Cancer genome map key gene positioning method

Abstract

The invention discloses a cancer genome map key gene positioning method, which comprises the following steps of: 1) for each case in a case set, generating pathological image embedding and text embedding of the case; for each gene in a gene expression matrix corresponding to the case set, a standardized expression vector of the gene is spliced with pathological image embedding and text embedding of each case into a unified state vector, and the unified state vector is used as initial observation of an intelligent agent in a reinforcement learning algorithm; (2) each agent generates a key gene subset of current positioning according to the action adopted by the initial random strategy of the agent; 3) each agent encodes the current key gene subset into a current state input DQN to obtain a corresponding action update key gene subset code and award; updating the next state of the intelligent agent; the environment knowledge explored by the intelligent agent meets a set condition; and 4) when each agent meets a set convergence condition, taking the currently positioned key gene subset as a key gene set.

Description

Technical Field

[0001] This invention belongs to the fields of multi-agent reinforcement learning, large-scale multimodal fusion, bioinformatics, and tumor omics analysis. Specifically, it relates to a method for locating key genes using RNA-seq data from The Cancer Genome Atlas (TCGA). This method utilizes a multi-agent reinforcement learning (MARL) framework to transform whole-slice images (WSI) and clinical text into semantic embeddings via a large multimodal model. These embeddings are then injected into the state representation of each agent, achieving high-precision, interpretable key gene localization under image-text-gene joint decision-making. This significantly improves the performance of downstream tasks such as tumor subtyping and prognosis prediction. Background Technology

[0002] Single-cell RNA sequencing data refers to data obtained by sequencing the RNA of a single cell. This technology can detect RNA expression in a single cell, including gene expression levels, transcript abundance, splicing variations, and other information.

[0003] This technology isolates and captures individual cells, then uses high-sensitivity sequencing technology to extract and sequence RNA from each cell. Single-cell RNA sequencing is of great significance for studying cell development, differentiation, disease occurrence and development, and can be used for a series of downstream tasks (such as cell clustering) to reveal heterogeneity in cell populations, including differences in gene expression among different cell types and states. However, due to the large number of genes in a cell, the number of features in single-cell RNA sequencing data is far greater than the number of samples. This high-throughput data (feature count far exceeding sample count) can affect the accuracy of downstream tasks (such as cell clustering), and consequently affect the analysis of cellular heterogeneity.

[0004] With the rapid development of machine learning technology, feature selection for high-throughput data to improve the performance of downstream tasks has become an important research area. Currently, a series of studies on gene (feature) selection for single-cell RNA sequencing data can be divided into three categories: filtering, wrapping, and embedding. These studies have achieved some results, but still have limitations. For example, filtering ignores the feature dependencies and interactions between feature selection and predictor variables; wrapping methods must search 2N candidate feature subspaces (where N is the number of features); and embedding methods are affected by the strong structured assumptions of the predictive model.

[0005] The emergence of reinforcement learning offers a new solution to the aforementioned problems. Reinforcement learning (RL) is a branch of machine learning, a trial-and-error-based learning method designed to teach an agent how to act in interactions with its environment to maximize rewards or achieve specific goals. The agent observes its environment and performs different actions to receive feedback (rewards or penalties), learning and adjusting its behavioral strategies through this feedback. Reinforcement learning algorithms update the agent's strategy based on historical experience and environmental rewards, hoping to obtain more rewards in future interactions. Depending on the number of agents involved, reinforcement learning can be divided into single-agent learning and multi-agent learning. Multi-agent learning combines deep learning methods to better adapt to complex environments and tasks.

[0006] Compared to filtering methods, multi-agent reinforcement learning-based methods use deep neural network models to better extract the correlation between feature selection and predictive variables; compared to wrapping methods, multi-agent learning replaces feature space search, reducing the size of candidate feature space and achieving automated feature subspace exploration; compared to embedding methods, it decouples from downstream task models, enhancing model flexibility.

