Pan-cancer cell recognition model training method and device, equipment and storage medium

By constructing an initial dataset and using a multi-task optimization method, a pan-cancer cell identification model is generated, which solves the problem of insufficient accuracy of single-cancer models in identifying unknown cancer types in existing technologies, and achieves efficient pan-cancer cell identification.

CN122455098APending Publication Date: 2026-07-24JIYINJIA BIOMEDICAL TECHNOLOGY (SHAOXING) CO LTD +3
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610865083.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-15
Publication Date
2026-07-24

AI Technical Summary

Technical Problem

Existing single-cancer models are not accurate enough in data analysis of tumors with unknown primary lesions, rare cancers, or cancers that metastasize across organs. Furthermore, existing identification methods rely on sensitive reference sample selection, consume a lot of computational resources, and have complex processes, making it difficult to achieve efficient identification of pan-cancer types.

Method used

By constructing an initial dataset, selecting highly variable genes as training features, using a pre-trained pedestal model and attaching a binary classification head, and combining classification cross-entropy loss and gene reconstruction loss for multi-task optimization, parameter fine-tuning is performed to generate a pan-cancer cell recognition model.

Benefits of technology

It achieves a balance between recall and precision in pan-cancer cell identification, improves the accuracy of identifying tumors with unknown primary lesions, rare cancer types, or metastatic cancers across organs, and reduces computational resource consumption.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122455098A_ABST
    Figure CN122455098A_ABST
Patent Text Reader

Abstract

Embodiments of the present application disclose a pan-cancer cell recognition model training method, device, equipment and storage medium. The method comprises selecting transcriptome test data of a plurality of preset cancer types according to a preset ratio of normal cells and malignant cells to construct an initial data set; obtaining a preset number of high variable genes in the transcriptome test data of the initial data set as training features to obtain a training set; based on a binary classification label, performing label annotation on the transcriptome test data in the training set to generate a binary classification label system test data set; taking the pre-training weight of a preset base model as an initialization parameter, mounting a binary classification head at the top layer of a transformer layer, and performing parameter fine-tuning on the preset base model based on the binary classification label system test data set to obtain a pan-cancer cell recognition model. The scheme of the embodiments of the present application can realize accurate identification of benign and malignant cells in pan-cancer single-cell transcriptome data.
Need to check novelty before this filing date? Find Prior Art