CN110797080A - Prediction of synthetic lethal genes based on cross-species transfer learning - Google Patents

Abstract

The invention belongs to the field of bioinformatics, and particularly relates to a method for predicting synthetic lethal gene based on cross-species transfer learning. The method of the present invention migrates the synthetic lethal gene learned by Saccharomyces cerevisiae into human to predict the synthetic lethal gene of human. The method consists of two basic steps. First, manifold feature learning is performed, learning new feature representations of the two species. Then, the relative importance of the edge distribution and the conditional distribution is quantitatively evaluated by adopting a dynamic distribution alignment method, and the difference of the edge distribution and the conditional distribution between the two species is adaptively minimized. Finally, a domain invariant synthetic lethal gene classifier is learned by summarizing these two steps. The invention can be used to predict synthetic lethal genes in humans.

Description

Prediction of synthetic lethal genes based on cross-species transfer learning

technical field

The invention belongs to the field of bioinformatics, in particular to a method for predicting synthetic lethal genes based on cross-species transfer learning.

Background technique

The current screening methods for potential synthetic lethality (SL) gene pairs can be classified into three categories.

The first is a model organism-based approach. Their genomes are small and easily mutated and matched; thus, gene silencing techniques are easier to perform in model organisms. However, as with the homology inference method for all model organisms, most of the genes in the model organism SL gene pairs have no homologous genes in the human genome. Although homologous genes can be found in the human genome, their functions have changed dramatically and cannot be directly translated into SL genes.

The second screening method is the mammalian gene silencing method, and two gene silencing methods have been developed. One is speculation based on prior knowledge. Potential SL gene pairs contained two genes, the mutated cancer gene and the SL partner gene. Therefore, SL partner genes should be directly knocked out and tested one by one. The other is based on high-throughput experimental techniques for unbiased screening of the entire genome. Ultimately, siRNA and CRISPR screens proved to be the most reliable methods to detect SL genes against s15. However, compared with model genetic systems, human cell systems face greater challenges in genome-wide siRNA or CRISPR screening. Also, these methods are much more expensive, labor-intensive and time-intensive, so many of the essential genes discovered are either limited to these cell line models or are often overexpressed in cancer.

The third is the computing method based on big data and data mining. This data-driven approach includes biological network topology, data mining and statistical screening. Computational approaches are an attractive alternative to genome-wide siRNA or CRISPR-based human cell line screening methods that can help identify and prioritize potential SL genes for further experimental validation. These methods include inferring human homologous SL genes from yeast SL genes; evaluating the importance of gene pairs using robust features of tumor PPI networks; using statistical models of gene mutation/transcription expression data for mutual exclusion calculations; SL (DAISY) data-driven detection of cell copy number alterations, siRNA screening, cell survival and gene co-expression information has achieved good results in data-driven SL genes; and a learning-based training and prediction pipeline that covers mutation coverage, drives The three features, mutation probability and network information centrality, are combined into a manifold ranking model that generates a ranked list of potential SL pairs.

To sum up, the existing methods for predicting human synthetic lethal genes are costly and require a lot of labor and time.

SUMMARY OF THE INVENTION

Aiming at the problem that the utility of the existing supervised learning methods is limited and the amount of synthetic lethal gene data for humans is small, the present invention proposes to predict synthetic lethal genes based on cross-species transfer learning. Obtain abundant, experimentally validated synthetic lethal effects from model organisms such as yeast and mice to predict synthetic lethal genes in humans. The described method steps include:

1. Data collection stage

We generated PPI networks from data collected from the BioGrid protein interaction database, with each node representing a protein and each edge representing interactions between proteins. The source and target species genes obtained from the PPI network were then classified using a trained classifier, with gene pairs with synthetic lethality being the positive dataset, and gene pairs not having synthetic lethality being the negative dataset. The known synthetic lethality between two genes is represented by a binary matrix Ys,Yt, with 1 for synthetic lethality and 0 for not synthetic lethality.

