An environment DNA data dimension reduction method and system based on a hierarchical autoencoder

Abstract

The application provides an environmental DNA data dimension reduction method based on a hierarchical autoencoder, and belongs to the fields of bioinformatics and data processing technology. The method comprises the following steps: S1, data preprocessing: dividing environmental DNA data into environmental DNA abundance data and environmental DNA sequence data; S2, denoising the environmental DNA abundance data; S3, length balancing and translation of the environmental DNA sequence data; S4, representing sequence information of the sample; S5, compressing the environmental DNA sequence data; S6, dimension reduction by using a differential autoencoder VAE; the denoised environmental DNA abundance data obtained in S2 and the compressed and encoded environmental DNA sequence data obtained in S5 are used as inputs of the differential autoencoder VAE, and the dimension of the sample is further reduced to a two-dimensional latent space; and S7, drawing a visualization result graph of the two-dimensional space. The method can better preserve information of the environmental DNA data and obtain more accurate dimension reduction and visualization results.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and data processing technology, and in particular to a method and system for dimensionality reduction of environmental DNA data based on a hierarchical autoencoder. Background Technology

[0002] Human activities have led to the destruction of ecosystems, severe degradation of ecological functions, and impacts on human health. Convenient, accurate, and reliable ecological monitoring technologies are crucial for protecting the ecological environment. Environmental DNA technology is a highly efficient molecular-based method for monitoring biodiversity, with immense application potential in ecological restoration and natural resource protection. Environmental DNA is DNA extracted from environmental samples (water, soil, sediment, air, mixtures, etc.), and is a mixture of DNA from various organisms. By extracting DNA from the environment, biodiversity can be detected. However, due to the high dimensionality and complexity of environmental DNA data, dimensionality reduction and other operations must be performed to extract the useful information contained within the environmental DNA data.

[0003] Currently, most methods for dimensionality reduction of environmental DNA data are traditional, such as Principal Component Analysis (PCA), t-distributed random neighborhood embedding (t-SNE), Nonmetric Multidimensional Scaling (NMDS), and Uniform Manifold Approximation (UMAP). However, unlike other datasets, the number of taxa in the spectral matrix of environmental DNA data far exceeds the number of samples. Furthermore, due to limitations in sequencing technology, there is a high probability of non-zero values ​​being classified as zero, resulting in a large number of false zeros. Given these characteristics, traditional dimensionality reduction methods cannot effectively preserve the data structure, leading to significant information loss after dimensionality reduction of environmental DNA data.

[0004] Therefore, there is a need for a method for dimensionality reduction of environmental DNA data that can better preserve the information in the environmental DNA data and obtain more accurate dimensionality reduction data. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for dimensionality reduction of environmental DNA data based on hierarchical autoencoders, which reduces the high noise and sparsity of environmental DNA data, thereby obtaining more accurate dimensionality reduction data.

[0006] Therefore, the present invention provides the following technical solution:

[0007] A method for dimensionality reduction of environmental DNA data based on hierarchical autoencoders, the method comprising the following steps:

[0008] S1. Data preprocessing, including:

[0009] Environmental DNA data is divided into environmental DNA abundance data and environmental DNA sequence data;

[0010] The rows of environmental DNA abundance data with all zeros were filtered out, and Min-Max normalization and log-2 logarithmic transformation were performed to obtain the preprocessed environmental DNA abundance data and environmental DNA sequence data.

[0011] S2. Input the preprocessed environmental DNA abundance data into the depth counting autoencoder (DCA) for denoising to obtain the denoised environmental DNA abundance data.

[0012] S3. Length balancing and translation of environmental DNA sequence data, including:

[0013] The environmental DNA sequence data was padded with the nucleotide encoding N in the IUPAC nucleotide encoding to equalize the length;

[0014] Each typical base of the environmental DNA sequence after equalization of length: A, C, T, G, and the IUPAC ambiguous code are translated into four typical bases: A, T, C, G, and a four-dimensional probability distribution is applied. The nucleotide sequence is then represented as a two-dimensional array with the sequence length multiplied by 4.

