Quantum Mechanics-Based Multi-Tensor AI/ML As Predictor of Patients' Overall Survival, Gene Targets, and Drug Responses from Their Glioblastoma Tumors' Whole Genomes.

3020 Background: The drug failure rate has increased to ~95%, despite the growth in targeted therapies. As clinical trials demonstrated, a targeted gene alone does not predict whether patients have longer life expectancy in response to the drug. As studies with model organisms showed, the effect of the drug, and the mechanisms underlying it, depend on the entire multi-ome. But multi-omic data are small-cohort, noisy, and high-dimensional, i.e., extremely difficult to model. Methods: We have developed our artificial intelligence and machine learning (AI/ML) to overcome these challenges [doi: 10.1073/pnas.0530258100, 10.1158/1538-7445.AM2025-CT227]. We demonstrated our algorithms in the unsupervised modeling of, e.g., whole genomes of 85 astrocytoma patients. Mechanistic interpretation showed that the modeling blindly removed batch effects, separated normal demographic variations, and discovered a disease-specific genome-wide pattern of DNA copy-number alterations. This pattern was used to derive an actionable predictor of patients’ overall survival (OS) and gene targets to sensitize their tumors. We computationally validated both the predictor and the modeling in federated studies of mutually-exclusive sets of 59–251 patients. The modeling repeatedly discovered a representation of the predictor in every study, across astrocytoma grades II, III, and IV, i.e., glioblastoma (GBM), patients. We experimentally validated the predictor in a clinical trial of 79 GBM patients, initially retrospectively, and, in a four-year follow up, also prospectively [doi: 10.1063/1.5142559, 10.1145/3624062.3624078, 10.1200/JCO.2024.42.16_suppl.e14028]. In all the cohorts, the predictor, with 75–95% concordance with OS, was more accurate than all standard-of-care indicators. With 100% reproducibility among Complete Genomics, Illumina, and Ultima whole-genome sequencing, and > 99% when including Affymetrix and Agilent DNA microarrays, the predictor was also the most precise. Results: Here, we describe functional genomic experimental validation of both a predicted gene target and the predicted tumors’ responses to the targeting. Guide RNAs were designed and a lentiviral CRISPR-Cas9 all-in-one vector was utilized to knock out the modeling-predicted target METTL2A . Knockout validation at the protein level was performed using Western blot. Knockout in the patient-derived GBM cell lines U-87 MG and U-118 MG resulted in significantly attenuated cell viability and proliferation. The level of attenuation was significantly different between the cell lines, consistent with their whole genome-based predicted responses. Conclusions: Our quantum mechanics-based multi-tensor AI/ML solved the 75-year-old problem of correctly predicting — patients’ OS, drug responses, and gene targets — from their GBM tumors' whole genomes.