A double-adversarial autoencoder cancer subtype identification method based on multi-omics data

Abstract

The present application relates to the field of cancer subtype identification in bioinformatics, and particularly relates to a double-adversarial autoencoder cancer subtype identification method based on multi-omics data. The invention mainly includes: (1) a new model of double-adversarial MultiDAAE network is constructed; (2) representative omics data is selected, and the data is divided into a training set and a test set, and data preprocessing is performed; (3) the effect is evaluated by using P value and cluster number. Finally, the performance of MultiDAAE on most data sets exceeds other commonly used cancer subtype identification algorithms, thereby verifying the effectiveness of the method.

Description

Technical Field

[0001] This invention relates to the field of cancer subtype identification in bioinformatics, specifically to a dual adversarial autoencoder method for cancer subtype identification based on multi-omics data. Background Technology

[0002] Cancer is one of the leading causes of death worldwide, and due to its high mortality rate, high incidence rate, and long latency period, it has always been a hot topic and key research area for biomedical scientists. Cancer is activated through multiple pathways and different perturbations; depending on the activation pathway, patient survival rates vary greatly, and the efficacy of different drugs also differs. Cancer subtype identification is often a complex problem, and in most cases, it requires the fusion of omics data at different levels to achieve accurate subtype identification. Accurate subtype identification helps to understand the pathogenesis of cancer, promotes clinical treatment, improves patient survival rates, and is also one of the prerequisites for achieving the goals of precision medicine.

[0003] However, fusing multi-omics data is no easy task. The high dimensionality of the data makes it difficult for some traditional clustering methods to achieve satisfactory results. In recent years, the rapid development of deep learning has led to the use of this emerging method to solve the problem of high-dimensional data fusion. A series of models, such as neural networks, autoencoders, variational autoencoders, noisy autoencoders, and adversarial autoencoders, have been used to attempt to solve the problem of multi-omics cancer subtype identification. However, these methods all have some shortcomings. For example, using autoencoders to minimize reconstruction costs, while simple, is prone to overfitting and can also cause different training points to be encoded into non-overlapping regions, resulting in the encoder not providing a useful representation.

[0004] Therefore, Gregory et al. attempted to leverage the excellent results achieved by Variational Autoencoders (VAEs) in image generation for multi-omics data fusion; however, the final results were not particularly ideal. Building upon this, Yang introduced Generative Adversarial Networks (GANs) to form adversarial autoencoders for cancer subtype identification. While the model improved performance compared to VAEs by introducing a discriminator, it did not address a key issue: if the posterior and prior distributions differ significantly, the discriminator's inclusion could negatively impact performance, a problem already demonstrated by several researchers. Therefore, we propose a novel dual adversarial autoencoder (MultiDAAE) method for cancer subtype identification based on multi-omics data. This method attempts to address the problems encountered by traditional adversarial autoencoders in multi-omics data tasks by utilizing two discriminators. Through systematic methodological comparison, this approach achieves the best results on multiple cancer datasets compared to the most popular and state-of-the-art cancer subtype identification methods. Summary of the Invention

[0005] This invention proposes a novel method for cancer subtype identification based on a dual adversarial autoencoder using multi-omics data. The model construction steps are as follows:

[0006] We assume that the multi-omics data are First of all The input is fed into the encoder, and the encoder will... Encoding as a latent representation , and then The input is fed into the decoder, and the decoder will... Decoding to reconstruct data :

[0007] in This refers to the encoder's density function, while Let x represent the density function of the decoder. Then, the mean squared error is used as the reconstruction term for the encoder-decoder pair, following the convention of autoencoder models and their variants, to minimize the Euclidean distance between x and x'.