[0015] S4. Represent the sequence information of the sample; combine the two-dimensional array representing the nucleotide sequence information obtained in S3 with the presence or absence of each operational classification unit in each sample to represent it as a three-dimensional array;

[0016] The presence of each operational classification unit: nucleotide sequence information translated from the sequence of the detected OTU;

[0017] The absence of each operation classification unit: if no OTU is detected, the matrix is ​​zero;

[0018] S5. Input the three-dimensional array representing the sample sequence information into the autoencoder AE to compress the environmental DNA sequence data;

[0019] S6. Dimensionality reduction is performed using a differential autoencoder (VAE). The denoised environmental DNA abundance data obtained in S2 and the compressed encoded environmental DNA sequence data obtained in S5 are used together as the input of the differential autoencoder (VAE) to further reduce the dimension of the sample to a two-dimensional latent space.

[0020] S7. Draw a visualization of the two-dimensional space.

[0021] Further, the step of inputting the preprocessed environmental DNA abundance data into a depth counting autoencoder (DCA) for denoising to obtain denoised environmental DNA abundance data includes:

[0022] The specific formula for the depth counting auto encoder (DCA) architecture is as follows:

[0023] E = ReLU(XW) E );

[0024] B = ReLU(EW) B );

[0025] D = ReLU(BW) E );

[0026] M = diag(s) j )exp r(DW μ );

[0027] Π=sigmoid(DW π );

[0028] θ = exp(DW θ );

[0029] Where E, B, and D represent the encoder layer, latent layer, and decoder layer, respectively, and M, θ, and Π represent the matrix forms of the mean, dispersion, and loss probability estimates, respectively. j The size factor is used; the activation function chosen for the mean and dispersion is exponential; the activation function for the dropout probability estimation is sigmoid, representing the dropout probability;

[0030] The distribution of environmental DNA data is described using the zero-expanded negative binomial distribution ZINB, where the zero-expanded negative binomial distribution ZINB is:

[0031]

[0032] ZINB(x;π,μ,θ)=πδ(x)+(1-π)NB(x;μ,θ);

[0033] Where π, μ, and θ are parameters of the ZINB distribution, representing the probability, mean, and dispersion of the lost event, respectively; x represents the raw read count, and the autoencoder estimates parameters v, μ, and θ based on the ZINB model.

[0034] Furthermore, the loss function of the depth counting autoencoder (DCA) is the negative logarithm of the ZINB likelihood, as shown in the following formula:

[0035] L ZINB =-log(ZINB(X|π,μ,θ)).

[0036] Furthermore, the nucleotides that fill the gaps to balance the sequence length encode N are translated as [0.25, 0.25, 0.25, 0.25].

[0037] Furthermore, the step of inputting a three-dimensional array representing sample sequence information into the autoencoder AE to compress the environmental DNA sequence data specifically involves:

[0038] The self-encoder AE includes a fully connected layer and a latent layer;

[0039] Each fully connected layer uses ReLU activation and employs a mirror architecture as the decoder;

[0040] The loss function of the autoencoder AE is binary cross-entropy:

[0041] L AE =-y log(p(y))-(1-y)log(1-p(y));

[0042] Where y is a binary label 0 or 1, and p(y) is the probability that the output belongs to the label y.

[0043] Furthermore, the differential autoencoder (VAE) includes:

[0044] The system consists of an input layer, an encoder network, a latent sampling layer, a decoder network, and an output layer.

[0045] Furthermore, the latent sampling layer uses reparameterization techniques for sampling and employs gradient descent to optimize the likelihood function.

[0046] Furthermore, the loss function of the differential autoencoder (VAE) is composed of the VAE loss function: KL divergence, reconstruction error, and binary cross-entropy of different combinations of potential nucleotide sequence encodings.

[0047]

[0048] N(0,1)])+α(-y log(p(y))-(1-y)log(1-p(y)));

[0049] Where x is the abundance input matrix of the differential encoder. It is the abundance output matrix of the differential autoencoder, N(u x , σ x ) represents the distribution that the potential layer z follows, y is a binary label 0 or 1, p(y) is the probability that the output belongs to the label y, and α is the weight of the loss function by the binary cross-entropy.

[0050] Furthermore, the depth counting autoencoder (DCA) and the autoencoder (AE) are trained separately, and then the VAE is trained.

[0051] A hierarchical autoencoder-based environmental DNA data dimensionality reduction system, the system comprising the following modules: data preprocessing module, denoising module, encoding module, dimensionality reduction module, and visualization module;

[0052] The data preprocessing module is used to divide environmental DNA data into environmental DNA abundance data and environmental DNA sequence data, filter out rows with all zeros in the abundance data, and perform Min-Max normalization and log-2 logarithmic transformation on the abundance data, and output the preprocessed environmental DNA abundance data and environmental DNA sequence data.

