Chemical and genetic perturbations, such as those induced by small molecules and CRISPR, effect complex changes in the molecular states of cells. Despite advances in high-throughput single-cell perturbation screening technology, the space of possible perturbations is far too large to measure exhaustively. Here, we introduce PerturbNet, a flexible deep generative model designed to predict the distribution of cell states induced by unseen chemical or genetic perturbations. PerturbNet accurately predicts gene expression changes in response to unseen small molecules based on their chemical structures while also accounting for key covariates such as dosage and cell type. Moreover, PerturbNet accurately predicts the distribution of single-cell gene expression states following CRISPR activation or CRISPR interference by leveraging gene functional annotations. Our approach significantly outperforms previous methods, particularly for predicting the effects of perturbing completely unseen genes. Finally, we demonstrate for the first time that amino acid sequence embeddings can be used to predict gene expression changes induced by missense mutations. We use PerturbNet to predict the effects of all point mutations in GATA1 and nominate variants that significantly impact the cell state distribution of human hematopoietic stem cells. Using a crystal structure of GATA1 bound to DNA, we validate that these large-effect variants occur in the core DNA-contact region of GATA1 and tend to involve large changes in amino acid side-chain volume.
1 September 2020 in Contemporary Clinical Trials Communications
Cluster randomized trials (CRTs) usually randomize groups of individuals to interventions, and outcomes are typically measured at the individual level. Marginal intervention effects are frequently of interest in CRTs due to their population-averaged interpretations. Such effects are estimated using generalized estimating equations (GEE), or a recent alternative called the quadratic inference function (QIF). However, the performance of QIF relative to GEE have not been extensively evaluated in the CRT context, especially when the marginal mean model includes additional covariates. Motivated by the HALI trial, we conduct simulation studies to compare the finite-sample operating characteristics of QIF and GEE. We demonstrate that QIF and GEE are equivalent under some conditions. When the marginal mean model includes individual-level covariates, QIF shows an efficiency improvement over GEE with overall larger power, but its test size may be more liberal than GEE and GEE achieves better coverage than QIF. The test size inflation may not by fully addressed from using finite-sample bias corrections. The estimates of QIF tend to be closer to GEE in the HALI data, although the former presents a small standard error. Overall, we confirm that the QIF approach generally has potentially better efficiency than GEE in our simulation studies but might be more cautiously used as a viable approach for the analysis of CRTs. More research is needed, however, to address the finite-sample bias in the variance estimation of the QIF to better control its test size.
Beginner's Guides each cover a key technique and offer the scientifically literate but not necessarily expert audience a background briefing on the underlying science of a technique that is (or will be) widely used in molecular bioscience. The series covers a mixture of techniques, including some that are well established amongst a subset of our readership but not necessarily familiar to those in different specialisms, or the reverse. This is our Beginner's Guide to single-cell transcriptomics.
Deep generative models, including variational autoencoders (VAEs) and generative adversarial networks (GANs), have achieved remarkable successes in generating and manipulating highdimensional images. VAEs excel at learning disentangled image representations, while GANs excel at generating realistic images. Here, we systematically assess disentanglement and generation performance on single-cell gene expression data and find that these strengths and weaknesses of VAEs and GANs apply to single-cell gene expression data in a similar way. We also develop MichiGAN 1 , a novel neural network that combines the strengths of VAEs and GANs to sample from disentangled representations without sacrificing data generation quality. We learn disentangled representations of two large singlecell RNA-seq datasets [13, 68] and use MichiGAN to sample from these representations. MichiGAN allows us to manipulate semantically distinct aspects of cellular identity and predict single-cell gene expression response to drug treatment.
11 August 2020 in Communications in Statistics - Simulation and Computation
The quadratic inference function approach is a popular method in the analysis of correlated data. The quadratic inference function is formulated based on multiple sets of score equations (or extended score equations) that over-identify the regression parameters of interest, and improves efficiency over the generalized estimating equations under correlation misspecification. In this note, we provide an alternative solution to the quadratic inference function by separately solving each set of score equations and combining the solutions. We provide an insight that an optimally weighted combination of estimators obtained separately from the distinct sets of score equations is asymptotically equivalent to the estimator obtained via the quadratic inference function. We further establish results on inference for the optimally weighted estimator and extend these insights to the general setting with over-identified estimating equations. A simulation study is carried out to confirm the analytical insights and connections in finite samples.