Dynamic quantification method for interaction of polypeptide molecules and biological membrane

Abstract

The invention discloses a dynamic quantification method for interaction of polypeptide molecules and a biological membrane, the method is applied to a post-processing process of molecular dynamics simulation, and the method comprises the following steps: preprocessing atomic-scale motion trail data obtained by molecular dynamics simulation, and then sampling by using a sliding window with a fixed length; performing feature extraction on the atomic-scale movement track data in each sliding window to obtain a total contact number signal, a unit lipid contact intensity signal, an insertion depth signal and an insertion intensity signal, and performing feature fusion on the multi-source signal to obtain a feature vector of each sliding window; and inputting the feature vectors of the sliding windows into a trained teacher-student model to output disturbance categories and confidence scores of the polypeptide molecules in the sliding windows on the biological membrane, and outputting a final result in a weighted voting mode. According to the method, low-frequency collective deformation and medium-high frequency local events can be represented at the same time, and predefined disturbance categories can be classified.

Description

Technical Field

[0001] This application belongs to the fields of biophysics and computer science, and particularly relates to a dynamic quantification method for the interaction between polypeptide molecules and biological membranes. Background Technology

[0002] Molecular dynamics (MD) simulation is one of the core technologies for revealing the dynamic behavior and interaction mechanisms of biomolecules at the atomic scale, and it has wide applications in drug design, materials science, and other fields. However, a single MD simulation can generate terabytes of trajectory data. How to efficiently and accurately extract valuable physicochemical information from this high-dimensional and complex time-series data is a key challenge in this field. Meanwhile, time-frequency analysis, especially continuous wavelet transform (CWT), is a powerful signal processing tool whose advantage lies in generating two-dimensional scale maps to accurately capture the instantaneous frequency characteristics of non-stationary signals. Existing techniques combine CWT with unsupervised clustering to identify cooperative motion patterns of atomic three-dimensional coordinate displacements in MD trajectories. Existing MD trajectory post-processing techniques mainly suffer from the following shortcomings: (1) Reliance on a single indicator leads to the risk of misjudgment: Traditional analytical methods often rely on a few pre-set statistical quantities, such as root mean square deviation (RMSD), contact number, or membrane thickness. Although these indicators are intuitive, they may exhibit non-monotonicity when describing complex and nonlinear biological processes, thus leading to misjudgment of physical mechanisms. For example, when studying the interaction between peptides and cell membranes, after the peptide adsorbs from the membrane surface and completely inserts into the membrane core, the number of contacts between it and the membrane head group may actually decrease. If only this indicator is used, the mechanism of "deep insertion" may be misjudged as "weak adsorption".

[0003] (2) Lack of multi-scale information integration for dynamic processes: Biological processes often involve dynamic events at different time scales. For example, rapid local vibrations of molecules (high-frequency events) and slow global conformational changes (low-frequency events) occur simultaneously. Traditional analytical methods, such as Fourier transform, can only provide the global frequency distribution of the signal over the entire time period and cannot reveal the time point of occurrence of specific dynamic events. Therefore, existing technologies cannot effectively distinguish between transient, high-frequency local membrane depressions caused by molecular insertion and the low-frequency collective fluctuations of the membrane itself.

[0004] (3) Low level of automation and difficulty in label generation: In recent years, supervised machine learning has been introduced into Markov trajectory analysis, but its application is limited by the acquisition of high-quality training data. For MD trajectories that often contain millions of frames, it is impractical to manually label the physical mechanism frame by frame. A more systematic approach is the Markov State Model (MSM) workflow. Although this process can automatically divide the states, it is complex, involving multiple steps such as PCA dimensionality reduction and clustering, and is sensitive to parameter selection.

