Method for accurately predicting d-protein multi-dimensional structure information
By constructing a multi-species mass spectrometry dataset and integrating a deep learning model, the problem of identifying D-type amino acid modifications in existing technologies has been solved, achieving efficient and accurate prediction of multi-dimensional structural information and improving the efficiency and accuracy of large-scale sample screening.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANKAI UNIV
- Filing Date
- 2026-02-09
- Publication Date
- 2026-06-05
Smart Images
Figure CN122157777A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of protein post-translational modification identification technology. More specifically, this invention relates to a method for accurate prediction of multidimensional structural information of D-proteins. Background Technology
[0002] In proteomics research, D-type amino acid modification is an important non-coding post-translational modification. This modification only changes the chirality of amino acid residues without altering their molecular weight, thus having a profound impact on protein structure and function and being associated with a variety of physiological and pathological processes.
[0003] Because stereoisomerization does not alter the molecular weight of amino acids, traditional mass spectrometry identification techniques relying on mass shift cannot directly identify D-type amino acid modifications. Although subsequent advancements such as thermal proteomics and ion mobility mass spectrometry have provided effective experimental screening pathways, enabling identification based on differences in peptide thermal stability or gas-phase conformation, these methods still face numerous limitations in practical applications. For example, thermal proteomics experiments are complex and have limited throughput; ion mobility mass spectrometry is constrained by instrument resolution, sensitivity, and interference from complex biological sample matrices, making large-scale, high-throughput, and accurate identification difficult. Furthermore, other solutions, such as those relying on derivatization chemical reactions or using special chiral chromatographic columns for physical separation of isomers followed by mass spectrometry analysis, also suffer from cumbersome procedures, low throughput, high cost, and reliance on specific reagents, making them unsuitable for systematic screening of large-scale, complex samples. Therefore, existing experimental techniques still face significant bottlenecks in achieving efficient and broad-spectrum identification of D-type amino acid modifications.
[0004] Besides direct experimental identification, another approach is to use computational prediction to assist or guide experiments. In the field of protein computation, most existing bioinformatics tools and prediction models focus on the sequence, structure, and function of natural L-proteins. For example, many mature algorithms exist for predicting the three-dimensional structure or physicochemical properties of proteins. However, these methods typically handle typical peptides and proteins composed of L-amino acids by default, and the datasets and extracted features on which their models are trained do not cover the chiral changes introduced by D-modification. Therefore, directly applying these models designed for L-proteins to predict D-modification results in insufficient specificity and accuracy.
[0005] Furthermore, from the perspective of mass spectrometry data analysis, retention time and collision cross-sectional area (CRI) are two key experimentally measurable parameters for peptide identification, playing a crucial role in this process. Existing technologies include several independent predictive models capable of estimating retention time or CRI based on peptide sequences. However, these models are also trained on a large amount of conventional L-type peptide data. When D-type amino acids are incorporated into a peptide, its spatial conformation and physicochemical properties change, causing the measured retention time and CRI to deviate from the model's predictions based on the L-type sequence.
[0006] In summary, the difficulties faced by existing technologies mainly lie in two aspects: first, experimental identification methods are cumbersome and limited; second, there is a lack of computational prediction tools specifically targeting the characteristics of D-type amino acid modifications, especially precise methods capable of comprehensively predicting their modification probability and key multidimensional physicochemical parameters. These problems arise because the characteristic that D-type modification does not alter mass exceeds the discriminative scope of conventional mass spectrometry identification procedures, and standardized mass spectrometry datasets specifically labeled with D-type modification information and suitable for machine learning are extremely scarce, making it difficult to train dedicated high-precision prediction models at the data foundation level. Summary of the Invention
[0007] One object of the present invention is to provide a method for accurate prediction of multidimensional structural information of D-proteins, so as to at least solve the above-mentioned problems.
[0008] To achieve the objectives and other advantages of this invention, a method for accurate prediction of multidimensional structural information of D-proteins is provided, comprising the following steps: S1. Construction and preprocessing of multi-species mass spectrometry dataset: Obtain raw proteomics mass spectrometry datasets containing at least two different species, and preprocess the raw datasets to construct labeled datasets; the preprocessing includes extracting peptide sequences from the raw data, screening for stereoisomeric peptides, determining retention time (RT) and collision cross-sectional area (CCS) information; S2. Integration of deep learning model construction: Construct a heterogeneous modification classification prediction model and a multidimensional physicochemical parameter prediction model; the classification prediction model is used to predict the probability that a peptide is a D-type amino acid modified peptide based on the peptide sequence; the physicochemical parameter prediction model is used to predict the retention time RT and collision cross-sectional area CCS based on the peptide sequence and its characteristics. S3. Model Training and Cross-Species Validation: The ensemble deep learning model constructed in step S2 is trained using the labeled dataset obtained in step S1, and the generalization ability of the model is evaluated by leaving out species cross-validation strategy. S4. Model performance verification based on independent experiments: The results of independent hot proteomics experiments are used as an external test set and compared with the prediction results of the integrated deep learning model to evaluate the prediction accuracy of the integrated deep learning model. S5. Model Application: The integrated deep learning model validated in steps S3 and S4 is used to identify D-type amino acid modifications in the target protein sample, and the output includes a list of modification probabilities, predicted RT values, and predicted CCS values.
[0009] Preferably, in step S2, the classification prediction model adopts a bidirectional LSTM neural network architecture, with its input being the one-hot encoding of the peptide sequence and its output being the binary classification probability.
[0010] Preferably, the bidirectional LSTM neural network architecture has the following characteristics: an input dimension of 20, a hidden layer dimension of 128, and a number of layers of 2.
