GPU and machine learning solutions for comparative genomics

Usman RoshanDepartment of Computer Science

New Jersey Institute of Technology

Outline

• Talk centered around problem of mapping DNA sequences to genome, analysis, and applications

• Prediction of chronic lymphocytic leukemia with whole exome sequences and machine learning– Data processing– Results

• Graphics Processing Unit program for mapping divergent reads to genomes and applications on real data– Overview of program– Results on simulated and real data

Disease risk prediction

• Prediction of disease risk with genome wide association studies has yielded low accuracy for most diseases.

• Family history competitive in most cases except for cancer (Do et. al., PLoS Genetics, 2012)

Disease risk prediction

• Our own studies have shown limited accuracy with various machine learning methods– Univariate and multivariate feature selection– Multiple kernel learning

• What accuracy can we achieve with machine learning methods applied to variants detected from whole exome data?

Chronic lymphocytic leukemia prediction with exome sequences and machine learning

• We selected exome sequences of chronic lymphocytic leukemia from dbGaP. Largest at the time of download in August 2013. 186 cases and 169 controls

• Case and control prediction accuracy with genetic variants unknown

• Same dataset previously studied in Wang et. al., NEJM, 2011 where new associated genes are reported but no risk prediction

What is whole exome data?Human genome sequence

Illumina 76bp short reads (exome data).In practice flanking regions are also sequencedand so some intronic regions are included.

Exons

Coding regionsIntrons

Obtain structural variants (1)

• Data of size 3.2 Terrabytes and 140X coverage• Mapped to human genome reference with

BWA MEM (popular short read mapper)

Human genome reference sequence

Short reads are aligned to human genome

Obtain structural variants (2)

• Obtained SNPs and indels from the alignments for each individual

ACCAGACCAGACCAG

ACCAG

ACCCGACCCGACCCG

Heterozygous SNP A/C

ATT--AATT--AATT--AATTGAATTGA

ATTGA

Heterozygous indel

ATTGAHuman genomereference

Short reads from a Single individual

Obtain structural variants (3)

A/C C/G A/C C/G C0 AA CC C0 0 0C1 AC CG C1 1 1C2 AA GG C2 0 2Co1 AC CG Co1 1 1Co2 CC CG Co2 2 1

• Combine variants from different individuals to form a data matrix• Each row is a case or control and each column is a variant• 180 cases and 155 controls after excluding very large files and problematic

datasets• 545,721 SNPs and indels (530,129 SNPs, 15,592 indels)

Numerically encoded

Perform cross-validation study

Training data

Validation data

1. Split rows randomly into training validation sets (90:10 ratio).

2. Rank all variants on training3. Learn support vector

machine classifer on training data with top k ranked variants

4. Predict case and control on validation data.

5. Compute error and repeat 100 timesFull dataset: each row

is a case or control individual and eachcolumn is a variant (SNP or indel)

0 0 1 2 0 . . . 0 2 2 2 1 . . . ...

Variant ranking

F0 F1 F2 F1 F2 F0C0 1 2 0 C0 2 0 1C1 1 2 1 C1 2 1 1C2 1 2 2 C2 2 2 1Co1 1 0 1 Co1 0 1 1Co2 2 0 0 Co2 0 0 2

Rank features

Different feature rankings

• Correlation coefficients between rankings on SNPs

F-score Chi-square

Pearson 0.99 0.37

F-score 0.37

Risk prediction with chi-square ranked SNPsMean accuracy of 85.7% with top 60 ranked SNPs (across 100 splits)

• Mean accuracy with significant SNPs only is 81% and significantly lower (Wilcoxon rank test p-value=10-14)

• Significant SNPs on chromosome 14 in IGH gene, predictive SNPs on chromosomes 2, 14, and 15 in intron and exons of IGK, IGH, and LOC642131.

• One predictive SNP has mutations only in case individuals. Previous genes not significant.