[0005] In summary, there is an urgent need in this field for a new method for MD trajectory analysis that can be automated, accurately quantified, and simultaneously provide process explanations and global conclusions, in order to overcome the limitations of existing technologies. Summary of the Invention

[0006] In view of this, the purpose of this application is to provide a dynamic quantification method for the interaction between polypeptide molecules and biological membranes. By fusing features from multiple sources, it can simultaneously characterize low-frequency collective deformation and mid-to-high-frequency local events, thereby enhancing the separability of non-stationary perturbation processes. It can classify and score predefined perturbation categories through semi-supervised self-training. Through a weighted voting mechanism of model confidence, it can derive global conclusions from the process description.

[0007] This application provides a dynamic quantification method for the interaction between polypeptide molecules and biological membranes. This method is applied to the post-processing of molecular dynamics simulations and includes: The atomic-level motion trajectory data obtained from molecular dynamics simulations are preprocessed to eliminate coordinate jumps caused by periodic boundary conditions, and then a fixed-length sliding window is used for sampling. Feature extraction is performed on atomic-level motion trajectory data within each sliding window to obtain the total contact signal of peptide atoms and lipid head base atoms, the unit lipid contact intensity signal of each lipid type, the insertion depth signal and insertion intensity signal of peptide molecules inserted into the lipid bilayer, and feature fusion of multi-source signals to obtain the feature vector of each sliding window. The feature vectors of each sliding window are input into the trained teacher-student model to output the perturbation category and confidence score of the polypeptide molecules in each sliding window on the biological membrane, and then the final result is output through weighted voting.

[0008] Furthermore, the preprocessing of the atomic-level motion trajectory data obtained from molecular dynamics simulations includes: The distance between any two atoms is corrected based on the principle of minimum mirror image. Alternatively, an incremental unwrapping method can be used to correct and accumulate the coordinates of consecutive frames across half-boxes to obtain a continuous trajectory. Alternatively, a local coordinate system can be constructed based on the central plane of the biofilm or a local neighborhood, so that the calculation of local geometric quantities is not affected by cross-boundary jumps.

[0009] Furthermore, the total contact signal between peptide atoms and lipid head group atoms, and the unit lipid contact intensity signal for each lipid type were obtained through the following methods: Within the sliding window, several polypeptide atoms and lipid head atoms at each time step are extracted. At the same time, the distance between each polypeptide atom and each lipid head atom is obtained. Based on a pre-set contact threshold, it is determined whether the two are in contact, and the total number of contacts within the sliding window is counted. For each time step, the number of contacting peptide-lipid head atoms in each lipid type is statistically analyzed to obtain the number of contacts for each lipid type. Based on the number of head atoms in each lipid type, the contact ratio of each lipid type is calculated. The contact ratio of each lipid type is normalized based on the total number of contacts to obtain the unit lipid contact intensity signal of each lipid type within the sliding window.

[0010] Furthermore, the insertion depth and insertion intensity signals of the peptide molecules into the lipid bilayer were obtained through the following methods: Within the sliding window, the coordinates of several polypeptide atoms and lipid head atoms at each time step are extracted. Lipid head atoms within a specific radius are selected with the projection point of the polypeptide centroid onto the central plane of the biomembrane as the center. The central plane of the biomembrane is parallel to the upper and lower lipid head layers. At each time step, the projections of the selected lipid head atoms onto the biomembrane normal direction are summed and averaged, and the distance between the selected lipid head atoms and the central surface of the biomembrane is calculated to determine the insertion depth signal of the polypeptide molecule into the lipid bilayer within the sliding window. At each time step, the distance between the projection of the polypeptide centroid onto the biomembrane normal direction and the central plane of the biomembrane is used as the insertion intensity signal of the polypeptide molecule into the lipid bilayer within the sliding window.

[0011] Furthermore, feature fusion is performed on the multi-source signals within each sliding window using the following method: Within the sliding window, a continuous wavelet transform is performed on the insertion depth signal to obtain the time-frequency coefficients and energy spectrum. The energy spectrum is divided into three frequency bands—low, medium, and high—according to the time scale or logarithmic scale to obtain the energy proportion of each frequency band. These are then used together with the energy spectrum as time-frequency energy features. The mean, standard deviation, quantile, and maximum value of the total contact signal, unit lipid contact intensity signal, insertion depth signal, and insertion intensity signal were calculated respectively, and linear fitting was performed to obtain their respective statistical characteristics and rate of change characteristics. The statistical characteristics and rate of change characteristics of the total number of contacts, unit lipid contact intensity, insertion depth, and insertion intensity signals are concatenated with the time-frequency energy characteristics to obtain the feature vector of the sliding window.

