dna2vec: Consistent Vector Representations of Variable-Length k-mers

dna2vec: Consistent Vector Representations of Variable-Length k-mers Patrick Ng

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

dna2vec: Consistent Vector Representations of Variable-Length k-mers 2017 arXiv 1701.06279 ng2017-dna2vec ng2017-dna2vec.txt, the project's staged copy 3cbc357752997b602198bbdaa197ea42d1ad2f2e arxiv-open genetics

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them method Quotation withheld because this record is not cleared for redistribution. distributed vector representations of DNA k-mers Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. mechanisms training word2vec (CBOW/skip-gram) on overlapping variable-length DNA k-mers Quotation withheld because this record is not cleared for redistribution. The standard encoding asserts that every pair of sequences is equally unlike every other, which is false of the sequences themselves. what this limits: any similarity computed over one-hot k-mer encodings: the distance between any arbitrary pair is equidistant, even though ATGGC should be closer to ATGGG than CACGA Quotation withheld because this record is not cleared for redistribution. Every similarity result in the paper is measured against one stipulated alignment scoring, including a gap penalty of zero. what this limits: the reported correlation between dna2vec cosine similarity and sequence similarity: the Needleman-Wunsch scores use a match score of 1, mismatch of 0 and gap penalty of 0, so the correlation is with that scoring and not with alignment in general Quotation withheld because this record is not cleared for redistribution. The method's justification is a hypothesis about words, carried over to nucleotides, and the paper marks the inspiration as loose. what this limits: the claim that dna2vec vectors carry biological meaning: the training objective is warranted by the distributional hypothesis about words appearing in the same context, which the paper says it is loosely inspired by Quotation withheld because this record is not cleared for redistribution. the word2vec word-embedding model, applied to DNA k-mers Quotation withheld because this record is not cleared for redistribution. a null distribution built by scoring pairs of RANDOM 8-mers, against which the real scores are read eliminates: similarity scores that look high simply because any two short sequences share structure - the null is constructed data rather than a reshuffling of the observed data · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. k-mer k-mer Quotation withheld because this record is not cleared for redistribution. distributed representation distributed representation Quotation withheld because this record is not cleared for redistribution. one-hot vector one-hot vector Quotation withheld because this record is not cleared for redistribution. dna2vec dna2vec Quotation withheld because this record is not cleared for redistribution. word2vec word2vec Quotation withheld because this record is not cleared for redistribution. Needleman-Wunsch Needleman-Wunsch Quotation withheld because this record is not cleared for redistribution. context - AS PHYSICAL ADJACENCY ALONG THE SEQUENCE: the set of k-mers next to the target within a fragment, which is what the word means here, not linguistic co-occurrence Quotation withheld because this record is not cleared for redistribution. the distributional hypothesis - AS THE TRANSFERRED WARRANT: a 1954 claim about words in the same context having similar meaning, doing the justificatory work for a model of nucleotides Quotation withheld because this record is not cleared for redistribution. k-mer length - AS A RANDOM VARIABLE: the window size is drawn from a discrete uniform distribution for each position, so the same sequence yields different k-mers on different runs Quotation withheld because this record is not cleared for redistribution. builds-on → word2vec (from NLP/linguistics) Quotation withheld because this record is not cleared for redistribution. builds-on → word2vec (Mikolov et al., 2013a) Quotation withheld because this record is not cleared for redistribution. extends → BioVec (Asgari and Mofrad, 2015) and seq2vec (Kimothi et al., 2016) Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · k-mer —operationalised-as→ one-hot vector k-mer one-hot vector EVIDENCE CORRECTED: RE-GROUNDED. The old span named only 'one-hot vector'. This sentence names both and states the encoding explicitly - 'of k-mer AS a…'. — strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. 100 — dimensionality of the continuous dna2vec vector space Quotation withheld because this record is not cleared for redistribution. two-layer — depth of the shallow neural network (as in word2vec) Quotation withheld because this record is not cleared for redistribution. dna2vec trains distributed representations of variable-length DNA k-mers, based on the word embedding model word2vec. Quotation withheld because this record is not cleared for redistribution. Summing dna2vec vectors approximates nucleotide concatenation, and cosine similarity of the vectors correlates with Needleman-Wunsch alignment similarity. Quotation withheld because this record is not cleared for redistribution. DNA k-mer DNA k-mer Quotation withheld because this record is not cleared for redistribution. cosine similarity of dna2vec vectors cosine similarity of dna2vec vectors Quotation withheld because this record is not cleared for redistribution. k-mer k-mer Quotation withheld because this record is not cleared for redistribution. one-hot vector one-hot vector Quotation withheld because this record is not cleared for redistribution. correlation correlation Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://arxiv.org/abs/1701.06279.