[0011] Preferably, when training the classification prediction model, a cross-entropy loss function with class weights is used, where the class weights are calculated based on the reciprocal of the number of peptide classes in the training set; the AdamW optimizer is used with an initial learning rate of 0.001 and a weight decay of 0.0002, and optimization is performed in conjunction with a StepLR learning rate scheduler and an early stopping mechanism.
[0012] Preferably, in step S2, the physicochemical parameter prediction model is a hybrid neural network architecture based on Transformer, CNN, and LSTM, comprising at least: a Transformer encoder branch for processing one-hot encoded features of the peptide sequence to capture global dependencies; a Convolutional Neural Network (CNN) branch for extracting local pattern features from the peptide sequence features; a Long Short-Term Memory (LSTM) branch for processing the sequence information of the peptide sequence; and a fully connected network branch for processing the numerical features of the peptide; the output features of the four branches are concatenated and output as regression prediction values for RT or CCS via a fully connected layer.
[0013] Preferably, the Transformer encoder branch includes: a linear layer for mapping one-hot encoded features of peptide sequences to the embedding space, a custom positional encoding, and at least one TransformerEncoder layer, wherein the embedding space has a dimension of 96, the TransformerEncoder layer has 16 heads, and the feedforward network has a dimension of 128; the CNN branch includes two convolutional layers with kernel sizes of (1, 23) and (3, 1), respectively; the LSTM branch is a two-layer unidirectional LSTM with a hidden layer dimension of 128; the fully connected network branch includes at least one linear layer and an ELU activation function; the outputs of each branch are concatenated and then sequentially passed through two fully connected layers with a hidden layer dimension of 128 and an ELU activation function to finally output the regression prediction value.
[0014] Preferably, when training the physicochemical parameter prediction model, the mean squared error (MSE) is used as the loss function; the AdamW optimizer is used with an initial learning rate of 0.005 and a weight decay of 0.0002, combined with the ReduceLROnPlateau learning rate scheduler, wherein the ReduceLROnPlateau learning rate scheduler decays the learning rate to 50% of its original value when the loss has not decreased for at least 10 consecutive training epochs.
[0015] Preferably, in step S3, the cross-species validation specifically includes: S31, single-species training-cross-species testing: training the ensemble deep learning model using the labeled dataset of one species, and testing the performance of the ensemble deep learning model on the labeled datasets of other species; S32, mixed-species training of a general model: training the ensemble deep learning model after mixing the labeled datasets of all species, and evaluating the performance of the ensemble deep learning model on the independent labeled datasets of each species; the performance metrics of the ensemble deep learning model include classification accuracy, and the mean absolute error (MAE) and coefficient of determination (R²) of the RT and CCS predictions. 2 Correlation coefficient with Pearson.
[0016] Preferably, in step S4, the model performance verification based on independent experiments specifically involves: using the list of D-type modified proteins and peptides identified by the thermal proteomics experiment as the gold standard, calculating the consistency between the prediction results of the integrated deep learning model and the results of the thermal proteomics experiment at the protein and peptide levels; and simultaneously listing the targets predicted as positive by the integrated deep learning model but not captured by the thermal proteomics experiment as potential candidates, and feeding them back to the experiment for verification.
[0017] The present invention has at least the following beneficial effects: First, the method for accurate prediction of D-protein multidimensional structural information of the present invention achieves multidimensional and highly reliable computational prediction of D-type amino acid modifications through integrated design. This method first constructs a labeled dataset containing mass spectrometry information from multiple species, ensuring the diversity and biological relevance of the training data; then, it constructs and trains a deep learning model with both classification and regression functions, capable of simultaneously predicting the probability of stereoisomerization modifications and their key physicochemical parameters from peptide sequences; further, through a dual validation strategy of cross-species validation and independent experiments, it effectively confirms the robustness and generalization ability of the model; finally, it forms a systematic computational tool that can provide key auxiliary information in actual proteomics identification. This method reduces the dependence on complex chemical derivatization or special chromatographic systems in traditional experimental methods, providing a feasible computational path for large-scale screening of D-type modifications in samples.
[0018] Secondly, by further defining the classification prediction model as a bidirectional LSTM neural network architecture, this architecture, through bidirectional encoding of peptide sequences, can simultaneously capture upstream and downstream contextual information of amino acid residues, thereby providing a more comprehensive understanding of the impact of the local sequence environment on the probability of D-type modifications. Compared to models that only consider unidirectional sequences, bidirectional LSTM can more effectively identify potential patterns and dependencies related to chiral inversion in sequences, improving the model's ability to discriminate modification sites and the learning depth of sequence features, thus laying a more reliable model foundation for subsequent accurate classification.
[0019] Third, by limiting the specific dimensions and number of layers of the bidirectional LSTM, the model is provided with moderate representation ability and complexity. This configuration can effectively capture sequence features while better controlling the parameter scale and computational cost of the model. It helps to balance fitting ability and generalization performance during training, and avoids underfitting due to the model being too simple or overfitting due to the model being too complex. This ensures the stable and efficient execution of the classification prediction task.
[0020] Fourth, by limiting the loss function, optimizer, and related strategies in the training of the classification prediction model, it is beneficial to converge to a better solution more stably during the training process, improve the efficiency and effectiveness of model parameter updates, and terminate training in a timely manner when the model performance no longer improves, effectively preventing overfitting and ensuring that the final classification model has stronger robustness.
[0021] Fifth, the physicochemical parameter prediction model adopts a hybrid architecture of Transformer, CNN, and LSTM to extract and fuse diverse features of peptide sequences from different levels: the Transformer encoder is good at capturing long-range dependencies of the sequence globally; CNN can effectively extract local sequence patterns and motifs; and LSTM models the dynamic changes in sequence order. This multi-branch feature fusion strategy enables the model to comprehensively utilize the global context, local conformation, and sequence evolution information of the sequence, thereby achieving more comprehensive and accurate regression prediction of the two key physicochemical parameters, retention time and collision cross-sectional area, which are affected by multiple factors.