[0012] Furthermore, the teacher-student model is pre-trained in the following ways: Based on the quantiles of key indicators within the sliding window, the perturbation category of polypeptide molecules on the biological membrane is determined according to multiple pre-set limiting conditions, and only sliding windows that meet the limiting conditions are labeled; wherein, the key indicators include: the quantiles of the total number of contacts signal, the insertion depth signal, the insertion intensity signal, and the high-frequency energy ratio; Semi-supervised self-training of the teacher-student model is performed using feature vectors from labeled and unlabeled sliding windows.

[0013] Furthermore, the step of determining the type of disturbance to the biological membrane by the polypeptide molecule based on multiple pre-set limiting conditions includes: When the quantiles of the total number of contacts, insertion depth, and insertion intensity signals within the sliding window are all less than or equal to their respective 30th quantiles, the disturbance situation is determined to be disturbance-free. When the quantiles of the total number of contacts and the insertion intensity signal within the sliding window are both greater than or equal to their respective 70th quantiles, the disturbance is determined to be surface adsorption. When the quantiles of the insertion depth signal and the insertion intensity signal within the sliding window are both greater than or equal to their respective 70th quantiles, the disturbance is determined to be a shallow insertion. When the regular perturbation score within the sliding window is greater than or equal to the 70th percentile and the high-frequency energy percentage is greater than or equal to the 75th percentile, the perturbation is determined to be membrane perforation; wherein, the regular perturbation score is obtained by quantile normalization for each key indicator.

[0014] Furthermore, the semi-supervised self-training of the teacher-student model using feature vectors from labeled and unlabeled sliding windows includes: The labeled and continuous feature vector of the sliding window is used as the input to the teacher model, and the label is used as the output to train the teacher model. Based on the feature vector of each unlabeled sliding window, the probability and confidence score of the perturbation category of each unlabeled sliding window are predicted using the trained teacher model. The student model is trained using feature vectors from unlabeled sliding windows with high confidence scores and labeled, continuous sliding windows, until iterations converge to new samples.

[0015] The dynamic quantification method for the interaction between polypeptide molecules and biological membranes provided in this application has the following advantages: by fusing features of multi-source signals, it can simultaneously characterize low-frequency collective deformation and mid-to-high-frequency local events, enhancing the separability of non-stationary perturbation processes; it can classify and score predefined perturbation categories through semi-supervised self-training; and it can derive global conclusions from process descriptions through a weighted voting mechanism of model confidence. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the dynamic quantification method for the interaction between polypeptide molecules and biological membranes provided in the embodiments of this application is shown. Figure 2 An example diagram provided in the embodiments of this application is shown. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this technical solution clearer, the following detailed description, in conjunction with specific embodiments, further illustrates this technical solution. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this technical solution.

[0018] Example 1: Please see as follows Figure 1 The flowchart illustrates a method for dynamically quantifying the interaction between polypeptide molecules and biological membranes. This method is applied to the post-processing of molecular dynamics simulations. The method includes: S101. Preprocess the atomic-level motion trajectory data obtained from molecular dynamics simulation to eliminate coordinate jumps caused by periodic boundary conditions, and then use a fixed-length sliding window for sampling.

[0019] In this step, atomic-level motion trajectory data is obtained through molecular dynamics simulation, and the frame number, time step, and simulation box size of the trajectory data are automatically parsed to provide a unified time index and coordinate reference for subsequent feature calculations. Furthermore, to reduce coordinate jumps caused by periodic boundary conditions during the simulation, robust preprocessing based on periodic boundaries is performed on the trajectory data.

