Machine Learning-based Multi-omics Multimodal Fusion Analysis For Predicting Brain Metastasis Risk In Lung Cancer

Objective: Lung cancer,as the leading cause of malignancy-related death,has been a major concern for clinicians due to its high incidence,metastasis rate and mortality rate.Although lung cancer treatment methods have been increasing in recent years,the prognosis of lung cancer patients is still not optimistic.Among them,tumour metastasis is the main factor affecting the prognosis of lung cancer.Current studies on the mechanism of metastasis show that metastasis is not random but organically predisposed.The "seed and soil" theory of metastasis reveals that the coordination between tumour cells and the tumour microenvironment is the basis for the metastasis of tumour cells to specific organs and the formation of metastases.Therefore,the prediction of organ-specific metastasis is an important research direction in the field of oncology.Lung cancer has a tendency to metastasize to the brain.Brain metastases occur in20%-65% of lung cancer patients during the disease progression and the prognosis is extremely poor,with untreated survival time usually only 1-2 months.If brain metastases can be predicted in advance,patients can be assessed for hazard and risk stratification,which in turn can guide the selection of appropriate drugs and preventive measures in advance,thus providing better control over the entire treatment process.Research on the mechanism of brain metastasis can help to achieve this goal.Brain metastasis has unique mechanisms.Firstly,there are tight junctions between capillaries of endothelial cells in brain tissue,which form the blood-brain barrier with the peripheral astrocyte foot,and disruption of the tight junctions(ocludin,claudins,ZO-1,ZO-2,ZO-3,etc.)of the blood-brain barrier can promote brain metastasis;secondly,tumour cells can inhibit the production of fibrin by astrocytes to initiate metastasis.Finally,astrocytes and microglia,by regulating the secretion of various factors,can promote brain metastasis.Therefore,the analysis of brain metastasis mechanisms of lung cancer at the cellular-molecular as well as genetic levels may be more conducive to the prediction of brain metastasis.There is no definitive method for predicting brain metastasis from lung cancer.There are specific alterations in primary lung cancer foci that lead to brain metastasis,including tumour-initiating metastatic cells and TME alterations,as well as specific genes and pathway alterations that predispose to brain metastasis.Genetic information related to brain metastasis from lung cancer may be found by high-throughput sequencing of primary foci.Medical imaging can detect microscopic tumour alterations and assess tumour heterogeneity.However,due to the small proportion and number of metastatic cells in the primary focus,it seems difficult to predict brain metastases from primary focus imaging features alone,whereas clinical features may carry relevant information,so a model combining genetic,imaging and clinical features is needed to solve the prediction problem.Methods:1.Through literature screening,we identified genetic alterations in the primary foci that are prone to brain metastasis and divided the downloaded TCGALUAD patients into a characteristic gene positive group and a characteristic gene negative group,and then obtained Bulk RNA-seq results from both groups,which were combined with single cell sequencing data from the GEO database by the Scissor algorithm to finally classify the primary foci tumour cells into those that are prone to brain metastasis(Scissor-c+)and those that are not prone to brain metastasis(Scissor＿c-).The Scissor＿c+ cell population was compared with the Scissor＿c-cell population to find the differential genes,which were analysed for GO,KEGG,GSEA and GSVA functional enrichment.Mesenchymal cells in the primary foci of TME were also grouped by the Scissor algorithm into those that were prone to promote brain metastasis(Scissor+)and those that were not prone to promote brain metastasis(Scissor-).GO and KEGG functional enrichment analyses were performed for up-and down-regulated genes in two subpopulations of mesenchymal cells including B cells,T cells,myeloid cells,endothelial cells,fibroblasts and epithelial cells,respectively.Further,intercellular communication analysis was performed by Nichenet on Scissor+mesenchymal cells and Scissor-c+ cancer cells in TME to