Cnmf порнорассказ

Layers, while the other was mainly expressed in deeper layers. This suggests that an anatomical or developmental factor may underlie variability in the response. While commonly used approaches based on clustering or pseudotemporal ordering of cells are poorly suited to achieve such insights from single-cell data, these findings emerge naturally from our matrix factorization approach.We have made our tools and analyses easily accessible so that researchers can readily use cNMF and further develop on the approach. We have deposited all the cNMF code on Github https://github.com/dylkot/cNMF/ (Kotliar, 2019; copy archived at https://github.com/elifesciences-publications/cNMF) and have made available all of the analysis scripts for figures contained in this manuscript on Code Ocean (https://doi.org/10.24433/CO.9044782e-cb96-4733-8a4f-bf42c21399e6) for easy exploration and re-execution.As others apply this approach, one key consideration will be the choice of the three input parameters required by cNMF: the number of components to be found (K), the percentage of replicates to use as nearest neighbors for outlier-detection, and a distance threshold for defining outliers. While the choice of K must ultimately reflect the resolution desired by the analyst, we propose two simple decision aids based on (1) considering the trade-off between factorization stability and reconstruction error and (2) looking at the proportion of variance explained by K principal components to estimate the dimensionality of the data (Figure 2—figure supplement 3, Figure 3—figure supplement 1, Figure 4—figure supplement 1). In addition, we noticed that choosing consecutive values of K primarily influenced individual components at the margin, suggesting that cNMF may be robust to this choice within a reasonable range of options (Figure 5 and ‘Choosing the number of components’ section of the Materials and methods). Robustness of cNMF to the number of components (K). Line plots of the maximum Pearson correlation between each of the cNMF components presented in the main analysis, and the cNMF components that result from multiple choices of K. For the simulated data, for which we have access to ground truth, we plot the correlation between the inferred components for each choice of K and the ground truth 14 components. We highlight components corresponding to activity GEPs with distinct colors and denote the number of identity GEPs contained on the same plot in parenthesis in the legend. A dashed line indicates the K choice that was presented in the main analysis. Pearson correlations are computed considering only the 2000 most over-dispersed genes and on vectors normalized by the computed sample standard deviation of each gene. https://doi.org/10.7554/eLife.43803.024 The additional two parameters allow users to optionally identify outlier replicates to filter before averaging across replicates. This improves overall accuracy by removing infrequent solutions that often represent merges or splits of the true GEPs. Using 30% of the number of replicates as nearest neighbors worked well for all datasets we analyzed, and an appropriate outlier distance threshold was clear in our applications based on the long tail in the distance distribution (Figure 2—figure supplement 3, Figure 3—figure supplement 1, Figure 4—figure supplement 1).Our approach is an initial step toward disentangling identity and activity

Approaches performed worse as they inappropriately assigned activity GEP genes to these identity programs, resulting in an elevated FDR. This illustrates how matrix factorization can outperform clustering for inference of the genes associated with activity and identity GEPs.We decided to proceed with cNMF to analyze the real datasets due its accuracy, processing speed, and interpretability. First, it yielded the most accurate inferences in our simulated data. Second, NMF was the fastest of the basic factorization algorithms considered, which is especially useful given the need to run multiple replicates and given the growing sizes of scRNA-Seq datasets (Figure 2—figure supplement 6). Third, the non-negativity assumption of NMF naturally results in usage and component matrices that can be normalized and interpreted as probability distributions—that is, where the usage matrix reflects the probability of each GEP being used in each cell, and the component matrix reflects the probability of a specific transcript expressed in a GEP being a specific gene. The other high-performing factorization method, cICA, produced negative values in the components and usages which precludes this interpretation.Beyond identifying the activity program itself, we found that cNMF could also accurately infer which cells expressed the activity program and what proportion of their expression was derived from the activity program (Figure 2f). With an expression usage threshold of 10%, cNMF accurately classified 91% of cells expressing the activity program and 94% of cells that did not express the program. Moreover, we observed a high Pearson correlation between the inferred and simulated usages in cells that expressed the program (R = 0.74 for all simulations combined, R = 0.68 for the example simulation in Figure 2a). Thus, cNMF can be used both to infer which cells express the activity program, as well as what proportion of their transcripts derive from that program.We further demonstrated that cNMF was robust to the presence of doublets—instances where two cells are mistakenly labeled as a single cell. Due to limitations in the current tissue dissociation and single-cell sequencing technologies, some number of ‘cells’ in an scRNA-Seq dataset will actually correspond to doublets. Several computational methods have been developed to identify cells that correspond to doublets, but this is still an important artifact in scRNA-Seq data (McGinnis et al., 2018; Wolock et al., 2018). We found that cNMF correctly modeled doublets as a combination of the GEPs for the two combined cell types (Figure 2g). Moreover, we found that cNMF could accurately infer the GEPs even in a simulated dataset composed of 50% doublets (Figure 2—figure supplement 7). This illustrates another benefit of representing cells in scRNA-Seq data as a mixture of GEPs rather than classifying them into discrete clusters.In all the simulations described above, the 13 cell-types occurred at uniform frequencies. This allowed us to treat all identity programs as replicates of each other for evaluating inference accuracy, rather than having to separately consider rare GEPs which should, all else equal, be harder to infer than common ones. However, this is an approximation of reality where cell-type proportions can