2. Data preprocessing stage

The PPI network topological similarity measurement is performed on the source species and the target species to obtain the topological similarity matrix Ns∈Rn×k, Nt∈Rm×k, where k is the network parameter of the gene pair. Measure the GO semantic similarity between the source species and the target species to obtain the semantic similarity matrix Gs∈Rn×d, Gt∈Rm×d, where d is the number of methods to calculate the GO similarity. Then, based on the linear combination of the topological similarity matrix of the PPI network and the semantic similarity matrix based on the GO method, the characteristic matrices Xs, Xt of the source species and the target species are obtained, as follows:

X _s =[N _s G _s ]

X _t =[N _t G _t ]

The cross-species transfer learning method consists of two basic steps. First, manifold feature learning is performed to learn new feature representations for both species. Second, using a dynamic distribution alignment approach, the relative importance of marginal and conditional distributions was quantitatively evaluated, and the marginal and conditional distribution differences between the two species were adaptively minimized. Finally, a domain-invariant synthetic lethal classifier f can be learned by summarizing these two steps. Formally, the manifold feature learning function is represented by g( ), and the objective function is expressed as follows:

where the first term represents the loss of the data sample. is the square norm of f. Df(·,·) denotes dynamic distribution alignment. Rf(·,·) is the Laplace regularization, and η, λ and ρ are the corresponding regularization parameters.

3. Manifold feature learning stage

The goal of manifold feature learning is to determine a new feature space in which the source and target species exhibit common features. The new feature representation of common features is domain-invariant, thus enabling the transfer of classifiers from source species to target species. We embed the source and target datasets into the Grassmann manifold method G(d), which can be viewed as the set of all d-dimensional subspaces {Φ(T): 0≤t≤1}. For the D-dimensional eigenvectors of the two original gene pairs xi and xj, we compute Φ(T)tx, which is the projection of an eigenvector x in this subspace, for a continuous t from 0 to 1, and assign All projections are concatenated into infinite-dimensional eigenvectors zi and zj. The inner product of the eigenvectors zi and zj produces a positive semidefinite geodesic flow kernel function as:

Therefore, by transforming the source feature space into a Grassmann manifold feature space with z=g(X)=√gx, G can be efficiently calculated by singular value decomposition, and the objective function can be expressed as:

Then the structure of Ds is minimized:

where is the Frobenius norm. K∈r(Nm)×(Nm) is the kernel matrix, Kij=k(zi, ZJ), A∈r(Nm)×(Nm) is a diagonal matrix, if i∈Ds, then Aii=1, otherwise Aii=0. y=y1, y2, ., y(Nm) is the label matrix of S. cerevisiae and target species. tr( ) is a trace operation.

4. Dynamic distribution alignment stage

The main purpose of dynamic distribution alignment is distribution adaptation to minimize distribution differences between domains. Using dynamic distribution alignment, the importance of marginal distribution (P) and conditional distribution (Q) between two species was quantitatively evaluated. To this end, an adaptive factor μ is introduced, and the dynamic distribution alignment function is defined as:

(1) Measurement of distribution divergence

The maximum mean deviation MMD between the marginal distribution P and the conditional distribution Q is defined as:

Therefore, the dynamic distribution alignment function can be expressed as:

where the first term represents the marginal distribution bias between species and the second term represents the conditional distribution bias. By further utilizing the concrete theorem and the kernel trick, the dynamic distribution alignment function in the above equation can be transformed into:

(2) Adaptive factor μ

A-distance is used as a basic measure to obtain the adaptation factor. Define A-distance as the error in building a linear classifier to distinguish two domains. ε(H) represents the error of the linear classifier h discriminating the two regions Ds and Dt. A-distance is defined as follows:

d _A (D _s , D _t )=2(1-2ε(h))

Then μ can be estimated as:

where dM represents the marginal distribution of the c-th class A-distance, and dC represents the conditional distribution of A-distance.

