cfRNA pan-cancer early screening method and system based on large-scale deep model

Abstract

The present invention discloses a cfRNA pan-cancer early screening method and system based on a large-scale deep model, wherein the cfRNA-based pan-cancer early screening method includes: obtaining raw data and cancer labels, the raw data includes: gene expression data, patient basic information data and patient case data; preprocessing a number of raw data to obtain preprocessed data; the preprocessed data includes a training set and a test set; building a deep learning model, inputting the training set and the cancer label into the deep learning model for training, obtaining an optimized deep learning model, inputting the test set into the trained deep learning model, and obtaining a cancer classification result. The cfRNA-based pan-cancer early screening system includes an acquisition module, a preprocessing module, a model building module and a classification module. The method based on a deep neural network model proposed in the present invention can dynamically summarize the types of information of various cancers from the data and make accurate classifications of cancers.

Description

Technical Field

[0001] The present invention relates to the field of cancer classification, and in particular to a cfRNA pan-cancer early screening method and system based on a large-scale deep model. Background Art

[0002] Free cell-free RNA omics provides a new direction for cancer data classification. The use of cfRNA omics for early disease detection has become a new important direction in academia, and several international journals based on cfRNA for early disease detection have been published. Among them, cfRNA omics for the detection of diseases during pregnancy (preeclampsia) was published in the 2021 issue of Nature, and the detection of neurodegenerative diseases was published in Science Advance in 2020. Grail expects to publish its second international journal paper on the use of cfRNA groups for early screening of lung cancer in 2023, and its academic poster has been published at the top international conference "Biology of Genomes" in 2022. The cfRNA-based lung cancer early screening method published by Grail uses a large combination (large panel) of 466 RNA markers. Preliminary data suggest that in patients with a high tumor fraction ratio (TF) in the peripheral blood free cfDNA group, the sensitivity of lung cancer early screening is as high as 95% with a specificity of 100%, and in patients with low TF, the sensitivity is about 25%8. This report shows that the classification performance of cfRNA genomics cancer early screening data is clearly inclined towards high specificity, which meets the actual clinical needs of early screening.

[0003] The trend of using artificial intelligence in the field of life sciences is increasing year by year, among which language artificial intelligence models have made rapid breakthroughs in recent years. In 2017, the Transformer artificial intelligence model was published at the top conference "NIPS". In 2022, OpenAI used the underlying algorithm based on Transformer to superimpose and reorganize the generative pre-trained transformer model (Generative Pre-trained Transformer; GPT), and published the commercial product ChatGPT. This product has become a major breakthrough in the field of artificial intelligence with its high intelligence, strong understanding ability, and generative dialogue, allowing artificial intelligence to quickly become popular among the general public. Since 2022, the Transformer system has been widely used in the research and development of the biological sciences. For example, in 2022, the MetaAI (formerly Facebook) team published the model ESMFold in the bioRxiv preprint, and in 2023, the entrepreneurial team published the model Progen in "Nature Biotechnology". By using the language artificial intelligence model, it has achieved protein structure prediction capabilities comparable to Alphafold2, and has shown higher effects in specific fields such as predicting unknown proteins. Summary of the invention

[0004] To solve the problems of the prior art, the present invention provides a cfRNA pan-cancer early screening method and system based on a large-scale deep model. The present invention consists of a highlighting module and a deep information extraction module constructed by a deep neural network; the expression information of each gene fragment in cfRNA is used as input, and the cancer label is used as output to train the neural network, thereby achieving early screening of pan-cancer species.

[0005] In order to achieve the above technical objectives, the present invention provides the following technical solutions: a cfRNA pan-cancer early screening method based on a large-scale deep model, comprising:

[0006] Obtaining raw data and cancer labels, wherein the raw data includes: gene expression data, patient basic information data and patient case data;

[0007] Preprocessing the raw data to obtain preprocessed data; wherein the preprocessed data includes a training set and a test set;

[0008] A deep learning model is constructed, the training set and the cancer label are input into the deep learning model for training to obtain an optimized deep learning model, and the test set is input into the trained deep learning model to obtain a cancer classification result.

[0009] Preferably, the raw data are preprocessed into vector matrices based on standard bioinformatics processing techniques to obtain preprocessed data.

[0010] Preferably, the deep learning model includes: a Linformer module and a DRN module.

