Low signal-to-noise ratio single molecule force spectroscopy signal recovery enhancement method

Abstract

The application discloses a low signal-to-noise ratio single molecule force spectroscopy signal recovery enhancement method. The method comprises the following steps: 1) constructing a data simulation model by fitting a real single molecule force spectroscopy signal curve to generate a simulation data set; 2) inputting the simulation data set generated in step 1) into a residual neural network model as a training set and a verification set to learn and train, so as to construct a mapping relationship among high signal-to-noise ratio mechanical signals, low signal-to-noise ratio distance signals and noise-free distance signals, and finally generate a residual neural network model with a signal recovery enhancement function; and 3) inputting single molecule force spectroscopy signal data collected in an experiment into the trained residual neural network model as a test set to test the mechanical and distance signals, and obtain high signal-to-noise ratio distance signals after recovery and enhancement. The method realizes synchronous improvement of the smoothness of the distance signal curve and the accuracy of the signal at the folding position, and greatly improves the quality and accuracy of the single molecule force spectroscopy signal.

Description

Technical Field

[0001] This invention relates to the field of single-molecule detection, and more specifically to a method for recovering and enhancing single-molecule force spectrum signals with low signal-to-noise ratio. Background Technology

[0002] Single-molecule force spectroscopy optical tweezers utilizes optical gradient forces to optically capture and manipulate small dielectric particles. By capturing micro- and nano-sized particles with a highly focused laser beam, optical tweezers can apply piconewton-level forces to the captured particles. This is a key technology for measuring and analyzing the mechanical properties of single molecules, playing a particularly important role in the study of protein folding.

[0003] In single-molecule force spectroscopy, when a molecule is stretched by an external force to a certain threshold, its folded structure rapidly unfolds. This change manifests as a step change in the distance signal of a single-molecule force spectrum. Especially for DNA molecules, the amplitude of this step change during folding and unfolding is typically between a few piconewtons and tens of piconewtons.

[0004] In single-molecule force spectroscopy experiments, the random noise level (10-20 nm) generated by the Brownian motion and environmental influences of the DNA-linked beads is comparable to the magnitude of the step signal when DNA folds and unfolds, resulting in an extremely low signal-to-noise ratio of the acquired signal, making it difficult to accurately determine the start and end positions of the folding and unfolding.

[0005] Common methods such as mean filtering, median filtering, Kalman filtering, Fourier transform, and wavelet transform cannot effectively reduce noise and preserve signal accuracy. Furthermore, breakpoint detection techniques in the R language used for data processing cannot accurately identify folded and unfolded positions. Summary of the Invention

[0006] This invention addresses the problem of low signal-to-noise ratio in existing single-molecule force spectroscopy signals by proposing a method for recovering and enhancing such signals. By using a residual neural network and combining two signal channels—mechanical and distance—the single-molecule force spectroscopy signal is recovered and enhanced, achieving simultaneous improvement in the smoothness of the distance signal curve and the accuracy of the signal at folds, thus greatly improving the quality and accuracy of the single-molecule force spectroscopy signal.

[0007] This invention is achieved through the following technical solution:

[0008] 1) Based on the high signal-to-noise ratio mechanical signal and low signal-to-noise ratio distance signal curves of the single-molecule force spectrum collected in the experiment, a data simulation model is constructed, and the triplet simulation dataset is generated using the model.

[0009] The higher signal-to-noise ratio (SNR) mechanical signal is a mechanical signal with an average SNR of 20.45 dB or higher, with sampling time (1 / 100 second) as the x-axis and tensile force (in cows) as the y-axis; the lower signal-to-noise ratio (SNR) distance signal is a distance signal with an average SNR of 8.82 dB or lower, with sampling time (1 / 100 second) as the x-axis and displacement (in micrometers) as the y-axis.

[0010] The triplet simulation data is a data set composed of high signal-to-noise ratio mechanical simulation signals, low signal-to-noise ratio distance simulation signals, and noise-free distance simulation signals as elements.

[0011] The data simulation model is constructed as follows: First, smooth curves for a high signal-to-noise ratio (SNR) mechanical signal and a low SNR distance signal are fitted using polynomials. Then, the force value of the smooth curve of the mechanical signal and the distance value of the smooth curve of the distance signal are multiplied by a random number. Folded simulation signals are then applied to the smooth curves of the mechanical signal and the distance signal, respectively. The resulting distance signal curve is then used as a noise-free distance simulation signal. Finally, corresponding normally distributed simulation noise is added to the smooth curves of the mechanical signal and the distance signal after the folded simulation signals are applied.