[0053] The denoising module is used to denoise the preprocessed environmental DNA abundance data and output the denoised environmental DNA abundance data.

[0054] The encoding module is used to balance the length of environmental DNA sequence data and convert the balanced sequence into the probability of four typical bases, representing the nucleotide sequence as a two-dimensional array with the sequence length multiplied by 4; the sequence information combined with OTU presence / deletion encoding is encoded into a three-dimensional array and input into the autoencoder AE to compress and encode the environmental DNA sequence data, and output the compressed sample sequence information data;

[0055] The dimensionality reduction module combines two inputs, the denoised environmental DNA abundance data and the compressed encoded environmental DNA sequence data, to perform dimensionality reduction processing on the environmental DNA data, reducing the sample dimension to two-dimensional space.

[0056] The visualization module visualizes the data after it has been reduced to two dimensions.

[0057] Advantages and positive effects of the present invention:

[0058] Because environmental DNA abundance data has a highly sparse distribution, this invention uses a depth counting autoencoder (DCA) to denoise the environmental DNA abundance data, which can better capture the high-dimensional structural information of the environmental DNA data, thereby effectively reducing the impact of high noise and sparsity of the environmental DNA data.

[0059] The depth counting autoencoder (DCA) in this invention replaces the traditional mean squared error (MSE) loss function with a loss function based on the zero-inflated negative binomial (ZINB) model. This can better characterize the distribution of environmental DNA data, reduce the impact of "dropout" events on data dimensionality reduction, and effectively characterize discrete, over-dispersed, and zero-inflated counting data.

[0060] This invention simultaneously employs the abundance and sequence of environmental DNA data, and utilizes a differential autoencoder for encoding dimensionality reduction, which can better preserve the information of environmental DNA data and obtain more accurate dimensionality reduction and visualization results.

[0061] In this invention, the depth counting autoencoder (DCA) and autoencoder (AE) are trained separately, and then the VAE is trained, which reduces the model's runtime. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart of the method of the present invention;

[0064] Figure 2 This is a system structure diagram of the present invention;

[0065] Figure 3 This is a diagram showing the dimensionality reduction result of this embodiment. Detailed Implementation

[0066] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0067] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0068] This invention provides a dimensionality reduction method for environmental DNA data based on hierarchical autoencoders. Considering the high sparsity of the distribution of environmental DNA abundance data, a depth counting autoencoder (DCA) is used to denoise the environmental DNA abundance data, which can better capture the high-dimensional structural information of the environmental DNA data. Furthermore, for DCA, this method uses a loss function based on a zero-inflated negative binomial (ZINB) model instead of the traditional mean squared error (MSE) loss function, which can better characterize the distribution of environmental DNA data, reduce the impact of "dropout" events on data dimensionality reduction, and effectively characterize discrete, overly dispersed, and zero-inflated count data. Simultaneously, the sequence information of the environmental DNA data is encoded by an autoencoder (AE), and the abundance and sequence of the environmental DNA data are combined as inputs. A differential autoencoder (VAE) is then used for encoding dimensionality reduction. Compared with other methods that only utilize abundance data, this method can better preserve the information of the environmental DNA data and obtain more accurate dimensionality reduction and visualization results.

[0069] Combination Figure 1 The process shown illustrates this method, with specific steps described below:

[0070] Step 1: Divide the acquired marine environmental DNA dataset into environmental DNA abundance data and environmental DNA sequence data.

[0071] In this embodiment, the environmental DNA sequence data includes representative sequences of 237 OTU operation taxonomic units, and the environmental DNA abundance data includes abundance data of 334 samples × 237 OTU operation taxonomic units.

[0072] Step 2: Filter out all zero rows in the environmental DNA abundance data and perform Min-Max normalization and log-2 log transformation to obtain a preprocessed environmental DNA abundance dataset containing 321 samples × 237 OTU operational taxonomic units.

[0073] Step 3: Input the preprocessed environmental DNA abundance data into a depth counting autoencoder (DCA) for denoising to obtain denoised environmental DNA abundance data. The size of the abundance data does not change after denoising.

[0074] Step 4: Fill each sequence in the environmental DNA sequence data with the nucleotide encoding N from the IUPAC nucleotide encoding to equalize the length. After length equalization, the length of each sequence is uniformly set to the length of the longest sequence, 151.

