Tumor Mutation Burden–Related Histopathologic Features for Predicting Overall Survival in Gliomas Using Graph Deep Learning

Tumor mutation burden (TMB) is a potential biomarker for evaluating the prognosis and response to immune checkpoint inhibitors, but its costly and time-consuming method of measurement limits its widespread application. This study aimed to identify the TMB-related histopathologic features from hematoxylin and eosin slides and explore their prognostic value in gliomas. TMB-related features were detected using a graph convolutional neural network from whole-slide images of patients from The Cancer Genome Atlas data set (619 patients), and the correlation between features and TMB was evaluated in an external validation set (237 patients). TMB-related features were used for predicting overall survival (OS) of patients to investigate whether these features have potential for prognostic prediction. Moreover, biological pathways underlying the prognostic value of the features were further explored. Histopathologic features derived from whole-slide images were significantly associated with patient TMB (P = 0.007 in the external validation set). TMB-related features showed excellent performance for OS prediction, and patients with lower-grade gliomas could be further stratified into different risk groups according to the features (P = 0.00013; hazard ratio, 4.004). Pathways involved in the cell cycle and execution of immune response were enriched in patients with higher OS risk. The TMB-related features could be used to estimate TMB and aid in prognostic risk stratification of patients with glioma with dysregulated biological pathways. Tumor mutation burden (TMB) is a potential biomarker for evaluating the prognosis and response to immune checkpoint inhibitors, but its costly and time-consuming method of measurement limits its widespread application. This study aimed to identify the TMB-related histopathologic features from hematoxylin and eosin slides and explore their prognostic value in gliomas. TMB-related features were detected using a graph convolutional neural network from whole-slide images of patients from The Cancer Genome Atlas data set (619 patients), and the correlation between features and TMB was evaluated in an external validation set (237 patients). TMB-related features were used for predicting overall survival (OS) of patients to investigate whether these features have potential for prognostic prediction. Moreover, biological pathways underlying the prognostic value of the features were further explored. Histopathologic features derived from whole-slide images were significantly associated with patient TMB (P = 0.007 in the external validation set). TMB-related features showed excellent performance for OS prediction, and patients with lower-grade gliomas could be further stratified into different risk groups according to the features (P = 0.00013; hazard ratio, 4.004). Pathways involved in the cell cycle and execution of immune response were enriched in patients with higher OS risk. The TMB-related features could be used to estimate TMB and aid in prognostic risk stratification of patients with glioma with dysregulated biological pathways. Gliomas are the most common type of primary malignant central nervous system cancer.1Miller K.D. Ostrom Q.T. Kruchko C. Patil N. Tihan T. Cioffi G. Fuchs H.E. Waite K.A. Jemal A. Siegel R.L. Barnholtz-Sloan J.S. Brain and other central nervous system tumor statistics, 2021.CA Cancer J Clin. 2021; 71: 381-406Crossref PubMed Scopus (319) Google Scholar Poor outcomes in patients with glioma are mainly affected by the tumor microenvironment (TME) and genetic heterogeneity.2Ceccarelli M. Barthel F.P. Malta T.M. Sabedot T.S. Salama S.R. Murray B.A. Morozova O. Newton Y. Radenbaugh A. Pagnotta S.M. Molecular profiling reveals biologically discrete subsets and pathways of progression in diffuse glioma.Cell. 2016; 164: 550-563Abstract Full Text Full Text PDF PubMed Scopus (1382) Google Scholar,3Weller M. Stupp R. Hegi M.E. Van Den Bent M. Tonn J.C. Sanson M. Wick W. Reifenberger G. Personalized care in neuro-oncology coming of age: why we need MGMT and 1p/19q testing for malignant glioma patients in clinical practice.Neuro Oncol. 