[0012] The polynomial is a 7th-degree polynomial; the random number ranges from 0.5 to 1.5;

[0013] The folding simulation signal is set as follows: First, 1 to 4 points are randomly selected on the curve as the positions of the folding signal. The signal amplitude of these points is determined by the signal magnitude generated when a single molecule unfolds. For the mechanical signal, the force value after these points decreases by a random value of 5 to 50 piconewtons. For the distance signal, the distance value after these points decreases by a random value of 4 to 30 nanometers.

[0014] The normally distributed simulated noise has a mean of 0 and a standard deviation calculated based on the average of the standard deviations of the mechanical and distance signal noises of single-molecule force spectra obtained from 100 sets of experiments. The noise amplitude distribution is normally distributed.

[0015] 2) Input the triplet simulation dataset generated in step 1) into the residual neural network for learning and training, thereby constructing the mapping relationship between high signal-to-noise ratio mechanical signals, low signal-to-noise ratio distance signals and noise-free distance signals, and finally generating a residual neural network model with signal recovery and enhancement functions;

[0016] The residual neural network consists of two one-dimensional convolutional layers and five residual blocks to extract feature information from high signal-to-noise ratio (SNR) mechanical signals and low SNR distance signals, which is used to restore and enhance the low SNR distance signals. The signal restoration and enhancement function transforms the low SNR distance signals into high SNR distance signals while preserving signal accuracy.

[0017] The residual neural network consists of two one-dimensional convolutional layers: a data input convolutional layer with 2-channel input and 64-channel output, and a kernel size of 3; and a data output convolutional layer with 64-channel input and 1-channel output, also with a kernel size of 3. The residual neural network also contains five residual blocks, each consisting of two one-dimensional convolutional layers with 64-channel input and output. The kernel sizes for each residual block are 3, 5, 7, 9, and 15, respectively, and skip connections are established between the residual blocks. The layers are activated using the ReLU activation function.

[0018] The data input convolutional layer has 2 channels, which receive high SNR mechanical and low SNR distance signals from the single-molecule force spectrum. This layer increases the 2-channel input data to 64 channels using a convolutional kernel of size 3, corresponding to the 64-channel input of the subsequent residual block. The data output convolutional layer has 1 channel, which outputs the recovered and enhanced high SNR distance signal from the single-molecule force spectrum. This layer reduces the 64-channel data output from the residual block to 1 channel using a convolutional kernel of size 3.

[0019] The skip connection structure can directly pass feature information to subsequent layers, ensuring that gradients can propagate effectively even when the network depth is very large, and retaining more feature information.

[0020] As a preferred approach, 30,000 sets of simulation data are generated, of which 27,000 sets are used as the training set for neural network training, and the remaining 3,000 sets are used as the validation set for verifying the neural network's performance.

[0021] 3) Single-molecule force spectrum data were collected through experiments, and the high signal-to-noise ratio mechanical signal and low signal-to-noise ratio distance signal data were input into the trained residual neural network model for testing to obtain the restored and enhanced high signal-to-noise ratio distance signal;

[0022] The high signal-to-noise ratio distance signal refers to a distance signal with an average signal-to-noise ratio of 45.30 dB or higher.

[0023] Compared with the prior art, the present invention has the following beneficial technical effects:

[0024] Compared with commonly used filtering methods, this invention adopts a low signal-to-noise ratio single-molecule force spectrum signal recovery and enhancement method, which simultaneously improves the smoothness of the distance signal curve and the signal accuracy at the fold, greatly improving the quality and accuracy of the single-molecule force spectrum signal. Attached Figure Description

[0025] Figure 1 This is a structural diagram of the residual neural network of the present invention;

[0026] Figure 2 This is a structural diagram of the residual block in the residual neural network of the present invention;

[0027] Figure 3 It refers to the high signal-to-noise ratio mechanical signal and the low signal-to-noise ratio distance signal described in this invention, as well as the fitted smooth curve;

[0028] Figure 4 This is a comparison diagram of the single-fold distance signal restored and enhanced by the present invention and the actual distance signal (acquired by optical tweezers);

[0029] Figure 5 This is a comparison diagram of the restored and enhanced multi-folded distance signal and the simulated distance signal of this invention; Detailed Implementation