find brain metastasisassociated target genes.Finally,the cell trajectory relationship between the primary foci of brain metastasis-prone cell populations and cancer cell populations of brain metastasis was determined by the pseudotime analysis.2.246 patients who attended Shengjing Hospital of China Medical University with pathological diagnosis of lung adenocarcinoma and all of them had completed Bulk DNA gene testing were collected in this study,and whether brain metastasis occurred in the patients was clarified through long-term follow-up.3.The study also collected imaging data and general clinical data(including gender,age,smoking history,stage and tumour markers)from 178 patients with a pathological diagnosis of lung adenocarcinoma and EGFR mutation who attended Shengjing Hospital of China Medical University and the First Affiliated Hospital of China Medical University.The prediction models were constructed based on imaging features,clinical features and imaging combined clinical features,including random forest,support vector machine,logistic regression and Light GBM,and then the prediction effect of the different models was evaluated;then deep learning prediction models were constructed based on single and multiple images,including Resnet18+Ranger21 model,Resnet34+Ranger21 model and Resnet50+Ranger21model,were then used to evaluate the prediction effect of the different deep learning network models.4.The clinical data,imaging data and genetic information of 87 patients out of 178 patients for whom genetic testing data of lung primary foci were available were counted.The above four types of radiomics prediction models were constructed based on image and gene features,clinical and gene features,and radiogenomics and clinical multimodal fusion,respectively,so as to evaluate the prediction effects of different models.Results:1.In this study,eight landmark genes,namely EGFR mutation,ALK mutation,MYC amplification,TERT amplification,SPOCK1 mutation,KIFC1 mutation,CDH2 mutation and CDKN2 A mutation,were identified as the genetic alteration characteristics of primary foci prone to brain metastasis through literature screening,and bioinformatics was used to fuse the TCGA-LUAD Bulk RNA sequencing data with single-cell sequencing data from the GEO database(GSE131907)revealed that(1)there were 399 differential genes in the two groups of primary foci prone to brain metastasis and less prone to brain metastasis,including 115 up-regulated genes and 151down-regulated genes;(2)GO analysis showed enrichment of differential genes in cell tropism,local adhesions,cell-matrix junctions,and calnexin binding;(3)KEGG analysis showed that the differential genes were mainly enriched in glucose metabolism,lipid metabolism,protein metabolism and glycolysis pathways;(4)GSEA analysis showed that the differential genes were mainly enriched in NF-κB signaling pathway;(5)GSVA analysis showed that the gene sets of brain metastasis-prone cancer cells were mainly enriched in Wnt＿β＿catenin,TNFα＿NFκB,IL6＿JAK＿STAT3,PI3K＿AKT＿m TOR,MYC and the APICAL＿SURFACE signaling pathway associated with EMT;(6)Different mesenchymal cells in the primary foci were divided into metastasis-prone and non-metastasis-prone mesenchymal cell populations,and GO and KEGG analyses showed that mesenchymal differential genes were mainly enriched in ribosomes and their subunits,cytoplasmic translation;(7)The results of intercellular communication analysis showed that the ligand ITGAM in the primary foci of myeloid cells binds to the cancer cell receptor ICAM1/PLAUR and activates the target gene PTGS2,which may be a potential mechanism to promote the development of brain metastasis;(8)The results of the pseudotime analysis showed that the cancer cells in the primary foci that are prone to brain metastasis and the cancer cells in the brain metastasis foci have cell trajectory relationship;2.Analysis of clinical data from the246 patients included showed that(1)patients with EGFR mutation were prone to brain metastases(P=0.035);(2)patients aged ≤60 years were prone to brain metastases(P=0.042);(3)EGFR accompanied by other driver genes(including KRAS/BRAF/ERBB2/c-MET/ PIK3CA/AKT1)may tend to develop brain metastases(P=0.05);subgroup analysis of 189 patients with NGS in the primary lung foci showed that(1)patients with EGFR mutation were prone to develop brain metastases(P=0.015);(2)patients with EGFR alterations in the above six driver genes were prone to brain