[0022] Sixth, by clarifying the specific implementation parameters of each branch of the hybrid architecture, such as embedding dimension, number of attention heads, convolution kernel size and network layer dimension, the abstract architecture is transformed into a model that can be specifically built and trained, so that each branch can remain efficient in its specific feature extraction task, while ensuring that the final spliced fused features have sufficient expressive power.
[0023] Seventh, by limiting the specific training details of the physicochemical parameter prediction model, including using mean squared error (MSE) as the loss function, using the AdamW optimizer and setting specific initial learning rate and weight decay, and combining the ReduceLROnPlateau learning rate scheduling strategy, a systematic and robust model optimization scheme was formed. This ensured that the complex hybrid neural network architecture could be trained fully and stably, ultimately achieving a more reliable and accurate prediction capability for the two key parameters of peptide retention time and collision cross-sectional area.
[0024] Eighth, by employing two strategies—"single-species training-cross-species testing" and "mixed-species training of a general model"—and evaluating the model across different species using multiple statistical indicators, we systematically tested the model's generalization ability and improved its predictive performance for new species samples outside the training data.
[0025] Ninth, by directly comparing the prediction results of the deep learning model with the experimental results based on hot proteomics, the accuracy of the model is rigorously tested in the closest real-world application scenario. By calculating the consistency at different levels and sorting out potential candidate targets, not only can the predictive performance of the model be objectively evaluated, but a closed loop of "computation prediction - experimental verification - feedback optimization" can also be formed. This helps to discover blind spots in the model and provide direction for further optimization, greatly enhancing the reliability and persuasiveness of the conclusions of the entire method.
[0026] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description
[0027] Figure 1 The figure shows the variation curves of validation set loss and accuracy of the heterogeneous modification classification prediction model with training rounds during species-specific training in one embodiment of the present invention. In the figure, (a) validation set loss and validation set accuracy of HeLa; (b) validation set loss and validation set accuracy of Yeast; (c) validation set loss and validation set accuracy of Drosophila; and (d) validation set loss and validation set accuracy of C. elegans. Figure 2 In one embodiment of the present invention, during mixed species training, the validation set loss and validation set accuracy of the heteromorphic modification classification prediction model change with the training rounds (a) and the identification results of the heteromorphic modification classification prediction model on the test set for different species (b). Figure 3 The figure shows the accuracy of the multidimensional physicochemical parameter prediction model in predicting RT / CCS in one embodiment of the present invention. In the figure, (a) the retention time prediction result of the full L-type peptide; (b) the average relative error distribution of the stereoisomer-modified peptide (STEP) retention time prediction; (c) the Pearson correlation coefficient of the stereoisomer-modified peptide (STEP) retention time prediction; (d) the collision cross-sectional area prediction result of the full L-type peptide; (e) the average relative error distribution of the stereoisomer-modified peptide (STEP) collision cross-sectional area prediction; and (f) the Pearson correlation coefficient of the stereoisomer-modified peptide (STEP) collision cross-sectional area prediction. Figure 4 The number of stereoisomeric modified proteins and cross Venn diagrams screened by integrating the deep learning model AlphaSTEP with the hot proteomics TPP&STEPSearch method in one embodiment of the present invention. Figure 5 This is a cross-Venn diagram showing the number of stereoisomeric modified peptides screened by integrating the deep learning model AlphaSTEP with the thermal proteomics methods TPP & STEPSearch in one embodiment of the present invention. Detailed Implementation
[0028] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings, so that those skilled in the art can implement it based on the description.
[0029] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.
[0030] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.
[0031] According to one embodiment of the present invention, a method for accurately predicting the multidimensional structural information of D-proteins includes: Step 1: Construction and Preprocessing of a Multi-Species Mass Spectrometry Dataset. Specifically, raw proteomic mass spectrometry data from at least two different species are required. Species may include, but are not limited to, humans, mice, yeast, or *E. coli*. The raw data are sourced from public databases or raw mass spectrometry files provided by collaborating laboratories. The preprocessing process first uses analytical software to extract peptide sequences, experimentally measured retention times, and collision cross-sectional areas (CCIs) from the raw data. Subsequently, the peptide sequences are standardized, for example, by converting them to uppercase letters and filtering out sequences containing non-standard amino acids. Retention times and CCI values are normalized or standardized to eliminate systematic biases between different experimental batches. Finally, the processed peptide sequences are correlated with their corresponding measured retention times, CCI values, and experimentally validated D-type modification annotation information to construct a structured labeled dataset for subsequent model training. This step, by integrating multi-species, multi-dimensional experimental data, provides a necessary and diverse data foundation for training a prediction model specifically targeting D-type modifications, helping to improve the model's generalization ability and biological relevance.
[0032] Step 2: Construction of the integrated deep learning model. This integrated model comprises two independent but collaborative sub-models. The first is a heterogeneous modification classification prediction model, which employs a bidirectional long short-term memory network architecture. The input is a one-hot encoding of the peptide sequence, with an encoding dimension of 20, corresponding to 20 common amino acids. The network structure may include an embedding layer, two bidirectional LSTM layers with a hidden layer dimension of 128, and a fully connected layer and activation function for outputting binary classification probabilities. The second is a multidimensional physicochemical parameter prediction model, which is a hybrid neural network architecture designed to simultaneously predict retention time and collision cross-sectional area from peptide sequences. This multidimensional physicochemical parameter prediction model incorporates multiple feature extraction branches: a Transformer encoder branch for capturing global sequence dependencies (with an embedding dimension of 96 and attention heads of 16); a convolutional neural network branch for extracting local sequence patterns (using two convolutional layers with kernel sizes of (1,23) and (3,1); a unidirectional LSTM branch for processing sequence order information (using a 2-layer hidden layer structure with a dimension of 128); and a fully connected branch for handling possible additional numerical features. The feature vectors extracted from each branch are concatenated and then fused and regressed through several fully connected layers, ultimately outputting the predicted values for retention time and collision cross-sectional area. This step utilizes a targeted deep learning architecture to determine the probability of D-type modification and to predict key physicochemical parameters, providing computational tools for mining chiral-related multidimensional features from sequence information.