Example 1

[0030] This invention provides a method for recovering and enhancing low signal-to-noise ratio (SNR) single-molecule force spectrum signals. First, a smooth curve is obtained by fitting a 7th-order polynomial to both a high SNR mechanical signal (sampling time on the x-axis, tensile force on the y-axis, average SNR 20.45 dB) and a low SNR distance signal (sampling time on the x-axis, displacement on the y-axis, average SNR 8.82 dB). Figure 3As shown, the force value of the smoothed curve of the force signal and the distance value of the smoothed curve of the distance signal are multiplied by a random number ranging from 0.5 to 1.5, respectively. Folded simulation signals are then applied to both the smoothed curves of the force signal and the distance signal, resulting in a noise-free distance simulation signal. Next, normally distributed simulation noise is added to the smoothed curves of the force signal and the distance signal after the folded simulation signals are applied, ultimately forming a high signal-to-noise ratio (SNR) force simulation signal and a low SNR distance simulation signal. This constructs a data simulation model that can generate triplet simulation data of high SNR force signals, low SNR distance signals, and noise-free distance signals. Then, the data simulation model generates 30,000 sets of simulation data as a dataset, of which 27,000 sets are used as a training set for neural network training, and the remaining 3,000 sets are used as a validation set for verifying the neural network's performance. After generating the dataset, simulation data of high signal-to-noise ratio mechanical signals and low signal-to-noise ratio distance signals are used as inputs, and simulation data of noise-free distance signals are used as target data. These are then input into the residual neural network for learning and training, thereby constructing the mapping relationship between high signal-to-noise ratio mechanical signals, low signal-to-noise ratio distance signals and noise-free distance signals, and finally generating a residual neural network model with signal recovery and enhancement functions.

[0031] The folding simulation signal setup method is as follows: First, randomly select one point on the curve as the position of the folding signal. The signal amplitude at this point is determined by the signal magnitude generated when a single molecule unfolds. For the mechanical signal, the force value after these points decreases by a random value of 5 to 50 pN. For the distance signal, the distance value after these points decreases by a random value of 4 to 30 nanometers. The simulation noise is normally distributed with a mean of 0, and the standard deviation is calculated based on the average of the standard deviations of the mechanical and distance signal noises from 100 sets of experiments. The noise amplitude distribution is normally distributed.

[0032] The residual neural network structure of this invention is as follows: Figure 1 As shown, it consists of two one-dimensional convolutional layers and five residual blocks. The two one-dimensional convolutional layers are: a data input convolutional layer with 2-channel input and 64-channel output, and a kernel size of 3; and a data output convolutional layer with 64-channel input and 1-channel output, and a kernel size of 3. The five residual blocks in the residual neural network each consist of two one-dimensional convolutional layers with 64-channel input and output, as shown below. Figure 2As shown, the kernel sizes for each residual block are 3, 5, 7, 9, and 15, with skip connections between residual blocks. ReLU activation functions are used to activate the layers. The data input to the convolutional layer consists of 2 channels: a high SNR mechanical signal and a low SNR distance signal from the single-molecule force spectrum. This layer increases the input 2 channels to 64 channels using a kernel of size 3, corresponding to the 64-channel input of subsequent residual blocks. The data output to the convolutional layer consists of 1 channel: the recovered and enhanced high SNR distance signal from the single-molecule force spectrum. This layer reduces the 64-channel output of the residual block to 1 channel using a kernel of size 3. This skip connection structure allows feature information to be directly passed to subsequent layers, ensuring effective gradient propagation even at high network depths and preserving more feature information.

[0033] The network was tested by collecting single-molecule force spectrum data experimentally. The high signal-to-noise ratio (SNR) mechanical signal and low SNR distance signal data were then input into a trained residual neural network model for testing, yielding a restored and enhanced high SNR distance signal. The single-molecule force spectrum distance signal from DNAHairpin (LUMICKS, Netherlands) was used as an example. Taking dual-optical tweezers acquisition as an example, after restoration and enhancement via a residual neural network, the signal is compared with the original distance signal. This invention's method effectively filters out interference signals and noise. Figure 4 As shown, a comparison between the restored and enhanced single-fold distance signal and the true distance signal (acquired by optical tweezers) clearly demonstrates that this invention effectively filters out overall noise and different interference signals at the folds. It achieves a simultaneous improvement in the smoothness of the distance signal curve and the accuracy of the signal at the folds, significantly enhancing the quality and accuracy of the single-molecule force spectrum signal.

