Knowledge of immune cell phenotypes in the tumor microenvironment is essential for understanding mechanisms of cancer progression and immunotherapy response. We profiled 45,000 immune cells from eight breast carcinomas, as well as matched normal breast tissue, blood, and lymph nodes, using single-cell RNA-seq. We developed a preprocessing pipeline, SEQC, and a Bayesian clustering and normalization method, Biscuit, to address computational challenges inherent to single-cell data. Despite significant similarity between normal and tumor tissue-resident immune cells, we observed continuous phenotypic expansions specific to the tumor microenvironment. Analysis of paired single-cell RNA and T cell receptor (TCR) sequencing data from 27,000 additional T cells revealed the combinatorial impact of TCR utilization on phenotypic diversity. Our results support a model of continuous activation in T cells and do not comport with the macrophage polarization model in cancer. Our results have important implications for characterizing tumor-infiltrating immune cells.

Recent single-cell analysis technologies offer an unprecedented opportunity to elucidate developmental pathways. Here we present Wishbone, an algorithm for positioning single cells along bifurcating developmental trajectories with high resolution. Wishbone uses multi-dimensional single-cell data, such as mass cytometry or RNA-seq data, as input and orders cells according to their developmental progression by pinpointing bifurcation points and labeling each cell as pre-bifurcation or as one of two post-bifurcation cell fates. Using 30-channel mass cytometry data, we show that Wishbone accurately recovers the known stages of T cell development in the mouse thymus, including the bifurcation point. We also apply the algorithm to mouse myeloid differentiation and demonstrate its generalization to additional lineages. A comparison of Wishbone to diffusion maps, SCUBA and Monocle shows that it outperforms these methods both in the accuracy of ordering cells and in the correct identification of branch points.

SUMMARYKnowledge of the phenotypic states of immune cells in the tumor microenvironment is essential for understanding immunological mechanisms of cancer progression and immunotherapy responses, as well as the development of novel treatments. By combining single-cell RNA-seq data from over 45,000 immune cells collected from eight primary breast carcinomas, as well as matched normal breast tissue, peripheral blood, and lymph node, we created an immune map of breast cancer. We developed a preprocessing pipeline, SEQC, and a Bayesian clustering and normalization method, Biscuit, to address the computational challenges inherent to single-cell RNA-seq data, enabling integration of data across patients. This atlas revealed significant similarity between normal and tumor tissue resident immune cells. However, we observed continuous tumor-specific phenotypic expansions driven by environmental cues. Our results argue against discrete activation states in T cells and the polarization model of macrophage activation in cancer, with important implications for characterizing tumorinfiltrating immune cells.

We consider the problem of learning high-level controls over the global structure of sequence generation, particularly in the context of symbolic music generation with complex language models. In this work, we present the Transformer autoencoder, which aggregates encodings of the input data across time to obtain a global representation of style from a given performance. We show it is possible to combine this global embedding with other temporally distributed embeddings, enabling improved control over the separate aspects of performance style and and melody. Empirically, we demonstrate the effectiveness of our method on a variety of music generation tasks on the MAESTRO dataset and a YouTube dataset with 10,000+ hours of piano performances, where we achieve improvements in terms of log-likelihood and mean listening scores as compared to relevant baselines. * Work completed during internship at Google Brain.

No abstract