Short-read RNA sequencing cuts every RNA molecule into 100–150-letter fragments. That makes it impossible to tell which distant splicing choices sit on the same molecule, so full-length gene variants go uncounted. Nanopore direct RNA sequencing avoids this. It threads an intact native RNA molecule through a pore and records the electrical current, and a neural network called a basecaller converts that current trace into bases. The approach only pays off if the basecaller is accurate. Its errors are not uniform either: they shift with sequencing chemistry, organism, and chemically modified bases. Today's best basecallers look at the trace in short fixed windows, so they cannot see properties that belong to the molecule as a whole.
We propose to give the basecaller a cheap view of the entire molecule. The model computes one averaged "summary" vector over the whole read at linear cost, borrowed from linear-time speech encoders, and every local window is conditioned on it. When a new domain arrives, adaptation then touches mainly this summary pathway. This mirrors speaker adaptation in speech recognition for atypical voices. The test that matters is downstream: whether the adapted basecaller recovers full-length splice-junction chains more accurately on a domain it was not trained on.
Core Claim
Basecalling errors in direct RNA sequencing have a large per-molecule component. Part of the error comes from properties of the whole read, such as how fast the molecule moves through the pore, drift in the pore signal, chemistry, and density of modified bases. A window-local model can only estimate these properties from its 4,096 samples.
If this is true, a whole-read summary vector has two effects. First, it should reduce error on the training distribution. Second, and more importantly, it should give a small, well-placed set of parameters that adapts to a new domain from little labelled data, much better than fine-tuning a chunked model end to end.
Three constraints shape the design. Quadratic attention makes full-read context expensive. Speech work shows that a mean-based summary is enough in encoders but harmful in decoders. Speech adaptation work shows that frozen general-purpose features fail under domain shift.
Research Program
The method comes from SummaryMixing. That work replaces self-attention in speech encoders with an average of per-frame projections, $\bar{s} = \frac{1}{T}\sum_{t=1}^{T} f(x_t)$, where $x_t$ is the frame at time $t$ and $T$ is sequence length. The average is broadcast back to each frame. The paper matched or beat attention on LibriSpeech while roughly halving memory.
We make one change: $\bar{s}$ is computed over the entire read, not the current chunk. Computing $\bar{s}$ is a running mean, so it can be done causally over visible samples, as in the streaming SummaryMixing Conformer. Each chunk's encoder then receives it through a small modulation:
$$h_t = \gamma(\bar{s}) \odot \mathrm{Enc}(x_{t-w:t+w}) + \beta(\bar{s})$$
Here $\mathrm{Enc}$ is the unchanged local chunk encoder over a window of half-width $w$, and $\gamma, \beta$ are small learned maps that output a per-channel scale and shift.
The local encoder and decoder follow Melchior, a hybrid Mamba-Transformer basecaller on 4,096-sample chunks. Local base decoding keeps full local attention or Connectionist Temporal Classification (CTC), a loss that aligns frame outputs to a shorter label sequence. We do not use the summary in the decoder because SummaryMixing's authors report that mean-based mixing fails there.
Adaptation borrows from clinical speech recognition. Zheng et al. adapted wav2vec 2.0 to Parkinsonian speech and found that jointly predicting impairment severity helped the most severe speakers. We add an analogous auxiliary head on $\bar{s}$ that predicts read-level attributes available for free: chemistry kit, species, and, on synthetic constructs with known modifications, whether the molecule is modified.
For a new domain we compare three regimes, each given the same small labelled budget. The first adapts only $f$, $\gamma$ and $\beta$. The second fine-tunes everything. The third freezes the model entirely.
The first step is diagnostic and can kill the idea cheaply. On the Melchior five-species benchmark, we basecall with the published chunked model and decompose per-chunk error rate into between-read and within-read variance (an intraclass correlation). If read identity explains a trivial share of error, whole-read conditioning has nothing to exploit, and we stop.
Baselines
The naive reference is the chunked basecaller applied unchanged to the new domain, with no whole-read context and no adaptation. It tells us how much domain shift costs.
The strongest existing method on the benchmark we can reproduce is Melchior. Its authors report 94.40% percent identity on human reads and a 1.17% human mismatch rate on its five-species R9.4/R9.4.1 benchmark (116,072 training reads, 100,000 test reads). We retrain it from the released architecture on identical splits and chunking. It then receives the same labelled target-domain data and the same GPU-hour and hyperparameter-search budget as our model.
Coral is the strongest reported accuracy overall. Its authors report about 97% median read accuracy on RNA001/002, 99.52% on RNA004, and a 26% gain in annotated isoform detection over Dorado. We treat it as an external reference on isoform detection. We will not claim to beat it unless we can run it on our splits. No reported Dorado accuracy figure appears in our sources; only Coral's relative improvement over it does.
Beating Melchior means lower mismatch and indel rates on the held-out domain at matched adaptation budget. It also means more correctly recovered junction chains downstream. Improvements must hold across three seeds.