[0075] Step 5: Translate each typical base (A, C, T, G) and IUPAC ambiguous code of the balanced-length sequence into an appropriate four-dimensional probability distribution on the four typical bases (A, T, C, G), including uncertain base sequences W and S, etc. For example, base A is translated as [1, 0, 0, 0], W is translated as [0.5, 0, 0, 0.5], and the nucleotide encoding N added to balance the sequence length is translated as [0.25, 0.25, 0.25, 0.25]. Each translated sequence is represented as a 151×4 two-dimensional array.

[0076] Step 6: Combine the two-dimensional array representing sequence information with the presence / absence of each OTU operation classifier in each sample to represent it as a three-dimensional array. Each sample in this dataset contains 237 OTU operation classifiers, so the sequence information of each sample is represented as a three-dimensional array of size 237×151×4, and the entire dataset size is 321×237×151×4.

[0077] Step 7: Input the three-dimensional array representing the sample sequence information into the autoencoder (AE) and compress it to 100 dimensions. The size of the sequence encoding information of the entire dataset is 321 (number of samples) × 100 (sequence information dimension).

[0078] The autoencoder (AE) consists of 6 fully connected layers and 1 latent layer. Each fully connected layer uses ReLU activation, employs a mirror architecture as the decoder, and uses sigmoid in the output layer to make the output within [0,1].

[0079] Step 8: The denoised environmental DNA abundance data obtained in Step 3 and the compressed encoded environmental DNA sequence data obtained in Step 7 are used together as input to the differential autoencoder (VAE) to further reduce the dimension of the sample to a two-dimensional latent space.

[0080] The Differential Autoencoder (VAE) can be divided into an input layer, an encoder network, a latent sampling layer, a decoder network, and an output layer. The VAE has two input layers, taking abundance data and sequence-encoded data as inputs, respectively. A dropout layer with a dropout rate of 0.5 is added after the input layers. This layer sets some features to zero during the encoding phase to improve model learning performance. The encoder network has two fully connected neural network layers with dimensions of 150 and 50, respectively. Instead of using an activation function, the weights of the first fully connected neural network are regularized using the l1 norm, reducing the sparsity of the model. The second layer uses ReLU activation, which makes the model's output sparse and stable. The latent sampling layer has a dimension of 2. It compresses the data dimension to two dimensions through the encoder network and uses a reparameterization technique to sample the data. The sampled result z is used as the input to the subsequent decoder network. In this layer, the final dimensionality reduction result, reducing the data to two dimensions, is obtained. The decoder network uses the sampling results obtained through reparameterization techniques to recover the original expression matrix. It also has two fully connected neural network layers with hidden unit dimensions of 50 and 150, respectively, and uses ReLU activation. Finally, the output layer uses sigmoid to control the output within [0,1].

[0081] To reduce model runtime, the depth counting autoencoder (DCA) and autoencoder (AE) in this method are trained separately, and then the VAE is trained. The RMSprop method is chosen for training the VAE, with a learning rate set to 0.0001 to ensure convergence on all test datasets. Training is stopped if the training loss does not decrease significantly within 50 iterations.

[0082] Step 9: Draw a visualization of the two-dimensional result. The result of this dataset being reduced to two-dimensional space is shown below. Figure 3 As shown, the coordinate points are colored according to the logarithm of OTU abundance.

[0083] Corresponding to the environmental DNA data dimensionality reduction method based on hierarchical autoencoders of the present invention, the present invention also provides an environmental DNA data dimensionality reduction system based on hierarchical autoencoders for performing the aforementioned method.

[0084] Combination Figure 2 This system includes: a data preprocessing module, a noise reduction module, an encoding module, a dimensionality reduction module, and a visualization module; the specific process is as follows:

[0085] The data preprocessing module is used to divide environmental DNA data into environmental DNA abundance data and environmental DNA sequence data, filter out rows with all zeros in the abundance data, and perform normalization and logarithmic transformation on the abundance data, outputting preprocessed environmental DNA abundance data and environmental DNA sequence data.

[0086] In this embodiment, the environmental DNA sequence data includes representative sequences of 237 OTU operation taxonomic units, and the environmental DNA abundance data includes abundance data of 321 samples × 237 OTU operation taxonomic units.

[0087] The denoising module is used to denoise the preprocessed environmental DNA abundance data and output the denoised environmental DNA abundance data. The size of the abundance data does not change after denoising.