[0007] The TCGA dataset provides RNA-seq, digitized pathological images (WSI), and structured clinical text (including staging, survival, and treatment records). Existing gene selection methods are mainly geared towards single-modality RNA-seq and have the following drawbacks: (1) they ignore the tumor microenvironment and morphological information in pathological images; (2) they ignore higher-order semantics such as phenotype and prognosis in clinical text; (3) the gene dimension of high-throughput is much larger than the sample size, making it prone to overfitting; and (4) traditional filtering / wrapping / embedding methods cannot capture cross-modal nonlinear interactions. In recent years, multimodal large models (such as GPT-5, BioVIL, and Clinical-BERT) have made breakthroughs in understanding pathological images and medical texts, but no work has yet introduced their semantic representation into the reinforcement learning gene localization process. Therefore, there is an urgent need for a key gene localization technology that can jointly utilize RNA-seq, pathological images, and clinical texts. Summary of the Invention

[0008] To address the problems existing in the prior art, the present invention aims to provide an RNA-seq key gene localization method for the TCGA cancer genome atlas. This method uses a multi-agent reinforcement learning framework, introducing semantic embeddings of pathological images and clinical text generated by a multimodal large model to form a "image-text-gene" ternary collaborative decision-making mechanism. This achieves cross-modal, high-precision, and interpretable key gene localization, significantly improving the performance of downstream tasks such as tumor prognosis prediction and molecular subtyping. The core idea of ​​this invention is: each gene is treated as an agent, and the agent's action is "selecting / abandoning" the gene; the agent's state is jointly constituted by the expression matrix of the current gene subset, the pathological image embedding extracted by the multimodal large model, and the clinical text embedding; the reward function simultaneously considers downstream prognostic prediction performance and gene subset sparsity. Through multi-agent collaborative exploration and strategy improvement, a set of key genes highly correlated with image-text phenotypes is automatically discovered.

[0009] The technical solution of this invention is as follows: A method for locating key genes in a cancer genome map, comprising the following steps: 1) For each case in the case set, extract a representative image of the digital pathological image WSI from the case and input it into the multimodal large model to generate a pathological image embedding. Concatenate the clinical text of the case into a fixed-length sequence and input it into the multimodal large model to generate a text embedding. For each gene in the gene expression matrix corresponding to the case set, concatenate its standardized expression vector with the pathological image embedding and text embedding of each case into a unified state vector, which is used as the initial observation of an agent in the reinforcement learning algorithm to initialize the agent. 2) In reinforcement learning algorithms, each agent generates a subset of key genes for its current location based on the actions it takes according to its initial random policy. 3) Each agent encodes the current key gene subset as the current state input to DQN to obtain the corresponding action update key gene subset encoding and calculate the corresponding reward for the agent; then, based on the updated key gene subset encoding, the agent's next state is obtained, and (current state, action, reward, next state) is put into the agent's memory; 4) Repeat step 3) multiple times until the environmental knowledge explored by the agent meets the set conditions; 5) Each agent obtains its state vector based on the key gene subset located and the pathological image embedding and text embedding of each case, and then takes an action according to its own policy; and puts (current state, action, reward, next state) into memory; then extracts a batch of data from memory to update the corresponding agent's policy, so as to obtain the maximum long-term reward. 6) When each agent meets the set convergence conditions, the currently located subset of key genes is taken as the key gene set.

[0010] Preferably, each agent has one memory slot to store a quadruple of (current state, action, reward, and next state); where the state is a one-dimensional vector obtained by inputting a subset of the currently located key genes into the autoencoder, the action is selecting a gene as a key gene or discarding a gene, and the reward is calculated using the following formula: NMI is an evaluation metric used to assess downstream cell clustering tasks. The number of genes used to locate prior knowledge. This represents the number of genes currently located.

[0011] Preferably, in step 5), each agent extracts several quadruplets from its memory and updates its own strategy in real time according to the Bellman equation in order to obtain the maximum long-term reward.

[0012] Preferably, the convergence condition is that the average reward of each agent does not increase within a set time period.

[0013] Preferably, the clinical text includes TNM stage, age, gender, and smoking history.

[0014] Preferably, the case set is the TCGA dataset.

[0015] Preferably, the multimodal large model is GPT-5.

[0016] A server is characterized by comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the methods described above.

[0017] A computer-readable storage medium having a computer program stored thereon, characterized in that the computer program implements the above-described method when executed by a processor.