[0033] Step 3: Model Training and Validation Strategy. The model is trained using the previously constructed labeled dataset. During training, the classification prediction model employs a cross-entropy loss function with class weights to address data imbalance, using the AdamW optimizer with an initial learning rate of 0.001, and a learning rate scheduling strategy. The physicochemical parameter prediction model uses mean squared error as the loss function, employing the AdamW optimizer with an initial learning rate of 0.005. To evaluate the model's generalization ability, a cross-species validation strategy is adopted, implementing two schemes: first, training the model using single-species data and then testing its performance on data from other species; second, training a general model using data from all species and then testing it on independent data from each species. Performance evaluation metrics include classification accuracy, precision, recall, and mean absolute error, coefficient of determination, and Pearson correlation coefficient for retention time and collision cross-sectional area predictions. Furthermore, the model's prediction results are externally validated using hot proteomics experiments independent of the training set as the gold standard, calculating the consistency between peptide and protein levels. This step involves systematic training and rigorous cross-species and independent experimental verification, which allows for an objective assessment of the model's predictive accuracy, robustness, and applicability to new data, ensuring the model's reliability before it is put into application.
[0034] Step 4: Practical Application of the Validated Model. The fully trained and validated ensemble deep learning model is deployed on the computing platform. For the target protein sample to be analyzed, a list of possible peptide sequences is first generated through enzyme digestion simulation. These peptide sequences are then input into the ensemble model. The model processes each peptide sequence; the classification prediction model outputs the probability value of the peptide being modified by D-amino acids, and the physicochemical parameter prediction model outputs the predicted retention time and predicted collision cross-sectional area of the peptide. The system integrates these three prediction results for all peptides to generate a structured list. This list can serve as preliminary information for downstream experimental identification. For example, in mass spectrometry identification, peptides with high modification probabilities and their predicted retention times and collision cross-sectional areas can be prioritized or used to optimize identification parameters. In this step, computational prediction capabilities are transformed into a practically usable auxiliary identification tool, providing a reference list containing quantified probabilities and key physicochemical parameter predictions for rapid, multi-dimensional screening of D-amino acid modifications in target samples. This helps narrow the scope of experimental validation and improves research efficiency.
[0035] According to another embodiment of the present invention, a classification prediction model employing a bidirectional LSTM neural network architecture and its workflow is described. First, the peptide sequence is converted into a fixed-length numerical representation, specifically using one-hot encoding, where each amino acid residue corresponds to a sparse vector of length 20, representing 20 standard amino acids. Subsequently, this encoded sequence is input into a bidirectional long short-term memory network. The network contains two LSTM layers, forward and backward, scanning from the N-terminus to the C-terminus and C-terminus to the N-terminus, respectively. At each time step, a network unit receives the embedding representation of the current amino acid and the hidden state from the previous time step, calculating and updating the unit state and hidden state through input gates, forget gates, and output gates. The final hidden state of the forward and backward LSTM layers, typically aggregated from the hidden states of all time steps (e.g., taking the state from the last time step or performing pooling), is concatenated into a fused feature vector. This feature vector is then nonlinearly transformed through one or more fully connected layers, ultimately outputting a two-dimensional Logits vector representing the classification category. This vector is then processed using a cross-entropy loss function with class weights to achieve binary classification prediction of whether the peptide contains D-type amino acid modifications. During implementation, network hyperparameters such as the hidden layer dimension can be selected as 64, 128, or 256, and the number of network layers can be selected as 1, 2, or 3. The specific values are determined through grid search or by tuning based on the performance of the validation set.
[0036] In this embodiment, the local environmental information upstream and downstream of each residue in the peptide sequence is captured simultaneously by a bidirectional loop structure, thereby more comprehensively modeling the influence of sequence context on the possibility of stereoisomer modification and improving the model's ability to distinguish modification patterns.
[0037] According to another embodiment of the present invention, the specific architectural parameters of the bidirectional LSTM neural network are further defined. The input dimension is 20, directly corresponding to the length of the vector at each position in the one-hot encoding of the peptide sequence. The hidden layer dimension is set to 128, meaning that the forward and backward LSTMs will output a 128-dimensional hidden state vector at each time step. The number of network layers is set to 2, i.e., it contains two stacked bidirectional LSTM layers, with the output sequence of the first layer serving as the input to the second layer. During operation, the one-hot encoding of the peptide sequence is first fed into the first bidirectional LSTM layer, which learns the first-order feature representation of the sequence; its output hidden state sequence is then input into the second bidirectional LSTM layer, which captures higher-order sequence dependencies at a more abstract level. After two layers of processing, the final forward and backward hidden state features are merged. This parameter combination (20-128-2) is set based on a balance between model capacity and training data scale, aiming to provide sufficient expressive power to capture complex patterns while controlling model complexity to prevent overfitting on limited data.
[0038] In this embodiment, by limiting the structural parameters of the bidirectional LSTM neural network architecture, it is ensured that the model can learn effectively during training and maintain a certain level of generalization performance when faced with new sequences.