[0088] The encoding module is used to balance the length of environmental DNA sequence data;

[0089] In this embodiment, the length of each sequence is uniformly set to the longest sequence length of 151. The sequences after equalization are converted into the probabilities of four typical bases. Finally, the sequence information is combined with the OTU presence / deletion encoding into a three-dimensional array and input into the autoencoder (AE) to realize the compressed encoding of environmental DNA sequence data. The compressed sample sequence information data is output with a size of 321 (number of samples) × 100 (sequence information dimension).

[0090] The dimensionality reduction module combines two inputs, environmental DNA abundance data and environmental DNA sequence data, to perform dimensionality reduction processing on the environmental DNA data, reducing it to a two-dimensional space.

[0091] The visualization module visualizes the data after it has been reduced to two dimensions.

[0092] Compared with the prior art, the present invention has the following advantages:

[0093] The method and system provided by this invention take into account the high sparsity of environmental DNA abundance data distribution. By using a depth counting autoencoder (DCA) to denoise the environmental DNA abundance data, it can better capture the high-dimensional structural information of the environmental DNA data and effectively reduce the impact of high noise and sparsity. In addition to utilizing the abundance information of the environmental DNA data, the sequence information of the environmental DNA data is also encoded. Combining the abundance and sequence inputs of the environmental DNA data, a differential autoencoder is used for dimensionality reduction. Compared with other methods that only utilize abundance data, this method can better preserve the information of the environmental DNA data and obtain more accurate dimensionality reduction and visualization results.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hierarchical autoencoder-based environmental DNA data dimensionality reduction method, characterized in that, The method includes the following steps: S1. Data preprocessing, including: Environmental DNA data is divided into environmental DNA abundance data and environmental DNA sequence data; The rows of environmental DNA abundance data with all zeros were filtered out, and Min-Max normalization and log-2 logarithmic transformation were performed to obtain the preprocessed environmental DNA abundance data and environmental DNA sequence data. S2. Input the preprocessed environmental DNA abundance data into a depth counting autoencoder (DCA) for denoising to obtain denoised environmental DNA abundance data, including: The specific formula for the depth counting auto encoder (DCA) architecture is as follows: ； ； ； ； ； ； where E, B and D represent the encoder layer, the latent layer and the decoder layer, respectively, M, and Π represent the matrix form of the mean, the dispersion and the loss probability estimate, respectively, is the size factor; the activation function chosen for the mean and the dispersion is exponential; the activation function for the loss probability estimate is sigmoid, representing the dropout probability; The distribution of environmental DNA data is described using the zero-expanded negative binomial distribution ZINB, where the zero-expanded negative binomial distribution ZINB is: ； ； in These are the parameters of the ZINB distribution, representing the probability, mean, and dispersion of the missing event, respectively. Represents the raw read count, and the estimated parameters of the autoencoder based on the ZINB model. and ; S3. Length balancing and translation of environmental DNA sequence data, including: The environmental DNA sequence data was padded with the nucleotide encoding N in the IUPAC nucleotide encoding to equalize the length; Each typical base of the environmental DNA sequence after equalization of length: A, C, T, G, and the IUPAC ambiguous code are translated into four typical bases: A, T, C, G, and a four-dimensional probability distribution is applied. The nucleotide sequence is then represented as a two-dimensional array with the sequence length multiplied by 4. S4. Represent the sequence information of the sample; combine the two-dimensional array representing the nucleotide sequence information obtained in S3 with the presence or absence of each operational classification unit in each sample to represent it as a three-dimensional array; The presence of each operational classification unit: nucleotide sequence information translated from the sequence of the detected OTU; The absence of each operation classification unit: if no OTU is detected, the matrix is ​​zero; S5. Input the three-dimensional array representing the sample sequence information into the autoencoder AE to compress the environmental DNA sequence data; S6. Dimensionality reduction is performed using a differential autoencoder (VAE). The denoised environmental DNA abundance data obtained in S2 and the compressed encoded environmental DNA sequence data obtained in S5 are used together as the input of the differential autoencoder (VAE) to further reduce the dimension of the sample to a two-dimensional latent space. The loss function of the differential autoencoder (VAE) is composed of the VAE loss function: KL divergence, reconstruction error, and binary cross-entropy of different combinations of nucleotide sequence latent encodings. ； in, It is the abundance input matrix of the differential encoder. It is the abundance output matrix of the differential autoencoder. Let represent the distribution followed by the latent layer z, where y is a binary label 0 or 1, and p(y) is the probability that the output belongs to the label y. It is the weight of the loss function occupied by the binary cross-entropy; S7. Draw a visualization of the two-dimensional space.