2012; 14: iv100-iv108Crossref PubMed Scopus (172) Google Scholar The World Health Organization classification has undergone major restructuring since 2016, emphasizing that patients' treatment planning and prognosis assessment need to be combined with the histologic parameters and genetic features.4Louis D.N. Perry A. Reifenberger G. Von Deimling A. Figarella-Branger D. Cavenee W.K. Ohgaki H. Wiestler O.D. Kleihues P. Ellison D.W. The 2016 World Health Organization classification of tumors of the central nervous system: a summary.Acta Neuropathol. 2016; 131: 803-820Crossref PubMed Scopus (10718) Google Scholar,5Louis D.N. Perry A. Wesseling P. Brat D.J. Cree I.A. Figarella-Branger D. Hawkins C. Ng H.K. Pfister S.M. Reifenberger G. The 2021 WHO classification of tumors of the central nervous system: a summary.Neuro Oncol. 2021; 23: 1231-1251Crossref PubMed Scopus (3490) Google Scholar The combination of genetic features improves diagnostic objectivity and provides more precise treatment planning for patients with glioma, whereas traditional therapies have not significantly improved the overall survival (OS) of gliomas.6Kang K. Xie F. Wu Y. Wang Z. Wang L. Long J. Lian X. Zhang F. Comprehensive exploration of tumor mutational burden and immune infiltration in diffuse glioma.Int Immunopharmacol. 2021; 96107610Crossref Scopus (7) Google Scholar In recent years, immunotherapy has revolutionized treatment of many cancers. The success of immune checkpoint blockade, such as programmed cell death ligand 1 and cytotoxic T-lymphocyte–associated antigen 4, has shown that immunotherapies may also be effective in the treatment of gliomas.7Ahmed N. Brawley V. Hegde M. Bielamowicz K. Kalra M. Landi D. Robertson C. Gray T.L. Diouf O. Wakefield A. HER2-specific chimeric antigen receptor–modified virus-specific T cells for progressive glioblastoma: a phase 1 dose-escalation trial.JAMA Oncol. 2017; 3: 1094-1101Crossref PubMed Scopus (516) Google Scholar, 8Cui J. Zhang Q. Song Q. Wang H. Dmitriev P. Sun M.Y. Cao X. Wang Y. Guo L. Indig I.H. Targeting hypoxia downstream signaling protein, CAIX, for CAR T-cell therapy against glioblastoma.Neuro Oncol. 2019; 21: 1436-1446Crossref PubMed Scopus (46) Google Scholar, 9Xu S. Tang L. Li X. Fan F. Liu Z. Immunotherapy for glioma: current management and future application.Cancer Lett. 2020; 476: 1-12Crossref PubMed Scopus (301) Google Scholar Tumor mutation burden (TMB), a measure of the number of gene mutations occurring in the genome of cancer cells,10Fusco M.J. West H.J. Walko C.M. Tumor mutation burden and cancer treatment.JAMA Oncol. 2021; 7: 316Crossref PubMed Scopus (34) Google Scholar has been proven to be an effective biomarker for predicting the response to immunotherapy in lung cancer and melanoma.11Samstein R.M. Lee C.-H. Shoushtari A.N. Hellmann M.D. Shen R. Janjigian Y.Y. Barron D.A. Zehir A. Jordan E.J. Omuro A. Tumor mutational load predicts survival after immunotherapy across multiple cancer types.Nat Genet. 2019; 51: 202-206Crossref PubMed Scopus (2221) Google Scholar Recent studies have found that TMB is associated with response to immunotherapy and is negatively correlated with survival in diffuse glioma.12Wang L. Ge J. Lan Y. Shi Y. Luo Y. Tan Y. Liang M. Deng S. Zhang X. Wang W. Tumor mutational burden is associated with poor outcomes in diffuse glioma.BMC Cancer. 2020; 20: 1-12Crossref Scopus (9) Google Scholar, 13Capper D. Reifenberger G. French P.J. Schweizer L. Weller M. Toua M. Niclou S.P. Euskirchen P. Haberler C. Hegi M.E. Brandner S. Le Rhun E. Rudà R. Sanson M. Tabatabai G. Sahm F. Wen P.Y. Wesseling P. Preusser M. van den Bent M.J. EANO guideline on rational molecular testing of gliomas, glioneuronal and neuronal tumors in adults for targeted therapy selection.Neuro Oncol. 2023; 25: 813-826Crossref PubMed Scopus (13) Google Scholar, 14Brown M.C. Ashley D.M. Khasraw M. Low tumor mutational burden and immunotherapy in gliomas.Trends Cancer. 