[0039] According to another embodiment of the present invention, a specific training method for a classification prediction model is disclosed. The training process uses a labeled dataset constructed in step S1, where each sample contains a peptide sequence and its true modification label (0 indicates unmodified, 1 indicates D-type modification). Since modified peptides are typically far fewer than unmodified peptides in the data, a cross-entropy loss function with class weights is used to alleviate class imbalance. The class weights are calculated as follows: first, the number of positive (modified) and negative (unmodified) samples in the training set is counted; the weight of each class is set as the reciprocal of the proportion of samples in that class to the total number of samples, or calculated using other normalization methods, so that the loss function imposes a higher penalty on mispredictions of the minority class. The optimization process uses the AdamW optimizer, which combines the adaptive learning rate characteristic of the Adam algorithm with weight decay regularization. The initial learning rate is set to 0.001, and the weight decay coefficient is set to 0.0002. The learning rate scheduling adopts a StepLR strategy, for example, setting the learning rate to decay by multiplying it by 0.5 every 20 training epochs. Simultaneously, an early stopping mechanism is implemented: during training, the model's performance (such as classification accuracy or loss value) on an independent validation set is continuously monitored. When the validation set performance no longer improves within a set number of consecutive epochs (e.g., 15 epochs), training is terminated, and the model parameters are rolled back to the snapshot where the validation performance was optimal. This training configuration aims to achieve stable and efficient gradient descent and suppress overfitting through regularization and early stopping.
[0040] In this embodiment, a systematic training strategy is employed, including a loss function for handling imbalanced data, a regularized optimizer, and dynamically adjusted learning rate and early stopping mechanism. This ensures that the classification model can robustly learn discriminative features from the data, thereby achieving more reliable and generalized prediction performance.
[0041] According to another embodiment of the present invention, the specific architecture of the physicochemical parameter prediction model is a hybrid neural network. This model is designed to predict retention time and collision cross-sectional area from peptide sequences and their features. Its core computational logic includes four parallel processing branches. The first branch is a Transformer encoder branch, whose operation is as follows: the one-hot encoding of the input peptide sequence is first mapped to an embedding space of dimension d_model through a linear projection layer; then it is added to a custom positional encoding vector, which is used to inject sequence information into the position of each amino acid in the sequence; the resulting embedded sequence with positional information is fed into stacked Transformer encoder layers, each layer containing a multi-head self-attention mechanism and a feedforward neural network, which calculates the association weight between any two amino acids in the sequence through self-attention, thereby capturing global long-range dependencies; the encoded sequence dimension is then subjected to global average pooling as the global feature vector extracted by this branch. The second branch is a convolutional neural network branch, which receives the same peptide sequence one-hot encoding or initial embedding as input, and processes it through two convolutional layers. After two convolutional layers (kernel sizes (1,23) and (3,1) respectively), operations are performed on the feature dimension to fuse information from different feature maps. Each convolutional layer is followed by an activation function and possible pooling layers, ultimately flattening the output feature map into a local feature vector. The third branch is a long short-term memory network branch, which sequentially inputs the sequence one-hot encoding into a unidirectional multi-layer LSTM network. The LSTM units are iteratively updated step-by-step, modeling the sequence dependency and dynamic changes between amino acids through its gating mechanism. The hidden state of the last time step or the aggregation of hidden states from all time steps is taken as the sequential feature vector of this branch. The fourth branch is a fully connected network branch, which receives peptide-related numerical auxiliary features (such as peptide length, net charge, etc., if present) as input, and performs nonlinear transformations through several fully connected layers with activation functions to obtain a numerical feature vector. Finally, the feature vectors output from the four branches are concatenated along the feature dimension to form a comprehensive feature representation. This comprehensive feature is then processed by a regression head network consisting of several fully connected layers, and finally outputs a 1D regression result (training single-task models separately), retaining time or collision cross-sectional area.
[0042] This embodiment integrates multiple neural network components that can capture global dependencies, local patterns, sequential dynamics, and numerical features, enabling the model to comprehensively analyze peptide sequences from different angles and levels. This provides a more comprehensive feature basis for accurately predicting two key physicochemical parameters that are affected by multiple factors: retention time and collision cross-sectional area.
[0043] According to another embodiment of the present invention, the specific implementation parameters of each branch of the hybrid neural network are further defined. In the Transformer encoder branch, the dimension of the embedding space is set to 96; the custom positional encoding can be generated using sine and cosine functions or learnable positional embeddings; the number of TransformerEncoder layers can be set to 1, 2, or more, the number of heads in its internal multi-head self-attention mechanism is set to 16, and the inner layer dimension of the feedforward neural network is set to 128. In the convolutional neural network branch, the kernel size of the first convolutional layer is set to (1, 23); the kernel size of the second convolutional layer is set to (3, 1); and the ReLU activation function can be used after each convolution. In the Long Short-Term Memory network branch, two unidirectional LSTM layers are used, and the hidden state dimension of each LSTM layer is set to 128. In the fully connected network branch, at least one linear layer is included, followed by an ELU activation function to transform the numerical features. After feature concatenation, the processing flow is as follows: the concatenated comprehensive feature vector is first input into a fully connected layer with a hidden dimension of 128, and then passes through the ELU activation function; its output is then input into a fully connected layer with a hidden dimension of 128, and passes through the ELU activation function again; finally, a 1D regression result is output. These parameters, such as embedding dimension of 96, number of attention heads of 16, convolutional kernel height of 23, and LSTM hidden layers of 128, are typical configurations selected based on a trade-off between model expressive power, computational cost, and preliminary performance on the validation set, and can also be adjusted within a certain range.
[0044] In this embodiment, a clear and implementable parameter configuration is provided for the above-mentioned hybrid neural network architecture, which enables the complex model to be specifically constructed and trained, and each component to operate effectively at a specified complexity, working together to complete the regression prediction task.