[0008]

[0009] Next, we assume the latent representation Follow the prior distribution ,in From The sample drawn from the middle, and From The samples are extracted from the output of the latent representation. A discriminator D1 is introduced to form a generative adversarial network with the encoder to encourage... Matching occurs, and at this point, the encoder becomes the generator of the generative adversarial network. Let... Let D1 be the discriminator function. The encoder ensures that the aggregated posterior distribution can fool the discriminator D1, which assumes... The samples come from the prior distribution The discriminator D1 was trained to distinguish... and We use adversarial training based on a min-max game to update the parameters of the discriminator D1 and the encoder, as shown in the following formula:

[0010]

[0011] The autoencoder and discriminator D1 are jointly trained in two phases: reconstruction and regularization. In the reconstruction phase, the encoder and decoder are trained to update their parameters via backpropagation to reduce reconstruction error. In the regularization phase, the generator (i.e., encoder) parameters are first fixed, and then its discriminator is updated to distinguish between real samples. ) and generated samples ( Then, the adversarial network updates its generator to confuse the discriminator. The goal of this training process is to minimize the adversarial loss.

[0012]

[0013] in It is the loss of discriminator D1. It is the loss of the generator (encoder). The loss is composed of a discriminator D1 and an encoder in a GAN. The model structure is similar to that of an adversarial autoencoder, designed to address the prior distribution... and encoder distribution To address the potential performance drawback of introducing a discriminator when significant differences exist between datasets, we added a second discriminator, D2. This additional discriminator was trained to distinguish between real-world multi-omics data. and the reconstructed multi-omics data generated by the decoder .set up This is the discriminator function D2. This training process also aims to minimize adversarial loss.

[0014]

[0015] in It is the loss of discriminator D2. It is the loss of the generator (decoder). The loss is denoted by the discriminator D2 and the decoder in the GAN. By introducing D2, our architecture ensures that the distribution of the reconstructed samples is similar to that of the base real samples. Furthermore, the presence of the first discriminator D1 promotes the similarity between the latent space distribution and the prior distribution. We expect this modified AAE architecture to converge to an equilibrium state after training. The overall training of the model aims to minimize the total loss:

[0016]

[0017] in yes The weight, yes The weights are determined by the parameters. In practice, we use deep neural networks to parameterize the encoder, decoder, and discriminator, and utilize backpropagation and stochastic gradient descent to update the model parameters.

[0018] Table 1 lists the network structures of Q, G, D1, and D2. The last layer of the discriminator uses the Sigmoid function as the activation function, while the remaining fully connected layers use the ReLU function. To avoid overfitting and accelerate the neural network training process, a batch normalization layer is added between every two fully connected layers to normalize each batch, thus accelerating training and improving the model's generalization ability. The parameters of these four networks are repeatedly updated during model training until convergence is achieved. After model fitting, latent representation features are extracted, and then the K-means algorithm is used to cluster the samples.

[0019] Table 1 MultiDAAE Network Structure

[0020] Encoder Q 3 256 / 100 / 100 Decoder G 3 100 / 100 / 256 Discriminator D1 2 100 / 32 Discriminator D2 3 100 / 32 / 32

[0021] To optimize our model's performance, we conducted a comprehensive hyperparameter search, considering various combinations of batch size (32, 64, 128, 256), number of training epochs (10, 25, 50, 100, 200), learning rate (0.01, 0.001, 0.0001, 0.00001), and λ1 and λ2 (1, 0.1, 0.01, 0.001, 0.0001). We computed training and testing losses for each epoch, with the test set consisting of a randomly selected 20% subset of the entire dataset to ensure data-biased evaluation. Specifically, we used the Adam method to tune the model parameters, and ultimately, a batch size of 64, a learning rate of 0.0001, 50 epochs, λ1 of 0.01, and λ2 of 0.01 were considered optimal hyperparameters. It is worth noting that, considering the utilization of multiple cancer datasets, our parameter selection aims to achieve excellent performance on as many datasets as possible, rather than optimizing for a single dataset. Any parameter that produces excellent results on some cancer datasets but performs poorly on others will not be considered in our final hyperparameter selection. This strategy aims to avoid overfitting to specific datasets; we hope that the proposed method can generalize as much as possible rather than being limited to a single cancer dataset. Attached Figure Description

[0022] Figure 1 is a schematic diagram of the multi-omics dual adversarial autoencoder model method. Detailed Implementation

[0023] This invention relates to the field of cancer subtype identification in bioinformatics, specifically to a method for cancer subtype identification based on a dual adversarial autoencoder using multi-omics data. Specific embodiments of the invention are described below. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the invention and are not intended to limit the scope of protection of the invention.

