Loading…
Direct RNA nanopore basecaller pairing a densely connected 1D CNN and CTC decoder with a Random Forest classifier that demultiplexes 24 barcodes.
No providers recorded yet. Browse all providers
Nanopore direct RNA sequencing (DRS) reads native RNA strands end to end, leaving modifications and poly(A) tails in place instead of erasing them through reverse transcription. Two costs come with that. Basecalling native RNA is harder than DNA — several nucleotides sit in the pore at once and homopolymer runs barely disturb the current — so read accuracy has trailed DNA basecallers. And a DRS flow cell has conventionally carried one sample, which means comparing two conditions means two runs and the batch effects that follow.
DEMINERS, from Lu Chen's group at Sichuan University, addresses both. Its learned core is Densecall, a basecaller whose architecture is borrowed from DenseNet: every layer inside a dense block receives the concatenated feature maps of all layers before it. Reusing features instead of relearning them is what buys the headroom, and Densecall spends it on width — growing to 1,024 channels and a 99-sample convolution kernel in the final dense block, so one layer sees a long stretch of the current trace without the parameter cost width normally carries. A connectionist temporal classification (CTC) head then aligns the variable-length signal to bases with no per-timestep label.
Around that sit the parts that make multiplexing work: a wet-lab protocol ligating one of 24 custom RNA transcription adapters to each sample, and DecodeR, a Random Forest classifier that reads the barcode from the raw squiggle by segmenting the adapter region into 100 normalised current units. Where Melchior and SqueezeCall tackle basecalling alone, DEMINERS ships basecalling, demultiplexing and downstream isoform and modification analysis as one pipeline.
densecall command line, so existing pipelines can adopt them without rewiring.Densecall consumes 4,096-value signal chunks normalised by median absolute deviation and decodes with a beam search of width 5. Training used one million signal chunks, plus 100,000 for validation, drawn from the RODAN corpus — Arabidopsis thaliana, Homo sapiens, Caenorhabditis elegans and Escherichia coli — under PyTorch 2.0.1 at batch size 32 for 30 epochs; a memory-optimisation technique cut training GPU memory from 6,628 MB to 4,700 MB at that batch and chunk size. On a 10-species test set spanning two mammals, two plants, a fungus, the malaria parasite Plasmodium berghei and four RNA viruses, accuracy matched RODAN with fewer insertions and beat Guppy; per species it exceeded RODAN on yeast, Arabidopsis, poplar and human and fell short on mouse. Fine-tuning on 20,000 mouse signals closed that gap, raising mouse accuracy from 87.84% to 90.44% against RODAN's 88.16% and Guppy's 84.06%. DecodeR reached 99.4% classification accuracy at a 0.5 probability cutoff across 24 barcodes with 73.3% read recovery, and 99.3% accuracy at 89.2% recovery across 10.
The multiplexed workflow suits studies that need many RNA samples compared under identical chemistry: the authors profiled glioma transcriptomes for m6A-linked isoform diversity, tracked modification changes across blood stages of the malaria parasite, and ran clinical metagenomics on COVID-19 respiratory swabs, assembling RNA virus genomes and calling variants from pooled libraries. Because reads keep their native modifications and poly(A) tails, the same demultiplexed data supports isoform quantification, poly(A) length estimation and single-read modification calling without separate assays.
DEMINERS is one of the few nanopore efforts to treat throughput and accuracy as one problem, and the barcoding protocol is the more consequential half: 24-plex DRS changes the economics of experiments that were previously one sample per flow cell. The basecalling gain is real but modest — parity with RODAN on the general model, with the clear win arriving only after species-specific fine-tuning — and read recovery falls to 73.3% at full 24-plex, so throughput is traded against retained reads. The published checkpoints are trained on R9.4.1 direct RNA chemistry, which Oxford Nanopore has since superseded with RNA004, so applying them to current flow cells requires retraining on new signal. Code and analysis scripts are released under GPL-3.0 and the checkpoints under CC BY 4.0, with the raw sequencing data deposited under a public BioProject accession.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.