Example 2

[0034] This invention provides a method for recovering and enhancing low signal-to-noise ratio (SNR) single-molecule force spectrum signals. First, a smooth curve is obtained by fitting a 7th-order polynomial to both a high SNR mechanical signal (sampling time on the x-axis, tensile force on the y-axis, average SNR 20.45 dB) and a low SNR distance signal (sampling time on the x-axis, displacement on the y-axis, average SNR 8.82 dB). Figure 3As shown, the force value of the smoothed curve of the force signal and the distance value of the smoothed curve of the distance signal are multiplied by a random number ranging from 0.5 to 1.5, respectively. Folded simulation signals are then applied to both the smoothed curves of the force signal and the distance signal, resulting in a noise-free distance simulation signal. Next, normally distributed simulation noise is added to the smoothed curves of the force signal and the distance signal after the folded simulation signals are applied, ultimately forming a high signal-to-noise ratio (SNR) force simulation signal and a low SNR distance simulation signal. This constructs a data simulation model that can generate triplet simulation data of high SNR force signals, low SNR distance signals, and noise-free distance signals. Then, the data simulation model generates 30,000 sets of simulation data as a dataset, of which 27,000 sets are used as a training set for neural network training, and the remaining 3,000 sets are used as a validation set for verifying the neural network's performance. After generating the dataset, simulation data of high signal-to-noise ratio mechanical signals and low signal-to-noise ratio distance signals are used as inputs, and simulation data of noise-free distance signals are used as target data. These are then input into the residual neural network for learning and training, thereby constructing the mapping relationship between high signal-to-noise ratio mechanical signals, low signal-to-noise ratio distance signals and noise-free distance signals, and finally generating a residual neural network model with signal recovery and enhancement functions.

[0035] The method for setting up the folding simulation signal is as follows: First, randomly select 1 to 4 points on the curve as the positions of the folding signal. The signal amplitude at these points is determined by the signal magnitude generated when a single molecule unfolds. For the mechanical signal, the force value after these points decreases by a random value of 5 to 50 pN. For the distance signal, the distance value after these points decreases by a random value of 4 to 30 nanometers. The simulation noise is normally distributed with a mean of 0, and the standard deviation is calculated based on the average of the standard deviations of the mechanical and distance signal noises from 100 sets of experiments. The noise amplitude distribution is normally distributed.

[0036] The residual neural network structure of this invention is as follows: Figure 1 As shown, it consists of two one-dimensional convolutional layers and five residual blocks. The two one-dimensional convolutional layers are: a data input convolutional layer with 2-channel input and 64-channel output, and a kernel size of 3; and a data output convolutional layer with 64-channel input and 1-channel output, and a kernel size of 3. The five residual blocks in the residual neural network each consist of two one-dimensional convolutional layers with 64-channel input and output, as shown below. Figure 2As shown, the kernel sizes for each residual block are 3, 5, 7, 9, and 15, with skip connections between residual blocks. ReLU activation functions are used to activate the layers. The data input to the convolutional layer consists of 2 channels: a high SNR mechanical signal and a low SNR distance signal from the single-molecule force spectrum. This layer increases the input 2 channels to 64 channels using a kernel of size 3, corresponding to the 64-channel input of subsequent residual blocks. The data output to the convolutional layer consists of 1 channel: the recovered and enhanced high SNR distance signal from the single-molecule force spectrum. This layer reduces the 64-channel output of the residual block to 1 channel using a kernel of size 3. This skip connection structure allows feature information to be directly passed to subsequent layers, ensuring effective gradient propagation even at high network depths and preserving more feature information.

[0037] To test the network, 1000 sets of multi-fold simulation data were generated as a test set. Simulation data of high SNR mechanical signals and low SNR distance signals were input into the trained residual neural network model for testing, resulting in the restored and enhanced distance signal. The comparison results between the multi-fold distance signal restored and enhanced by the residual neural network and the simulated distance signal are as follows: Figure 5 As shown, this invention can effectively filter out overall noise and different interference signals at the folds. It achieves a simultaneous improvement in the smoothness of the distance signal curve and the accuracy of the signal at the folds, greatly enhancing the quality and accuracy of the single-molecule force spectrum signal.