[0024] Step 1: Omics Data Collection Phase. Three types of omics data were selected from the abundant data generated by TCGA: DNA methylation at the epigenomic level, and mRNA and miRNA expression at the transcriptomic level. These molecular features were chosen because of their potential in describing cancer subtypes and reporting treatment strategies. Epigenetic regulation of gene expression by DNA methylation has a profound impact on tumorigenesis and development. Simultaneously, mRNA expression profiles and miRNA expression can reflect the underlying biological processes and molecular pathways that define cancer subtypes. Level 3 data from HTSeq-FPKM and BCGSC miRNA Profiling miRNA-Seq data were downloaded from TCGA for mRNA and miRNA data, respectively. For DNA methylation data, we selected Level 3 data from TCGA's Illumina Human Methylation 450 dataset.

[0025] Step 2: Data Preprocessing. Based on the collected data, for each omics data type, we first filtered out features with more than 20% missing values ​​across all patients, and patients with more than 20% missing values ​​across all features. Then, we selected samples belonging to the same patient group as these three omics datasets. We used the K-nearest neighbor algorithm to calculate the remaining missing values. Finally, we normalized each dataset by calculating the z-score to eliminate differences caused by using different scales in these datasets. The final selection of data samples is shown in Table 2.

[0026] Table 2 Cancer Dataset

[0027] BLCA Bladder and urethral cancer 396 BRCA Invasive breast cancer 1030 KIRC Clear cell carcinoma of the kidney 488 LUAD lung cancer 481 SKCM melanoma 436 STAD Stomach cancer 402 UCEC Endometrial cancer 509

[0028] Step 3: Training and Evaluation Phase. To evaluate the effectiveness of our multiDAAE, using the data processed in Step 2, we compared it with nine other commonly used cancer subtype identification algorithms (NEMO, MCCA, PINS, SNF, K means, Spectral, AE, VAE, and SubtypeGAN) on seven cancer datasets. Due to the lack of labels for the omics data, we evaluated each algorithm based on the p-value of the Log-rank test. A p-value below 0.05 for observed subtypes was considered to indicate significant differences. For ease of comparison, the p-values ​​were transformed using negative numbers and logarithmic functions. We also used the C-index to evaluate the accuracy of the survival prediction model. Notably, for each method, we attempted to set the parameters according to their respective prescribed rules. However, we made slight adjustments to some methods to better suit our dataset. Since most clustering methods do not automatically select an appropriate number of clusters, and different methods have different criteria for determining the optimal number of subtypes, we did not require the number of subtypes to be the same for each method. We used the silhouette coefficient to determine the optimal number of clusters for each method. The silhouette coefficient measures both the density and dispersion of clusters, ranging from -1 to 1; a higher value indicates better clustering performance. Based on previous research, we define a reasonable range for the number of cancer subtypes as 2 to 8.

[0029] Table 3 Results of MultiDAAE and Comparison Methods

[0030]

[0031] As can be seen from the results in Table 3 (the first value is the -log10P value, the second is the number of clusters, and the best result is highlighted in bold), our proposed subtype identification method, MultiDAAE, outperforms other methods on most datasets.

[0032] Those skilled in the art will understand that the scope of protection of this invention is not limited to the specific embodiments described. Without departing from the principles of this invention, those skilled in the art can make equivalent modifications or substitutions to the relevant technical features. It should be noted that the modified or substituted technical solutions will all fall within the scope of protection of this invention.

