Past Talks

Logistic regression is a powerful tool that extends beyond statistical applications; it serves as a window to machine learning. During this week’s camp, we will explore the basics of logistic regression in comparison to linear regression. We will go through some detailed examples that can help researchers to implement logistic regression in medical research. If time allows, we will even be able to walk through some code with these examples.

Our favorite speaker Dr. Walid Abdelmoula will talk about the basic Bayesian statistics in an interactive way. Feel free to join us.

Lorena will present clustering based on this Nature Method article:

https://www.nature.com/articles/nmeth.4299

Almost everyone uses linear regression for their daily analysis. But do you know that there are several assumptions to check before you will apply linear regression model to your data? If your data don’t satisfy the assumptions, what can you do? What’s R² intuitively mean? etc.

The speaker Dr. Youngeun Hub is a Neurologist & Assistant Professor in South Korea. She is now a visting scholar in Dr. Clemens Scherzer’s lab at Brigham and Women’s Hospital, working on several interesting projects. She will review the following prime article in Nature Methods and cover the topic in various aspects. So it’s highly recommend to read the article before the talk.

https://www.nature.com/articles/nmeth.3627

Dr. Wenbo Li is invited as a Featured Speaker of 2019 in the Bioinformatics Club. Dr. Li published several landmark papers in the field of enhancer RNAs (eRNAs) when he was a postdoc in Dr. Michael G. Rosenfeld’s group. Now he is an Assistant Professor in the McGovern Medical School of the University of Texas Health Science Center, and CPRIT Scholar in the Cancer Prevention Research Institute Of Texas.

Here is the abstract of his talk:

The “recognition” between an enhancer and its cognate promoter in eukaryotes ensures proper gene transcription output. However, both the biological basis and the pathological involvement of enhancer-promoter recognition remain poorly understood. Here, using estrogen-regulated enhancer-promoter communication as an initial model, we report that loss of a preferred promoter can unexpectedly release its regulatory enhancer to loop to and activate an alternative promoter(s) in its chromosomal neighborhood. This target switching process of active enhancers was referred to as “enhancer release and retargeting” (ERR).  Based on additional analysis of disease risk loci, cancer mutation, and the GTEx database, putative ERR events represent a remarkably common, if overlooked, mechanism underlying alternative gene activation in cancer and neurodegeneration. Mechanistically, ERR may be controlled by a cohesin-mediated looping extrusion process but takes place in a much smaller genomic scale than topological domains or compartments. We expect that ERR plays a wide roles in determining gene-enhancer communication and gene expression deregulation in human disease.

EM algorithm is one of the core algorithms in statistics and machine learning. It has been widely used in many tools in the fields of bioinformatics, for example, motif discovery (MEME), comparative genomics (Phylo-HMM), multiple sequence alignment (ClustalW), NGS analysis (Cufflinks, RSEM, and eXpress), etc.  So, please join us to master the hardcore technique.

Dr. Kazuki Yoshida will review the following Nature Biotechnology prime article:

https://www.nature.com/articles/nbt1406

https://www.nature.com/articles/nmeth.4210

http://www.nature.com/nbt/journal/v24/n12/abs/nbt1206-1565.html

I would like to discuss basic data analysis methods for causal modeling.  Given my background originating in applied statistics for the behavioral sciences, where causal modeling of one form or another has been used for decades and given its more recent introduction to bio-medical statistics, I would like to introduce some basic aspects of it that may be unfamiliar to bio-medical researchers and even many biostatisticians.  As a backdrop, I will refer to a recent book which has created quite a bit of interest on this subject, “The Book of Why” (2018) by Judea Pearl, who claims that a “Revolution in Causality” is taking place in data analysis and science. (Reading the book is not a prerequisite for my talk). I want to emphasize that the purpose of my presentation is not to explain how to conduct any specific data analysis technique like structural equation modeling (SEM), path analysis, or confirmatory factor analysis (using e.g., the SAS Calis Procedure), though I will touch on that.  Rather I will try to show people how causal models, especially as explicated with causal diagrams, can be seen as underlying research questions tested by many varied statistical techniques and to hopefully provide a helpful way of making certain decisions about appropriate data analyses given the research questions and available data. Thus, I want to emphasize causal modeling as a methodology rather than a specific method.     

a 101 talk on how to use git / github / gitlab. More details to be announced.

TBD

Dr. Lorena from the Department of Statistics of Harvard T.H. Chan School of Public Health will talk about how to accurately detect miRNA isoforms (“isomiRs”) based on different sequencing platforms. She will introduce the popular toolkit, mirTOP, for miRNA variant analysis. In the end, she will give a useful guideline for how to choose platforms/protocols/sequencing depth/replication number etc. in order to detect the true miRNA variants.  Find more information from the slide attached.

Dr. Beatrice Weykopf from the Neurogenomics Lab of BWH will introduce the biology of iPSC and her research in Parkinson’s disease.

Dr. Ilya Korsunskiy from the Raychaudhuri Lab @ HMS, BWH will talk about how to analyze scRNAseq data on your laptop. Very cool! Sign up for more details.

Emily will introduce various pathway analysis methods, including IPA, GSEA, SPIA, Enrichr etc. with demo and examples. More details to be announced.

