CReadNet

CReadNet is a contrastive encoder that maps short DNA sequences — k-mers drawn from sequencing reads — into an embedding space organized by genomic position. Sequencing produces enormous numbers of reads carrying no positional information, and recovering where each came from is normally the job of an aligner such as BWA or Bowtie2, which indexes a linear reference and searches it. CReadNet reframes that lookup as representation learning: k-mers from nearby positions are trained to land close together, making a genomic region a cluster and a chromosome a trajectory through the latent space.

The training signal comes from the data's own structure rather than from task labels. Two k-mers taken from the same read, or within a short offset along a reference, form a positive pair; everything else in the batch is a negative. That objective never sees an alignment, a variant call, or a species label, yet the frozen embedding supports several downstream tasks through small prediction heads: read mapping, structural-variant detection, and metagenomic species identification. Because the encoder can be trained purely from reads, it also applies where no assembled reference exists — a setting where classical aligners have nothing to index.

The work is by Filip Thor and Carl Nettelblad of the Division of Scientific Computing at Uppsala University and SciLifeLab, posted to arXiv in May 2025 as a preprint not yet peer reviewed. It sits apart from DNA language models such as DNABERT and Nucleotide Transformer, which pretrain on masked-token objectives over reference sequence; CReadNet's objective is positional and its input is read-level.

#Key Features

• Position-structured embedding space: Training only to place co-located k-mers together produces a latent space in which genomic order appears as continuous trajectories, visible directly in UMAP projections.
• Supervised and self-supervised modes: With a linear reference, a distance threshold Γ promotes every k-mer within Γ base pairs of the anchor to a positive, weighted by genomic distance; without one, the loss falls back to pairs drawn from the same read.
• Damage-aware augmentation: A domain-specific noise model reproducing ancient-DNA deamination and fragmentation is applied during training, improving embedding robustness on degraded reads.
• Bitwise coordinate prediction: Prediction heads emit the digits of the coordinate in a small base rather than regressing it or binning it, turning one large-cardinality classification into a handful of small ones that scale logarithmically with genome size.
• Constant-cost inference: Throughput is essentially independent of reference length, whereas index-based aligners slow down as the reference grows.

#Technical Details

The encoder is a one-dimensional ConvNeXt-style residual network chosen for translation invariance over short sequences, with a stride-1 stem, regular rather than depthwise convolutions, and four compute stages. Three sizes are reported: CReadNet-T (16.8M parameters), CReadNet-S (29.8M), and CReadNet-B (66.5M), each ending in global average pooling and a linear layer that yields 256-dimensional L2-normalized embeddings. Prediction heads are three-layer MLPs of 2,048 units trained on the frozen representation, in three variants: MSE regression, an independent-digit classifier, and a one-block GPT that emits coordinate digits autoregressively. A cheap local alignment pass over a small window around the predicted coordinate recovers base-exact placement.

Experiments use the E. coli K-12 reference genome (NC_000913, 4.64 Mbp) with k = 30, evaluated on simulated 30 bp ancient-DNA reads generated with the Briggs damage model. Trained on the first 10% of the genome, CReadNet-S and CReadNet-B both map 98.78% of reads correctly against 98.02% for BWA-aln at its recommended aDNA setting of three allowed mismatches, with BWA-mem at 54.76% and Bowtie2 at 9.78%. On the full genome CReadNet-B reaches 97.47% with the bitwise head and 97.66% with the GPT head, against 97.08% for BWA-aln. CReadNet-T holds roughly 0.42 million reads per second across both settings while BWA-aln falls from 0.375 to 0.230. The same frozen embedding detects 1 kbp engineered inversions and separates species in a small metagenomic demonstration.

#Applications

Ancient DNA is the clearest fit: reads are short, chemically damaged, and poorly served by general-purpose aligners, and the method reaches gold-standard accuracy without an index. Metagenomics is the second target, since the embedding assigns reads to clusters rather than to one reference genome and inference cost does not grow with the reference collection. Structural-variant screening follows from embedding geometry, with abnormally large distances between k-mers flagging inversion boundaries and the neighbor distribution distinguishing an inversion from a repeat.

#Impact

The paper demonstrates that a purely positional contrastive objective yields a genomic representation general enough to be reused, and that bitwise coordinate prediction makes position regression scale. Its scope is genuinely narrow. All results come from a single 4.64 Mbp bacterial genome; performance on human-scale references or real metagenomic communities is argued from favorable scaling rather than measured, and mapping accuracy already declines from the 10% subset to the full E. coli genome. No code or trained weights accompany the preprint.