2022; 8: 345-346Abstract Full Text Full Text PDF PubMed Scopus (4) Google Scholar This is an encouraging sign that TMB may have potential as a prognostic biomarker in glioma. However, as the gold standard for the measurement of TMB, whole-exome sequencing (WES) cannot be widely employed in clinical practice because of its high cost and because it is time-consuming. There is a need for an alternative low-cost method to detect TMB and stratify patients using information related to TMB to optimize individualized treatment planning and help improve patient outcome benefits. Tumor evolution in gliomas correlates with changes in the TME, whereas tumor histopathology reflects the heterogeneous characteristics of the TME.15Wang Q. Hu B. Hu X. Kim H. Squatrito M. Scarpace L. DeCarvalho A.C. Lyu S. Li P. Li Y. Tumor evolution of glioma-intrinsic gene expression subtypes associates with immunological changes in the microenvironment.Cancer Cell. 2017; 32: 42-56Abstract Full Text Full Text PDF PubMed Scopus (952) Google Scholar There are studies demonstrating that the morphometric features extracted from whole-slide images (WSIs) are related to the DNA methylation of gliomas.16Zheng H. Momeni A. Cedoz P.-L. Vogel H. Gevaert O. Whole slide images reflect DNA methylation patterns of human tumors.NPJ Genomic Med. 2020; 5: 11Crossref PubMed Scopus (17) Google Scholar Quantitative features can quantify the morphometric changes and differences of cells, but the finite number and characterizing capabilities limit performance.17Liu Z. Wang S. Dong D. Wei J. Fang C. Zhou X. Sun K. Li L. Li B. Wang M. Tian J. The applications of radiomics in precision diagnosis and treatment of oncology: opportunities and challenges.Theranostics. 2019; 9: 1303-1322Crossref PubMed Scopus (483) Google Scholar Recent advances in artificial intelligence and digital pathology have made it possible to extract a variety of histopathologic features from WSI18Zhao L.-T. Liu Z.-Y. Xie W.-F. Shao L.-Z. Lu J. Tian J. Liu J.-G. What benefit can be obtained from magnetic resonance imaging diagnosis with artificial intelligence in prostate cancer compared with clinical assessments?.Mil Med Res. 2023; 10: 29Crossref PubMed Scopus (3) Google Scholar,19Lu M.Y. Chen T.Y. Williamson D.F.K. Zhao M. Shady M. Lipkova J. Mahmood F. AI-based pathology predicts origins for cancers of unknown primary.Nature. 2021; 594: 106-110Crossref PubMed Scopus (210) Google Scholar automatically. Most studies have employed weakly supervised deep learning strategies for the analysis of WSIs. Concretely, segment a whole WSI into small patches and consider each as an independent instance, and then use a pooling operation to obtain the predictive information at the WSI and patient level. Convolutional neural network (CNN),20Sun C. Li B. Wei G. Qiu W. Li D. Li X. Liu X. Wei W. Wang S. Liu Z. Tian J. Liang L. Deep learning with whole slide images can improve the prognostic risk stratification with stage III colorectal cancer.Comput Methods Programs Biomed. 2022; 221106914Crossref Scopus (12) Google Scholar,21Li B. Li F. Liu Z. Xu F. Ye G. Li W. Zhang Y. Zhu T. Shao L. Chen C. Sun C. Qiu B. Bu H. Wang K. Tian J. Deep learning with biopsy whole slide images for pretreatment prediction of pathological complete response to neoadjuvant chemotherapy in breast cancer：a multicenter study.Breast. 2022; 66: 183-190Abstract Full Text Full Text PDF PubMed Scopus (2) Google Scholar sparse autoencoder,22Liu X.-P. Jin X. Seyed Ahmadian S. Yang X. Tian S.-F. Cai Y.-X. Chawla K. Snijders A.M. Xia Y. van Diest P.J. Weiss W.A. Mao J.-H. Li Z.-Q. Vogel H. Chang H. Clinical significance and molecular annotation of cellular morphometric subtypes in lower-grade gliomas discovered by machine learning.Neuro Oncol. 2023; 25: 68-81Crossref PubMed Scopus (13) Google Scholar recurrent neural network,23Campanella G. Hanna M.G. Geneslaw L. Miraflor A. Werneck Krauss Silva V. Busam K.J. Brogi E. Reuter V.E. Klimstra D.S. Fuchs T.J. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images.Nat Med. 2019; 25: 1301-1309Crossref PubMed Scopus (1083) Google Scholar and other neural networks are widely used in intelligent analysis of WSIs. The convolution operation in CNN has distinct strengths in capturing local and spatial features of images, which is favorable for detecting the morphologic changes of the lymphocytes and the tumor cells.24Lazard T. Bataillon G. Naylor P. Popova T. Bidard F.-C. Stoppa-Lyonnet D. Stern M.