A Bayesian Framework for Detecting Gene Expression Outliers in Individual Samples

Abstract

Objective: Many antineoplastics are designed to target upregulated genes, but quantifying upregulation in a single patient sample requires an appropriate set of samples for comparison. In cancer, the most natural comparison set is unaffected samples from the matching tissue, but there are often too few available unaffected samples to overcome high inter-sample variance. Moreover, some cancer samples have misidentified tissues or origin, or even composite-tissue phenotypes. Even if an appropriate comparison set can be identified, most differential expression tools are not designed to accommodate comparing to a single patient sample. Materials and Methods:We propose a Bayesian statistical framework for gene expression outlier detection in single samples. Our method uses all available data to produce a consensus background distribution for each gene of interest without requiring the researcher to manually select a comparison set. The consensus distribution can then be used to quantify over-and under-expression. Results:We demonstrate this method on both simulated and real gene expression data. We show that it can robustly quantify overexpression, even when the set of comparison samples lacks ideally matched tissues samples. Further, our results show that the method can identify appropriate comparison sets from samples of mixed lineage and rediscover numerous known gene-cancer expression patterns. Conclusions:This exploratory method is suitable for identifying expression outliers from comparative RNA-seq analysis for individual samples and Treehouse, a pediatric precision medicine group that leverages RNA-seq to identify potential therapeutic leads for patients, plans to explore this method for processing their pediatric cohort.

N-of-one differential gene expression without control samples using a deep generative model

“…For most cancers, more than 80% of samples end up in the expected tissue component. These results are comparable to (Vivian et al 2020). Here it is worth noting that we do not necessarily expect a perfect match to the tissue of origin because, in metastatic cancers, the tumor may be more similar to the tissue of origin.…”

“…Probabilities for TCGA normal samples lie between the two, which may be explained by previous findings suggesting that adjacent normal tissue carries cancer traits (Aran et al 2017). A similar analysis has previously been performed using a Bayesian approach (Vivian et al 2020), but the separation seems much clearer (Vivian et al 2020), although a direct comparison is difficult. When analyzing the quality of the integration of cancer samples into latent space, we observe that most representations of cancer samples are consistent with the tissue of origin.…”

“…For an assessment of the model’s capability to assign independent samples to the correct tissue clusters according to the training data, we selected 14 cancers from TCGA that fulfilled the following two conditions: 1) the tissue corresponding to the cancer is present in GTEx, and 2) the tissue must have at least ten adjacent normal samples from each cancer type (Zeng et al 2019). We also included the Adrenal and Brain tumor samples, even though they lack adjacent normal samples, to compare our result with the work of Vivian et al (Vivian et al 2020). This results in 6111 TCGA tumor samples and 624 TCGA adjacent normal samples (Supplementary Table S2).…”

“…We also included the Adrenal and Brain tumor samples, even though they lack adjacent normal samples, to compare our result with the work of Vivian et al (Vivian et al 2020). This results in 6111 TCGA tumor samples and 624 TCGA adjacent normal samples (Supplementary Table S2).…”

“…Lastly, a general problem concerning bulk sequencing data is that samples differ in cell type composition. This problem may in part be alleviated by using weighted averages of the closest normal samples (Rapin et al 2014; Vivian et al 2020).…”