Two ablations isolate the key ingredient. The first computes $\bar{s}$ over the current chunk only rather than the whole read. This separates "a modulation layer" from "molecule-level context." The second replaces the learned summary with hand-computed per-read statistics: mean and spread of current, and dwell-time estimates. This tests whether the network needs to learn the summary at all.
Why not just…?
Why not just fine-tune the chunked basecaller end to end on the new domain? This is the real threat. Speech results show full fine-tuning works when enough target data exists: Sanguedolce et al. cut AphasiaBank word error rate from 35.8% to 21.5%. They also found small models transferred poorly, and they had about seven hours of data. Our claim is narrower. At small target budgets, adapting a molecule-level pathway should beat full fine-tuning and forget the source domain less. We sweep the target budget from a few hundred to tens of thousands of reads and plot both curves. If full fine-tuning matches us at every budget, the adaptation half of the claim is refuted.
Why not just use full attention over the whole downsampled read, as Coral does, or a Mamba backbone over longer chunks? Coral's encoder attention is quadratic in read length after a fixed 12x downsampling. Melchior's authors report that pure Mamba overfit, which is why they kept attention layers. A single mean vector is the cheapest possible carrier of whole-read context. We will report memory and throughput against a Coral-style long-context encoder on the longest reads. If long attention gives no accuracy gain over the summary, that supports the claim that per-molecule context, not pairwise long-range interaction, is what matters.
Why not just normalize each read with fixed statistics before basecalling? Per-read signal normalization already removes some per-molecule variation. The hand-crafted-statistics ablation tests exactly this. If it closes most of the gap, the learned summary is unnecessary, and we would report that honestly.
Why not freeze a pretrained basecaller and add a small output head? Balagopalan et al. found frozen speech embeddings discard the pathological cues needed for Alzheimer's detection. The Parkinson's study likewise needed encoder adaptation. The frozen regime is included in the budget sweep, and we expect it to be worst under shift.
Why not skip nanopores and sequence short reads deeper? RefShannon improves short-read assembly through sparse flow decomposition. However, multi-junction phasing stays underdetermined. The colorectal cancer long-read atlas found 299 dysregulated structures built from multiple distant splicing events, including splicing choices coupled over kilobases. LRGASP concluded that read length matters more than depth for identifying transcripts. This proposal accepts that premise and targets the accuracy of the long reads themselves.
Proposed Experiments
The primary experiment uses the Melchior five-species benchmark in a leave-one-species-out design. We train on four species and adapt to the fifth, with C. elegans and human as primary targets, then evaluate the full budget sweep. Metrics are percent identity, mismatch, insertion and deletion rates, homopolymer accuracy, and aligned-read counts. The Epinano synthetic constructs supply modification labels. We report error within a small window of known modified positions separately, since this is where molecule-level context should help most.
The downstream check, where the short-read motivation is actually tested, uses the ONT direct RNA data in LRGASP for human WTC11 and mouse, together with the SIRV spike-ins. We basecall with each system, then run the same isoform callers (IsoQuant, FLAIR, Bambu). The measure is accuracy of unique junction chains against SIRV ground truth and against the GENCODE-annotated held-out genes. LRGASP reports that isoforms supported by only one or two tools validated at just 50–90%, versus 100% for concordant calls. We therefore also measure whether better basecalling raises cross-tool concordance on novel junction chains.
A secondary check tests whether the mechanism is specific to nanopores. We apply the same summary-only adaptation to a SummaryMixing speech encoder on SAP-1005 and compare against full fine-tuning at matched budgets. Zheng et al. report 26.92% word error rate from standard fine-tuning of wav2vec 2.0 and 26.53% with their best weighted and multi-task setups. A matching pattern on speech would suggest the idea is general; a mismatch would localize it to nanopore physics.
The idea is decisively supported under three conditions. Read identity must explain a substantial share of error in the diagnostic. Whole-read summaries must beat both chunk-local summaries and hand-crafted statistics. Summary-only adaptation must beat full fine-tuning of Melchior at low budgets, with improved junction-chain accuracy on the SIRV spike-ins. Failure of the diagnostic or of either ablation refutes it.
References
- SummaryMixing: A Linear-Complexity Alternative to Self-Attention for Speech Recognition and Understanding
- Linear Time Complexity Conformers with SummaryMixing for Streaming Speech Recognition
- Melchior: A Hybrid Mamba-Transformer RNA Basecaller
- Fine-Tuning Automatic Speech Recognition for People with Parkinson's: An Effective Strategy for Enhancing Speech Technology Accessibility
- A dual context-aware basecaller for nanopore direct RNA sequencing
- When Whisper Listens to Aphasia: Advancing Robust Post-Stroke Speech Recognition
- Comparing Acoustic-Based Approaches for Alzheimer's Disease Detection
- RefShannon: A genome-guided transcriptome assembler using sparse flow decomposition
- An isoform-resolution transcriptomic atlas of colorectal cancer from long-read single-cell sequencing
- Systematic assessment of long-read RNA-seq methods for transcript identification and quantification