Abstract 5020: Quantum algorithms improve feature selection for multimodal cancer classification

Introduction: Cancer machine learning research is often limited by overparameterization and overfitting, which arise because cancer ‘omic’ variables significantly outnumber patient samples. Traditional feature selection methods for multimodal cancer data fail to completely capture complex biological relationships across genomic, transcriptomic, and pathomic data due in part to computational intractability—interaction information for 3rd-order and higher terms scale exponentially, infeasible for classical methods. We developed a novel quantum-classical hybrid algorithm to identify small, informative feature sets that incorporate higher-order correlations, potentially revealing previously undetected biological signals to improve cancer classification. Methods: We implemented a quantum-classical hybrid algorithm that frames feature selection as a polynomial constrained binary optimization problem (PCBO-Tournament). Our approach leverages quantum resources through the Recursive Quantum Approximate Optimization Algorithm to explore higher-order feature correlations. We evaluated the algorithm on The Cancer Genome Atlas (TCGA) pan-squamous cell carcinoma cohort (lung, cervical, and head and neck) (1,427 patients) with DNA mutational data, mRNA expression data, and pathomics data for tissue-of-origin classification. To reduce noise and enable efficient data loading, we binarized DNA single-nucleotide polymorphisms (2,720 loci), discretized normalized mRNA expression values into quantiles (18,879 genes), and processed pathomics data with the Slideflow platform, using the Prov-GigaPath foundation model to generate slide-level features (2,048) that we then discretized into quartiles. Feature sets were assessed in terms of the balanced accuracy of logistic regression models in cancer sub-type classification. Results: On the TCGA pan-squamous dataset, our quantum-enhanced algorithm selected feature sets of size 14 that achieved higher balanced accuracy in cancer classification (97%, P<0.01), compared to both models containing all features and those with classically selected feature sets. The algorithm identified biologically relevant features, including tumor suppressor genes such as TP53, along expression of tissue-specific markers (SFTPA1/A2/B), developmental regulators (TBX4/5), and epigenetic factors (HIRIP3). The selected features demonstrated robust performance across data modalities, with a variant of our full-qubo PCBO outperforming classical approaches in the small feature set regime (N≤20). Conclusion: Our quantum-enhanced PCBO feature selection method identifies small, biologically meaningful feature sets that outperform classical approaches, particularly in terms of incorporating higher-order correlations. This work illustrates the potential for quantum computing to advance multimodal biomarker discovery and precision oncology. Citation Format: Siddhi Ramesh, Teague Tomesh, Ryan Robinett, Dmitry Karpeyev, Colin Campbell, Bharath Thotakura, Frederic Chong, Samantha Riesenfeld, Alexander Pearson. Quantum algorithms improve feature selection for multimodal cancer classification [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5020.