Unraveling Drug Response Heterogeneity and Personalizing Treatment in Cancer

National Institute Of Environmental Health Sciences

Investigators

Abstract

We developed PHENO-DEX, an experimental and machine-learning framework designed to characterize cancer cell response spectrums using dynamic molecular phenotypes. PHENO-DEX is specifically applied to dexamethasone (Dex) treatment, a widely used glucocorticoid receptor (GR) agonist that influences gene transcription. While Dex is commonly used in breast cancer treatment, patient responses vary, underscoring the need for predictive models. PHENO-DEX employs two core algorithms: (1) DSFMix (Anchang et al. 2022) which is a tree-based model that identifies key regulators and maps responsive and non-responsive cell trajectories in single-cell transcriptomic data from Dex-treated breast cancer cells. (2) PHENOSTAMP (Karacosta , Anchang et al. 2020), a neural network algorithm that generates a Dex-response reference map, capturing distinct cell states with varying sensitivities. This map enables projection of new cancer cell lines or clinical samples, revealing their heterogeneous responses to Dex. This work is an ongoing collaboration with Dr. Archer’s lab. Previously, we treated T47D A1-2 human breast cancer cells with Dex for 1, 2, 4, 8, and 18 hours. UMAP analysis identified nine distinct clusters that varied over time. To map cell trajectories, we used DSFMix (Anchang et al. 2022) to select key genes accounting for distribution shape variations and time. Two major trajectory trees emerged, representing response and non-response pathways. Comparing GR regulatory networks (GRNs) across these trajectories revealed distinct differences. The trajectories link pre-existing clusters (time 0) to later time points, illustrating Dex-induced cell state transitions. We next employed PHONESTAMP to generate a Dex response reference map in, which allowed us to infer the presence of nine distinct Dex response states in T47D A1-2. Within these states, R1-4 represent cell subtypes responding to Dex at different stages, ranging from early response (R1) to medium (R2 and R3) to late response (R4). Conversely, five Dex non-response states are labeled with N1-5, each displaying unique gene expression profiles separating them into different clusters, despite their shared lack of response to Dex. When dexamethasone (Dex) binds to the glucocorticoid receptor (GR), it initiates a cascade of molecular events that regulate various cellular functions. To ensure the validity of these phenotypes, we investigated the expression levels of the glucocorticoid receptor(GR), encoded by the NR3C1 gene, across the identified states. Interestingly, we found no significant pattern or differences in NR3C1 and NR3C2 expression levels among the states, suggesting that the observed states are not solely determined by GR expression levels, but rather by other unidentified factors. To further validate the reference map, we examined the expression levels of the SMARCA4 gene, known to interact with chromatin and influence the glucocorticoid (GC) response. Our analysis revealed a correlation between SMARCA4 expression and Dex response states, confirming that our reference map accurately captures the characteristics of Dex response states. Additionally, the CDH1 gene, which encodes the E-cadherin protein, exhibited distinct expression patterns between response and non-response groups. These patterns suggest that during Dex response, cells may exist in different epithelial and mesenchymal states. Collectively, these findings validate that our reference map accurately captures the characteristics of Dex response, independent of receptor levels. To further confirm the validity of our reference map, we assessed cell heterogeneity at the protein level. We identified differentially expressed genes (DEGs) between response and non-response cell states at the 18-hour time point. We selected two transcription factors, CHEK2 and LENG1, for analysis using fluorescence-activated cell sorting (FACS). This approach successfully identified multiple cell groups based on these transcription factors. These findings further demonstrate that there are distinct cell states with different characteristics, reinforcing the capability of our reference map to capture cell heterogeneity beyond receptor level differences. After validating the alignment of the PHENO-DEX reference map with established knowledge, we explored the application of this cell-line based model to other breast cancer cell lines using single-cell transcriptomics data. We analyzed a dataset containing single-cell RNA-seq data from 31 breast cancer cell lines available in GSE173634 including T47D (Luminal A), MCF7 (Luminal A), and HS578T (triple-negative) cell lines. All were sequenced without treatment. To predict the Dex response states in each cell line, we initially filtered the data using the gene signatures identified by our pipeline. PHENO-DEX then determined the UMAP coordinates for each cell, projecting them onto the reference map. Based on their location, cells were assigned to one of 9 Dex response states described above. Our analysis revealed that all cell lines exhibited multiple response states evidenced by different cell state densities around various cluster centers, with significant heterogeneity observed across all lines. Given that one of dexamethasone’s (Dex) cellular functions is to promote proliferation, we compared our predicted response states with actual growth rates following Dex treatment. Growth rate data, downloaded from GEO including growth rate data from testing of the T47D A-2 cell line in our lab, allowed us to correlate the ratio of response to non-response states with laboratory growth rates. The results revealed a strong correlation (0.88) of cell lines with a higher percentage of response to non-response states exhibiting significantly higher growth rates. Thus the more cells response to Dex, the more they proliferate or grow. This also aligns with the GO terms from our enrichment of hormonal response phenotypes in the response compared to the non-response trajectories. These findings underscore the accuracy and reliability of our algorithm when validated against real laboratory data. Overall, our results confirm that the model is broadly applicable to other breast cancer cell lines, providing a robust tool for Dex response prediction. We next projected thirty-one breast cancer (BC) cell lines and fifty-nine clinical breast samples onto the reference map, effectively revealing their cell state heterogeneity in response to Dex. For example, TNBC cell lines showed the most diversity in terms of response. In summary, PHENO-DEX generates a unique reference map to predict both cell line and clinical samples in terms of Dex response by leveraging two innovative methods. Firstly, it captures Dex response states by tracking dynamic changes in cells during Dex treatment. Secondly, rather than categorizing individual cells as response or non-response, it projects cells onto our reference map, inferring cell states from it. Consequently, cell states are fluid, representing a more nuanced characterization that better encapsulates the complexity of clinical samples. Looking ahead, we aim to construct multiple reference maps with diverse perturbations. Mapping samples onto these multiple reference maps will enable us to pinpoint precise cell states within the sample and offer more tailored.