[0011] Preferably, the process of inputting the training set into the deep learning model for training includes:

[0012] Inputting the preprocessed data into the Linformer module;

[0013] The Linformer module highlights important information in the preprocessed data to obtain highlighted data;

[0014] Inputting the highlighted data into the DRN module to extract deep information and obtain feature information;

[0015] Obtaining a cancer classification prediction result based on the characteristic information;

[0016] Based on the error between the cancer classification prediction result and the true value, the Adam algorithm is used to optimize the model until the error is minimized to obtain a trained deep learning network.

[0017] Preferably, the Linformer module further includes a Dropout layer and a LayerNormalization layer.

[0018] Preferably, the DRN module is constructed based on a five-layer DRN;

[0019] Wherein, each residual block in the DRN module includes two convolutional layers and a shortcut connection;

[0020] The DRN module also includes a BatchNormalization layer and a ReLU activation function.

[0021] The present invention also discloses a cfRNA pan-cancer early screening system based on a large-scale deep model, comprising: an acquisition module, a preprocessing module, a model building module and a classification module; wherein the acquisition module, the preprocessing module, the model building module and the classification module are connected in sequence;

[0022] Wherein, the acquisition module is used to obtain raw data and cancer labels;

[0023] The preprocessing module is used to process the original data to obtain preprocessed data;

[0024] The model building module is used to build a deep learning model;

[0025] The classification module is used to classify cancer types.

[0026] The present invention has the following technical effects:

[0027] The traditional method based on RNA marker set (biomarker) is mainly based on human experience or statistical methods to screen markers. Since there are actually many gene fragments in cfRNA that can be used as markers, the screening process is very resource-intensive, and the markers screened in this way are also difficult to cover the cancer information of various situations in practice. In response to this deficiency, the method based on the deep neural network model proposed in the present invention can dynamically summarize the types of information possessed by various cancers from the data. The highlighting module based on Linformer can effectively distinguish the importance of each gene fragment and highlight it, thereby forming the "soft" biomarker of the present invention. Then, the subsequent DRN module can be used to make accurate judgments on the types of cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0029] Figure 1 A system architecture diagram in an embodiment of the present invention;

[0030] Figure 2 The third batch of sample clinical evaluation graphs in the embodiment of the present invention, wherein graph (a) is the quantitative difference between cancer and healthy controls in artificial intelligence, and graph (b) is the tissue traceability accuracy of esophageal cancer, lung cancer, liver cancer, and gastric cancer;

[0031] Figure 3 This is a clinical evaluation diagram of the fourth batch of samples (pancreatic cancer) in the embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] Embodiment 1

[0034] This embodiment discloses a cfRNA pan-cancer early screening method based on a large-scale deep model. Figure 1 As shown, this embodiment constructs an architecture based on deep learning technology for cancer prediction of cfRNA. This architecture uses Linformer and deep residual network (DRN) as the basic architecture. On the basis of highlighting the sequences useful for cancer prediction, the deep residual neural network is used to screen the features of the highlighted gene sequences and perform multi-classification judgments on various cancers and health. The data set used in this study is the human gene transcription spectrum after bioinformatics analysis, which is imported into the Linformer and DRN models in the form of vector matrices (60,000 gene types / person, 0-1000 gene expression levels / type) for artificial intelligence training.

[0035] The architecture of this embodiment is divided into three modules. Module one is mainly used for the preprocessing of gene expression data, converting it into a vector matrix suitable for model processing. This module adopts standard bioinformatics processing technology. Module two is mainly used to extract more important information from the gene expression data. This embodiment uses the Linformer module to complete this task. Linformer is homologous to Transformer and can capture the correlation between them in the sequence of gene vectors. Linformer is more suitable for modeling longer input sequences than Transformer. After the transformation of this embodiment, Linformer can successfully capture the key information in the gene expression sequence of 60,000 in length and highlight it. Specifically, this embodiment converts the input gene expression data into corresponding sorting information data, and converts each sequence number in the data into a corresponding embedding representation. The self-attention mechanism in the Linformer module is replaced by a low-rank learnable attention mechanism, and a Dropout layer and a LayerNormalization layer are added to reduce the risk of overfitting.

[0036] Module three is mainly used to establish an accurate mapping relationship between the information vector given by Linformer and the classification results of cancer types. In order to retain the complete information from shallow to deep in the deep network as much as possible, this embodiment uses DRN for modeling. Specifically, this embodiment uses a 5-layer DRN, in which each residual block includes two convolutional layers and a shortcut connection. This embodiment also uses BatchNormalization and ReLU activation functions to improve the performance of the model.