Principal component analysis of SNP data

PCA plot of all 530,129 SNPs PCA plot of top 60 chi-square ranked SNPs

Summary

• Our predictive could be used for prognosis but replication in a different sample is first required.

• Better alignments may yield more predictive variants. NextGenMap has a better mapping rate than BWA but is much slower

• Would our pipeline work other cancers?

Mapping divergent short reads to genomes

• Recall the problem of mapping short read to genomes• Methods based on hash-tables and Burrows-Wheeler transform are fast but

accuracy falls quickly at divergence increases• High performance Smith-Waterman implementations like CUDASW++ and SSW

take long to finish (even for bacterial genome mapping)• Our objective: Align divergent reads faster than Smith-Waterman and more

accurate than hash-tables and Burrows-Wheeler transform.

Human genome reference sequence

Short reads are aligned to human genome

MaxSSmap algorithm

• Thread number i maps the read to fragment i.• Threads run in parallel on a GPU (or CPU with many

cores)• We also account for junctions between fragments

Input: Whole genome and a short read

Thread 0 Thread 1 Thread 2 Thread 3 Thread 4 Thread 5

Genome fragments of same length

Experimental study

Genome sequence

Align reads with NextGenMap

Some reads are not mapped due to mismatches and gaps. We realign them with MaxSSmap and Smith-Waterman

Simulation studyDiv. BWA

(multi-core)

NextGenMap(GPU)

NextGenMap+MaxSSmap_fast

NextGenMAp+MaxSSmap

NextGenMap+CUDASW++

30% with gaps

0.5 (0) 19 (0.4) 82 (2.9) 90.5 (3.5) 92.5 (1.6)

Timemins

0.4 2.1 162 222 1528

• Simulated 1 million 251 bp E.coli reads with Stampy and aligned to Ecoli genome (approximately 4.6 million base pairs). We know the true positions of the reads.

• Shown above are percentage of reads that were correctly mapped by each program (incorrect in parenthesis)

Ancient DNA mapping

• Aligned 100,000 76bp ancient horse DNA reads to the horse genome (approximately 2.3 billion base pairs). Measure number of reads that were mapped.

• Shown above are percentage of reads that were mapped by each program

• MaxSSmap alignments contain 39% mismatches on the average

Mapping paired reads

Genome sequence

Reads come in pairs.We align them withNextGenMap and expect them to be mapped within 500 base pairs

We realign pairs 1. where both are mapped farther than 500 base pairs2. where at least one read in the pair is unmapped

Realigning paired reads to human genome

• Align 100,000 101 bp paired reads from NA18278 in 1000 genomes to human genome reference (3 billion base pairs).

• Shown here are percent of paired reads whose mapped positions are within 500 base pairs (also known as concordant reads).

• In MaxSSmap we realign discordant reads from NextGenMap as well.• MaxSSmap alignments have 19% mismatches on the average• Variant detection not performed yet

Summary

• Better accuracy and mapping rate than NextGenMap and BWA

• Runtime for large genomes still very high relative to NextGenMap but faster than Smith-Waterman (speedup increases with number of reads).

• More analysis needed on real data

Software and acknowledgements

• Our software, data, and publications can be found at http://www.cs.njit.edu/usman

• Students: Bharati Jhadev, Nihir Patel, and Turki Turki• Dennis R. Livesay for GPU cluster at University of

North Caroline at Charlotte and Shahriar Afkhami for GPU machine at NJIT

• NJIT system admins David Perel, Kevin Walsh, and Gedaliah Wolosh for high performance computing support and storage of genomic data.

References

• Turki Turki and Usman Roshan, MaxSSmap: A GPU program for mapping divergent short reads to genomes with the maximum scoring subsequence (submitted)

• Bharati Jhadav, Nihir Patel, and Usman Roshan, Prediction of chronic lymphocytic leukemia with exome sequences, machine learning (in preparation for submission)

Thank you!

• Questions….