[0020] In practice, the atomic-level motion trajectory data obtained from molecular dynamics simulations are preprocessed using the following methods: Step 1011: Perform a correction calculation on the distance between any two atoms based on the principle of minimum mirror image.

[0021] Alternatively, step 1012 can be performed to correct and accumulate the coordinates of continuous frames across half-boxes using an incremental unwrapping method, in order to obtain a continuous trajectory.

[0022] Alternatively, perform step 1013 to construct a local coordinate system based on the central plane of the biofilm or a local neighborhood, so that the calculation of local geometric quantities is not affected by cross-boundary jumps.

[0023] It should be noted that periodic boundary conditions are a common setting in molecular dynamics simulations. This application uses their robust processing as a fixed step in the signal construction and discrimination process, thereby reducing spurious signals and improving the stability of subsequent time-frequency features and machine learning discrimination.

[0024] S102. In each sliding window, feature extraction is performed on the atomic-level motion trajectory data to obtain the total contact signal of peptide atoms and lipid head base atoms, the unit lipid contact intensity signal of each lipid type, the insertion depth signal and insertion intensity signal of peptide molecules inserted into the lipid bilayer, and feature fusion of multi-source signals to obtain the feature vector of each sliding window.

[0025] In this step, feature extraction is performed using atomic-level motion trajectory data to obtain a set of multi-source signals that can characterize the interaction between polypeptide molecules and biological membranes.

[0026] In practical implementation, the total contact signal between peptide atoms and lipid head group atoms, and the unit lipid contact intensity signal for each lipid type are obtained through the following methods: Step 1021: Extract several polypeptide atoms and lipid head atoms at each time step within the sliding window, obtain the distance between each polypeptide atom and each lipid head atom, determine whether they are in contact according to a preset contact threshold, and count the total number of contacts within the sliding window.

[0027] As an example, let the set of peptide atoms be P, and the set of lipid head atoms be H (preferably phosphorus atoms, but nitrogen or oxygen atoms are also acceptable). Based on the distance between the set of peptide atoms P and the set of lipid head atoms H, set a contact threshold r. c The total number of contacts, signal C, is obtained by counting the number of contact logs. total (t): ; (1) In the formula, d ij (t) represents the distance between the polypeptide atom and the lipid head group atom. For the indicator function, the contact threshold r c These are configurable parameters.

[0028] Step 1022: For each time step, count the number of contacting peptide-lipid head atoms for each lipid type to obtain the number of contacts for each lipid type, and calculate the contact ratio for each lipid type based on the number of head atoms for each lipid type.

[0029] Step 1023: Normalize the contact ratio of each lipid type based on the total number of contacts to obtain the unit lipid contact intensity signal of each lipid type within the sliding window.

[0030] As an example, to accommodate different biomembrane component systems, each lipid type is set as {L1,…,L...}. M}, and count the number of contacts C for each lipid type. Lm (t): ; (2) Then, based on the number of head group atoms in each lipid type, the contact ratio E of each lipid type was calculated. Lm (t) (e.g., the proportion of anionic lipids) is used to characterize selective adsorption and enrichment, and the contact ratio is normalized according to the total number of contacts to obtain the unit lipid contact intensity signal ρ. Lm (t): ; (3) In the formula, N Lm ε represents the number of head group atoms for any lipid type, with ε being a small amount to prevent the denominator from being zero.

[0031] In practical implementation, the insertion depth and insertion intensity signals of peptide molecules into the lipid bilayer are obtained through the following methods: Step 1024: Extract the coordinates of several polypeptide atoms and lipid head atoms at each time step within the sliding window. Select lipid head atoms within a specific radius, centered on the projection point of the polypeptide centroid on the central plane of the biomembrane. The central plane of the biomembrane is parallel to the upper and lower lipid head layers.

[0032] Step 1025: At each time step, sum and average the projections of the selected lipid head atoms in the normal direction of the biomembrane, and calculate the distance between them and the central surface of the biomembrane to determine the insertion depth signal of the polypeptide molecule into the lipid bilayer within the sliding window.

