Identifying high confidence structural variants in the human genome with Machine Learning

Lesley Chapman

Github lesleymaraina

Overview

• Background/Motivation
• Data Analysis Pipeline
• Results
• Future Direction

Background

Structural Variants

Baker, M. et al (2012) Nature Methods

SV Discovery Overview

Previous Studies

• Studies have shown that there are thousands of differences between variant calls from different whole human genome sequencing methods and bioinformatics methods
• High-confidence set of genome wide genotype calls can be used as a benchmark
• Zook et al (2014) generated a high confidence benchmark set of small variants (SNPS and indels) for NA12878

Central Aim

The goal of this project is to analyze Next Generation Sequencing (NGS) data in order to generate a high confidence list of structural variants (SVs) within the human genome.

Data

• SVs >= 20bp
• 300000+ candidate variants derived from 5 sequencing technologies

Dataset Sources

* Datasets optimized for SVVIZ

Variant Callers

PacBio

• MSPacMon
• Assemblytics (2)

CG (small)

10x(2)

Analysis Overview

• Assign labels (Genotype[+/+, +/-, -/-] or unknown) to 300000 candidate variants
• Generate labels for breakpoint and SV sequence accuracy

Strategy

• Sample dataset (5K INS and 5K Del)
• Goal 1: Generate labeled data
• Goal 2: Establish an analysis pipeline
• Assign labels to the remaining datapoints

Analysis Approach

Preliminary Results

Data Preprocessing

kNN and Missing Values

kNN Distributions

Deletions

Distribution comparison for Ill300x.alt_alnScore_std: Ks_2sampResult(pvalue=1.0000000000000002)

Insertions

Distribution comparison for Ill300x.alt_alnScore_std: Ks_2sampResult(pvalue=0.99999999999999989)

EDA

t-Distributed Stochastic Neighbor Embedding(tSNE)

Dimensionality Reduction Overview

• Data visualization techniques used to display the structure of the data
• methods that preserve a significant structure of a high-dimensional dataset in a low-dimensional space
• Linear and Non-linear dimensionality reduction techniques
• Linear: maintain large pairwise distances in a low dimensional space [i.e.: Principal Components Analysis (PCA) and Multidimensional Scaling(MDS)]
• Non-linear: preserves small pairwise distances - points and their nearest neighbors in a low dimensional space [i.e.: stochastic neighbor embedding(SNE) and t-Distributed SNE]
• PCA/MDS > Isomap > SNE > tSNE

Linear vs. Non-Linear Reduction

tSNE Overview

tSNE: Approach

• Truncated Singular Value Decomposition(SVD)
• tSNE

SVD Explained Variance

Deletions

SVD Explained Variance

Insertions

Pair Plot

Deletions

Insertions

TSNE: Deletions

GTCons

Sample

TSNE: Deletions

GTCons

Size Range

TSNE: Insertions

GTCons

Sample

TSNE: Insertions

GTCons

Size Range

Next Steps

Future Directions

• Process additional datapoints (5000 total)
• Trial manual curation: develop best practices for analyzing images
• Manual Curation: Develop an efficient way to display and distribute genomic images for analysis
• Javascript Based App: summer student
• Collect labeled datapoints
• Semi-supervised machine learning: use labeled datapoints to train semi-supervised model and assign labels to remaining datapoints
• Semi-supervised machine learning strategy: stratify based on technology?
• Include additional features: CNVthresher, Parliament, Mendelian

End

Sample Presentation

HTML Presentations Made Eas y

Created by Hakim El Hattab / @hakimel

<section> 
<h3>Overview</h3>
<ul> <li>Background/Motivation</li> 
<li>Data Analysis Pipeline</li> 
<li>Results</li> 
<li>Future Direction</li> 
</section>