Beyond the Alphabet: Deep Signal Embedding for Enhanced DNA Clustering

TL;DR

This work addresses the clustering bottleneck in DNA storage by shifting from alphabet-based read clustering to embedding raw Nanopore signals. It introduces a deep neural embedding built on a truncated Dorado CRF feature extractor trained with ArcFace loss to produce discriminative signal embeddings for clustering before basecalling. Across three real datasets, the approach achieves near-optimal discrimination (e.g., AUC up to $0.99$) and orders-of-magnitude faster computation than read-based methods, while preserving all ground-truth clusters and enabling robust reconstruction. The results suggest that raw-signal analysis can substantially improve the efficiency and accuracy of the DNA storage pipeline, with potential applicability to broader sequencing and bioinformatics tasks.

Abstract

The emerging field of DNA storage employs strands of DNA bases (A/T/C/G) as a storage medium for digital information to enable massive density and durability. The DNA storage pipeline includes: (1) encoding the raw data into sequences of DNA bases; (2) synthesizing the sequences as DNA \textit{strands} that are stored over time as an unordered set; (3) sequencing the DNA strands to generate DNA \textit{reads}; and (4) deducing the original data. The DNA synthesis and sequencing stages each generate several independent error-prone duplicates of each strand which are then utilized in the final stage to reconstruct the best estimate for the original strand. Specifically, the reads are first \textit{clustered} into groups likely originating from the same strand (based on their similarity to each other), and then each group approximates the strand that led to the reads of that group. This work improves the DNA clustering stage by embedding it as part of the DNA sequencing. Traditional DNA storage solutions begin after the DNA sequencing process generates discrete DNA reads (A/T/C/G), yet we identify that there is untapped potential in using the raw signals generated by the Nanopore DNA sequencing machine before they are discretized into bases, a process known as \textit{basecalling}, which is done using a deep neural network. We propose a deep neural network that clusters these signals directly, demonstrating superior accuracy, and reduced computation times compared to current approaches that cluster after basecalling.

Figures (6)

• Figure 1: The stages in the DNA storage pipeline.
• Figure 2: Deep signal embedding scheme, using a trimmed version of the Dorado model with an appended linear layer, using the ArcFace Loss function
• Figure 3: The test set is visualized using tSNE, with each color representing a distinct cluster. The presence of 500 clusters makes it difficult to distinguish similar colors visually. Each dot-like shape in the Figure represents a single cluster that is primarily well-separated from the other clusters.
• Figure 4: Signal similarity vs. sequence similarity across the different experiments
• Figure 5: Compute times for three different experiments. For the signal-model it includes both the embeddings and the cosine similarity computation times; for the different strand-$k=xx$-models it includes edit-distance similarity matrix computation time. The red cross indicates the omitted entries in which the running time is larger than 30 hours