Research On Proteomics Molecular Subtyping Of Gastric Cancer Based On Deep Neural Network Feature Extraction

Gastric cancer is one of the most common malignancies worldwide and a leading cause of cancer-related death.The high morbidity and mortality of gastric cancer reflect the insufficiency of its diagnosis and treatment methods.Tumor resection is currently the only potential curative treatment for gastric cancer,but even with radical surgical resection with adjuvant chemotherapy,the prognosis remains unsatisfactory.With the development of precision medicine,although there are clinically available targeted drugs for gastric cancer,only trastuzumab is approved for first-line treatment,and it can only be used in HER2-positive patients.The standard treatment for gastric cancer patients with HER2-negative is still chemotherapy,but not all patients benefit from it.Studies have shown that molecular subtyping can indicate which patients may benefit from chemotherapy,and in recent years,deep learning has shown great potential in cancer omics prognostic research.In order to solve the precision chemotherapy of gastric cancer,this study applied deep learning to gastric cancer proteomics for the first time,and carried out the following researches:1.Based on the deep neural networks,we established a proteomics molecular subtyping framework.In this study,we used 833 formalin-fixed,paraffin-embedded surgically resected TNM stage II/III gastric cancer samples,and established a feature extracted deep learning-based proteomic molecular subtyping framework.The framework includes a consensus clustering workflow based on deep learning autoencoder feature extraction and a random forest classifier.Using the workflow we identified two subtypes with survival differences in the discovery set: S-I and S-II.Moreover,patients in the S-I who received adjuvant chemotherapy had a significant improvement in 5-year overall survival rate compared with patients who received surgery alone(65.3% vs 52.6%,log-rank P-value=0.014),but no improvement was observed in S-II(54% vs 51%,logrank P-value=0.96).Subsequently,we validated the two subtypes on the independent validation set using a random forest classifier,with an average AUC value of 0.92 for the model combined with ten-fold cross-validation on the training set and 0.91 on the test set.2.Demonstrate the advantages of features extracted by autoencoder in predicting prognosis and chemotherapy benefit based on two alternative approaches.In this study,Principal Component Analysis and Uniform Manifold Approximation and Projection for Dimension Reduction were used to replace the autoencoder extracted features in the consensus clustering workflow,and found that the subtypes identified based on the autoencoder had more significant survival differences,which proved that the features extracted by the autoencoder had advantages in predicting the prognosis of gastric cancer and the chemotherapy benefit.3.The consensus clustering workflow established in this study has scalability.The consensus clustering workflow based on deep learning autoencoder feature extraction established in this study was tested on two clinically publicly available gastric cancer datasets,and subtypes with survival differences were successfully identified,proving that the consensus clustering workflow has extensibility.