[0038] 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 method for recovering and enhancing single-molecule force spectral signals with low signal-to-noise ratio, characterized in that, The method includes the following steps: 1) Based on the high signal-to-noise ratio mechanical signal and low signal-to-noise ratio distance signal curves of the single-molecule force spectrum collected in the experiment, a data simulation model is constructed, and the triplet simulation dataset is generated using the model. 2) Input the triplet simulation dataset generated in step 1) into the residual neural network for learning and training, thereby constructing the mapping relationship between high signal-to-noise ratio mechanical signals, low signal-to-noise ratio distance signals and noise-free distance signals, and finally generating a residual neural network model with signal recovery and enhancement functions. 3) Collect single-molecule force spectrum data through experiments, and input the high signal-to-noise ratio mechanical signal and low signal-to-noise ratio distance signal data into the residual neural network model trained in step 2) for testing, to obtain the restored and enhanced high signal-to-noise ratio distance signal; In step 1), the higher signal-to-noise ratio (SNR) mechanical signal refers to a mechanical signal with an average SNR of 20.45 dB or higher; the lower SNR distance signal refers to a distance signal with an average SNR of 8.82 dB or lower. The data simulation model described in step 1) is constructed as follows: First, a smooth curve for a high signal-to-noise ratio (SNR) mechanical signal and a smooth curve for a low SNR distance signal are fitted using a polynomial. Then, the force value of the smooth curve of the mechanical signal and the distance value of the smooth curve of the distance signal are multiplied by a random number. Folded simulation signals are then set on the smooth curves of the mechanical signal and the distance signal, respectively. The resulting distance signal curve is used as a noise-free distance simulation signal. Next, corresponding normally distributed simulation noise is added to the smooth curves of the mechanical signal and the distance signal after setting the folded simulation signal, respectively, to finally form a high SNR mechanical simulation signal and a low SNR distance simulation signal. The triplet simulation dataset mentioned in step 1) is a data set composed of high signal-to-noise ratio mechanical simulation signals, low signal-to-noise ratio distance simulation signals, and noise-free distance simulation signals as tuples; The high signal-to-noise ratio distance signal mentioned in step 3) refers to a distance signal with an average signal-to-noise ratio of 45.30 dB or higher.

2. The method for recovering and enhancing low signal-to-noise ratio single-molecule force spectrum signals according to claim 1, characterized in that, The residual neural network described in step 2) consists of two one-dimensional convolutional layers and five residual blocks to extract feature information from the high signal-to-noise ratio (SNR) mechanical signal and the low SNR distance signal, which is used to restore and enhance the low SNR distance signal.

3. The method for recovering and enhancing low signal-to-noise ratio single-molecule force spectroscopy signals according to claim 2, characterized in that, The polynomial is a 7th-order polynomial; the random number ranges from 0.5 to 1.5; the folding simulation signal is set as follows: first, 1 to 4 points are randomly selected on the curve as the positions of the folding signal, and the signal amplitude of these points is determined by the signal magnitude generated when a single molecule unfolds; for the mechanical signal, the force value after these points decreases by a random value of 5 to 50 piconewtons; for the distance signal, the distance value after these points decreases by a random value of 4 to 30 nanometers; the normal distribution simulation noise has a mean of 0, and the standard deviation is calculated based on the average of the standard deviations of the mechanical and distance signal noises of the single-molecule force spectrum obtained from 100 sets of experiments, respectively, and the noise amplitude distribution is a normal distribution.

4. The method for recovering and enhancing low signal-to-noise ratio single-molecule force spectroscopy signals according to claim 3, characterized in that, The residual neural network consists of two one-dimensional convolutional layers: a data input convolutional layer with 2-channel input and 64-channel output, and a kernel size of 3; and a data output convolutional layer with 64-channel input and 1-channel output, and a kernel size of 3. The residual neural network also contains five residual blocks, each consisting of two one-dimensional convolutional layers with 64-channel input and output. The kernel sizes for each residual block are 3, 5, 7, 9, and 15, respectively, and skip connections are established between the residual blocks.

5. The method for recovering and enhancing low signal-to-noise ratio single-molecule force spectroscopy signals according to claim 4, characterized in that, The data input is the input of channel 2 of the convolutional layer, which is a high signal-to-noise ratio (SNR) mechanical signal and a low SNR distance signal of the single-molecule force spectrum; the data output is the output of channel 1 of the convolutional layer, which is the high SNR distance signal of the recovered and enhanced single-molecule force spectrum.

Images