Discovery from single-cell RNA sequencing profiles of 18 CPTAC glioblastoma patients and validation in bulk profiles of 138 TCGA patients and two human-derived cell lines of a whole-transcriptome predictor of overall survival and drug targets by using quantum mechanics–based AI/ML.

3019 Background: Single-cell, more than bulk, multi-omic data, are small-cohort, noisy, and high-dimensional, and extremely difficult to model. We have developed artificial intelligence and machine learning (AI/ML) to overcome these challenges [doi: 10.1073/pnas.0530258100]. We have shown that our algorithms are uniquely able to discover accurate, precise, actionable, and mechanistically interpretable predictors, applicable to the general population, from the bulk multi-omes of as few as 19 patients [doi: 10.1158/1538-7445.AM2025-CT227]. We have demonstrated that these predictors of overall survival (OS) and drug targets consistently validate across laboratories and sometimes across indications, in federated and imbalanced studies and over time, and outperform all others where they exist [doi: 10.1063/1.5099268, 10.1200/JCO.2024.42.16_suppl.10043]. Here, we demonstrate our AI/ML in single-cell data. Methods: We used our algorithms to derive models from the single-cell RNA sequencing profiles of 18 glioblastoma (GBM) Clinical Proteomic Tumor Analysis Consortium (CPTAC) patients, and tested them in the bulk profiles of 138 patients in the Cancer Genome Atlas (TCGA). CPTAC and TCGA both profiled the core primary tumor of each patient, but CPTAC also sampled the peritumoral brain tissue. The two cohorts are indistinguishable in terms of their gender, age, and OS distributions, but they significantly differ in terms of race. Results: A whole-transcriptome model was discovered that is correlated with OS among the 18 CPTAC patients, with a Kaplan-Meier median OS difference of 30 months between the two groups of predictor-stratified patients, and a Cox hazard ratio of 4.6 and a concordance index of 87% (log-rank and Wald P-values < 5.0×10-2). When tested in the TCGA cohort, the model was a similarly significant predictor of OS. In both cohorts, despite the differences in profiling protocols and cohort demographics, the predictor outperformed the best standard-of-care indicator of OS in GBM, i.e., age. Consistent with a previous whole-genome predictor, which was derived from bulk DNA profiles, and validated in a clinical trial [doi: 10.1063/1.5142559], shorter OS was associated with overexpression of such anterior/posterior pattern specification genes as the Notch ligand DLL3 . We experimentally validated, in two human-derived cell lines, that its putative activator, METTL2A , is required for GBM cells' proliferation and viability [doi: 10.1158/1538-7445.AM2025-3686]. Conclusions: Our algorithms can discover predictors of patients’ OS and drug targets, in real-world small-cohort, noisy, and high-dimensional — single-cell — in addition to bulk multi-omic data, and the predictors experimentally validate.