes of adjacent frames is adjusted by using an exponential decay factor, and the deviation of topological parameters is measured under the Mahalanobis distance framework to avoid the problems of inconsistent optimization directions and mutual interference of gradients when calculated independently.

[0070] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A machine learning-driven method for reconstructing the multi-scale dynamic conformation of DNA from angular correlation data, characterized in that, include: Time-resolved angular correlation scattering data of DNA molecules is obtained, wherein the time-resolved angular correlation scattering data includes scattering intensity curves of multiple time frames; The scattering intensity curves of each time frame are preprocessed, and the one-dimensional scattering feature vectors of each scattering intensity curve are extracted and arranged in chronological order to form a temporal feature matrix; The temporal feature matrix is ​​input into a multi-scale coding network to extract global topological temporal features and local conformational temporal features and map them to the latent space to obtain temporal latent variables. The temporal latent variables are input into the temporal generation network to generate three-dimensional all-atom conformations for each time frame, thus obtaining the initial conformation trajectory. The full atomic space coordinates of each time frame are input into the scattering forward model. The first difference between the reconstructed scattering intensity and the scattering intensity curve of the corresponding time frame, the second difference between the coordinate changes of adjacent time frames, the third difference between the global topological morphology parameters of the initial conformation trajectory and the global topological temporal features, and the fourth difference between the local fine conformation parameters and the local conformation temporal features are calculated. A comprehensive loss function is constructed based on the first difference, the second difference, the third difference, and the fourth difference. The initial conformational trajectory is iteratively optimized based on the comprehensive loss function until the comprehensive loss function converges, and the optimized conformational trajectory is output as the result of the multi-scale dynamic conformational reconstruction of the DNA molecule.

2. The method for reconstructing DNA multi-scale dynamic conformation of machine learning-driven angular correlation data according to claim 1, characterized in that, The multi-scale coding network includes a first coding branch and a second coding branch set in parallel. The receptive field of the convolution kernel of the first coding branch is larger than that of the convolution kernel of the second coding branch. The first coding branch outputs the global topological temporal features, and the second coding branch outputs the local conformational temporal features. The global topological temporal features and the local conformational temporal features are spliced ​​and fused in the latent space to obtain the temporal latent variables.

3. The method for reconstructing DNA multi-scale dynamic conformation of machine learning-driven angular correlation data according to claim 1, characterized in that, The time-series generation network is a recursive generation network based on an autoregressive strategy. When the recursive generation network generates the three-dimensional all-atom conformation in each time frame, it uses the spatial coordinates of the last base generated in the previous time frame as the anchor point of the starting conformation of the current time frame generation process.

4. The method for reconstructing DNA multi-scale dynamic conformation of machine learning-driven angular correlation data according to claim 2, characterized in that, The kernel size of the first coding branch is larger than that of the second coding branch. The first coding branch and the second coding branch each contain a batch normalization layer and a nonlinear activation layer. The splicing and fusion is to connect the global topological temporal features and the local conformational temporal features along the feature dimension.

5. The method for reconstructing DNA multi-scale dynamic conformation of machine learning-driven angular correlation data according to claim 2, characterized in that, Before extracting the global topological temporal features and the local conformational temporal features, the multi-scale coding network first encodes the temporal position information of each time frame in the temporal feature matrix through a position coding layer, and then superimposes the encoded temporal position information into the temporal feature matrix.

6. The method for reconstructing DNA multi-scale dynamic conformation of machine learning-driven angular correlation data according to claim 1, characterized in that, The comprehensive loss function is a weighted linear combination of the first difference, the second difference, the third difference, and the fourth difference. The weight coefficients corresponding to the first difference, the second difference, the third difference, and the fourth difference change according to a predetermined scheduling strategy during the iterative optimization process.

7. The method for reconstructing DNA multi-scale dynamic conformation of machine learning-driven angular correlation data according to claim 6, characterized in that, The predetermined scheduling strategy is that the weight coefficient corresponding to the second difference decreases as the number of iterations increases, and the weight coefficient corresponding to the first difference increases as the number of iterations increases. The scheduling strategy keeps each weight coefficient constant after the number of iterations reaches a predetermined step threshold.

8. The method for reconstructing DNA multi-scale dynamic conformation of machine learning-driven angular correlation data according to claim 1, characterized in that, The steps of iteratively optimizing the initial conformation trajectory based on the comprehensive loss function include: A gradient descent optimizer is used to update the all-atomic space coordinates of each time frame in the initial conformation trajectory, and a rigid constraint correction is applied to the all-atomic space coordinates after each iteration update. The rigid constraint correction is used to keep the bond lengths and bond angles between adjacent atoms in the DNA molecule within a preset physical range.

9. The method for reconstructing DNA multi-scale dynamic conformation of machine learning-driven angular correlation data according to claim 8, characterized in that, The first difference between the reconstructed scattering intensity and the scattering intensity curve of the corresponding time frame is calculated according to a first difference formula that includes logarithmic space residuals and solvent plate scattering correction. The first difference formula is: ; Where L1 represents the first difference, T is the total number of frames of the plurality of time frames, t is the time frame index, q is the scattering vector amplitude index, Q is the total number of sampling points of the scattering vector amplitude, and I rec (q,t) represents the reconstructed scattering intensity calculated using a forward scattering model from the full atomic space coordinates of time frame t in the initial conformation trajectory. exp (q,t) represents the intensity value of the scattering intensity curve at time frame t in the time-resolved angularly correlated scattering data at the scattering vector amplitude index q. β is a numerical stability constant, and β is a solvent plate scattering correction coefficient. The value of the solvent plate scattering correction coefficient β is related to the solvent electron density of the DNA molecule in the solution environment.

10. The method for reconstructing DNA multi-scale dynamic conformation of machine learning-driven angular correlation data according to claim 9, characterized in that, The second difference in the coordinate changes of adjacent time frames and the third difference in the global topological morphological parameters of the initial conformation trajectory and the global topological temporal features are calculated uniformly according to a joint constraint formula. The calculation result of the joint constraint formula is the joint constraint value, which is used as a unified substitute for the second difference and the third difference, respectively. The joint constraint formula is as follows: ; Among them, L 2,3 N represents the joint constraint value of the second difference and the third difference. t The total number of time frames is t, where t is the time frame index, N is the total number of atoms in the DNA molecule, i is the atom index, and r is the atom index. i (t) is the full atomic space coordinate vector of atom i at time frame t, r i (t-1) is the all-atom spatial coordinate vector of atom i at time frame t-1, σ is the scale normalization factor, and P topo (t) is the global topological morphological parameter vector of time frame t in the initial conformational trajectory. The global topological morphological parameter vector is composed of the bending angle, twisting angle, and end-to-end distance length of the DNA molecule. latent,t Let C be the global topological temporal feature vector corresponding to time frame t in the temporal latent variables, and let C be the covariance matrix of the global topological temporal feature vector in the latent space. The covariance matrix C is a positive definite matrix. -1 is the inverse of the covariance matrix.

Images