DeepSME

Nanopore sequencers read DNA by measuring the ion current disturbance as a strand threads through a protein pore, and a neural basecaller converts that squiggle into letters. Every production basecaller is trained on canonical DNA, so a strand in which all cytosines have been replaced by 5-hydroxymethylcytosine (5hmC) is effectively unreadable: the modified bases perturb the current in ways the model has never seen, and 67.2% to 100% of reads fail outright. DeepSME, developed at Southern University of Science and Technology, is a basecaller trained from scratch for exactly that chemistry.

The application the authors build around it is private communication. Because heavy 5hmC substitution renders a strand illegible to any standard basecaller, the modification acts as a molecular key: the message is physically present in the DNA but cannot be decoded without the matching model. This inverts the usual framing of modified-base detection, where methylation is signal rather than cipher. It also distinguishes DeepSME from nanopore-signal models aimed at biological readouts, such as DamageFormer, which localizes DNA lesions in native samples.

The harder problem is bootstrapping. Training a basecaller normally requires reference alignments, unavailable when no existing model can read the data. DeepSME solves this with a three-stage, alignment-free pipeline that builds a k-mer current model from a small set of known synthetic sequences, expands it in silico, then fine-tunes on real modified bacterial genomic DNA. The work was published in Nature Communications in 2025.

#Key Features

• Purpose-built for heavy modification: Handles strands where 5hmC substitution is complete and motif-insensitive, rather than the sparse, sequence-context-dependent methylation that modification-aware basecallers target.
• Alignment-free bootstrap: The first stage derives a 6-mer current dictionary from eleven known synthetic sequences, removing the need for a reference-quality basecaller that does not yet exist for the chemistry.
• In silico k-mer expansion: Simulated signal from 50 microbial genomes lifts coverage from 4,096 six-mer contexts to the full 262,144 nine-mer space, turning a toy dictionary into a general basecaller.
• End-to-end message recovery: At 16x coverage the pipeline reconstructs a fully concealed text file and image, where commodity basecallers recover almost nothing from the same reads.
• Deployable on existing tooling: The released checkpoints are Bonito- and Dorado-compatible, dropping into standard nanopore workflows and onto embedded GPUs such as the Jetson Xavier AGX.

#Technical Details

DeepSME modifies the Bonito encoder–decoder basecaller. The first stage uses a CTC decoding layer with the convolutional stride reconfigured to 1 for higher temporal resolution; the later stages adopt the conditional random field decoder of Bonito's dna_r9.4.1_e8_sup@v3.3 model. Stage one, 6-mer QC DeepSME, is trained on eleven fully 5hmC-substituted synthetic sequences of 1,145–1,341 nt sequenced on an R9.4.1 MinION flow cell, and observes 4,093 of the 4,096 possible 6-mers. Stage two, Enhanced DeepSME, trains on 118,299 chunks simulated with Squigulator across 50 microbial genomes using that dictionary, lifting precision from effectively zero to 82.87%. Stage three, Reinforced DeepSME, fine-tunes on 118,096 chunks of real 5hmC-substituted genomic DNA from Pseudomonas aeruginosa PAO1, Aeromonas hydrophila BJ054, and Vibrio cholerae E1, reaching 92.99% precision and 92.93% recall.

The full schedule takes 266.4 minutes and 19.9 GB of VRAM on an RTX 3090. On the DNA-storage benchmark — 487 single-stranded 243 nt sequences encoding a 978-byte text file and a 7,775-byte image — DeepSME reaches an F1 of 0.864 and decodes the text completely at 16x coverage, rising from 52.73% of strands at 4x and 94.5% at 8x. Guppy 6.0 in high-accuracy mode decodes 5.45% of strands (F1 0.025) and Bonito/Dorado SUP 3.64% (F1 0.040) on the same data.

#Applications

The immediate use is steganographic DNA data storage, where chemical modification puts access control in the molecule rather than in software. Beyond that, the pipeline is a recipe: any lab facing a nucleotide chemistry no existing basecaller supports can follow the same three stages from a handful of known synthetic sequences. That matters for synthetic biology and xenonucleic-acid work, and for epigenetic assays where 5hmC density breaks canonical basecalling.

#Impact

DeepSME's contribution is less a checkpoint than a demonstration that basecallers for novel nucleotide chemistries can be bootstrapped without reference alignments, at modest compute on hardware a single lab can afford. The peer-reviewed release ships the final Reinforced weights and raw pod5 training signal on Zenodo alongside Apache-licensed code; the paper is open access but non-commercial, no-derivatives, and the companion DNA-storage decoding scripts carry no license. Only the stage-one 6-mer QC training and k-mer extraction are scripted, the Enhanced and Reinforced stages fall back to stock Bonito training, and neither intermediate checkpoint is deposited. The scope is narrow by construction: trained for R9.4.1 chemistry and complete 5hmC substitution, it is not a drop-in modification detector for native samples with partial or mixed methylation, and its 92.99% precision stays below what canonical basecallers achieve on unmodified DNA.