Algorithms for Next-gen Sequence Analysis

Algorithms for Next-gen Sequence Analysis


The field of next-gen sequence analysis is advancing so rapidly that new algorithms come out almost every day. Here we provide a broad categorization for such algorithms and describe critical challenges for each category. This will help us understand the approaches presented in various algorithms, when we look into each one in more detail.

Usually a scientist submits DNA samples for an organism to a core facility for sequencing and receives a large disk full of sequence data. The first categorization comes from whether the organism has a reference sequence. Second categorization comes from whether the sample is genomic or transcriptomic.

As a 2x2 matrix, the categories are

A. Organism has reference genome and sample is genomic

B. Organism does not have reference genome and sample is genomic

C. Organism has reference genome and sample is transcriptomic

D. Organism does not have reference genome and sample is transcriptomic

Challenges in Category A

  • Alignment of short reads on to reference genome.

  • Identification of SNPs.

  • Identification of insertion-deletions.

Challenges in Category B

  • Genome assembly from short reads.

Challenges in Category C

  • Alignment of transcriptomic reads on to reference genome.

  • Expression profiling.

  • Alternate splicing.

Challenges in Category D

  • Transcriptome assembly (works differently from genome assembly).

  • Expression profiling without reference genome.

  • Alternate splicing.

We prefer this categorization, because each category presents its unique challenges.

The primary focus of algorithms developed for category A is speed. Usually millions, and sometimes billions, of reads need to be rapidly mapped on to reference genome for further analysis. Any improvement in speed can make major improvement in time for analysis.

The amount of computer memory (RAM) needed for assembly is the main concern of category B algorithms. Conceptually, all reads need to be loaded on to memory and compared before an assembly method joins them into large scaffolds. Therefore, assembly of large genomes is memory intensive.

Category C has similarities with category A with added difficulty from intron- exon junctions. Algorithms that map contiguous genomic reads do not handle split junctions and alternate splicing properly.

Category D cannot directly use category B algorithms, because genome assembly programs assume frequency uniformity of reads from all parts of genome. That is a bad assumption fortranscriptomic reads, because expression levels of different genes can vary wildly.

We will keep the above categories in mind in our future discussions on various algorithms.


Written by M. //

Web Statistics