Abstract 2397: Order parameter methods for interpretable comparisons across tumors and cell lines

Abstract As precision cancer medicine develops, new bioinformatic tools that are both more granular and interpretable become imperative. In particular, when working with cancer cell lines to identify and validate potential treatment targets in cohorts of interest it is necessary to understand whether and precisely how cancer cell lines recapitulate tumor biology. As part of the Cancer Dependency Map project we are working to develop approaches for relating cell lines to patient tumors. Here we demonstrate a simple supervised classifier inspired by order-parameter methods from spin-glass physics that uses a pre-determined set of tissue type labels to produce scores per-gene and per-type that can be evaluated on both tumors and cell lines. The model described here complements this previous work, producing scores that qualitatively agree with the unsupervised model, while also enabling a quantitative, gene-level understanding of the origin of the classification result. Further, this method allows multiple cohorts of tumors and cell lines to be compared directly, relative to a common basis. We describe both the method itself, a variation of which has been deployed in single-cell developmental biology, as well as the metadata processing necessary for its effective use in comparing models to tumors. We developed an algorithm that defines our scope of tissue types maximizing granularity while minimizing collinearity and accounting for gaps and overlaps in annotation between tumor and cell line datasets, with over 35 total disease types assessed. We show using several examples from cell lines that this simple method can identify not only which type of tumor each cell lines most closely resembles, but also extract the most important genes involved in that determination. Our model, using 75 samples per tissue type, achieves >95% top-3 accuracy on a test dataset of over 10k tumor samples. The matchup between annotated type and predicted type is weaker in cell lines and also highly type-specific, ranging from 95% in some cohorts of 20% in others, and we explore the reasons for this. In one example of lung cell line models our method identifies cell line cohorts that match up to adenoid as well as squamous tumors, and another cohort that doesn’t match with either set of tumors. Based on the genes identified by the model as important for classifying these lines we describe their phenotype as mesenchymal. Additionally, we show how this method can be augmented with normal tissue data from GTEx to foreground signals that are specific to cancer relative to related normal tissues. Citation Format: Lauren Golden, Naima Abdirahman, Xiaomeng Zhang, Maria Yampolskaya, Barbara DeKegel, Pankaj Mehta, Francisca Vazquez, Catarina D. Campbell. Order parameter methods for interpretable comparisons across tumors and cell lines [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 2397.