[0033] As an example, define the central plane z of the biofilm. mid (t) (which can be constructed from the mean of the upper and lower leaflet head bases z or by using other stable definitions), with the projection point of the polypeptide centroid onto the central plane of the biomembrane as the center, within a specific radius R D The insertion depth signal Depth(t) is calculated by selecting the nearest neighbor set of lipid head atoms Near(t): ; (4) In the formula, z j (t) represents the projection of lipid head atoms selected within a specific radius onto the normal direction of the biomembrane.

[0034] Simultaneously output the number of nearest neighbors n near (t)=|Near(t)| and auxiliary information such as blade markings are used to support the analysis of the perturbation differences between the upper and lower blades. In addition, R D These are configurable parameters.

[0035] Step 1026: At each time step, the distance between the projection of the polypeptide centroid in the normal direction of the biomembrane and the central plane of the biomembrane is used as the insertion intensity signal of the polypeptide molecule into the lipid bilayer within the sliding window.

[0036] As an example, to adapt to different biomembrane component systems, the projection z of the polypeptide centroid (or key atom) in the biomembrane normal direction is used. pep (t) and the central plane of the biomembrane z mid Using the generalized definition of "relative reference surface", the insertion intensity signal Ins(t) is calculated: ; (5) In the formula, z ref (t) represents the position relative to the reference surface, which can be the central surface of the biofilm, the interfacial density extremum surface, or any stable reference layer.

[0037] Furthermore, the normal direction of the biofilm can be specified or determined by trajectory statistics, without limiting the orientation of the coordinate axes.

[0038] In practice, feature fusion of multi-source signals within each sliding window is performed using the following method: Step 1027: Perform continuous wavelet transform on the insertion depth signal within the sliding window to obtain the time-frequency coefficients and energy spectrum. Divide the energy spectrum into three frequency bands (low, medium, and high) according to the time scale or logarithmic scale to obtain the energy proportion of each frequency band. Then, use the energy spectrum together as the time-frequency energy feature.

[0039] As an example, to characterize the multi-scale dynamic events of perturbations between polypeptide molecules and biological membranes, the depth signal Depth(t), which can characterize biological membrane deformation, is used as the selected signal x(t). A continuous wavelet transform (CWT) is performed within a sliding window to obtain the time-frequency coefficients and energy spectrum. Specifically, this embodiment uses the Morlet wavelet basis, which can be expressed as: ; (6) In the formula, ω0 is the center frequency parameter, which is configurable; Furthermore, the process of performing CWT on signal x(t) can be expressed as: ; (7) In the formula, a represents the scale, b represents the time shift, and * represents the complex conjugate.

[0040] And the energy spectrum can be represented as: (8) Within each sliding window, the energy spectrum is divided into low, medium, and high frequency bands according to either a time scale or a logarithmic scale. Integrating the energy across the entire frequency band yields the following representation: ; (9) In the formula, E total (k) represents the full-frequency energy, E low (k) represents low-frequency energy, Emid (k) represents the mid-frequency energy, E high (k) represents the high-frequency energy.

[0041] Furthermore, the energy proportions of each frequency band are constructed as follows: ; (10) In the formula, Frac low (k) represents the proportion of low-frequency energy, Frac mid (k) represents the proportion of mid-frequency energy, Frac high (k) represents the proportion of high-frequency energy.

[0042] Here, time-frequency fingerprint features are used to characterize steady-state disturbances dominated by low frequencies and local disturbance patterns enhanced by mid-to-high frequency events, and to enhance the separability of non-stationary processes.

[0043] Step 1028: Calculate the mean, standard deviation, quantile, and maximum value of the total contact signal, unit lipid contact intensity signal, insertion depth signal, and insertion intensity signal, respectively, and perform linear fitting to obtain their corresponding statistical characteristics and rate of change characteristics.

[0044] Step 1029: Concatenate the statistical characteristics and rate of change characteristics of the total number of contacts, unit lipid contact intensity, insertion depth, and insertion intensity signals with the time-frequency energy characteristics to obtain the feature vector of the sliding window.

