Blind Sequence Denoising with Self-Supervised Set Learning
Abstract: Denoising discrete-valued sequences typically relies on training a supervised model on ground-truth sources or fitting a statistical model of a noisy channel. Biological sequence analysis presents a unique challenge for both approaches, as obtaining ground-truth sequences is resource-intensive and the complexity of sequencing errors makes it difficult to specify an accurate noise model. Recent developments in DNA sequencing have opened an avenue for tackling this problem by producing long DNA reads consisting of multiple subreads, or noisy observations of the same sequence, that can be denoised together. Inspired by this context, we propose a novel method for denoising sets of sequences that does not require access to clean sources. Our method, Self-Supervised Set Learning (SSSL), gathers subreads together in an embedding space and estimates a single set embedding as the midpoint of the subreads in both the latent space and sequence space. This set embedding represents the “average” of the subreads and can be decoded into a prediction of the clean sequence. In experiments on simulated long-read DNA data, SSSL-denoised sequences contain 31% fewer errors compared to a traditional denoising algorithm based on a multi-sequence alignment (MSA) of the subreads. When very few subreads are available or high error rates lead to poor alignment, SSSL reduces errors by an even greater margin. On an experimental dataset of antibody sequences, SSSL improves over the MSA-based algorithm on two proposed self-supervised metrics, with a significant difference on difficult reads with fewer than ten subreads that comprise over 75% of the test set. SSSL promises to better realize the potential of high-throughput DNA sequencing data
Submission Length: Regular submission (no more than 12 pages of main content)
Previous TMLR Submission Url: https://openreview.net/forum?id=go1hTFGbft&referrer=%5BTMLR%5D(%2Fgroup%3Fid%3DTMLR)
Changes Since Last Submission: We apologize for the formatting errors. The font has been changed and the style file import has been moved to the bottom to ensure clashing packages are overridden.
Assigned Action Editor: ~Andriy_Mnih1
Submission Number: 549
Loading