ChimeraLM

A long read whose two halves align to different genomic loci is the primary evidence a structural-variant caller uses to declare a translocation or a large deletion. When the library was built by whole-genome amplification — the standard route for single cells and any other low-input sample — some of those reads are not variants at all: the amplifying polymerase switched templates mid-extension and welded two unrelated loci into one molecule. No alignment threshold separates the two cases: the artifact generates exactly the signal the caller looks for.

ChimeraLM, from Rendong Yang's group at Northwestern University Feinberg School of Medicine, moves the decision off the alignment and onto the sequence. It reads a mapped BAM, extracts every primary read carrying a supplementary-alignment (SA) tag — precisely the reads that go on to become SV evidence — and classifies each one as biological or as an amplification chimera from its nucleotides. Predicted artifacts are dropped; the surviving BAM goes to whatever SV caller the pipeline uses. The premise is that template switching leaves a compositional trace a genomic language model can learn even though no alignment rule captures it.

It is the DNA counterpart of DeepChopper from the same lab: the same HyenaDNA backbone and the same goal of removing library-preparation chimeras before they are read as biology, but a different artifact and a different granularity — DeepChopper labels individual bases as adapter in nanopore direct-RNA reads, while ChimeraLM emits one label per read. It was released as open-source software in October 2025 with a pretrained checkpoint on HuggingFace. There is no accompanying paper; the authors ask that it be cited as software.

#Key Features

• Read-level verdict from sequence alone: Each read is scored whole and returned as 0 (biological) or 1 (chimeric artifact), from base composition rather than alignment geometry, mapping quality or base-quality strings.
• Scoped to the reads that matter: BAM parsing keeps only mapped, primary reads bearing an SA tag, so inference runs on the chimeric alignments that would otherwise become SV calls instead of on the entire library.
• Fixed published checkpoint: chimeralm predict loads yangliz5/chimeralm from HuggingFace by default, with no training or per-dataset fitting required; the local-checkpoint override is a hidden flag, not part of the documented workflow.
• Filtering that returns an analysis-ready BAM: chimeralm filter applies the predictions and emits a sorted, indexed BAM, which drops into a shell, Nextflow or Snakemake pipeline between alignment and SV calling.
• Runs anywhere, with a visual mode: Inference works on CPU, CUDA or Apple Silicon, and a Gradio interface — a HuggingFace Space, or chimeralm ui locally — scores single sequences with confidence values.

#Technical Details

ChimeraLM is a 5.1-million-parameter binary classifier on the hyenadna-small-32k-seqlen checkpoint of HyenaDNA, a genomic language model whose implicit long convolutions replace attention and give it a 32,768-token context at single-nucleotide resolution — long enough to hold a typical long read end to end. The backbone's 256-dimensional per-position hidden states are collapsed into one read-level vector by a learned attention-pooling layer, then passed through a two-layer, 512-unit GELU head with residual connections and 0.1 dropout to a two-way softmax. The backbone is not frozen: it is fine-tuned with the head under cross-entropy loss, using AdamW at 1e-4 with weight decay 0.01 and ReduceLROnPlateau scheduling, and validation F1 selects the checkpoint.

The released weights were trained on real whole-genome-amplified long-read data; that corpus is not distributed, and no benchmark results have been published. The repository also carries training configurations for CNN, Mamba, transformer and Evo 2 backbones alongside the HyenaDNA one, but the shipped checkpoint is the HyenaDNA variant. Distribution is a pip install chimeralm package with Rust extensions for BAM parsing; the bundled example is 175 reads from the PC3 prostate cancer cell line, sequenced on a Nanopore MinION Mk1C after amplification.

#Applications

Single-cell and low-input DNA sequencing studies benefit most, since amplification is unavoidable there and the false structural variants it creates are a dominant noise source: single-cell cancer genomics tracing subclonal rearrangements, preimplantation and forensic genetics, and work on uncultured microbes. Because it consumes and produces ordinary BAM files, it is a drop-in quality-control stage between alignment and any existing SV caller rather than a replacement for one.

#Impact

ChimeraLM treats an amplification artifact as a sequence-recognition problem rather than an alignment-filtering one, and shows that a small fine-tune of a pretrained genomic language model suffices to make the call — a pattern this lab has now applied to two chimera problems in two sequencing modalities. Its limits are equally clear: it targets whole-genome-amplification chimeras specifically, decides only for reads that already carry a supplementary alignment, and returns a binary label with no breakpoint-level explanation. Adoption is early and no published evaluation covers libraries unlike its training data, so groups deploying it should measure retention and artifact rates themselves.