[0018] Compared to previous methods, the key gene localization method of multi-agent reinforcement learning TCGA based on multimodal large model semantic embedding of the present invention has the following advantages and contributions: (1) For the first time, the semantic embedding of pathological images and clinical texts was introduced into the multi-agent reinforcement learning gene localization framework to achieve cross-modal collaborative decision-making and significantly improve the accuracy of downstream prognosis prediction.

[0019] (2) The reward function simultaneously optimizes prediction performance and gene sparsity, and the results are biologically interpretable and easy to verify in subsequent experiments.

[0020] (3) The framework design of the present invention is highly versatile and can be seamlessly extended to other cancer types or omics data, with good flexibility and scalability. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0023] The goal of this invention is to leverage the TCGA cancer genome atlas, utilizing multimodal large-scale model semantic embedding and multi-agent reinforcement learning, to achieve key gene localization through a "image-text-gene" ternary collaborative approach, thereby improving the dataset's performance on downstream tasks and better revealing heterogeneity within cell populations. This embodiment uses the TCGA-BRCA dataset as an example, which includes single-cell RNA sequencing data (each row represents one cell (sample), each column represents one gene (feature), i.e., a two-dimensional table), WSI pathological images, and clinical text information (TNM stage, age, gender, smoking history, etc.). The method flow of this invention is as follows: Figure 1 As shown.

[0024] Step 1: Multimodal Semantic Embedding Generation and State Initialization. For each TCGA case, extract representative images from the corresponding WSI and generate pathological image embeddings using a multimodal large model (such as GPT-5). Concatenate clinical text (TNM stage, age, gender, smoking history, etc.) into a fixed-length sequence and generate text embeddings using a multimodal large model (such as GPT-5). For each gene, concatenate its standardized expression vector with the image and text embeddings to form a unified state vector, which serves as the initial observation for the corresponding agent. Each gene corresponds to one agent, and all agents are initialized using DQN (Deep Q-Network). For example, if a dataset contains 5 genes (columns), then 5 DQNs are created to represent the 5 genes respectively. Each agent has one memory slot to store a quadruple of (current state, action, reward, next state). The state refers to the one-dimensional vector obtained by inputting the subset of genes located by all agents into the autoencoder, the action refers to selecting a gene as a key gene or discarding a gene, and the reward is calculated using the formula... NMI (Normalized Mutual Information) is a commonly used evaluation metric for clustering algorithms, used to evaluate downstream cell clustering tasks. The number of genes used to locate prior knowledge. This represents the number of genes currently located. The agent's policy refers to the mapping from state to action. This method uses DQN to fit this mapping; that is, the DQN input is the agent's current state, and the output is the action to be taken.

[0025] Step Two: Multi-Agent Exploration. Each agent first employs a random strategy (randomly locating or discarding genes), encoding the located gene subset into the current state using an autoencoder. DQN is then used as input to obtain the corresponding action (the currently located gene subset), and the corresponding reward is calculated. The autoencoder is then used to encode the current gene subset into the next state, and the (current state, action, reward, next state) is stored in memory. For example, in a single-cell RNA sequencing dataset with 6 cells, the current state is `state`, and the gene subsets located by the agent are 1, 3, and 5. The autoencoder takes the selected 3 feature subsets (a 6×3 matrix) as input to obtain a one-dimensional state vector `state_next`, calculates the reward corresponding to that feature subset (1 represents selection, 0 represents discard), and stores the quadruple `(state, 1, reward, state_next)` in memory for agents 1, 3, and 5, and the quadruple `(state, 0, reward, state_next)` for other agents. Finally, the `state` is updated using the `state_next` value. Repeat the above process multiple times until the agent has explored enough environmental knowledge.

[0026] Step 3: Multi-Agent Policy Enhancement. Each agent still takes an action based on its own policy (DQN) according to the current state, and stores (current state, action, reward, next state) in memory, just like in Step 2. However, unlike the exploration phase, each training session randomly extracts a batch of data from memory and applies it according to the Bellman equation. The strategy is updated in real time (DQN) to maximize long-term rewards. This represents the current state of the i-th agent. This represents the current action of the i-th agent. This represents the reward obtained by the i-th agent for its current state. This represents the next state and the corresponding action. For network parameters, This is to decay the weights. For example, for an agent, 32 quadruplets (state, action, reward, next state) are randomly taken from its memory, the Q-value is calculated according to the Bellman equation, and then backpropagation is performed using the mini-batch gradient descent algorithm to fit the Q-value, and the policy is updated in real time.

