Single molecule force spectroscopy classification method based on convolutional neural network
By preprocessing and data augmenting single-molecule force spectrum images using convolutional neural networks, a model is constructed and trained, solving the problem of low classification efficiency in existing single-molecule force spectrum technologies and achieving fast and accurate classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA JILIANG UNIV
- Filing Date
- 2022-11-01
- Publication Date
- 2026-06-09
AI Technical Summary
Existing single-molecule force spectroscopy classification methods are inefficient, slow, and time-consuming, making it difficult to meet the needs of rapid classification of large amounts of data.
A convolutional neural network was used to preprocess, augment, and manually annotate single-molecule force spectrum images. A convolutional neural network model was constructed, and the model was trained using the gradient descent method to achieve fast and accurate classification.
While ensuring classification accuracy, the time for single-molecule force spectrum classification was significantly reduced, thus improving classification efficiency.
Smart Images
Figure CN115641918B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of single-molecule detection, and more specifically to a single-molecule force spectrum classification method based on a convolutional neural network. Background Technology
[0002] The conformational changes within biomolecules and the mechanisms of intermolecular recognition can be studied qualitatively and quantitatively using mechanical properties. Single-molecule force spectroscopy utilizes moment data generated by single-molecule techniques to reveal important biophysical information about biomolecules, such as the number of folding and unfolding transitions in protein conformational changes, as well as their relative stability and approximate kinetics. Single-molecule techniques, including optical tweezers and magnetic tweezers, generate massive and complex data signals during single-molecule force spectroscopy testing, accompanied by significant noise interference and randomness.
[0003] The massive data volume and reproducibility requirements of single-molecule force spectroscopy (SFSS) make the classification and filtering of SFSS data a significant factor limiting its development. Bosshart PD et al., in their paper "Reference-free alignment and sorting of single-molecule force spectroscopy data" published in the *Biophysical Journal*, grouped the data by the number of force spectral unfolding events, then subclassified the SFSS using cross-correlation-based ranking, and further refined the classification by extracting unfolding paths and peak positions through principal component analysis and clustering. However, this type of single-molecule force spectroscopy classification method is inefficient, slow, and time-consuming, and with the increasing data throughput, further manual subdivision and filtering are still necessary. Summary of the Invention
[0004] This invention addresses the problems of low efficiency, slow speed, and long time consumption in existing single-molecule force spectrum classification methods by proposing a single-molecule force spectrum classification method based on convolutional neural networks. By using convolutional neural networks, the method classifies folding events in single-molecule force spectrum images, achieving rapid classification while maintaining classification accuracy.
[0005] The technical solution of this invention is a single-molecule force spectrum classification method based on convolutional neural networks, comprising the following steps:
[0006] 1) Obtain single-molecule force spectrum images.
[0007] The single-molecule force spectrum image was obtained by performing protein folding experiments using optical tweezers, and the experimental data was converted into a single-molecule force spectrum image using the Matplotlib plotting library.
[0008] 2) Perform preprocessing, data augmentation, and manual labeling on the single-molecule force spectrum images in sequence to obtain manually labeled single-molecule force spectrum images, and create a dataset containing training, validation, and test sets.
[0009] The preprocessing involves converting all single-molecule force spectrum images into grayscale images with a pixel size of 256×256.
[0010] The data augmentation process involves cropping the preprocessed single-molecule force spectrum image from nine positions: top left, top center, top right, left center, center, center right, bottom left, bottom center, and bottom right. The cropped image has a pixel size of 224×224.
[0011] The manual labeling category refers to classifying all single-molecule force spectrum images into category 1, category 2, category 3, category 4, and category 5 based on the different numbers of folding events in the data-enhanced single-molecule force spectrum images. These categories correspond to single-molecule force spectrum images with 0, 1, 2, 3, and 4 folding events, respectively.
[0012] Preferably, the dataset contains no fewer than 7,000 single-molecule force spectrum images, with 80% of the dataset serving as the training set, 10% as the validation set, and 10% as the test set.
[0013] 3) Construct a convolutional neural network model, and use the training set to train the constructed convolutional neural network model to generate a trained convolutional neural network model.
[0014] The convolutional neural network model is composed of an input module, a first pooling module, a first convolutional module, a second pooling module, a second convolutional module, a third pooling module, a third convolutional module, a fourth pooling module, a fully connected module, and an output module, which are cascaded in sequence.
[0015] The input module and the first convolutional module consist of two convolutional layers, which are cascaded sequentially with identical parameters. The kernel size in each convolutional layer is 3*3, the stride is 1, and the expansion edge is 1.
[0016] The second and third convolutional modules consist of three convolutional layers, which are cascaded sequentially with identical parameters. The kernel size in each convolutional layer is 3*3, the stride is 1, and the expansion margin is 1.