Stimulus (Figure 5—figure supplement 1a - left). By contrast, there was significant variability between organoids in the Quadrato et al. (2017) data that was primarily associated with the bioreactors in which the organoids were grown (Figure 5—figure supplement 1b - left). This variability was discussed in the original manuscript and validated using immunohistochemistry, and thus represents true biological signal that we would hope for cNMF to discern.We also considered whether any GEPs could be attributed to just one or a small number of replicates which could suggest that they are not reproducible within the experiment. We therefore looked at what percentage of the aggregate usage of a GEP derived from cells in each replicate. We found that each GEP contributed to cells from multiple independent replicates in both datasets (Figure 5—figure supplement 1, right panels). No GEP derived more than 15% of its usage from a single replicate in the visual cortex data or more than 45% of its usage from a single replicate in the organoid data. Furthermore, each organoid GEP was the maximum contributing GEP for a cell in at least six distinct organoid replicates, and each visual cortex GEP was the maximum contributor for a cell in at least 10 distinct mouse replicates. This supports our conclusion that the inferred GEPs represent reproducible signals within the primary organoid and visual cortex datasets.We also analyzed a human pancreatic islet scRNA-Seq dataset where variability between four donors resulted in more substantial batch-effects to see how that would impact the behavior of cNMF (Baron et al., 2016). Applied to this dataset of 10,939 cells, cNMF identified 16 GEPs that corresponded well with the cell-type clusters described in the initial publication (Figure 5—figure supplement 2). Our application of cNMF failed to identify GEPs corresponding to a few cell-types described in Baron et al. (2016) (e.g. cells distinguished as delta and gamma cell-types were assigned the same GEP). However, many of the cell types that were missed by cNMF were only distinguished through iterative sub-clustering in the initial publication, which we did not attempt.Notably, we identified multiple GEPs for many cell-type clusters that corresponded to ‘donor of origin.’ For example, we identified separate GEPs corresponding to acinar cells derived from donors 1 and 3, and acinar cells derived from donors 2 and 4, and similarly for alpha, ductal, and stellate cells. One potential contributor to the batch-effect could be that donors 1 and 3 were male and donors 2 and 4 were female. Consistent with this, we noticed that among the genes that were most differentially expressed between donors 1 and 3 compared to donors 2 and 4 in alpha, beta, and acinar cells were XIST on the X chromosome and RPSY1 on the Y chromosome (linear regression F-test p-values−243 for for XIST and p-values−145 for RPSY1 for all 3 cell-types tested). But in general, the fact that cNMF is discerning multiple GEPs for the same cell-types suggests that technical sources of variation such as batch-effect can confound the identification of identity

GEPs in scRNA-seq data. We evaluated this in simulated data of 15,000 cells composed of 13 cell types, one cellular activity program that is active to varying extents in a subset of cells of four cell types, and a 6% doublet rate (Figure 2A). We generated 20 replicates of this simulation, each at three different ‘signal to noise’ ratios, in order to determine how matrix factorization accuracy varies with noise level (Materials and methods). cNMF infers identity and activity expression programs in simulated data. (a) t-distributed stochastic neighbor embedding (tSNE) plot of an example simulation showing different cell types with marker colors, doublets as gray Xs, and cells expressing the activity gene expression program (GEP) with a black edge. (b) Pearson correlation between the true GEPs and the GEPs inferred by cNMF for the simulation in (a). (c) Same tSNE plot as (a) but colored by the simulated or the cNMF inferred usage of an example identity program (left) or the activity program (right). (d) Percentage of 20 simulation replicates where an inferred GEP had Pearson correlation greater than 0.80 with the true activity program for each signal to noise ratio (parameterized by the mean log2 fold-change for a differentially expressed gene). (e) Receiver Operator Characteristic (except with false discovery rate rather than false positive rate) showing prediction accuracy of genes associated with the activity GEP. (f) Scatter plot comparing the simulated activity GEP usage and the usage inferred by cNMF for the simulation in (a). For cells with a simulated usage of 0, the inferred usage is shown as a box and whisker plot with the box corresponding to interquartile range and the whiskers corresponding to 5th and 95th percentiles. (g) Contour plot of the true GEP usage on the Y-axis and the second true GEP usage for doublets or the second highest GEP usage inferred by cNMF for singletons for the simulation in (a). 1000 randomly selected cells are overlayed as a scatter plot for each group. https://doi.org/10.7554/eLife.43803.003 We first analyzed the performance of ICA, LDA, and NMF and noticed that they yielded different solutions when run several times on the same input simulated data. We ran each method 200 times and assigned the components in each run to their most correlated ground-truth program. We saw that there was significant variability among the components assigned to the same program -- particularly for NMF and LDA (Figure 2—figure supplement 1). Unlike PCA, which has an exact solution, these factorizations use stochastic optimization algorithms to obtain approximate solutions in a solution space including many local optima. We observed that such local optima frequently corresponded to solutions where a simulated GEP was split into multiple inferred components and/or multiple GEPs were merged into a single component (Figure 2—figure supplement 2a). This variability reduces the interpretability of the solutions and may decrease the accuracy as well.To overcome the issue of variability of solutions, we employed a meta-analysis approach, which we call consensus matrix factorization, that averages over multiple replicates to increase the