[0045] As an example, the total number of contacts signal C within the sliding window total (t), unit lipid contact intensity signal ρ Lm The mean, standard deviation, quantiles, and maximum values ​​of the insertion depth signal Depth(t) and insertion intensity signal Ins(t) are calculated respectively to obtain their respective statistical features stats(Ct), stats(ρ), stats(Depth), and stats(Ins). The rate of change feature slope(*) is obtained by linear fitting. These features, together with the time-frequency energy features Frac* and E*, constitute the feature vector of the sliding window.

[0046] Here, the components in the feature vector can be appropriately expanded or reduced according to the component system of the biofilm, and this application does not impose any limitations on this.

[0047] S103. Input the feature vectors of each sliding window into the trained teacher-student model to output the perturbation category and confidence score of the polypeptide molecules in each sliding window on the biological membrane, and then output the final result through weighted voting.

[0048] In this step, the feature vectors of each sliding window are input into the trained teacher-student model, which can output the perturbation category of each sliding window for classification and scoring, thereby describing the perturbation process of peptides and biomembranes. Furthermore, through the weighted voting mechanism of model confidence, global conclusions can be derived from the process description.

[0049] As an example, please refer to Figure 2 The example diagram is shown. Figure 2 (A) is a snapshot of the simulated conformation of the peptide and DOPE / POPG (3:1) bilayer at a certain moment, where blue represents the peptide, red represents the lipid head region, and green represents the lipid tail region. Figure 2 (B) is the Ct(t) graph, which shows the total number of contacts between polypeptide atoms and lipid head group atoms. Figure 2 (C) is the Ins(t) plot, which shows the insertion intensity signal of the polypeptide molecule into the lipid bilayer. Figure 2 (D) is a unit lipid contact intensity signal ρ(t) diagram for each lipid type, used to characterize the proportion of peptide preference for POPG lipids. If ρPOPG(t)>1, it means that POPG is "enriched" in contact and the peptide has a preference for POPG; if ρPOPG(t)≈1, it means that there is no obvious preference; if ρPOPG(t)<1, it means that the peptide prefers DOPE or non-POPG. Figure 2 (E) shows the Morlet-CWT and band energy percentage strip diagram obtained based on the insertion depth signal of peptide molecules into the lipid bilayer; Figure 2 (F) is a time series diagram showing the types of perturbations of biological membranes by polypeptide molecules; Figure 2 (G) is a weighted voting result diagram, where I, II, III, and IV represent the four disturbance categories, respectively.

[0050] In addition, to avoid large-scale manual annotation when training the teacher-student model, this application adopts a combination of rule-based weak annotation and confidence screening methods to construct the training set.

[0051] In practice, the teacher-student model is pre-trained using the following methods: Step 201: Based on the quantiles of key indicators within the sliding window, determine the type of disturbance of the biomembrane by the polypeptide molecule according to multiple pre-set limiting conditions, and only label the sliding window that meets the limiting conditions.

[0052] The key indicators include: total number of contacts signal, insertion depth signal, quantile of insertion intensity signal, and high frequency energy percentage.

[0053] In practice, the type of disturbance to the biological membrane by polypeptide molecules is determined based on several pre-set limiting conditions using the following methods: Step 2011: When the quantiles of the total number of contacts, insertion depth, and insertion intensity signals within the sliding window are all less than or equal to their respective 30th percentiles, the disturbance situation is determined to be disturbance-free.

[0054] Step 2012: When the quantiles of the total number of contacts signal and the insertion intensity signal within the sliding window are both greater than or equal to their respective 70th percentiles, the disturbance is determined to be surface adsorption.

[0055] Step 2013: When the quantiles of the insertion depth signal and the insertion intensity signal within the sliding window are both greater than or equal to their respective 70th quantiles, the disturbance is determined to be shallow insertion.

[0056] Step 2014: When the regular disturbance score within the sliding window is greater than or equal to the 70th percentile and the high-frequency energy percentage is greater than or equal to the 75th percentile, the disturbance condition is determined to be membrane perforation; wherein, the regular disturbance score is obtained by quantile normalization for each key indicator.

