Abstract
Sequence comparison is the basis of various applications in bioinformatics. Recently, the increase in the number and length of sequences has allowed us to extract more and more accurate information from the data. However, the premise of obtaining such information is that we can compare a large number of long sequences accurately and quickly. Neither the traditional dynamic programming-based algorithms nor the alignment-free algorithms proposed in recent years can satisfy both the requirements of accuracy and speed. Recently, in order to meet the requirements, researchers have proposed a data-dependent approach to learn sequence embeddings, but its capability is limited by the structure of its embedding function. In this paper, we propose a new embedding function specifically designed for biological sequences to map sequences into embedding vectors. Combined with the neural network structure, we can adjust this embedding function so that it can be used to quickly and reliably predict the alignment distance between sequences. We illustrated the effectiveness and efficiency of the proposed method on various types of amplicon sequences. More importantly, our experiment on full length 16S rRNA sequences shows that our approach would lead to a general model that can quickly and reliably predict the pairwise alignment distance of any pair of full-length 16S rRNA sequences with high accuracy. We believe such a model can greatly facilitate large scale sequence analysis.
Competing Interest Statement
The authors have declared no competing interest.
Footnotes
We made changes to the algorithm designed. Include gap cost as a parameter to be learnt in the training process. Also the we use the differentiable version of dynamic programming to get a more robust approximation of the gradient. Also we added experiment result on v3 and v4 region of the 16s sequences to the paper.