• home

Zoom link: https://uvic.zoom.us/j/83114822200?pwd=bDY1RnFmb05wZXJRZk52THBGbDFYZz09

ABSTRACT: Although log-likelihood is widely used in model selection, the log-likelihood ratio has had few applications in this area. In this talk, I present a log-likelihood ratio based method for selecting regression models which focuses on the set of models deemed plausible by the likelihood ratio test. I show that when the sample size is large and the significance level of the test is small, there is a high probability that the smallest model in the set is the true model; thus, the method selects this smallest model. The significance level of the test serves as a tuning parameter that controls the balance between the false active rate and false inactive rate of the selected model. I consider three levels of this parameter in a simulation study and compare this method with the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) to demonstrate its excellent accuracy and adaptability to different sample sizes.

Model selection is an active area of research with a long history, a wide range of perspectives, and a rich collection of methods. For students unfamiliar with this area, this talk includes a review of key methods including the AIC, BIC and modern Lp penalty methods. The new method presented in this talk offers a frequentist perspective on the model selection problem. It is an alternative and a strong competitor to the AIC and BIC for selecting regression models.

Abstract: Content Moderation is the key and fundamental component for any content services and platforms to operate and to offer friendly, meaningful and non-toxic content for consumers to enjoy and engage with, and for users to build an online community to interact and communicate as well. Moderation service is the safety gatekeeper for other features to build on top of, such as content recommendations and personalization, targeted advertisement and so on. However, there are many big practical challenges we are facing on a daily basis. In this talk, we will present the trending solutions for content moderation in the Tech Industry, by leveraging both Artificial Intelligence (AI) and Human Intelligence (HI) to overcome multi-dimensional obstacles and to achieve the goals from multiple perspectives in real practices.

Abstract: Single-cell RNA sequencing (scRNA-seq) has transformed our understanding of cellular heterogeneity in health and disease, but the lack of physical relationships among dissociated cells has limited its applications. In this talk, we present CeLEry, a supervised deep learning algorithm to recover the spatial origins of cells in scRNA-seq by leveraging gene expression and spatial location relationships learned from spatial transcriptomics. CeLEry has a data augmentation procedure via a variational autoencoder to enlarge the training sample size, which improves the robustness of the method and overcomes noise in scRNA-seq. CeLEry can infer the spatial origins of cells in scRNA-seq at multiple levels, including 2D location as well as the spatial domain or tissue layer of a cell. CeLEry also provides uncertainty estimates for the recovered locations. This framework can be applied to study the changing of cell distribution in cerebral cortex layers during the progression of Alzheimer's disease.

Zoom link: https://uvic.zoom.us/j/83114822200?pwd=bDY1RnFmb05wZXJRZk52THBGbDFYZz09

Abstract: Like traditional meta-analysis that pools effect sizes across studies to improve statistical power, it is of increasing interest to conduct clustering jointly across datasets to identify disease subtypes for bulk genomic data and discover cell types for single-cell RNA-sequencing (scRNA-seq) data. Unfortunately, due to the prevalence of technical batch effects among high-throughput experiments, directly clustering samples from multiple datasets can lead to wrong results. The recent emerging meta-clustering approaches require all datasets to contain all subtypes, which is not feasible for many experimental designs.

In this talk, I will present our Batch-effects-correction-with-Unknown-Subtypes (BUS) framework. BUS is capable of correcting batch effects explicitly, grouping samples that share similar characteristics into subtypes, identifying features that distinguish subtypes, and enjoying a linear-order computational complexity. We prove the identifiability of BUS for not only bulk data but also scRNA-seq data whose dropout events suffer from missing not at random. We mathematically show that under two very flexible and realistic experimental designs—the “reference panel” and the “chain-type” designs—true biological variability can also be separated from batch effects. Moreover, despite the active research on analysis methods for scRNA-seq data, rigorous statistical methods to estimate treatment effects for scRNA-seq data—how an intervention or exposure alters the cellular composition and gene expression levels—are still lacking. Building upon our BUS framework, we further develop statistical methods to quantify treatment effects for scRNA-seq data.

Abstract: Statistical detection of a rare class of objects in a two-class classification problem can pose several challenges. Because the class of interest is rare in the training data, there is relatively little information in the known class response labels for model building. At the same time the available explanatory variables are often moderately high dimensional. In the four assays of our drug-discovery application, compounds are active or not against a specific biological target, such as lung cancer tumor cells, and active compounds are rare. Several sets of chemical descriptor variables from computational chemistry are available to classify the active versus inactive class; each can have up to thousands of variables characterizing molecular structure of the compounds. The statistical challenge is to make use of the richness of the explanatory variables in the presence of scant response information. Our algorithm divides the explanatory variables into subsets adaptively and passes each subset to a base classifier. The various base classifiers are then ensembled to produce one model to rank new objects by their estimated probabilities of belonging to the rare class of interest. The essence of the algorithm is to choose the subsets such that variables in the same group work well together; we call such groups phalanxes.

Zoom link: https://uvic.zoom.us/j/83114822200?pwd=bDY1RnFmb05wZXJRZk52THBGbDFYZz09

ABSTRACT: The advent of high-throughput next-generation sequencing (NGS) technologies has transformed our understanding of cell biology and human disease. As NGS has been adopted earliest by the scientific community, its use has now become widespread, and the technology has improved rapidly. At present, it is now common for laboratories to assay genome-wide transcriptomes of thousands of cells in a single scRNA-seq experiment. In addition, technologies that enable the measurement of new information, for example, chromatin accessibility, protein quantification, and spatial location, have been developed. In order to take full advantage of the multi-modality information when analyzing NGS data, new methods are demanded. This seminar will introduce several machine learning algorithms for NGS data analysis with different aims, including cell type classification, spatial domain detection, and tumor microenvironment annotation.

KEYWORDS: single cell RNA sequencing (scRNA-seq), Spatial transcriptomics (ST), tumor microenvironment, machine learning

Zoom link.

Abstract: Modern biomedical studies often collect multi-view data, that is, multiple types of data measured on the same set of objects. A typical approach to the joint analysis of two high-dimensional data views/sets is to decompose each data matrix into three parts: a low-rank common-source matrix that captures the shared information across data views, a low-rank distinctive-source matrix that characterizes the individual information within each single data view, and an additive noise matrix. Existing decomposition methods often focus on the orthogonality between the common-source and distinctive-source matrices, but inadequately consider the more necessary orthogonal relationship between the two distinctive-source matrices. The latter guarantees that no more shared information is extractable from the distinctive-source matrices. We propose a novel decomposition method that defines the common-source and distinctive-source matrices from the L2 space of random variables rather than the conventionally used Euclidean space, with a careful construction of the orthogonal relationship between distinctive-source matrices. The proposed estimators of common-source and distinctive-source matrices are shown to be asymptotically consistent and have reasonably better performance than some state-of-the-art methods in both simulated data and the real data analysis.