Abstract The understanding of the proteomic landscape of IDH mutant astrocytoma has been limited in previous studies. To address this knowledge gap, we developed an innovative approach using artificial intelligence (AI) to generate proteomic profiles of tumors based on their transcriptome data. Particularly, we employed a combination of Variational Auto-Encoder and Contrastive Language Image Pretraining techniques to develop an AI model. The model was trained on publicly available databases containing matched RNA and protein data. Leveraging this extensive dataset, we applied the trained model to infer proteomic profiles for over one thousand adult IDH mutant astrocytoma specimens using their transcriptome sequencing data. This novel approach allowed us to explore and identify previously undiscovered subtypes based on the predicted proteomic profiles. As a result, we identified seven distinct clusters: IDHmut-noncodel-Proliferative, IDHmut-noncodel-Mesenchymal, IDHmut-noncodel-Mitochondrial, IDHmut-NFkB-signal, IDHmut-codel-Proliferative, IDHmut-Metabolism, and IDHmut-Neuronal. Of particular clinical significance, the IDHmut-noncodel-Proliferative and IDHmut-noncodel-Immune-Mesenchymal subtypes displayed the worst prognoses, while the IDHmut-Neuronal, IDHmut-NFkB-signal, and IDHmut-Metabolism subtypes exhibited relatively better survival outcomes. Remarkably, within the IDHmut-noncodel-Proliferative cluster, we observed a high prevalence of CDKN2A homozygous deletion, providing a potential explanation for its poor survival prognosis. Conversely, the IDHmut-noncodel-Mesenchymal cluster did not demonstrate any significant genetic events but exhibited elevated expression of mesenchymal marker genes. Furthermore, this cluster displayed notable characteristics, including the presence of gemistocytic differentiated tumor cells and infiltration of activated CD8 T lymphocytes. Moreover, we encompassed an evolutionary analysis of 189 initial-recurrent IDH mutant astrocytoma pairs, and discovered that gemistocytic differentiation represents an early event in tumor progression. In summary, we introduced a novel method utilizing an AI model to generate proteomic profiles, thus enabling the discovery of previously unknown subtypes of IDH mutant glioma with clinical relevance. These findings contribute to a deeper understanding of the molecular landscape of IDH mutant astrocytoma, potentially leading to improved prognostication and personalized treatment strategies.
Read full abstract