[0045] According to another embodiment of the present invention, a specific optimization method for training a physicochemical parameter prediction model is provided. The training process uses a labeled dataset, where each sample contains a peptide sequence and its corresponding measured value as the regression target. The loss function for model training is the mean squared error (MSE), which is calculated by averaging the squared difference between the model's predicted value and the measured value for each sample across all samples. The optimizer used is AdamW, with an initial learning rate set to 0.005 and a weight decay coefficient set to 0.0002. During training, the ReduceLROnPlateau learning rate scheduling strategy is employed: at the end of each training cycle, the loss value of the model on an independent validation set is calculated; this validation loss is monitored, and if it does not show a decrease (i.e., does not reach a new lower point) for N consecutive training cycles (N is at least 10, for example, it can be set to 10 or 15), learning rate decay is triggered, multiplying the current learning rate by a factor γ (γ is set to 0.5) to reduce it; after the learning rate decreases, the scheduler continues to monitor the validation loss, and this process can be repeated multiple times until the learning rate falls below a preset lower limit or training is terminated by an early stopping mechanism. The entire training process typically sets a maximum number of training epochs and may simultaneously employ early stopping mechanisms to prevent overfitting. This training configuration aims to directly optimize the closeness of predicted values to true values through mean squared error, efficiently update parameters using the AdamW optimizer, and fine-tune model parameters in the later stages of training by dynamically reducing the learning rate based on validation set performance, thereby promoting model convergence to a better local optimum.
[0046] In this embodiment, by specifying a set of typical training hyperparameters and scheduling strategies for regression tasks, a stable and effective optimization scheme is provided for training the above-mentioned physicochemical parameter prediction model, which helps to improve the accuracy and stability of the model in predicting retention time and collision cross-sectional area.
[0047] According to another embodiment of the present invention, cross-species validation includes single-species training-cross-species testing and mixed-species training of a general model. The single-species training-cross-species testing process involves selecting data from a specific species (e.g., humans) as the training set from a constructed multi-species mass spectrometry labeled dataset for de novo training or fine-tuning of the ensemble deep learning model. After training, the trained model is fixed, and its parameters are no longer updated. Subsequently, independent labeled datasets from one or more species not encountered during training (e.g., mice, yeast) are used as the test set. Peptide sequences from the test set are input into the trained model to obtain the model's prediction results for each peptide, including modification classification probability, retention time prediction, and collision cross-sectional area prediction. These prediction results are then compared and analyzed with known labeled information in the test set. This strategy can be repeated cyclically, for example, sequentially using each species as the training source and the remaining species as the test targets, to comprehensively evaluate the model's ability to transfer knowledge learned from single-species data to other species.
[0048] The implementation process of training a general model using hybrid species involves merging all labeled datasets from all available species (e.g., humans, mice, yeast, and E. coli) to form a hybrid training dataset with diverse species origins. This hybrid dataset is used to uniformly train the ensemble deep learning model, aiming to enable the model to learn sequence feature patterns shared across species and related to D-type modifications and physicochemical parameters. After training, the performance of the general model needs to be evaluated. Evaluation is not performed on the hybrid dataset itself or its randomly partitioned test set, but rather by separately splitting the datasets of each original species as independent test sets. The test data for each species is then input into the trained general model, and the model's performance metrics on that species' data are calculated. This strategy is used to verify whether the model can maintain good predictive performance for data from each individual species after fusing information from multiple species.
[0049] The performance evaluation metrics used in the above validation strategy are as follows: For classification tasks, i.e., predicting whether a peptide is D-modified, the evaluation metrics include accuracy, precision, and recall. These metrics are calculated based on the model's predicted binary classification results (by setting a probability threshold, such as 0.5, to convert probability values into class labels) and the true labels. For regression tasks, i.e., predicting retention time and collision cross-sectional area, the evaluation metrics include mean absolute error, coefficient of determination, and Pearson correlation coefficient. The mean absolute error is calculated as the average of the absolute differences between the predicted and true values for all test samples, used to measure the average deviation of the predictions. The coefficient of determination is obtained by performing linear regression analysis on the predicted and true values, used to quantify the proportion of variation in the true values explained by the predicted values. The Pearson correlation coefficient directly calculates the degree of linear correlation between the predicted and true values. These metrics together provide a multi-dimensional measure of the model's predictive accuracy, stability, and consistency.
[0050] In this embodiment, by designing and implementing a systematic cross-species validation process and a multi-dimensional performance evaluation system, the generalization ability and robustness of the integrated deep learning model when faced with new data from different biological sources and with potential distributional differences can be objectively and comprehensively evaluated, thus providing empirical evidence for judging whether the model is applicable to a wider range of unknown samples.
[0051] According to another embodiment of the present invention, model performance validation based on independent experiments includes: using the identification results of independent thermal proteomics experiments as the gold standard to quantitatively evaluate the model's prediction accuracy. The implementation process is as follows: First, a list of D-type modified proteins and their specific modified peptides identified through independently completed thermal proteomics experiments is obtained; this list serves as the "gold standard" dataset for external validation. This experimental data originates from independent studies using different sample pretreatment procedures or different biological samples to ensure no overlap with the model training dataset. Subsequently, the ensemble deep learning model to be evaluated is applied to the original samples that generated the "gold standard" list or the corresponding protein sequence database to obtain a list of the model's prediction results. Evaluation is performed at both the protein and peptide levels. At the peptide level, the overlap ratio between peptides predicted as positive by the model and peptides in the "gold standard" list is calculated, along with the proportion of peptides in the "gold standard" list correctly predicted as positive by the model. At the protein level, based on the peptide-to-protein mapping relationship, the overlap between proteins predicted by the model as containing modifications and modified proteins in the "gold standard" is calculated. These calculations can be performed by constructing a confusion matrix and deriving metrics such as precision and recall, thereby quantifying the degree of consistency between model predictions and experimental observations.