-H. Decencière E. Walter T. Vincent-Salomon A. Deep learning identifies morphological patterns of homologous recombination deficiency in luminal breast cancers from whole slide images.Cell Reports Med. 2022; 3100872Abstract Full Text Full Text PDF PubMed Scopus (9) Google Scholar Thus, CNN-based methods have shown exemplary performance in glioma classification,25Barker J. Hoogi A. Depeursinge A. Rubin D.L. Automated classification of brain tumor type in whole-slide digital pathology images using local representative tiles.Med Image Anal. 2016; 30: 60-71Crossref PubMed Scopus (154) Google Scholar molecular classification,26Jin L. Shi F. Chun Q. Chen H. Ma Y. Wu S. Hameed N.U.F. Mei C. Lu J. Zhang J. Aibaidula A. Shen D. Wu J. Artificial intelligence neuropathologist for glioma classification using deep learning on hematoxylin and eosin stained slide images and molecular markers.Neuro Oncol. 2021; 23: 44-52Crossref PubMed Scopus (49) Google Scholar and prognostic assessment.27Mobadersany P. Yousefi S. Amgad M. Gutman D.A. Barnholtz-Sloan J.S. Velázquez Vega J.E. Brat D.J. Cooper L.A.D. Predicting cancer outcomes from histology and genomics using convolutional networks.Proc Natl Acad Sci U S A. 2018; 115: E2970-E2979Crossref PubMed Scopus (538) Google Scholar Moreover, several studies have also verified the feasibility of evaluating TMB of cancer by analyzing pathologic images using CNN-based methods.28Jain M.S. Massoud T.F. Predicting tumour mutational burden from histopathological images using multiscale deep learning.Nat Mach Intell. 2020; 2: 356-362Crossref Scopus (43) Google Scholar, 29Shimada Y. Okuda S. Watanabe Y. Tajima Y. Nagahashi M. Ichikawa H. Nakano M. Sakata J. Takii Y. Kawasaki T. Histopathological characteristics and artificial intelligence for predicting tumor mutational burden-high colorectal cancer.J Gastroenterol. 2021; 56: 547-559Crossref PubMed Scopus (18) Google Scholar, 30Liu X. Liu Z. Yan Y. Wang K. Wang A. Ye X. Wang L. Wei W. Li B. Sun C. He W. Zhu X. Liu Z. Liu J. Lu J. Tian J. Development of prognostic biomarkers by TMB-guided WSI analysis: a two-step approach.IEEE J Biomed Heal Informatics. 2023; 27: 1780-1789Crossref Scopus (4) Google Scholar However, it is difficult for CNN to distinguish whether immune cells are tumor-infiltrating lymphocytes or an adjacent inflammatory response because this depends on lymphocytes interacting with tumor or stromal cells.31Saltz J. Gupta R. Hou L. Kurc T. Singh P. Nguyen V. Samaras D. Shroyer K.R. Zhao T. Batiste R. Spatial organization and molecular correlation of tumor-infiltrating lymphocytes using deep learning on pathology images.Cell Rep. 2018; 23: 181-193Abstract Full Text Full Text PDF PubMed Scopus (547) Google Scholar,32Diao J.A. Wang J.K. Chui W.F. Mountain V. Gullapally S.C. Srinivasan R. Mitchell R.N. Glass B. Hoffman S. Rao S.K. Human-interpretable image features derived from densely mapped cancer pathology slides predict diverse molecular phenotypes.Nat Commun. 2021; 12: 1613Crossref PubMed Scopus (77) Google Scholar Therefore, a new network structure is needed that can simultaneously detect the local features of cell morphology and the contextual features of the TME. Recent studies have suggested that graph convolutional neural networks (GCNNs) may be an appropriate approach to overcome this challenge.33Chen R.J. Lu M.Y. Shaban M. Chen C. Williamson D.F.K. Mahmood F. Whole slide images are 2D point clouds: context-aware survival prediction using patch-based graph convolutional networks. Medical Image Computing and Computer Assisted Intervention–MICCAI. Lecture Notes in Computer Science. LNCS 12908..in: de Bruijne M. Cattin P.C. Cotin S. Padoy N. Speidel S. Zheng Y. Essert C. Cham. Springer, Switzerland2021: 339-349Google Scholar GCNNs use graph structures to construct topological connections between small images and their neighbors, which enables message passing between images and extracting contextual features. For some studies that cannot be directly evaluated on the basis of local morphologic features, such as genetic mutation34Ding K. Zhou M. Wang H. Zhang S. Metaxas D.N. Spatially aware graph neural networks and cross-level molecular profile prediction in colon cancer histopathology: a retrospective multi-cohort study.Lancet Digit Heal. 2022; 4: e787-e795Abstract Full Text Full Text PDF PubMed Scopus (12) Google Scholar and survival prediction,35Lee Y. Park J.H. Oh S. Shin K. Sun J. Jung M. Lee C. Kim H. Chung J.