Claims

1. A cancer subtype identification method based on a dual adversarial autoencoder using multi-omics data, characterized in that, The implementation steps are as follows: (1) Construct a dual adversarial autoencoder model MultiDAAE, which includes encoder Q, decoder G, first discriminator D1, and second discriminator D2. Input multi-omics data x into encoder Q and encode it into latent representation z. z is decoded into reconstructed data x' by decoder G. Construct reconstruction loss using mean squared error. Minimize the Euclidean distance between the multi-omics data x and the reconstructed data x'; construct the first generative adversarial network GAN1 by combining the encoder Q and the first discriminator D1, and perform minimum-maximum game adversarial training to match the encoder's posterior distribution with its prior distribution and construct the adversarial loss. The decoder G and the second discriminator D2 are combined to form a second generative adversarial network GAN2; adversarial training is used to match the distribution of the reconstructed data with the distribution of the real data and to construct an adversarial loss. Construct the total loss ,in , To represent respectively as resistance to loss and combat losses The weighting coefficients, and Backpropagation and stochastic gradient descent are used to update the model parameters until convergence. After the model converges, latent representation features are extracted, and K-means clustering is used to identify cancer subtypes. (2) Select and collect three types of omics data from TCGA: DNA methylation at the epigenome level, mRNA expression and miRNA expression at the transcriptome level; (3) Preprocess the collected omics data; (4) Evaluate the performance of the trained MultiDAAE model to verify the effectiveness of the method.

2. The cancer subtype identification method based on a dual adversarial autoencoder using multi-omics data according to claim 1, characterized in that, The specific methods for the MultiDAAE model construction phase are as follows: (1) A first generative adversarial network is formed by using encoder Q as generator and first discriminator D1 as discriminator. The parameters of encoder Q and first discriminator D1 are updated through adversarial training of mini-maximum game so that the posterior distribution Q(z|x) of encoder matches the prior distribution P(z') of latent representation z. A second generative adversarial network is formed by using decoder G as generator and second discriminator D2 as discriminator. The parameters of decoder G and second discriminator D2 are updated through adversarial training so that the distribution of reconstructed data x' output by decoder matches the distribution of real multi-omics data x, thereby realizing the dual adversarial construction of two discriminators. (2) The reconstruction term is the minimum Euclidean distance between the original multi-omics data x input by the autoencoder and the reconstructed data x' output by the decoder; (3) Regarding the batch size, training epochs, learning rate, and... A comprehensive hyperparameter search is performed to optimize model performance.

3. The cancer subtype identification method based on a dual adversarial autoencoder using multi-omics data according to claim 1, characterized in that... Omics data collection phase: (1) Select three types of omics data from the rich data, including DNA methylation at the epigenome level, and mRNA and miRNA expression at the transcriptome level; (2) Download the Level 3 data of HTSeq-FPKM data and BCGSC miRNA Profiling miRNA-Seq data from TCGA, and use them for mRNA and miRNA data respectively. For DNA methylation data, select the Level 3 data of Illumina HumanMethylation 450 from TCGA.

4. The cancer subtype identification method based on a dual adversarial autoencoder using multi-omics data according to claim 1, characterized in that... Data preprocessing stage: (1) Filter out features with more than 20% missing values ​​in all patients, and patients with more than 20% missing values ​​in all features; (2) Use the K-nearest neighbor algorithm to calculate the remaining missing values; (3) Normalize each dataset by calculating z-score to eliminate the differences caused by using different scales in these datasets.

5. The cancer subtype identification method based on a dual adversarial autoencoder using multi-omics data according to claim 1, characterized in that the training and evaluation phase includes: (1) The trained MultiDAAE model was compared with nine other commonly used cancer subtype identification algorithms, NEMO, MCCA, PINS, SNF, K-means, Spectral, AE, VAE and SubtypeGAN, on seven cancer datasets. (2) Evaluate each algorithm based on the p-value of the Log-rank test. If the p-value of the observed subtype is less than 0.05, they are considered to have significant differences. (3) The consistency index C-index is used to evaluate the accuracy of the MultiDAAE model, where, The C-index is used to characterize the consistency between the MultiDAAE model's ranking of patient survival risks and the actual ranking of survival outcomes. A larger C-index indicates higher prediction accuracy. (4) Use the silhouette coefficient to determine the optimal number of clusters for the MultiDAAE model.

Images