Neural Network Deconvolution Method for Resolving Pathway-Level Progression of Tumor Clonal Expression Programs With Application to Breast Cancer Brain Metastases

Abstract

Metastasis is the primary mechanism by which cancer results in mortality and there are currently no reliable treatment options once it occurs, making the metastatic process a critical target for new diagnostics and therapeutics. Treating metastasis before it appears is challenging, however, in part because metastases may be quite distinct genomically from the primary tumors from which they presumably emerged. Phylogenetic studies of cancer development have suggested that changes in tumor genomics over stages of progression often result from shifts in the abundance of clonal cellular populations, as late stages of progression may derive from or select for clonal populations rare in the primary tumor. The present study develops computational methods to infer clonal heterogeneity and dynamics across progression stages via deconvolution and clonal phylogeny reconstruction of pathway-level expression signatures in order to reconstruct how these processes might influence average changes in genomic signatures over progression. We show, via application to a study of gene expression in a collection of matched breast primary tumor and metastatic samples, that the method can infer coarse-grained substructure and stromal infiltration across the metastatic transition. The results suggest that genomic changes observed in metastasis, such as gain of the ErbB signaling pathway, are likely caused by early events in clonal evolution followed by expansion of minor clonal populations in metastasis, a finding that may have translational implications for early detection or prevention of metastasis.

Keywords: brain metastases; breast cancer; deconvolution; gene modules; matrix factorization; pathways; phylogenetics; transcriptome.

Figure 1

The pipeline of BrM phylogenetics using matched bulk transcriptome.

Figure 2

Method details. (A) Neural network architecture of NND. (B) Test errors of NND using 20-fold CV. Errors in unit of mean square error (MSE). (C) Illustration of a phylogeny with five extant nodes and three Steiner nodes.

Figure 3

Comparison of NND and other algorithms on the semi-simulated GSE11103 dataset shows the better accuracy and robustness of NND algorithm. (A–C) Accuracies in estimated C, F, and B with three different algorithms. We tested four noise levels and repeated the experiments for ten replicates. (D) Estimated $\widehat{\text{C}}$ and ground truth C using three different algorithms with noise level s of 1.3. (E) Estimated $\widehat{\text{F}}$ and ground truth F using three algorithms with noise level s of 1.3.

Figure 4

Results and analysis. (A) First three gene module/pathway representation PCA dimensions of matched primary and metastatic samples. Matched samples are connected. (B) Hierarchical clustering of tumor samples based on raw gene expressions (left panel) and compressed gene module/pathway representation (right panel). Metastatic samples are shown in red rectangles and primary ones in yellow. (C) Portions and changes of the five communities in primary and metastatic sites. Each gray line connects the portions of a community in the primary site (blue node) and metastatic site (red node) from the same patient. (D) Pathway strengths across cell communities. (E) Phylogeny of cell subcommunities.