We thank all reviewers for the valuable feedback.  Rev1,3 were concerned about the significance of the work. Below we explain the need for better clustering algorithms in the scRNAseq domain and demonstrate that scHDPMM is a substantial improvement over previous approaches. $
Importance of clustering: We currently possess limited knowledge of cell types in our body and the primary use of scRNAseq is to characterize an atlas of cell types through discovery of novel types (Gawad PNAS’14,Paul Cell’15). However, scRNAseq is riddled with technical noise that confound most clustering algorithms: cells are often grouped due to technical artifacts rather than true biology. There is a dearth of robust clustering algorithms that distinguish technical noise from biological signal. Our algorithm is the first to simultaneously correct these technical artifacts (without needing spike-in genes) while clustering and we demonstrate that such inference is superior to first normalizing and then clustering. Moreover we achieve this through a principled Bayesian framework.  

Comparing other methods: We evaluated performance using (Zeisel Science’14) as a real world dataset with ground truth annotation for cell types. We apologize the F-scores were accidentally omitted from our submission: scHDPMM outperforms competing methods (F=0.9127) compared to Phenograph(PG) (F=0.7417), HDPMM (F=0.7913), Spectral Clustering(SC) (F=0.3205) and the widely used Seurat algorithm (Satija NatBioTech’15) identifies 145 clusters (overclustering 7 cell types) with F=0.2486. 

Following Rev3, we compared performance to first normalizing with BASiCS (Vallejos’15) and subsequently clustering using PG(F=0.6173),SC(F=0.1425),HDPMM(F=0.7125). We will include these comparisons in the final version. This sequential approach of BASiCS and then clustering is often inferior to clustering without normalization for reasons below. 

BASiCS was originally tested on a cell-line dataset, comprising only one cell type. However, most single-cell datasets created today involve tissues consisting of multiple cell types with large differences in number of expressed transcripts (orders of magnitude) and other technical distinctions that require different normalization parameters for each (often unknown) cell type.  Normalization prior to clustering is problematic (e.g. expects all cells to express a similar number of transcripts) hence clustering following BASiCS is inferior to clustering with no normalization at all. scHDPMM overcomes this problem by learning both normalization and clustering in a combined model. Indeed, we do find dramatically different distributions for cell-specific normalization parms(alpha,beta) for different subtypes. 

Moreover, BASiCS requires measurement of spike-in genes and without them reduces to a Poisson estimation. To the contrary scHDPMM can be successfully applied to datasets lacking spike-ins. Spike-ins are undesirable due to: a) Cell-specific variations such as lysis efficiency accrue before introducing spike-ins and cannot be corrected with spike-ins limiting their normalizing potential. b) Spike-ins cost 10-15% of sequencing yield increasing the expense of each experiment (by~$3000). c) Technically spike-ins cannot be used at all in many recent promising technologies (drop-seq by Macosko’15,in-drop by Klein’15) that enable substantial scale up in cell number. 

For these reasons, many recent published datasets do not have spike-ins (Macosko Cell’15,Itzkovitz NatCellBio’11,Darmanis PNAS’15,Treutlein Nat’14,Paul Cell’15,Brennecke NatImm’15,Hashimshony Nat’15) and BASiCS cannot be used for these. Our method in contrast, is able to accomplish the same task without the technologically, economically burdensome spike-in requirement. 

Biological Impact: In summary scHDPMM substantially outperforms all methods tested and provides a near accurate inference of ground truth (F=0.9127). An algorithm that performs accurately on known cell types can then be reliably used to explore and characterize novel cell types. Re Rev1, in addition to the data presented, we have been approached by biologists with unpublished tissue data on fallopian tube and tumour infiltrating immune cells. Applying scHDPMM to these datasets, we observe clear clusters with biologically-valid interpretations and their covariance structures.Our collaborators are happy with scHDPMM clusters and are following up on novel cell types identified experimentally. 

Gaussian assumption: We successfully applied the Lilliefors Normality test in 3 different datasets (Klein Cell’15,Macosko Cell’15,Zeisel Science’14). Thus, the idea is if the dropouts are imputed, the distribution of each gene per cell type follows a Gaussian. Our mismatch experiments are on both continuous (StudT) and discrete distributions (NegBin: F-score fig in suppl). The log(counts) drawn from NegBin can be approximated by a Gaussian (Central limit) and also allows posterior conjugacy.