[0027] Step 4: Convergence Determination and Optimal Gene Subset Acquisition. When the average reward for each agent does not increase after 10 rounds, the policy is considered converged, and the gene located at this point is the key gene.

[0028] Using DQN as the agent in multi-agent reinforcement learning, the multimodal embeddings generated in step one are used. Sufficient environmental knowledge is acquired according to the method in step two, and training is performed according to the method in step three. By selecting fewer but more valuable features, the performance of downstream tasks is improved. In actual experiments, the high-throughput single-cell RNA sequencing dataset Ting with 187 samples and 21,583 gene features was tested. With 100 exploration iterations and 100 training iterations, a 24.37% improvement in NMI was achieved in the downstream single-cell clustering task, despite selecting fewer features than existing methods. Further testing on 13 TCGA datasets showed results superior to traditional methods, with improvements exceeding 5%. This fully demonstrates the effectiveness of this method in the key gene localization scenario of cancer genome atlas RNA-seq data. The experimental test results are shown in Table 1 below.

[0029] Table 1 Experimental Results The selection of the agent neural network model in step one above can be adjusted according to the actual results.

[0030] Although specific embodiments of the invention have been disclosed for illustrative purposes to aid in understanding and implementing the invention, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the invention should not be limited to the content disclosed in the preferred embodiments, and the scope of protection claimed by the invention is defined by the claims.

Claims

1. A method for locating key genes in cancer genome atlas, comprising the steps of: 1) for each case in the case set, extracting a representative picture of the digitized pathology image WSI of the case and inputting it into a multi-modal large model to generate a pathology image embedding, concatenating the clinical text of the case into a fixed-length sequence and inputting it into the multi-modal large model to generate a text embedding; for each gene in the gene expression matrix corresponding to the case set, concatenating its standardized expression vector with the pathology image embedding and the text embedding of each case into a unified state vector as the initial observation of an agent in a reinforcement learning algorithm, and initializing the agent; 2) each agent in the reinforcement learning algorithm generates a current set of key genes according to the action taken by its initial random strategy; 3) each agent encodes the current set of key genes into a current state input into a DQN to obtain an updated set of key genes corresponding to the action of the agent, and calculates the reward corresponding to the agent; then, according to the updated set of key genes, the next state of the agent is obtained, and (current state, action, reward, next state) is put into the memory of the agent; 4) repeat step 3) multiple times until the environmental knowledge explored by the agent meets the set conditions; 5) each agent obtains the state vector of the agent according to the located set of key genes and the pathology image embedding and the text embedding of each case, then takes an action according to its own strategy and puts (current state, action, reward, next state) into the memory; then, a batch of data is extracted from the memory to update the strategy of the corresponding agent, so that it obtains the maximum long-term reward; 6) when each agent meets the set convergence conditions, the current set of key genes is taken as the set of key genes. In step 5), each agent extracts several quadruples from its memory to update its own strategy in real time according to the Bellman equation, for obtaining the maximum long-term reward. The convergence condition is that the average reward of each agent does not improve within a set time period. The clinical text includes TNM stage, age, gender, and smoking history. The case set is a TCGA dataset. The multi-modal large model is GPT-5. A computer program is stored in a memory and executed by a processor, and the computer program includes instructions for executing the method of any one of claims 1 to 7.

2. The method of claim 1, wherein, Each agent has a memory to store the four-tuple (current state, action, reward, next state); wherein the state is a one-dimensional vector obtained by inputting the current positioning key gene subset into the autoencoder, the action is to select the gene as the key gene or discard the gene, and the reward is calculated according to the formula , NMI is an evaluation index for evaluating the downstream cell clustering task, is the number of genes positioned by prior knowledge, is the number of currently positioned genes.

3. The method of claim 1, wherein, The computer program is executed by the processor to implement the method of any one of claims 1 to 7.

4. The method according to claim 1 or 2 or 3, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1 to 7.

5. The method according to claim 1 or 2 or 3, characterized in that, ​ 6. The method according to claim 1 or 2 or 3, characterized in that, ​ 7. The method of claim 1, wherein, ​ 8. A server, characterized by ​ 9. A computer readable storage medium having stored thereon a computer program, characterized in that, ​

Images