5. Laplacian regularization Laplacian regularization is introduced to further exploit the similar geometric properties of adjacent points in the manifold method G. The pair-wise affinity matrix is as follows:

where sim(·,·) is a similarity function (such as cosine distance) that measures the distance between two points. Np(Zi) represents the nearest neighbor set of point Zi. P is a free parameter that must be set in this method. By introducing the Laplacian matrix L=D-W of the diagonal matrix, the final Laplacian regularization term of the equation is obtained.

The final objective function is expressed as:

得到解set derivative

get a solution

β ^* = ((A+λW+ρL)K+ηI) ^-1 AY ^T

Description of drawings

Figure 1: Similarity Metrics for Gene Pairs

Figure 2: Manifold feature matrix transformation

Figure 3: Dynamic Distribution Alignment

Figure 4: Two different target domains

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with experiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

1. Data collection

We obtained species-specific PPI networks from the BioGrid database, which included more than 740,000 protein interactions in Saccharomyces cerevisiae, more than 74,000 protein interactions in Schizosaccharomyces cerevisiae, and more than 470,000 protein interactions in humans. The BioGrid database also provides synthetic lethality among experimentally validated genes, including more than 14,000 genes with synthetic lethal effects in Saccharomyces cerevisiae, more than 900 genes with synthetic lethal effects in fission yeast, and more than 800 genes with synthetic lethal effects in humans Synthetic lethal genes. GO has three sub-ontologies, namely biological process (BP), molecular function (MF) and cellular component (CC). BP has 29660 items, MF has 11120 items, and CC has 4115 items. Among various computational methods based on GO semantic similarity, we used the protein semantic similarity tool proposed by Mazandu et al. The synthetic lethal prediction algorithm is as follows:

2. Prediction of synthetic lethal genes in fission yeast

We applied the TLSL model to S. cerevisiae and S. cerevisiae, with S. cerevisiae as the source species and Schizosaccharomyces cerevisiae as the target species. In Saccharomyces cerevisiae, we constructed a PPI network that includes 9000 experimentally derived synthetic lethalities. The PPI network consisted of five types of 904 synthetic lethality, 50 dose lethality, 200 negative genetics, 200 comprehensive growth defects, and 200 positive genetic interactions. In Saccharomyces cerevisiae, we considered 8500 pairs of synthetic lethal genes as the positive dataset and 18000 random pairs were generated in a connected component graph as the negative dataset. In fission yeast, 906 SLs were positive datasets and 8237 NSLs were negative datasets. Second, the topology-based PPI similarity matrix and the GO-based semantic similarity matrix are calculated respectively. Finally, the gene pairs with missing functional similarity were removed, and the feature matrices Xs ∈ R ^25039×35 and Xt ∈ R ^8463×35 were obtained by linear combination. Using the feature matrix as the input of the transfer learning model, the synthetic lethal prediction result Yt of fission yeast was obtained. To evaluate the performance of the proposed method, we employ a series of performance evaluation procedures to evaluate our model to predict SLS, including accuracy (ACC), sensitivity (Se), specificity (Sp), precision (Pr) , F1-measure (F1), G-mean (GM), Matthews correlation coefficient (MCC). TLSL identifies SLs that are missing in fission yeast, and we expected to find 177 SL pairs, but only 65 were found. Table 1 shows that the sensitivity of this method is 95.9%-80.5%, the specificity is 91.6%-89.7%, and the accuracy is 88.6%-85.1%.

Table 1. Performance comparison of fission yeast synthetic lethality prediction models

3. Human synthetic lethal gene prediction

We used Saccharomyces cerevisiae as the tagged source species and humans as the untagged target species. We used the source dataset in fission yeast synthetic lethal gene prediction. A human PPI network was constructed using the BiorGrid database, including 6645 genes and 17083 physical interactions. 803 SLs were randomly selected as the positive dataset and 6000 NSLs as the negative dataset. Second, the topology-based PPI similarity matrix and the GO-based semantic similarity matrix are calculated respectively. Finally, the gene pairs with missing functional similarity are removed, and the human feature matrix Xt∈R ^8463×35 is obtained by linear combination. Using the feature matrix as the input to the transfer learning model, the synthetic lethal prediction result Yt for humans is obtained. To evaluate the predictive performance of the TLSL method, its results were compared with the SINaTRA method. The results (Table 1) show that TLSL performs best on all metrics for human SL gene pair classification, with significant improvements in each.