-H. Moon K.C. Kwon S. Derivation of prognostic contextual histopathological features from whole-slide images of tumours via graph deep learning.Nat Biomed Eng. 2022; ([Epub ahead of print])https://doi.org/10.1038/s41551-022-00923-0Crossref Scopus (15) Google Scholar GCNNs exhibit better performance than CNNs. Previous studies only predicted high and low TMB; because there is no clinically significant optimal TMB threshold that could divide patients with glioma into high and low TMB subgroups, the prognostic risk stratification ability for patients is weak.28Jain M.S. Massoud T.F. Predicting tumour mutational burden from histopathological images using multiscale deep learning.Nat Mach Intell. 2020; 2: 356-362Crossref Scopus (43) Google Scholar,29Shimada Y. Okuda S. Watanabe Y. Tajima Y. Nagahashi M. Ichikawa H. Nakano M. Sakata J. Takii Y. Kawasaki T. Histopathological characteristics and artificial intelligence for predicting tumor mutational burden-high colorectal cancer.J Gastroenterol. 2021; 56: 547-559Crossref PubMed Scopus (18) Google Scholar Some studies attempted to classify the TMB of patients by a percentile or a specific 10 mutations per megabase,36Huang K. Lin B. Liu J. Liu Y. Li J. Tian G. Yang J. Predicting colorectal cancer tumor mutational burden from histopathological images and clinical information using multi-modal deep learning.Bioinformatics. 2022; 38: 5108-5115Crossref PubMed Scopus (8) Google Scholar,37He B. Dong D. She Y. Zhou C. Fang M. Zhu Y. Zhang H. Huang Z. Jiang T. Tian J. Chen C. Predicting response to immunotherapy in advanced non-small-cell lung cancer using tumor mutational burden radiomic biomarker.J Immunother Cancer. 2020; 8e000550Crossref Scopus (86) Google Scholar but the percentile was significantly affected by collection bias, and 10 mutations per megabase was not suitable for gliomas with overall low TMB values. Therefore, this study proposes a GCNN-based model to assess TMB in patients with glioma by analyzing WSIs. The aim of this study is to identify the histopathologic features that are significantly correlated with TMB. The extracted TMB-related features will then be used to stratify the OS risk of patients. In addition, RNA-sequencing data will be used to perform pathway analysis to understand the biological drivers underlying the different OS risks observed in patients with glioma. This retrospective study was approved by the ethics committee and waived the informed consent requirement. The design of this study consists of two parts: extracting TMB-related features from histologic images and evaluating the prognostic value of those features. The Cancer Genome Atlas (TCGA) data set was used to develop the association between histologic GCNN features and TMB. The inclusion criteria are summarized in Figure 1. In TCGA data set, WSI and exome sequencing data (level 2) were available, and the estimation of the TMB was collected from a previous study.12Wang L. Ge J. Lan Y. Shi Y. Luo Y. Tan Y. Liang M. Deng S. Zhang X. Wang W. Tumor mutational burden is associated with poor outcomes in diffuse glioma.BMC Cancer. 2020; 20: 1-12Crossref Scopus (9) Google Scholar The proprietary data set was used to externally validate the robustness of TMB-related features and their prognosis value, which was collected from Southwest Hospital and consisted of 237 patients diagnosed between 2015 and 2020. The mutational data of TCGA data set were obtained from cBioPortal.38Cerami E. Gao J. Dogrusoz U. Gross B.E. Sumer S.O. Aksoy B.A. Jacobsen A. Byrne C.J. Heuer M.L. Larsson E. The cBio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data.Cancer Discov. 2012; 2: 401-404Crossref PubMed Scopus (10589) Google Scholar,39Gao J. Aksoy B.A. Dogrusoz U. Dresdner G. Gross B. Sumer S.O. Sun Y. Jacobsen A. Sinha R. Larsson E. Integrative analysis of complex cancer genomics and clinical profiles using the cBioPortal.Sci Signal. 