More on sourmash gather vs lca sourmash gather

More on sourmash

'gather’ vs ‘lca’

’sourmash gather’ – iterative, greedy approach on genomes.

lca = “lowest common ancestor”, assigns taxonomic identity to k-mers.

’sourmash lca gather’ – iterative, greedy approach on taxonomic lineages. We understand this approach less well than ‘sourmash gather’ : )

Synthetic metagenomic data https: //genomebiology. biomedcentral. com/articles/10. 1186/s 13059 -017 -1299 -7 Emerging datasets from the U. S. National Institute of Standards and Technology (NIST) for benchmarking metagenomic classifiers https: //ftp-private. ncbi. nlm. nih. gov/nist-immsa/IMMSA Mc. Intyre et al. , 2017. https: //www. nist. gov/mml/bbd/immsa-missionstatement

sourmash gather performs as well or better than other classifiers in some circumstances This is a modified version of Mc. Intyre et al. 2017 Additional File 4: Table S 3 Phillip Brooks

<INSERT CONTROL MARKING HERE> HC 1, HC 2, & LC 1 - LC 8 Data from Mc. Intyre Publication • HC 1, HC 2, LC 1, LC 2, LC 3, LC 4, LC 5, LC 6, LC 7, and LC 8 datasets were simulated with known abundances of each species in the dataset (i. e. , number of reads) How many sequence reads are needed to differentiate true positives from false negatives? • These datasets are used to evaluate limits of detection Phillip Brooks

<INSERT CONTROL MARKING HERE> But sometimes… sourmash performs poorly Recall Pretty good Poor recall Precision Phillip Brooks

<INSERT CONTROL MARKING HERE> False Negatives Prevalent with <7, 000 reads Reads Per Species 30, 000 - 189, 000 20, 000 - 30, 000 16, 000 - 20, 000 15, 000 - 16, 000 14, 000 - 15, 000 13, 000 - 14, 000 12, 000 - 13, 000 11, 000 - 12, 000 10, 000 - 11, 000 9, 000 - 10, 000 8, 000 - 9, 000 7, 000 - 8, 000 6, 000 - 7, 000 5, 000 - 6, 000 4, 000 - 5, 000 3, 000 - 4, 000 2, 000 - 3, 000 1, 000 - 2, 000 0 - 1, 000 Correct Species 15 15 17 13 16 20 11 16 12 16 19 19 25 10 4 1 0 False Negatives 0 0 0 1 3 3 7 19 36 69 False negatives began appearing at <7, 000 reads per species. The exact number of reads and unique 51 -mers needed for a correct species call varies depending on the organism, reference database, and genomic regions covered by the reads. Phillip Brooks

<INSERT CONTROL MARKING HERE>

Subsampling via hash-based randomization.

Subsampling via hash-based randomization.

Subsampling via hash-based randomization.

This subsampling allows two estimators:

This is in some ways similar to capture/recapture analysis. Ondov et al. , 2016 – mash paper

Summary of approach: • Choose 1 / B words to be “special”, and subsample word collections down to those special words. • • If B = 1, you are picking all words! If B = 2, you are picking half the words! If B = 10, one tenth… Typically we choose B to be 1000, or 10, 000. • If “special” word is in a sample, it will always be in subsample (unlike Min. Hash). • Work only with these (much) smaller subsamples.

Features of approach: • Provides fast and lightweight estimation of similarity, containment; • B is tunable, with tradeoffs in resolution vs time & memory. • Subsampling to larger B is straightforward and does not involve revisiting raw data. • Incredibly computationally convenient! All of the analyses in this talk (over a Tbp of data) can be done on my laptop.

<INSERT CONTROL MARKING HERE>

Exemplar data set: Tara Oceans Approximately 1 TB of data, in ~250 different samples (“libraries”). Sunagawa et al. , 2015

Tara vs genbank+JGI, and extracted via assembly: Known words Total words (Billions) 10 or more samples: Reference set + Extracted 0. 47 1. 58 7. 16 6. 6% 22. 0% 2 or more samples: Reference set + Extracted 1. 18 4. 54 48. 65 2. 4% 9. 3% Tara ocean data, k=31.

Observation 1: Vast majority is unknown. Known words Total words (Billions) 10 or more samples: Reference set + Extracted 0. 47 1. 58 7. 16 6. 6% 22. 0% 2 or more samples: Reference set + Extracted 1. 18 4. 54 48. 65 2. 4% 9. 3% Tara ocean data, k=31.

Observation 2: Assembly is biased towards common words. Known words Total words (Billions) 10 or more samples: Reference set + Extracted 0. 47 1. 58 7. 16 6. 6% 22. 0% 2 or more samples: Reference set + Extracted 1. 18 4. 54 48. 65 2. 4% 9. 3% Tara ocean data, k=31.

<INSERT CONTROL MARKING HERE> Rarefaction curves on k-mers? TARA ocean results.