Table 2. Performance comparison of human synthetic lethality prediction models

4. Experiment and result analysis

Experimental results demonstrate that the transfer learning model outperforms contemporary state-of-the-art classifiers in the cross-species learning task of translocating synthetic lethal genes from Saccharomyces cerevisiae to synthetic lethal genes in humans. The empirical success of transfer learning models can be attributed to the following advantages. First, manifold feature learning of transfer learning models is able to learn new feature representations that are invariant to common features of both species. Therefore, shallow models that only focus on the covariance of observed variables, such as random forests and support vector machines, will have a hard time capturing this common feature between the two domains. Second, dynamic distribution alignment in transfer learning models takes into account marginal and conditional distributions across species and adaptively exploits the importance of each distribution. While traditional classifiers, such as random forests, often fail to capture distribution differences between domains, which limits their performance on cross-species tasks, resulting in poor performance.

Claims (6)

1. Predicting synthetic lethal gene based on cross-species transfer learning, it is characterized in that implementing steps: (1) Collect data to generate a PPI network from Saccharomyces cerevisiae, fission yeast and human protein interaction data collected from the BioGrid protein interaction database; (2) Data preprocessing, calculate PPI network topological similarity and GO-based semantic similarity measure for source species and target species, and obtain feature matrix through linear combination; (3) Manifold feature learning, embedding the source dataset and target dataset into the Grassmann manifold method, and then transforming the source feature space into a Grassmann manifold feature space; (4) Dynamic distribution alignment, using the dynamic distribution alignment method to quantitatively evaluate the importance of marginal distribution and conditional distribution between two species; (5) Laplacian regularization. Laplacian regularization is introduced to further utilize the similar geometric properties of adjacent points in the manifold method G to obtain the pair-wise affinity matrix. 2. The synthetic lethal gene prediction based on cross-species transfer learning according to claim 1 is characterized in that the data collection stage: Species-specific PPI networks, including Saccharomyces cerevisiae, Schizosaccharomyces cerevisiae, and humans, were obtained from the BioGird Protein Interaction Database. 3. The synthetic lethal gene prediction based on cross-species transfer learning according to claim 1, is characterized in that the data preprocessing stage: (1) Measure the topological similarity of the PPI network between the source species and the target species; (2) Measure the GO-based semantic similarity of source and target species; (3) Linearly combine the PPI network topological similarity matrix and the GO-based semantic similarity matrix: X _s =[N _s G _s ] X _t =[N _t G _t ] . 4. The synthetic lethal gene prediction based on cross-species transfer learning according to claim 1, is characterized in that the manifold feature learning stage: (1) We embed the source dataset and target dataset into the Grassmann manifold method; (2) Using the GFK (Geodesic Flow Kernel) algorithm, the inner product of the eigenvectors obtains the positive semi-definite geodesic flow kernel function:

(3) Convert the source feature space into a manifold feature space. 5. The synthetic lethal gene prediction based on cross-species transfer learning according to claim 1 is characterized in that the dynamic distribution alignment stage: (1) Calculate the maximum mean deviation MMD between the marginal distribution P and the conditional distribution Q:

(2) Measure the alignment divergence of the dynamic distribution:

(3) Calculate the adaptive factor μ:

. 6. The synthetic lethal gene prediction based on cross-species transfer learning according to claim 1, characterized in that the Laplacian regularization stage: (1) Calculate the Laplace regularization term:

(2) Set the derivative

get the solution: β ^* =((A+λM+ρL)K+ ^ηI ) ⁻¹ AYT.

Images