2013; 6: pl1Crossref PubMed Scopus (9776) Google Scholar The same approach was used in a recent study for the estimation of the TMB of TCGA set.40Chalmers Z.R. Connelly C.F. Fabrizio D. Gay L. Ali S.M. Ennis R. Schrock A. Campbell B. Shlien A. Chmielecki J. Analysis of 100,000 human cancer genomes reveals the landscape of tumor mutational burden.Genome Med. 2017; 9: 1-14Crossref PubMed Scopus (2130) Google Scholar For the external validation set, the DNA was extracted from the tissues according to the instructions of the QIAGEN DNA FFPE Extraction Kit GATK (Qiagen, Valencia, CA). Subsequently, the quality of DNA was accessed through spectrophotometry and absorbance, and quantification was performed using Qubit 3.0 (Thermo Fisher Scientific, Waltham, MA). To enrich the targeted DNA samples, 425 gene probes (Shihe Gene, Nanjing, China) were used. The target enrichment libraries were subjected to sequencing on the Nextseq CN500 platform (Illumina, San Diego, CA). Genomic data were processed using a proprietary bioinformatics platform and knowledge base, enabling the identification of various genomic abnormalities, including single base mutations and insertions/deletions, excluding driver mutations and tumor suppressor gene definite inactivating mutations, and calculating the number of somatic mutations per Mb. The TMB calculation was defined as the number of somatic mutations in the exon coding region per coding sequence length (0.946), where somatic mutations include synonymous mutations, nonsynonymous mutations, insertion and deletion mutations, and splice mutations. To mitigate the dimensional differences in TMB between the two data sets, the continuous TMB values were discretized into discrete variables using a quantile-based approach. Discretize variable into equal-sized buckets based on rank or based on sample quantiles. For example, 1000 values for 10 quantiles would produce a categorical object indicating quantile membership for each data point. The bin intervals obtained in the training set were subsequently used to discretize the TMB of the validation sets. The diagnostic slides were scanned with Pannoramic MIDI scanner (3DHISTECH, Budapest, Hungary) at ×20 magnification. For each slide, the CLAM toolbox was first used to automatically segment the tissue area.41Lu M.Y. Williamson D.F.K. Chen T.Y. Chen R.J. Barbieri M. Mahmood F. Data-efficient and weakly supervised computational pathology on whole-slide images.Nat Biomed Eng. 2021; 5: 555-570Crossref PubMed Scopus (408) Google Scholar More detailed steps, shown in Figure 2A, included the following: i) converting a down-sampled WSI to another color space (hue, saturation, value); ii) separating hematoxylin and eosin–stained tissue from the background using Otsu binarization on the saturation channel; and iii) patching 256 × 256 nonoverlapping images into a truncated ResNet-50 model (https://download.pytorch.org/models/resnet50-0676ba61.pth) trained on ImageNet (https://image-net.org). After the third residual block, the model extracts a 1024-dimensional feature vector h∈R1024. GCNN is designed to analyze graph-structured data using neural networks. A graph consists of two elements: node and edge. In this study, WSI is represented as graph-structured data in euclidean space, a patch is considered a node in the graph, and if the spatial coordinates of two patches are adjacent, one edge is constructed between the two nodes (Figure 2B). GCNN could directly analyze WSI-level information, avoiding information loss caused by pooling operations. For the construction of the WSI graph, the feature vector is packed into a node feature matrix Xi∈Rm×1024 from Mj total patches in Wj, where {Wj}j=1K∈P represents the set of all WSIs for patient P. The coordinate position of each patch in the Wj is saved for the construction of the adjacency matrix Aj via K-nearest neighbor (K = 8). After building a subgraph Gj=(Xj,Aj), a patient-level graph over all WSIs is constructed as G={Gj}j=1, named WSI graph. Finally, a WSI graph was generated containing all of the WSI information based on the patch topology that was used as the input to GCNN. The model iteratively aggregates node features in neighborhoods of different hidden layers through message passing. A graph convolution operation could be formulated as follows:Gl+1=Update(Aggregate(Gl,Wlagg),Wlupdate)+Gl(1) Where Gl and Gl+1 are the input and output graphs at the l-th layer, Wlagg and Wlupdate are weights of the aggregation and update functions. GCNN was implemented with two hidden layers. Similar to DenseNet42Huang G. Liu Z. Van Der Maaten