[0052] The discrepancies between model predictions and experimental results are then analyzed, and feedback validation is performed. After consistency comparison, a special category of targets is identified: peptides or proteins that the model predicts with high confidence to be D-type modified positive, but which were not captured or reported in the gold standard hot proteomics experimental data. These targets are listed as potential modification candidates. To explore the reasons for these discrepancies and potentially discover new modification sites, this list of potential candidates needs to be fed back into the experimental stage for targeted validation. Validation experiments may include, but are not limited to: reanalysis using mass spectrometry detection at higher depths or under different conditions; validation using standards synthesized for specific peptides; or designing targeted validation experiments based on chiral separation. Through this feedback validation, it can be confirmed whether some model predictions are true positives but were missed by previous experiments, thereby assessing the model's discovery potential.
[0053] In this embodiment, by introducing experimental data independent of model training as an objective evaluation benchmark, the real-world predictive performance of the model can be rigorously externally verified. At the same time, by systematically analyzing the differences between prediction and experiment and establishing a feedback verification mechanism, not only can the reliability of the model be evaluated, but a closed loop from computational prediction to experimental verification can also be formed, which helps to discover potential model blind spots or new biological clues, thereby enhancing the rigor and practicality of the entire methodology.
[0054] The following specific embodiment will further illustrate the method for accurate prediction of D-protein multidimensional structure information of the present invention.
[0055] Step 1: Construction and preprocessing of multi-species mass spectrometry dataset.
[0056] Raw mass spectrometry data from human HeLa cells, yeast, nematode (C. elegans), and fruit fly (Drosophila) were obtained from a four-species mass spectrometry dataset published by Matthias Mann's team. This dataset contains a total of 2,490,131 peptide segments and 426,845 unique peptide sequences, as shown in Table 1. Additionally, 5,149 protein sequences related to common human diseases were downloaded from the UniProt database for subsequent validation.
[0057] Table 1 The raw mass spectrometry data were preprocessed, including peptide sequence extraction, retention time (RT), and collision cross-sectional area (CCS) extraction and normalization. For potential D-type modifications, a stereoisomer screening strategy was employed: for peptides with the same amino acid sequence, if the difference in retention time was greater than 0.5% of the retention time of the high-intensity peak or the difference in collision cross-sectional area was greater than 1.5%, and the signal response intensity ratio was greater than 1%, they were considered candidate stereoisomer peptides (depending on the experiment). Finally, a labeled dataset containing sequence, RT, CCS, and modification tags (0 for no modification, 1 for D-modification) was constructed.
[0058] Step 2: Integrate deep learning model training.
[0059] An ensemble deep learning model (AlphaSTEP) was trained using the datasets of the four species mentioned above. The ensemble deep learning model consists of two sub-modules: 1. Heterogeneous Modification Classification Prediction Model: Used to predict whether a peptide is D-type modified. The input is a 20-dimensional one-hot encoding of the peptide sequence. The network structure is a 2-layer bidirectional LSTM with 128 hidden layers. It uses a cross-entropy loss function with class weights, the AdamW optimizer (learning rate 0.001, weight decay 0.0002), and a StepLR learning rate scheduler and early stopping mechanism.
[0060] 2. Multidimensional Physicochemical Parameter Prediction Model: Used to predict peptide retention time (RT) and collision cross-sectional area (CCS). The model employs a multi-branch architecture, including a Transformer encoder (embedding dimension 96, head number 16), a CNN branch (convolutional kernel sizes (1,23) and (3,1) respectively), an LSTM branch (2 unidirectional layers, 128 hidden layers), and a fully connected branch. Mean squared error (MSE) is used as the loss function, with the AdamW optimizer (learning rate 0.005, weight decay 0.0002) and the ReduceLROnPlateau learning rate scheduler.
[0061] An ensemble deep learning model was trained on a four-species dataset, and its generalization ability was evaluated using a species-leaving cross-validation strategy. The training results are as follows: Figure 1 , Figure 2 and Figure 3 The integrated deep learning models achieved an accuracy of over 92.04% on classification tasks, with the mean absolute error of both RT and CCS predictions below 2.5% and the Pearson correlation coefficient greater than 0.9.
[0062] Step 3: Experimental Validation Based on Thermal Proteomics To verify the model's performance in real-world applications, independent thermal proteomics experimental data were used as external tests. The thermal proteomics experimental dataset contained 24,318 input peptides, corresponding to 1,424 proteins. Changes in protein thermostability were detected using thermal proteomics to verify the consistency between the AlphaSTEP model's predictions and the results identified by the thermal proteomics experiments (TPP & STEPSearch).
[0063] The peptide sequences identified in the experiment were input into the AlphaSTEP model to screen for potential stereoisomer-modified peptides and their corresponding proteins. Subsequently, the model predictions were analyzed in conjunction with the results of thermal proteomics experiments. The results are as follows: Figure 4 , Figure 5 As shown: At the protein level, the overlap between the 298 modified proteins predicted by AlphaSTEP and the 194 stereoisomers identified in the experimental data was 74.74% (145 / 194). At the peptide level, the model-predicted modified peptides showed a high agreement of 93.59% (424 / 453) with the stereoisomers confirmed by the isomer search.
[0064] The two-way validation strategy combining experimentation and prediction significantly improved the coverage (by approximately 20%) and accuracy (false positive rate reduced to below 5%) of the identification of stereoisomer-modified proteins. Step 4: Model application and functional analysis.
[0065] The validated model was applied to human disease-related proteome data to predict D-type modifications. APM enzyme digestion experiments were used for validation, confirming the predicted modification sites. Further functional enrichment analysis revealed that these modified proteins were significantly enriched in pathways related to neurodegenerative diseases, such as Alzheimer's and Parkinson's diseases, providing new molecular targets for disease mechanism research.
[0066] This embodiment demonstrates the effective application of the described method for accurate prediction of D-protein multidimensional structural information in real biological samples. The ensemble deep learning model trained using mass spectrometry data from four species exhibits high accuracy, strong generalization ability, and good experimental consistency in both cross-species validation and independent cell experiments. This method can provide reliable computational support for the systematic identification and functional analysis of D-type modifications in complex diseases such as neurodegenerative diseases.
