Direct RNA nanopore basecaller that translates raw ionic-current signal into nucleotide sequence using a hybrid Mamba-Transformer backbone.
No providers recorded yet. Browse all providers
Nanopore sequencing reads a molecule by measuring the ionic current as it threads through a pore, and basecalling is the step that converts that current trace into letters. It is harder than it sounds: several nucleotides sit inside the pore at once, so no current level maps to a single base; homopolymer runs barely change the signal at all; and translocation speed varies, so trace length is a poor proxy for sequence length. Direct RNA basecalling has received far less attention than DNA, despite being the modality that reads native transcripts with their modifications intact.
Melchior, from Elon Litman at Stanford University, attacks this with an architecture argument. Convolutional basecallers see only a local receptive field and cannot reason about context hundreds of timesteps away; attention-only models can, but scale quadratically and empirically neglect fine local detail. Melchior alternates MambaVision Mixer blocks — a selective state-space token mixer with near-linear cost — with multi-head self-attention blocks, so each layer type compensates for the other's blind spot.
The design deliberately borrows from vision: the same hybrid pattern that set an accuracy-throughput frontier on ImageNet is repurposed for a 1D current trace, treating each timestep as a token. The related Caduceus DNA models showed selective state spaces work on nucleotide sequence; Melchior applies them one level upstream, to the signal that produces the sequence.
The network is isotropic: a convolutional stem, then a 20-block backbone at embedding dimension 768 with 8 attention heads and a 3072-dimensional MLP, then an adaptive average-pooling head that compresses 4096 input timesteps to 420 output positions before projecting to five classes (A, C, G, T, N) and a log-softmax. Training used connectionist temporal classification loss with label smoothing, a 5% linear warm-up to a peak learning rate of 6e-4 followed by cosine decay, batch size 32, dropout, and stochastic weight averaging, for 16 epochs over 223 hours on one A100-SXM4. The final model has 134 million parameters.
Data follows the RODAN benchmark: 116,072 training reads on R9.4/R9.4.1 flow cells spanning Arabidopsis thaliana, Homo sapiens, Caenorhabditis elegans, Escherichia coli, and Epinano synthetic constructs, chunked into 4096-value segments and normalised by median absolute deviation. Testing used 100,000 held-out reads from human, Arabidopsis, mouse, yeast, and poplar. Melchior posts the highest read identity on every species — 94.40% human, 94.13% Arabidopsis, 92.60% yeast, 92.42% poplar, 90.17% mouse — against GCRTcall, RODAN, and Guppy, with the lowest insertion and mismatch rates in nearly every case; the margin over the next-best method, GCRTcall, is a few tenths of a percentage point. Mean aligned reads reach 99,253, against 98,763 for GCRTcall and 98,034 for RODAN. Comparisons are against these three basecallers only, not against ONT's current Dorado models.
Basecalling is the first irreversible step of every nanopore experiment, so accuracy gains propagate into transcript assembly, isoform and splice-junction calling, poly-A tail measurement, and variant detection. The homopolymer improvement matters specifically for microsatellite regions and poly-A tracts, which are both biologically informative and the classic failure mode of nanopore reads. Because the model is a drop-in that maps FAST5 signal to FASTQ, it slots into existing minimap2-based pipelines without changing anything downstream, and the near-linear backbone makes it a candidate for portable field sequencing where compute is limited.
Melchior's broader argument is that basecalling should be treated as a signal-architecture problem rather than a scaling problem, and that state-space and attention layers are complementary rather than competing choices for it. Its transparency is unusual for the field: ONT's production basecallers are closed, and the open third-party alternatives generally ship code without checkpoints or logs, which makes independent comparison difficult. That said, this is a single-author preprint with a small community footprint, evaluated on R9-era flow cells rather than the newer RNA004 chemistry, and its accuracy claims rest on the author's own benchmark rather than third-party replication. Code is MIT-licensed and the weights carry an MIT model card on HuggingFace, so nothing blocks a group from checking the numbers.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.