DeepChopper

A nanopore direct-RNA read that spans two chromosomes looks exactly like a gene fusion. Sometimes it is one. More often the library preparation ligated a sequencing adapter between two unrelated RNA molecules, and the aligner faithfully reports the chimera as a chimeric alignment. No alignment rule separates the two cases — artifact and biology produce the same evidence — but the artifact carries a tell the biology does not: an adapter sitting in the middle of the read.

DeepChopper, from Rendong Yang's group at Northwestern University Feinberg School of Medicine, turns that tell into a per-base classification problem. It reads the base-called FASTQ and labels every nucleotide as adapter or non-adapter, then splits the read at the predicted adapter boundaries and trims the 3′ end. Because it works on sequence and quality strings alone, it needs neither the raw squiggle nor an alignment and slots in immediately after basecalling. The model is built on HyenaDNA — a long-context genomic language model whose implicit convolutions reach across 32 kb of single-nucleotide tokens — with a classification head and a quality-aware block on top.

The preprint appeared in October 2024 as Deep-Chopper; the peer-reviewed version followed in Nature Communications in January 2026, and the software and checkpoints use the one-word form. A second contribution came with it: internal adapters turn up in 0.33–1.62% of reads across cell lines, species and both the RNA002 and RNA004 chemistries — a systematic error mode not previously characterized in the peer-reviewed literature.

#Key Features

• Single-nucleotide labelling over full transcripts: Each base gets its own adapter/non-adapter probability across contexts up to 32,770 nucleotides, long enough to hold more than 99.97% of annotated human transcripts.
• Quality-aware prediction: A block of MLPs with residual connections folds z-normalized per-base Q-scores into the sequence embedding, which lifts the test F1 from 0.97 to 0.99 in an ablation and helps separate real adapters from adapter-like motifs in low-quality regions.
• Recovery rather than filtering: Chimeric reads are split into their component sub-reads at the predicted junction instead of discarded, raising usable read yield by 3% on a 9.2-million-read prostate cancer dataset.
• Alignment-free and signal-free: Operating on base-called FASTQ avoids mapping bias, and complements Dorado's terminal trimming rather than replacing it.
• Two-command workflow on a fixed checkpoint: deepchopper predict then deepchopper chop, with Rust internals and optional GPU inference; the Python API loads pretrained weights and needs no training step from the user.

#Technical Details

DeepChopper is deliberately small: 4.6 million trainable parameters, against the billions in general-purpose genomic language models such as Evo. A HyenaDNA-small-32k backbone maps single-nucleotide tokens to 256-dimensional features, the quality block mixes in standardized base qualities, and a two-layer head emits a softmax over adapter and non-adapter per position. A sliding-window majority vote then refines raw calls into contiguous, plausible adapter boundaries. Training used 600,000 synthetic reads built from six SG-NEx human cell lines, split 480,000/60,000/60,000, over 60 epochs on two A100 GPUs.

On the held-out synthetic test set, precision, recall and F1 all exceed 0.99, while Pychopper, Porechop and Porechop_ABI — none designed for direct-RNA data — score near zero. On real VCaP prostate cancer dRNA-seq held out of training, DeepChopper reduced unsupported chimeric alignments by roughly 95% and raised the fraction of chimeric alignments corroborated by matched direct-cDNA sequencing from 5.8% to 48.7%; Breakinator, an alignment-rule chimera caller, cut artifacts by 62% but lost biologically supported signal. Runtime scales near linearly — 23 million reads in about 10.6 hours on two A100s. A separate checkpoint fine-tuned on VCaP RNA004 data adds a further 3–4% over the zero-shot RNA002 model.

#Applications

Any Oxford Nanopore direct-RNA experiment can use this as a preprocessing step. The clearest case is cancer transcriptomics, where adapter-bridged chimeras masquerade as fusion transcripts: on VCaP, DeepChopper-corrected reads produced 89% fewer FusionSeeker calls, and the removed calls were unsupported by matched short-read Arriba results and enriched for ribosomal protein genes. The same cleanup improves transcript annotation, novel isoform discovery, alternative splicing analysis and expression quantification. A HuggingFace Space runs single reads for inspection; production use is the pip-installable CLI.

#Impact

The contribution is as much diagnostic as computational: adapter-bridged chimeras are a routine feature of direct-RNA libraries, surviving both the RNA004 chemistry and Dorado's own trimming, and affecting tens to hundreds of thousands of reads per experiment. It also shows a 4.6M-parameter fine-tune of a genomic language model beating rule-based tooling on a task defined by fuzzy, error-corrupted motifs. The scope is narrow by design: it is not a basecaller, it targets direct-RNA rather than cDNA or PacBio data, and it cannot address reverse-transcription chimeras, which carry no internal adapter to detect. The authors note incomplete 3′-end detection in multi-adapter reads, and recommend pairing Dorado's terminal trimming with DeepChopper's internal-adapter calls.