metastases(P=0.04),further propensity matching score comparion studies with large sample data were still needed to clarify that patients with EGFR concomitant with specific gene alterations were more prone to brain metastases and help stratify the risk of brain metastases in patients with EGFR mutation;The median time to develop brain metastasis in patients with EGFR mutation was 17.6m;except for EGFR,there was no statistical difference between the two groups with/without brain metastasis for other driver genes and clinical features alterations in this study,which needs to be validated by large sample data;3.Predictive model of radiomics was constructed for 178 patients by LASSO on 1334 imaging features and 10 clinical features were downscaled,and the results showed that(1)it seemed difficult to predict brain metastasis based on 11 imaging features of the primary focus,among which the logistic regression model worked best with AUC=0.58;(2)the prediction ability of the model was improved based on three clinical features,among which the Light GBM model had the best prediction effect with AUC=0.74;three clinical features were screened out including NSE,extrathoracic metastasis,N stage(N0/N3),while NSE,extrathoracic metastasis and N3 were associated with the occurrence of brain metastasis,with NSE and extrathoracic metastases were more influential,the higher the NSE,the higher the risk of brain metastases,and N0 was associated with no brain metastases;statistical analysis of the clinical information of the 178 patients included revealed that the baseline NSE(P=0.019)and extrathoracic metastases(P=0.042)clinical characteristics of patients were closely related to the occurrence of brain metastases;(3)the prediction effect of the fusion model based on 18 imaging features and four clinical features(including NSE,extrathoracic metastasis,N stage(N0/N3)and smoking history)was better,among which the Light GBM model had the best prediction effect with AUC=0.84;4.The results of constructing prediction models based on deep learning showed(1)5696CTs obtained from a single image by data enhancement and poor prediction model based on Res Net18+Ranger21 with an average AUC=0.58;(2)incorporating 2 images by data enhancement for a total of 8256 CTs to build a model based on Res Net18+Ranger21 with a slightly improved prediction,mean AUC=0.60;(3)8256CTs to build a prediction model based on Res Net34+Ranger21 with mean AUC=0.60;(4)8256 CTs to build a prediction model based on Res Net50+Ranger21 with mean AUC=0.62,with a slightly improved prediction effect compared to Res Net18+Ranger21 and Res Net34+Ranger21;5.The results of the prediction model construction based on radiogenomics multimodality for 87 of these patients with genetic test results showed that after filtering features by LASSO dimensionality reduction,(1)based on 9 imaging features combined with 4 genetic features(ERBB2/KRAS/MET/PI3KCA),the random forest model was the best,with AUC=0.68;(2)prediction models based on 3 clinical features(NSE,extrathoracic metastases,N3)combined with 4 genetic features,the random forest model and support vector machine model were the best,with AUC=0.78;(3)prediction models based on7 imaging features,4 genetic features and two clinical features(NSE and extrathoracic metastasis),with the random forest model having the best prediction effect,AUC=0.85.Conclusion:1.The set of genes predisposing to brain metastasis in primary lung adenocarcinoma foci is mainly enriched in the NF-κB signaling pathway;2.PTGS2 gene in primary foci may act as a downstream gene of the pathway where EGFR and other driver genes are located,and is one of the potential mechanisms promoting the development of brain metastasis;3.Lung adenocarcinoma patients aged ≤60 years,with mutation in EGFR,or EGFR accompanied by KRAS/BRAF/ERBB2/cMET/PIK3CA/AKT1 mutations are prone to brain metastases,with the four genes KRAS/ERBB2/c-MET/PIK3 CA appearing to have a greater impact.4.Radiomics models based on imaging combined with clinical features have good predictive performance for brain metastasis in EGFR-positive lung adenocarcinoma patients;5.Deep learning models based on imaging features have poor predictive performance;6.Radiogenomics combined with clinical features multimodal fusion models have better predictive performance,which is expected to stratify the risk of brain metastasis for patients and thus provide more rational treatment decisions for patients.