Sequencing methods have become commonplace. Many laboratories generate tens of samples in a day and hundreds of libraries worth of data every month. Sequence data, however, as opposed to previous technologies, significant computational power and knowledge required for a scientific discovery. The current bottleneck for many laboratories is the processing and analysis of data in a timely manner. As sequencing becomes both cost effective and inexpensive, experiments increase in complexity. A single sequencing run usually includes many different conditions and replicates designed to answer a specific question. To process all samples in a study as a unit, automated data processing pipelines are necessary by leveraging the power of parallel nature of current computer environments called high performance computing (HPC) systems for a faster outcome. To address data processing and analysis, we implemented a highly parallel, platform, called DolphinNext specifically designed to process sequencing data generated towards answering a specific question. Once the data is processed, the result is usually a count table that specifies the estimated number of reads that originate from genomic loci. Differential analysis to determine which loci have different cellular activity in different conditions is based on this count tables and it requires a common, iterative, cycle of data assessment, data preparation and differential analysis. We developed DEBrowser as an R bioconductor project, to interactively visualize every step of the differential analysis of count data without requiring any programming expertise. It is based on a shiny infrastructure, which offers an interactive, web based graphical user interface for R packages. We leveraged the reactive programming of shiny to visualize the data at all stages of the analysis.

High-throughput sequencing technologies have been widely used in modern genomics studies. Due to experimental/technical issues, diverse sources of biases can be introduced into the sequencing data and subsequently disturb the downstream analysis. In the talk, I will use ChIP-seq as an example to demonstrate how we improved the downstream analysis by accounting for biases. I will discuss the effects of GC-content bias in large ChIP-seq datasets and our proposed statistical method to account for this bias in protein binding quantification. 

Note the changed time due to the snowstorm.

Path Analysis (also referred to as Structural Equation Modeling, SEM, or causal modeling) is a method for systematically studying the possible causal models that may underlie an observed correlation matrix of variables. In path analysis, various explicitly detailed models of causal dynamics are hypothesized. Then the fit of the observed correlation matrix to what would be the correlation matrix predicted by a given hypothetical causal model is assessed. This is done in order to obtain support and greater specification for the proposed model and rule out as statistically improbable alternative models. Path analysis can be viewed as a specific statistical analysis method or as a more general methodological approach to research, i.e., an “algebra of causality”. I will discuss both points of view, but will emphasize more the latter. Path analysis methodology allows for a more fundamental and flexible approach to understanding some statistical techniques currently considered useful and important such as mediation/moderation analysis, factor analysis, latent variable analysis, multiple regression, partial correlation, and analysis of covariance.

For decades, many statisticians have criticized the inappropriate use and overuse of null hypothesis significance testing (NHST) especially in medical and behavioral science research. This controversy has now especially intensified partly as a reaction to a recent ban on publishing results of NHST in the psychology journal Basic & Applied Social Psychology (BASP).  The American Statistical Association has also recently taken the unusual step of publishing a formal policy statement cautioning against misuse of NHST.  The NHST controversy is also intimately connected to other widely recognized problems in science publishing such as publication bias and “the crisis of irreproducibility”.  Dr. Locascio will try to lay out the basic issues in these controversies and suggested solutions to these problems including one of his own just published in BASP. 

Most commercial small RNA library preparation kits currently select a few RNA classes, rather than profile the full-coverage without bias. Dr. Wei will discuss the chemistry behind it, and options to reduce bias.

The Partners Clinical Image Bank (CIB) portal enables researchers to access registries of expertly curated, phenotypically characterized medical images and associated clinical data extracted from the Partners electronic medical records.

      Access to the Clinical Image Bank is available to all with a Partners logon who have registered to use the Research Patient Data Registry (RPDR) and who accept the terms of the Data Use Agreement.  The portal enables full exploration and download of de-identified clinical information extracted from the RPDR and quantitative MR imaging metrics. Users with IRB approval who wish to download the fully identifiable clinical data and native DICOM medical images may do so easily.

     The first registry to be made available is a collection of pediatric brain MR images that includes a cohort of neonates with clinically confirmed hypoxic ischemic encephalopathy, a cohort of “normative” children imaged between the ages of 0-6 years of age with no known CNS pathology and a cohort of children imaged between the ages of 0-6 years of age that have a neuropsychiatric disorder of indeterminate severity.  In addition to secure access to the DICOM image data, the portal allows interaction with and download of extracted quantitative metrics from the Apparent Diffusion Coefficient (ADC) maps calculated from the Diffusion Weighted Images for many of the cases.  Each HIE case also has detailed perinatal clinical course information including maternal risk factors and clinical outcomes.

    Dr. Gollub, one of the developers of the CIB will demonstrate the use of this new software tool and provide a guided tour of the available data.  She would like to gather your input on features and data sets that could be added that would be of greatest value to you in your work.

The IDEA Analytics platform is designed to perform complex analytics and data transformations across large and diverse datasets in a secure and reliable way. The platform provides research teams access to a range of open-source tools, including Hadoop and Spark, hosted on a highly scalable infrastructure specifically designed for large analytics workloads. In addition to well-known Hadoop applications such as Pig, Hive, and HBase, specialized machine learning and natural language processing tools are also integrated into the solution. The platform will be presented together with a demo on how to use it with PySpark in Jupiter and SparklyR in RStudio.