class: center, middle, inverse, title-slide # preciseTAD: A machine learning framework for precise 3D domain boundary prediction at base-level resolution ## #EuroBioc2020 workshop ### Spiro Stilianoudakis, PhD Candidate; Mikhail Dozmorov, PhD ### Virginia Commonwealth University ### December 16th, 2020 --- <!-- layout: true <div class="my-footer"> <img src="img/logo_bioconductor.gif"> </div> --- --> ## preciseTADworkshop resources - [The slides](https://stilianoudakis.github.io/slides_preciseTADworkshop/#1): https://stilianoudakis.github.io/slides_preciseTADworkshop/#1 - [Github repo](https://github.com/dozmorovlab/preciseTADworkshop): https://github.com/dozmorovlab/preciseTADworkshop - [Docker image](https://hub.docker.com/repository/docker/stilianoudakis/precisetadworkshop): https://hub.docker.com/repository/docker/stilianoudakis/precisetadworkshop - [Pkgdown website](https://dozmorovlab.github.io/preciseTADworkshop/): https://dozmorovlab.github.io/preciseTADworkshop/ --- ## The 3D structure of the genome - .small[ Human genome is big - ~3.2 billion base pairs ] - .small[ Chromosome structure and function is organized at multiple scales ] - .small[ At kilobase (kb) scale chromatin loops form regions that are highly self-interacting called Topologically Associated Domains (TADs) ] .center[<img src="img/3d_chromatin.png" width = 700>] --- ## Why study TADs? - TADs are stable and dynamic structures + Highly conserved across species and cell lines + Associated with gene regulation - Disruption of the boundaries that demarcate TADs has been shown to lead to rare diseases and even cancer - Understanding the mechanisms through which TADs are formed remains a top priority in our goal to fully understand the functionality of the human genome <!--- Can we leverage functional genomic elements in order to more precisely identify TAD boundaries?--> --- ##Hi-C data as a matrix .pull-left[ - The genome (chromosome) is split into equally sized regions controlled by sequencing depth (resolution) - Data is represented by a symmetric matrix of contacts `\(C_{ij}\)` where entry `\(ij\)` corresponds to the number of times region `\(i\)` comes into contact with region `\(j\)` (interaction frequency) - Power-law decay of interactions with increasing **distance** ] .pull-right[ <img src="img/contact_matrix.png" width = 450> ] --- ## TAD-calling .pull-left[ - “TAD-callers” have been proposed to identify the start and end coordinates of TAD boundaries - Heavily reliant on Hi-C data “resolution” - Ignore the fundamental role of functional genomic data (ChIP-seq data) in chromatin folding - Our aim: Can we leverage functional genomic elements in order to more precisely identify TAD boundaries? ] .pull-right[ <img src="img/TAD_detection.png" width = 650> ] --- ## Turning TAD calling into a supervised ML framework .pull-left[ - Defining a resolution-specific response vector Y by shifted binning ] .pull-right[ <img src="img/response_vector.png" width = 650> ] --- ## Turning TAD calling into a supervised ML framework .pull-left[ - Defining the feature space `\(\textbf{X} = \{X_1, X_2, \cdots, X_p \}\)` ] .pull-right[ <img src="img/feature_space.png" width = 650> ] --- ## Precisely predicting TAD boundaries at base-level resolution using *preciseTAD* .pull-left[ What it does… + Leverages a predictive model in conjunction with: - Density-based spatial clustering of applications with noise (DBSCAN) - Partitioning around mediods (PAM) to precisely predict the most biologically meaningful domain boundaries ] .pull-right[<img src="img/psuedocode.png" width = 500>] --- ## Evaluating precision of boundary prediction - Enriched heatmaps - Signal profile plots - Distance to nearest ChIP-seq peak region - Venn diagrams/Jaccard indices --- ## preciseTADworkshop resources - [The slides](https://stilianoudakis.github.io/slides_preciseTADworkshop/#1): https://stilianoudakis.github.io/slides_preciseTADworkshop/#1 - [Github repo](https://github.com/dozmorovlab/preciseTADworkshop): https://github.com/dozmorovlab/preciseTADworkshop - [Docker image](https://hub.docker.com/repository/docker/stilianoudakis/precisetadworkshop): https://hub.docker.com/repository/docker/stilianoudakis/precisetadworkshop - [Pkgdown website](https://dozmorovlab.github.io/preciseTADworkshop/): https://dozmorovlab.github.io/preciseTADworkshop/