[0017] The convolutional layer extracts feature vectors from corresponding regions of the input single-molecule force spectrum image or feature map by performing convolution operations between the convolution kernel and a portion of the input single-molecule force spectrum image or feature map. Each convolutional layer is followed by LReLU nonlinear activation and batch normalization.
[0018] The first, second, third, and fourth pooling modules are composed of a max pooling layer with a pooling kernel size of 2*2 and a step size of 2.
[0019] The max pooling layer extracts the maximum eigenvalue from the feature vectors within the pooling region using a pooling kernel.
[0020] The fully connected module consists of a fully connected layer and a dropout layer. The dropout layer will randomly select some nodes in the fully connected layer for output and use the softmax activation function to generate probability distributions for the five output categories.
[0021] The output module consists of an output layer containing 5 neurons.
[0022] The training refers to training the parameters of the convolutional neural network using the gradient descent method, so that the cross-entropy loss function converges. The optimizer function based on the gradient descent method adopts the Adam optimizer function with an initial learning rate of 0.001.
[0023] 4) Input the single-molecule force spectrum images from the test set into the trained convolutional neural network model to obtain the number and category information of folding events in the single-molecule force spectrum.
[0024] Compared with the prior art, the present invention has the following beneficial technical effects:
[0025] The above scheme, targeting the characteristic features of single-molecule force spectrum images, employs a convolution kernel padding operation to maintain the original size of the feature map and preserve features after passing through the convolutional layer, and uses the LReLU activation function to involve more nodes in training, achieving fast and accurate classification of folding events in single-molecule force spectrum images. Compared with traditional single-molecule force spectrum classification methods, the single-molecule force spectrum classification method using convolutional neural networks in this invention significantly reduces classification time while maintaining accuracy. Attached Figure Description
[0026] Figure 1 This is a flowchart illustrating the single-molecule force spectrum classification method based on convolutional neural networks of this invention.
[0027] Figure 2 These are the single-molecule force spectrum images and data-enhanced images of the present invention.
[0028] Figure 3 This is a structural diagram of the convolutional neural network of the present invention.
[0029] Figure 4 This is information on the number and category of folding events in some single-molecule force spectrum images obtained by the present invention. Detailed Implementation
[0030] The present invention will now be described in detail with reference to the accompanying drawings, but the present invention is not limited thereto.
[0031] This invention provides a single-molecule force spectrum classification method based on convolutional neural networks, such as... Figure 1 The diagram shows a flowchart of the single-molecule force spectrum classification method based on convolutional neural networks of the present invention. The specific method includes the following steps:
[0032] 1) Protein folding experiments were performed using m-trap optical tweezers from Lumicks, obtaining 860 sets of single-molecule force spectrum data. The Matplotlib plotting library in Python was used to convert the single-molecule force spectrum data into single-molecule force spectrum images;
[0033] 2) Preprocess the obtained single-molecule force spectrum images sequentially. Use the torchvision utility library in Python to convert the single-molecule force spectrum images into grayscale images with a pixel size of 256×256. Then, use data augmentation to increase the sample size, such as... Figure 2 As shown, the torchvision utility library in Python was used to crop the preprocessed single-molecule force spectrum images from nine directions: top left, top center, top right, left center, center, center right, bottom left, bottom center, and bottom right. The cropped image size was 224×224 pixels. The data-augmented images were then manually labeled into categories based on the number of folding events. These categories were categorized into five classes: category 1, category 2, category 3, category 4, and category 5, corresponding to folding events of 0, 1, 2, 3, and 4, respectively. A total of 7740 manually labeled single-molecule force spectrum images were generated, with 80% used for the training set, 10% for the validation set, and 10% for the test set.
[0034] 3) Construct a convolutional neural network model, such as Figure 3As shown, the convolutional neural network model consists of an input module, first to third convolutional modules, first to fourth pooling modules, a fully connected module, and an output module. The input module, first pooling module, first convolutional module, second pooling module, second convolutional module, third pooling module, third convolutional module, fourth pooling module, fully connected module, and output module are cascaded sequentially. The input module consists of two convolutional layers, cascaded sequentially with identical parameters. The kernel size in each convolutional layer is 3*3, the stride is 1, and the padding margin is 1. The first convolutional layer in the input module takes a single-molecule force spectrum image as input. It performs convolution operations with a portion of the 224×224 pixel single-molecule force spectrum image, extracting feature vectors from the corresponding regions of the single-molecule force spectrum image or feature map. After each convolutional layer, LReLU nonlinear activation and batch normalization are applied. The second convolutional layer further extracts features, and after convolution, LReLU activation and batch normalization are applied. The first pooling module consists of a single max-pooling layer with a 2x2 kernel and a stride of 2. The first convolutional module consists of two convolutional layers cascaded sequentially with identical parameters. The kernel size in each convolutional layer is 3x3 with a stride of 1 and a margin of 1. LReLU activation is applied after each convolution, followed by batch normalization. The second pooling module consists of a single max-pooling layer with a 2x2 kernel and a stride of 2. The second convolutional module consists of three convolutional layers cascaded sequentially with identical parameters. The kernel size in each convolutional layer is 3x3 with a stride of 1 and a margin of 1. LReLU activation is applied after each convolution, followed by batch normalization. The third pooling module consists of a single max-pooling layer with a 2x2 kernel and a stride of 2. The third convolutional module consists of three convolutional layers, cascaded sequentially with identical parameters. The kernel size is 3x3, the stride is 1, and the padding margin is 1. After each convolution, the LReLU activation function is applied, followed by batch normalization. The fourth pooling module consists of a single max-pooling layer with a 2x2 kernel and a stride of 2. The fully connected module consists of fully connected layers and dropout layers. After the fully connected layers, LReLU non-linear activation and batch normalization are applied. Then, the dropout layer randomly selects some nodes from the fully connected layers for output, and the softmax activation function is used to generate probability distributions for the five output categories. The output layer consists of five neurons for each of the five output categories.