[0037] The training method adopts an end-to-end approach. In this embodiment, 60% of the data set is used as a training set, 10% as a validation set, and 30% as a test set. During the model training process, this embodiment uses the back propagation algorithm to update the parameters of the model, and uses the cross entropy loss function to measure the difference between the model's predicted results and the actual results. This embodiment uses the ensemble method to integrate 10 models for training together. During the training process, this embodiment uses the Adam algorithm to optimize the parameters of the model, setting the learning rate to 0.001 and the momentum to 0.9. This embodiment uses the RTX3090 GPU for training to speed up the training.

[0038] From the results, if Figure 2As shown, the pan-cancer (all cancer types) achieved 90% sensitivity, 100% specificity, and AUC = 0.981. Among them, the early screening accuracy of single cancer types was esophageal cancer AUC = 0.986, lung cancer AUC = 0.952, liver cancer AUC = 0.983, gastric cancer AUC = 0.964 and colorectal cancer AUC = 0.89. Jindu Medical quantitative analysis of the differences between cancer and healthy people Figure 2 (a) As shown. Through the dimensionality reduction method, artificial intelligence predicts and scores healthy people (theoretical 0 points) and cancer patients (theoretical 1 point). The AI ​​prediction values ​​show that there is a large quantitative difference between healthy controls and cancer patients, and the scores of cancer patients converge significantly. Only a small number of healthy people and cancer patients have overlapping AI prediction values, which has room for further optimization. Figure 2 (b) As shown. By further training artificial intelligence to classify different types of cancer, when it is known that the patient has changed cancer, the accuracy of different cancers is: liver cancer is 91%, stomach cancer is 74%, esophageal cancer is 68%, and lung cancer is 62%. Based on preliminary data, the accuracy of tissue tracing is higher than that of other similar products. The number of training sets is currently small, and there is a lot of room for optimization in the future.

[0039] like Figure 3 Shown are the clinical evaluation results of the fourth batch of samples (pancreatic cancer) in this example.

[0040] As a preferred embodiment of this embodiment, the information processed by the present invention is based on the expression sequence of each gene fragment of cfRNA, and the ultimate goal is to provide the cancer type of this cfRNA by analyzing the expression sequence;

[0041] The main technology of the present invention is to perform the discrimination of cancer types through neural network technology. The first algorithm uses deep learning technical means including the architecture of Linformer (Transformer variant) and deep residual network (DRN) based on the industry's cutting-edge. The first module extracts important information from the gene expression data and can capture the correlation between them in the sequence of gene vectors. After the modification, Linformer can already capture the key information in the gene expression sequence of 60,000 in length and highlight it. The second module, DRN, can retain complete information from shallow to deep in the deep network, and establish an accurate mapping relationship between the information vector given by Linformer and the classification results of cancer types.

[0042] As a preferred solution of this embodiment, this embodiment uses cfRNA for pan-cancer early screening based on a deep neural network, which is composed of a highlighting module and a deep information extraction module constructed by a deep neural network; the expression information of each gene fragment in cfRNA is used as input, and the cancer label is used as output to train the neural network, thereby achieving early screening of pan-cancer species;

[0043] The gene sequence of the cfRNA is expression data, and the highlighting module and the deep information extraction module are both neural network models.

[0044] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A cfRNA pan-cancer early screening method based on a large-scale deep model, characterized in that: include: Obtaining raw data and cancer labels, wherein the raw data includes: gene expression data, patient basic information data and patient case data; Preprocessing the raw data to obtain preprocessed data; wherein, the raw data are preprocessed into vector matrix based on standard bioinformatics processing technology to obtain preprocessed data; the preprocessed data includes a training set and a test set; Constructing a deep learning model, inputting the training set and the cancer label into the deep learning model for training to obtain an optimized deep learning model, and inputting the test set into the trained deep learning model to obtain a cancer classification result; The deep learning model includes: a Linformer module and a DRN module; the Linformer module also includes a Dropout layer and a Layer Normalization layer; The process of inputting the training set into the deep learning model for training includes: Inputting the preprocessed data into the Linformer module; The Linformer module highlights important information in the preprocessed data to obtain highlighted data; Inputting the highlighted data into the DRN module to extract deep information and obtain feature information; Obtaining a cancer classification prediction result based on the characteristic information; Based on the error between the cancer classification prediction result and the true value, the Adam algorithm is used to optimize the model until the error is minimized to obtain a trained deep learning network.

2. The cfRNA pan-cancer early screening method based on a large-scale deep model according to claim 1, characterized in that: Constructing the DRN module based on the five-layer DRN; Wherein, each residual block in the DRN module includes two convolutional layers and a shortcut connection; The DRN module also includes a Batch Normalization layer and a ReLU activation function.

Images