2. The method for dimensionality reduction of environmental DNA data based on hierarchical autoencoders according to claim 1, characterized in that, The loss function of the depth counting autoencoder (DCA) is the negative logarithm of the ZINB likelihood, and the specific formula is as follows: 。 3. The environmental DNA data dimensionality reduction method based on hierarchical autoencoders according to claim 1, characterized in that, The nucleotides padding to balance the sequence length encoded N were translated as [0.25, 0.25, 0.25, 0.25].

4. The environmental DNA data dimensionality reduction method based on hierarchical autoencoders according to claim 1, characterized in that, The step of inputting a three-dimensional array representing sample sequence information into an autoencoder AE to compress environmental DNA sequence data specifically involves: The self-encoder AE includes a fully connected layer and a latent layer; Each fully connected layer uses ReLU activation and employs a mirrored architecture as the decoder; The loss function of the autoencoder AE is binary cross-entropy: ； Where y is a binary label 0 or 1, and p(y) is the probability that the output belongs to the label y.

5. The environmental DNA data dimensionality reduction method based on hierarchical autoencoders according to claim 1, characterized in that, The differential autoencoder (VAE) includes: The system consists of an input layer, an encoder network, a latent sampling layer, a decoder network, and an output layer.

6. The environmental DNA data dimensionality reduction method based on hierarchical autoencoders according to claim 5, characterized in that, The latent sampling layer uses reparameterization techniques for sampling and employs gradient descent to optimize the likelihood function.

7. The method for dimensionality reduction of environmental DNA data based on hierarchical autoencoders according to claim 1, characterized in that, The depth counting autoencoder (DCA) and the autoencoder (AE) are trained separately, and then the VAE is trained.

8. A dimensionality reduction system for environmental DNA data based on a hierarchical autoencoder, characterized in that, The system includes the following modules: data preprocessing module, noise reduction module, encoding module, dimensionality reduction module, and visualization module; The data preprocessing module is used to divide environmental DNA data into environmental DNA abundance data and environmental DNA sequence data, filter out rows with all zeros in the abundance data, and perform Min-Max normalization and log-2 logarithmic transformation on the abundance data, and output the preprocessed environmental DNA abundance data and environmental DNA sequence data. The denoising module is used to denoise the preprocessed environmental DNA abundance data and output denoised environmental DNA abundance data, including: The specific formula for the DCA architecture of a depth counting autoencoder is as follows: ； ； ； ； ； ； Where E, B, and D represent the encoder layer, latent layer, and decoder layer, respectively, and M, Π and Π represent the matrix forms of the mean, dispersion, and loss probability estimates, respectively. The size factor is used; the activation function chosen for the mean and dispersion is exponential; the activation function for the dropout probability estimation is sigmoid, representing the dropout probability; The distribution of environmental DNA data is described using the zero-expanded negative binomial distribution ZINB, where the zero-expanded negative binomial distribution ZINB is: ； ； in These are the parameters of the ZINB distribution, representing the probability, mean, and dispersion of the missing event, respectively. Represents the raw read count, and the estimated parameters of the autoencoder based on the ZINB model. and ; The encoding module is used to balance the length of environmental DNA sequence data and convert the balanced sequence into the probability of four typical bases, representing the nucleotide sequence as a two-dimensional array with the sequence length multiplied by 4; the sequence information combined with OTU presence / deletion encoding is encoded into a three-dimensional array and input into the autoencoder AE to compress and encode the environmental DNA sequence data, and output the compressed sample sequence information data; The dimensionality reduction module uses a differential autoencoder (VAE) for dimensionality reduction. It combines two inputs, the denoised environmental DNA abundance data and the compressed encoded environmental DNA sequence data, to perform dimensionality reduction on the environmental DNA data, reducing the sample dimension to two-dimensional space. The loss function of the differential autoencoder (VAE) is composed of the VAE loss function: KL divergence, reconstruction error, and binary cross-entropy of different combinations of nucleotide sequence latent encodings. ； in, It is the abundance input matrix of the differential encoder. It is the abundance output matrix of the differential autoencoder. Let represent the distribution followed by the latent layer z, where y is a binary label 0 or 1, and p(y) is the probability that the output belongs to the label y. It is the weight of the loss function occupied by the binary cross-entropy; The visualization module visualizes the data after it has been reduced to two dimensions.

Images