[0035] The constructed convolutional neural network model is trained using the training set in the dataset. Gradient descent is employed to train the convolutional neural network parameters, leading to convergence of the cross-entropy loss function. The optimizer function based on gradient descent uses the Adam optimizer function with an initial learning rate of 0.001. After training, a trained convolutional neural network model is generated.
[0036] 4) The single-molecule force spectrum images from the test set were input into the trained convolutional neural network model to obtain the number and category information of folding events in 700 sets of single-molecule force spectrum images. Of these, 83 sets were accurately identified as category 1, 263 sets as category 2, 152 sets as category 3, 109 sets as category 4, 56 sets as category 5, and 37 sets were incorrectly identified. Figure 4 The image shows the number and category information of folding events in some of the obtained single-molecule force spectrum images.
[0037] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can make appropriate changes or modifications within the scope of the technology disclosed in the present invention, and such changes or modifications should be covered within the scope of protection of the present invention.
Claims
1. A single-molecule force spectrum classification method based on convolutional neural networks, characterized in that, The method includes the following steps: 1) Obtain single-molecule force spectrum images; 2) Perform preprocessing, data augmentation, and manual labeling on the single-molecule force spectrum images obtained in step 1) to obtain manually labeled single-molecule force spectrum images, and create a dataset containing training set, validation set, and test set. 3) Construct a convolutional neural network model. Use the training set created in step 2) to train the constructed convolutional neural network model to generate a trained convolutional neural network model. 4) Input the single-molecule force spectrum images from the test set into the convolutional neural network model trained in step 3) to obtain the number and category information of folding events in the single-molecule force spectrum; In step 2), the preprocessing involves converting all single-molecule force spectrum images into grayscale images with a pixel size of 256×256. In step 2), data augmentation involves cropping the preprocessed single-molecule force spectrum image from nine positions: top left, top center, top right, left center, center, center right, bottom left, bottom center, and bottom right. The cropped image has a pixel size of 224×224. In step 2), manual labeling refers to classifying all single-molecule force spectrum images into categories 1, 2, 3, 4, and 5 based on the number of folding events in the data-enhanced single-molecule force spectrum images. These categories correspond to single-molecule force spectrum images with 0, 1, 2, 3, and 4 folding events, respectively. In step 3), the convolutional neural network model is composed of an input module, a first pooling module, a first convolutional module, a second pooling module, a second convolutional module, a third pooling module, a third convolutional module, a fourth pooling module, a fully connected module, and an output module, which are cascaded in sequence. The input module and the first convolutional module consist of two convolutional layers, cascaded sequentially with identical parameters. The kernel size in each convolutional layer is 3x3, with a stride of 1 and an expansion margin of 1. After the convolutional layer, LReLU nonlinear activation and batch normalization are performed. The second and third convolutional modules consist of three convolutional layers, cascaded sequentially with identical parameters. The kernel size in each convolutional layer is 3x3, with a stride of 1 and an expansion margin of 1. After the convolutional layer, LReLU nonlinear activation and batch normalization are performed. The first, second, third, and fourth pooling modules consist of a single max-pooling layer with a 2x2 kernel and a stride of 2. The fully connected module consists of a fully connected layer and a dropout layer. The output module consists of an output layer containing 5 neurons.
2. The single-molecule force spectrum classification method based on convolutional neural networks according to claim 1, characterized in that, The dropout layer will randomly select some nodes from the fully connected layer for output and use the softmax activation function to generate probability distributions for the five output categories.
3. The single-molecule force spectrum classification method based on convolutional neural networks according to claim 1, characterized in that, In step 3), training refers to training the parameters of the convolutional neural network using the gradient descent method to make the cross-entropy loss function converge. The optimizer function based on the gradient descent method uses the Adam optimizer function with an initial learning rate of 0.001.