[0057] As an example, by applying quantile thresholds to key indicators, an adaptive boundary is formed as follows: ; (11) In the formula, The upper and lower boundaries of the total number of contact signal quantiles, To define the upper and lower boundaries of the quantiles of the interpolated intensity signal, These are the upper and lower boundaries of the insertion depth signal quantile.

[0058] Next, only windows that meet the following constraints will be labeled with the corresponding disturbance category label; the remaining windows will remain unlabeled: Undisturbed: ;(12) Surface adsorption: ; (13) Shallow insertion: ;(14) Membrane perforation: ; (15) The rule perturbation score is defined as follows: (16) In the formula, For rank normalization or quantile normalization, For sigmoid mapping, α i For fixed or configurable weights.

[0059] Step 202: Use the feature vectors of labeled and unlabeled sliding windows to perform semi-supervised self-training on the teacher-student model.

[0060] In practice, the teacher-student model is semi-supervised self-trained using feature vectors from labeled and unlabeled sliding windows in the following manner: Step 2021: Use the labeled and continuous feature vector of the sliding window as the input to the teacher model, and use the label as the output to train the teacher model.

[0061] Step 2022: Based on the feature vector of each unlabeled sliding window, use the trained teacher model to predict the probability and confidence score of the perturbation category of each unlabeled sliding window.

[0062] Step 2023: Train the student model using the feature vectors of unlabeled sliding windows with high confidence scores and labeled and continuous sliding windows, until the new samples converge.

[0063] The above content is only a preferred embodiment of the present invention. For those skilled in the art, many changes can be made in the specific implementation and application scope based on the ideas of the present invention. As long as these changes do not depart from the concept of the present invention, they all fall within the protection scope of the present invention.

Claims

1. A method for dynamically quantifying the interaction between polypeptide molecules and biological membranes, characterized in that, The method is applied to the post-processing of molecular dynamics simulations, and the method includes: The atomic-level motion trajectory data obtained from molecular dynamics simulations are preprocessed to eliminate coordinate jumps caused by periodic boundary conditions, and then a fixed-length sliding window is used for sampling. Feature extraction is performed on atomic-level motion trajectory data within each sliding window to obtain the total contact signal of peptide atoms and lipid head base atoms, the unit lipid contact intensity signal of each lipid type, the insertion depth signal and insertion intensity signal of peptide molecules inserted into the lipid bilayer, and feature fusion of multi-source signals to obtain the feature vector of each sliding window. The feature vectors of each sliding window are input into the trained teacher-student model to output the perturbation category and confidence score of the polypeptide molecules in each sliding window on the biological membrane, and then the final result is output through weighted voting.

2. The method as described in claim 1, characterized in that, The preprocessing of atomic-level motion trajectory data obtained from molecular dynamics simulations includes: The distance between any two atoms is corrected based on the principle of minimum mirror image. Alternatively, an incremental unwrapping method can be used to correct and accumulate the coordinates of consecutive frames across half-boxes to obtain a continuous trajectory. Alternatively, a local coordinate system can be constructed based on the central plane of the biofilm or a local neighborhood, so that the calculation of local geometric quantities is not affected by cross-boundary jumps.

3. The method as described in claim 1, characterized in that, The total contact signal between peptide atoms and lipid head group atoms, and the unit lipid contact intensity signal for each lipid type were obtained through the following methods: Within the sliding window, several polypeptide atoms and lipid head atoms at each time step are extracted. At the same time, the distance between each polypeptide atom and each lipid head atom is obtained. Based on a pre-set contact threshold, it is determined whether the two are in contact, and the total number of contacts within the sliding window is counted. For each time step, the number of contacting peptide-lipid head atoms in each lipid type is statistically analyzed to obtain the number of contacts for each lipid type. Based on the number of head atoms in each lipid type, the contact ratio of each lipid type is calculated. The contact ratio of each lipid type is normalized based on the total number of contacts to obtain the unit lipid contact intensity signal of each lipid type within the sliding window.

