This study introduces a multivariate integration of multi-omics data for risk stratification of patients with newly diagnosed, treatment-naïve IDH-wildtype GBM. We used information collected from several sources of data to build multiple models, each with different layers of prognostic information for disease stratification. Starting from the base model including the least available information for the patients with GBM, i.e., clinical variables of age, gender, and extent of resection, we showed incremental value of radiomics, MGMT methylation, and genomic data obtained by NGS sequencing of the tumor samples, resulting in a multi-omics model with superior performance.
Despite marked variation in treatment responsiveness and outcomes, nearly all patients with newly diagnosed GBM still receive the same standard radiation and temozolomide-based therapy, underscoring a lack of personalized treatment in GBM relative to many other solid tumors 37. Moreover, interpatient heterogenetiy in both tumor biology and clinical characteristics has made it difficult to accurately determine the efficacy of experimental treatments, particularly in newly diagnosed GBM trials with progression-free survival (PFS) or OS endpoints 38,39. Our findings suggest that multi-omics or multi-scale fusion of data can provide a comprehensive portrayal of tumor biology and the patient’s likely response to the treatment. If validated in a larger, multi-center study, this survival prediction tool has the potential for widespread clinical and research use, allowing for more personalized treatment in routine practice and more efficient clinical trial design in the newly diagnosed setting.
In the present study, the image-based predictive model of OS builds upon and reinforces prior studies 20,23 by employing an advanced computational methodology based on high-throughput data, i.e., conventional and deep radiomic features. We showed generalizability of our non-invasive prognostic approach to unseen data, evidenced by the classification performance of our ML models. An important strength of our proposed signature is that it is derived from pre-operative conventional MRI scans which are acquired as a part of standard-of-care diagnosis of patients with GBM. In both of our ML classifiers, i.e., high-risk and low-risk models, most of the selected features were among deep features derived from T2 and T2-FLAIR scans, emphasizing the significance of these features in quantifying higher-order patterns. GBM tumors exhibit spatial heterogeneity arising from variation in cellular density, vascularization, and necrosis across their area, and deep learning features enable detection of phenotypic characteristics within pre-operative images that may not be captured comprehensively by only conventional radiomic features.
The survival prediction index (SPIradiomics) was used to fuse with other omics data for risk stratification of patients with GBM. As the results suggest, SPIradiomics provided additive value over the clinical and MGMT data for risk assessment, improved the concordance index, and provided more distinctive separation of the low, medium, and high-risk groups. Genetics further improved risk stratification in synergy with other omics. Added value of a radiomics signature to clinical and MGMT methylation in prediction of OS and PFS in patients with GBM has been reported in only a few studies with smaller cohorts 22,40. The current study contributes to the improvement of the interpretability, predictive applicability and computational efficiency of existing machine learning algorithms, thereby increasing the precision of population-based registries and validate current and future prediction tools.
In studies including only genomics, genetic classification of IDH-wildtype GBM 15 has not yielded prognostic information, although a recent transcriptomic classification based on developmental and metabolic axes has dissected apart subclasses of GBM with prognostic significance and contrasting metabolic vulnerabilities 41. Our approach revealed two genes, RB1 (retinoblastoma tumor suppressor protein) and NOTCH2, previously studied for their role in oncogenesis and, and as potential targets for glioma therapy. RB1 was previously discovered to be a tumor suppressor gene, like TP53; inactivation of RB1 in preclinical models suggested it plays a significant role in carcinogenesis 42. RB1 is involved in the cell cycling pathway, allowing progression from G1 to S phase, activated by phosphorylation and regulated by CDK4/CK6 protein kinases 43. In a study of CDK4/6 inhibitors in xenografts, RB1 was identified as a determinant of therapeutic efficacy 44. Loss of RB1 transcript expression was seen in 10.6% of GBMs in the TCGA analysis, and these tumors were highly enriched for the proneural subtype; loss of RB1 transcript expression was not associated with a difference in overall survival on univariate analysis 44. To date, a specific inhibitor of the RB1 signaling pathway has not been developed; however, palbociclib, an inhibitor of the CD4/CDK protein kinase pathways, was reported to be ineffective in a phase 2 trial against recurrent, RB1 positive, recurrent GBM 43.
Notch signaling, discovered in Drosophila nearly a century ago, is known to be involved in numerous cellular processes, including signaling in solid tumors 45. Notch signaling has also been implicated in angiogenesis as well as cancer drug chemo-resistance and radioresistance of cancer stem cells; a role for Notch is well established in hematological malignancies, but in solid tumors, its role is highly contextual: an oncogene is some cancers but a tumor suppressor in others 45. Notch signaling occurs at the interface of the tumor-stromal microenvironment 46. NOTCH2 is a positive regulator of transcription 47; it is found as a somatic gene mutation in a small percentage (23%) of pleomorphic xanthoastrocytomas, along with other somatic mutations (BRAF V600E, NF1, etc.), and could be involved in the pathogenesis of PXAs 47 or in anaplastic astrocytomas, where it was found in 31% of these tumors 48.
This paper presents several contributions. First, it presents a comprehensive study of prognosis of OS leveraging machine learning-based integration of clinical, imaging, and genomic data drawing from a cohort totaling 516 patients, including an independent replication dataset. To our knowledge, this is the first study of this size to explore the synergistic value of genetic, imaging and clinical predictive factors. Second, it demonstrated that a combination of deep learning and conventional radiomics produces a strong predictive panel of radiomic features for OS using standard, routinely acquired MRI scans. Third, it demonstrates the relative value of each omics in the context of additive contribution of clinical, imaging, and genetic variables. Critically, proper integration of all these measures via machine learning produced high predictive value on an independent test cohort of 112 patients, thereby further bolstering our confidence in the reproducibility and clinical value of these emerging AI-based integrated precision diagnostic indices.
A limitation of our study is the single institutional data analysis. To ac hieve a generalizable method, it would be beneficial to explore data collected from multiple institutions. Furthermore, the sample size for risk stratification (n = 112) was a potential limitation, which was mainly due to the unavailability of all omics information for all patients in our cohort. In particular, as broader molecular datasets become available on more patients, the prognostic implications of the molecular changes in the context of other omics may become clearer. Future studies would benefit from multi-institutional, prospective, and larger study population, a goal which the ReSPOND consortium is aiming to achieve 49.
In conclusion, the present study fuses multiple omics data, namely clinical information, MGMT methylation, radiomics, and genetics, to accurately model clinical outcome in patients with newly diagnosed GBM. Accurate stratification of risk groups may facilitate improved patient management through personal optimization of treatment decisions, as well as effective risk stratification of patients for newly diagnosed GBM clinical trials.