1. Quality Control & Preprocessing of Metagenomic Data
SDSU
Robert Schmieder –

2. Need for automated approach
Metagenomic datasets contain 100,000s (454) or 1,000,000s (Illumina) of sequences
IlluminaHiSeq 2000: currently 300 GB of data
soon 2,000 GB (≈33 human genomes with 20x coverage in single sequencing run)
Can not just read sequence by sequence to get an idea of your data

3. Basic data analysis Perform similarity search New dataset Assemble

4. Bad data analysis

5. Bad data analysis

6. Bad data analysis

7. Bad data analysis

8. Bad data analysis

9. Bad data analysis

10. Good data analysis
New
dataset

11. Good data analysis
New
dataset
Quality control
& Preprocessing

12. Good data analysis New dataset Quality control & Preprocessing
Similarity
search
Assembly

13. Good data analysis New dataset Quality control & Preprocessing
Similarity
search
Assembly

14. 3 Tools for metagenomic data

15. Quality control and data preprocessing

16. Number and Length of Sequences

17. Number/Length of sequences
Bad
Reads should be approx. same length (same number of cycles)
 Short reads are likely lower quality
Good

18. Quality of Sequences

19. Linearly degrading quality across the read
Trim low quality ends

20. Quality filtering
Any region with homopolymer will tend to have a lower quality score
Huseet al. found that sequences with an average score below 25 had more errors than those with higher averages
Huse et al.: Accuracy and quality of massively
parallel DNA pyrosequencing. Genome Biology (2007)

21. Low quality sequence issue
Most assemblers or aligners do not take into account quality scores
Errors in reads complicate assembly, might cause misassembly, or make assembly impossible

22. What if quality scores are not available ?
Alternative:
Infer quality from the percent of Ns found in the sequence
Removes regions with a high number of Ns
Huseet al. found that presence of any ambiguous base calls was a sign for overall poor sequence quality
Huse et al.: Accuracy and quality of massively parallel
DNA pyrosequencing. Genome Biology (2007)

23. Ambiguous bases
If you can afford the loss, filter out all reads containing Ns
Assemblers (e.g. Velvet) and aligners (SHAHA2, BWA, …) use 2-bit encoding system for nucleotides
some replace Ns with random base, some with fixed base (e.g. SHAHA2 & Velvet = A)
2-bit example: 00 – A, 01 – C, 10 – G, 11 - T

24. Sequence duplicates

25. Real or artificial duplicate ?
Metagenomics = random sampling of genomic material
Why do reads start at the same position?
Why do these reads have the same errors?
No specific pattern or location on sequencing plate
11-35%
Gomez-Alvarez et al.: Systematic artifacts in metagenomes from
complex microbial communities. ISME (2009) 25

27. One micro-reactor – Many beads
Martine Yerle (Laboratory of Cellular Genetics, INRA, France)

28. Impacts of duplicates False variant (SNP) calling
Require more computing resources
Find similar database sequences for same query sequence
Assembly process takes longer
Increase in memory requirements
Abundance or expression measures can be wrong

29. Impacts of duplicates False variant (SNP) calling
Require more computing resources
Find similar database sequences for same query sequence
Assembly process takes longer
Increase in memory requirements
Abundance or expression measures can be wrong

30. Depends on the experiment
In contrast, for Illumina reads with high coverage:
eliminating singletons is an easy way of dramatically reducing the number of error- prone reads

31. Tag Sequences

32. No tag
MID tag
WTA tags

33. Detect and remove tag sequences

34. Fragment-to-fragment concatenations

35. Concatenated fragments in assembled contigs

37. Data upload
Tag sequence definition

38. Tag sequence prediction

39. Parameter definition
Download results

40. Sequence Contamination

41. Principal component analysis (PCA) of dinucleotide relative abundance
Microbial metagenomes
Viral metagenomes

42. Identification and removal of sequence contamination

43. Contaminant identification
Current methods have critical limitations
Dinucleotide relative abundance uses information content in sequences  can not identify single contaminant sequences
Sequence similarity seems to be only reliable option to identify single contaminant sequences
BLAST against human reference genome is slow and lacks corresponding regions (gaps, variants, …)
Novel sequences in every new human genome sequenced*
* Li et al.: Building the sequence map of the human
pan-genome. Nature Biotechnology (2010)

44. DeconSeq web interface
Two types of reference databases
Remove
Retain

45. DeconSeq web interface (cont.)

46. Human DNA contamination identified in 145 out of 202 metagenomes

47. Conclusions
Quality control and data preprocessing are very important to increase quality of downstream analysis
Preprocessing depends on the experiment