4. The method as described in claim 1, characterized in that, The insertion depth and insertion intensity signals of peptide molecules into the lipid bilayer were obtained using the following methods: Within the sliding window, the coordinates of several polypeptide atoms and lipid head atoms at each time step are extracted. Lipid head atoms within a specific radius are selected with the projection point of the polypeptide centroid onto the central plane of the biomembrane as the center. The central plane of the biomembrane is parallel to the upper and lower lipid head layers. At each time step, the projections of the selected lipid head atoms onto the biomembrane normal direction are summed and averaged, and the distance between the selected lipid head atoms and the central surface of the biomembrane is calculated to determine the insertion depth signal of the polypeptide molecule into the lipid bilayer within the sliding window. At each time step, the distance between the projection of the polypeptide centroid onto the biomembrane normal direction and the central plane of the biomembrane is used as the insertion intensity signal of the polypeptide molecule into the lipid bilayer within the sliding window.

5. The method as described in claim 1, characterized in that, Feature fusion of multi-source signals within each sliding window is performed using the following method: Within the sliding window, a continuous wavelet transform is performed on the insertion depth signal to obtain the time-frequency coefficients and energy spectrum. The energy spectrum is divided into three frequency bands—low, medium, and high—according to the time scale or logarithmic scale to obtain the energy proportion of each frequency band. These are then used together with the energy spectrum as time-frequency energy features. The mean, standard deviation, quantile, and maximum value of the total contact signal, unit lipid contact intensity signal, insertion depth signal, and insertion intensity signal were calculated respectively, and linear fitting was performed to obtain their respective statistical characteristics and rate of change characteristics. The statistical characteristics and rate of change characteristics of the total number of contacts, unit lipid contact intensity, insertion depth, and insertion intensity signals are concatenated with the time-frequency energy characteristics to obtain the feature vector of the sliding window.

6. The method as described in claim 1, characterized in that, The teacher-student model can be pre-trained using the following methods: Based on the quantiles of key indicators within the sliding window, the perturbation category of polypeptide molecules on the biological membrane is determined according to multiple pre-set limiting conditions, and only sliding windows that meet the limiting conditions are labeled; wherein, the key indicators include: the quantiles of the total number of contacts signal, the insertion depth signal, the insertion intensity signal, and the high-frequency energy ratio; Semi-supervised self-training of the teacher-student model is performed using feature vectors from labeled and unlabeled sliding windows.

7. The method as described in claim 6, characterized in that, The determination of the type of disturbance to the biological membrane by polypeptide molecules based on multiple pre-set limiting conditions includes: When the quantiles of the total number of contacts, insertion depth, and insertion intensity signals within the sliding window are all less than or equal to their respective 30th quantiles, the disturbance situation is determined to be disturbance-free. When the quantiles of the total number of contacts and the insertion intensity signal within the sliding window are both greater than or equal to their respective 70th quantiles, the disturbance is determined to be surface adsorption. When the quantiles of the insertion depth signal and the insertion intensity signal within the sliding window are both greater than or equal to their respective 70th quantiles, the disturbance is determined to be a shallow insertion. When the regular perturbation score within the sliding window is greater than or equal to the 70th percentile and the high-frequency energy percentage is greater than or equal to the 75th percentile, the perturbation is determined to be membrane perforation; wherein, the regular perturbation score is obtained by quantile normalization for each key indicator.

8. The method as described in claim 6, characterized in that, The semi-supervised self-training of the teacher-student model using feature vectors from labeled and unlabeled sliding windows includes: The labeled and continuous feature vector of the sliding window is used as the input to the teacher model, and the label is used as the output to train the teacher model. Based on the feature vector of each unlabeled sliding window, the probability and confidence score of the perturbation category of each unlabeled sliding window are predicted using the trained teacher model. The student model is trained using feature vectors from unlabeled sliding windows with high confidence scores and labeled, continuous sliding windows, until iterations converge to new samples.

Images