Quantum mechanics–based multi-tensor AI/ML discovery and validation of actionable and mechanistically interpretable whole-transcriptome predictors of survival in response to immunotherapy from real-world clinical trial data.

Abstract

2621 Background: Prediction in cancer remains limited, and 90% of drugs continue to fail trials and post-market validation. The entire multi-ome affects the disease. Previously, we developed quantum mechanics-based multi-tensor AI/ML to overcome the limitations of typical AI/ML, e.g., neural networks and deep learning, in small-cohort, noisy, high-dimensional, multi-omic clinical data [doi: 10.1073/pnas.0530258100 , 10.1145/3624062.3624078 ]. We have demonstrated the algorithms in the discovery and validation of whole-genome and -chromosome predictors of survival and response to treatment in, e.g., brain, lung, ovarian, and uterine cancers [doi: 10.1063/1.5142559 , 10.1200/JCO.2024.42.16_suppl.10043 ]. Methods: Here, we use the algorithms to discover two whole-transcriptome predictors of OS in response to atezolizumab PD-L1 inhibitor immunotherapy in a 348-patient, multi-center, single-arm, bladder cancer clinical trial, and validate the predictors in the 401-patient bladder cancer cohort in the Cancer Genome Atlas (TCGA). Results: The algorithms discovered the two predictors in the open-source, pre-atezolizumab, locally advanced or metastatic disease profiles of the 348 patients alone. By incorporating the patient labels, both predictors were found to outperform the best indicator of response to the treatment to date, i.e., the tumor mutation burden (TMB): The Cox proportional hazards model ratios, i.e., the corresponding relative risks, of 2.7 and 1.7, and concordance indices, i.e., accuracies, of 0.70 and 0.63, of each predictor, are greater than the ratio, of 1.3, and index, of 0.61, of TMB (Wald P -values=2.0×10 -7 and 8.2×10 -3 vs. 2.1×10 -5 ). The maximum Kaplan-Meier median OS difference of the two predictors together, of 22 months, is greater than that of TMB, of 16 months (log-rank P -values=3.6×10 -11 vs. 1.7×10 -4 ). One predictor is additionally correlated with the objective response rate (ORR), and the other – with the tissue of advanced or metastatic disease (Kruskal-Wallis P -values=9.8×10 -4 and 2.2×10 -3 ). Both predictors are similarly correlated with the OS of the 401 TCGA bladder cancer patients. Both are statistically independent of the imbalanced variations in the patient demographics, e.g., race, or the tissue batches, e.g., pre- vs. post platinum-based chemotherapy. By using the transcript labels, the predictors were interpreted in terms of known and new disease mechanisms and drug targets to sensitize the tumors to the treatment. Conclusions: Our multi-tensor AI/ML discovered and validated two whole-transcriptome predictors of OS in response to atezolizumab that outperform TMB. This further suggests that quantum mechanics-based algorithms can be used to derive predictors that are consistent across studies and over time.