L. Weinberger K.Q. Densely connected convolutional networks. IEEE, 2017: 4700-4708Google Scholar for CNNs, the output of each GCNN layer is passed to the last hidden layer of the GCNN for addition, to combine the patch features with the learned context features. The Aggregate and Update functions are applied from DeepGCN43Li G. Muller M. Thabet A. Ghanem B. Deepgcns: can gcns go as deep as cnns? IEEE, 2019: 9267-9276Google Scholar:Aggregate=p(l)({mvu(l):u∈N(v)})=∑u∈N(v)exp(βmvu(l))∑u∈N(v)exp(βmvu(l))·mvu(l)(2) Update=ζ(l)(hv(l),mv(l))=MLP(hv(l)+mv(l))→hv(l+1)(3) hv(l) represents the node feature of vertex v, hu(l) is the feature of neighboring node u∈N(v). mv(l) and mvu(l) represent the message passing. Furthermore, a global attention pooling operation is introduced to adaptively calculate the weighted sum of all graph node features, thus extending the aggregation function (2) calculation to all nodes in the graph.44Veličković P. Cucurull G. Casanova A. Romero A. Lio P. Bengio Y. Graph attention networks.arXiv. 2018; ([Preprint])https://doi.org/10.48550/arXiv.1710.10903Crossref Google Scholar The details of the GCNN model structure are shown in Supplemental Figure S1. The last fully connected layer of GCNN was considered as the extracted features, and the output layer was the predicted TMB obtained by performing nonlinear operations on the features. Label distribution smoothing was employed to increase the applicability of model across multiple data sets.45Yang Y. Zha K. Chen Y. Wang H. Katabi D. Delving into deep imbalanced regression. International Conference on Machine Learning.PMLR. 2021; (pp. 11842–11851)Google Scholar To train the model, Adam optimization was used with a default learning rate of 1 × 10−3, weight decay of 1 × 10−3, and focal mean squared error–based loss function, and trained for 100 epochs with early stopping. All computation is performed on two NVIDIA TITAN RTX graphics processing units (Nvidia, Santa Clara, CA) with batch size 1 and 16-step gradient accumulation. Two GCNN models were constructed to extract the pathologic features from the WSI at × 20 and ×10 magnifications. After obtaining two feature vectors using GCNN models with WSI graphs at two magnifications, a random forest regression model was trained to integrate the multiscale prediction information to obtain the final TMB prediction. The correlations between features and TMB were evaluated by calculating the Spearman correlation coefficient between predicted and true TMB. Previous studies have demonstrated that TMB is significantly associated with poor prognosis in lower-grade gliomas (LGGs), except for glioblastomas.12Wang L. Ge J. Lan Y. Shi Y. Luo Y. Tan Y. Liang M. Deng S. Zhang X. Wang W. Tumor mutational burden is associated with poor outcomes in diffuse glioma.BMC Cancer. 2020; 20: 1-12Crossref Scopus (9) Google Scholar,46Ding H. Zhao J. Zhang Y. Wang G. Cai S. Qiu F. Tumor mutational burden and prognosis across pan-cancers.Ann Oncol. 2018; 29: viii16-viii17Abstract Full Text Full Text PDF Scopus (0) Google Scholar The hypothesis of this study is that if TMB-related histopathologic features are valuable predictors of TMB, these features may also have prognostic value in LGGs. Accordingly, to evaluate whether TMB-related features had prognostic value, among the 2048-dimensional TMB-related features extracted by GCNN (connecting ×10 and ×20 features), least absolute shrinkage and selection operator with Cox proportional hazard regression was used to predict OS in patients with LGGs. After constructing the prediction model, the patients in the training set were classified into high and low OS risk by Youden index. The cutoff was subsequently applied in the internal and external validation sets. The Kaplan-Meier analysis and log-rank test were used to assess whether OS of the high and low risk had significant differences. The OS risk of patients with LGGs could be obtained by calculating the sparse weighted sum of TMB-related features. Different risks are determined by different expression levels of TMB-related features, which may have different biological meanings. To explore differences in biological pathways enriched by differentially expressed genes (DEGs) across OS risk groups, patients with both survival and RNA-sequenc