[0067] The number of devices and processing scale described herein are for simplification of the invention. Applications, modifications, and variations of the present invention's method for accurate prediction of D-protein multidimensional structure information will be readily apparent to those skilled in the art.
[0068] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.
Claims
1. A method for accurate prediction of D-protein multidimensional structural information, characterized in that, Includes the following steps: S1. Construction and preprocessing of multi-species mass spectrometry dataset: Obtain raw proteomics mass spectrometry datasets containing at least two different species, and preprocess the raw datasets to construct labeled datasets; The preprocessing includes extracting peptide sequences from the raw data, screening stereoisomer peptides, determining retention time RT and collision cross-sectional area (CCS) information; S2. Integration of deep learning model construction: Construct a heterogeneous modification classification prediction model and a multidimensional physicochemical parameter prediction model; the classification prediction model is used to predict the probability that a peptide is a D-type amino acid modified peptide based on the peptide sequence; the physicochemical parameter prediction model is used to predict the retention time RT and collision cross-sectional area CCS based on the peptide sequence and its characteristics. S3. Model Training and Cross-Species Validation: The ensemble deep learning model constructed in step S2 is trained using the labeled dataset obtained in step S1, and the generalization ability of the model is evaluated by leaving out species cross-validation strategy. S4. Model performance verification based on independent experiments: The results of independent hot proteomics experiments are used as an external test set and compared with the prediction results of the integrated deep learning model to evaluate the prediction accuracy of the integrated deep learning model. S5. Model Application: The integrated deep learning model validated in steps S3 and S4 is used to identify D-type amino acid modifications in the target protein sample, and the output includes a list of modification probabilities, predicted RT values, and predicted CCS values.
2. The method for accurate prediction of D-protein multidimensional structure information as described in claim 1, characterized in that, In step S2, the classification prediction model adopts a bidirectional LSTM neural network architecture, with its input being the one-hot encoding of the peptide sequence and its output being the binary classification probability.
3. The method for accurate prediction of D-protein multidimensional structure information as described in claim 2, characterized in that, The bidirectional LSTM neural network architecture is specifically designed with an input dimension of 20, a hidden layer dimension of 128, and 2 layers.
4. The method for accurate prediction of D-protein multidimensional structure information as described in claim 2, characterized in that, When training the classification prediction model, a cross-entropy loss function with class weights is used, where the class weights are calculated based on the reciprocal of the number of peptide classes in the training set; the AdamW optimizer is used with an initial learning rate of 0.001 and a weight decay of 0.0002, and is optimized in conjunction with a StepLR learning rate scheduler and an early stopping mechanism.
5. The method for accurate prediction of D-protein multidimensional structure information as described in claim 1, characterized in that, In step S2, the physicochemical parameter prediction model is a hybrid neural network architecture based on Transformer, CNN, and LSTM, including at least: a Transformer encoder branch for processing one-hot encoded features of peptide sequences to capture global dependencies; a Convolutional Neural Network (CNN) branch for extracting local pattern features from the peptide sequence features; a Long Short-Term Memory (LSTM) branch for processing the sequence information of the peptide sequence; and a fully connected network branch for processing the numerical features of the peptides; the output features of the four branches are concatenated and then output as regression predictions for RT or CCS through a fully connected layer.
6. The method for accurate prediction of D-protein multidimensional structure information as described in claim 5, characterized in that, The Transformer encoder branch includes: a linear layer for mapping one-hot encoded features of peptide sequences to the embedding space, a custom positional encoding, and at least one TransformerEncoder layer with a dimension of 96. The TransformerEncoder layer has 16 heads, and the feedforward network has a dimension of 128. The CNN branch contains two convolutional layers with kernel sizes of (1, 23) and (3, 1), respectively. The LSTM branch consists of two unidirectional LSTM layers with a hidden layer dimension of 128. The fully connected network branch contains at least one linear layer and an ELU activation function. The outputs of each branch are concatenated and then sequentially passed through two fully connected layers with a hidden layer dimension of 128 and an ELU activation function to finally output the regression prediction value.
7. The method for accurate prediction of D-protein multidimensional structure information as described in claim 5, characterized in that, When training the physicochemical parameter prediction model, the mean squared error (MSE) is used as the loss function; the AdamW optimizer is used with an initial learning rate of 0.005 and a weight decay of 0.0002, combined with the ReduceLROnPlateau learning rate scheduler, wherein the ReduceLROnPlateau learning rate scheduler decays the learning rate to 50% of its original value when the loss has not decreased for at least 10 consecutive training epochs.
8. The method for accurate prediction of D-protein multidimensional structure information as described in claim 1, characterized in that, In step S3, the cross-species validation specifically includes: S31, single-species training-cross-species testing: training the ensemble deep learning model using the labeled dataset of one species, and testing the performance of the ensemble deep learning model on the labeled datasets of other species; S32, mixed-species training of a general model: training the ensemble deep learning model after mixing the labeled datasets of all species, and evaluating the performance of the ensemble deep learning model on the independent labeled datasets of each species; the performance metrics of the ensemble deep learning model include classification accuracy, and the mean absolute error (MAE) and coefficient of determination (R²) of the RT and CCS predictions. 2 Correlation coefficient with Pearson.
9. The method for accurate prediction of D-protein multidimensional structure information as described in claim 1, characterized in that, In step S4, the model performance verification based on independent experiments specifically involves: using the list of D-type modified proteins and peptides identified by the thermal proteomics experiment as the gold standard, and calculating the consistency between the prediction results of the integrated deep learning model and the results of the thermal proteomics experiment at the protein and peptide levels; simultaneously, listing the targets predicted as positive by the integrated deep learning model but not captured by the thermal proteomics experiment as potential candidates, and feeding them back to the experiment for verification.