Metagenome misassembly detection and correction that renders read-to-contig alignments as multi-channel images scored by a contrastive CNN.
No providers recorded yet. Browse all providers
Metagenome-assembled genomes are only as good as the assemblies they come from, and short-read metagenome assembly is error-prone: strain variation, repeats shared between species, and uneven coverage all conspire to splice sequence from one organism onto another. A chimeric contig that survives into binning contaminates a MAG and, downstream, the biological conclusion drawn from it. Reference-free assembly evaluators exist, but the established tools rely on hand-crafted statistics over the read alignments, and their thresholds transfer poorly between datasets with different read lengths, insert sizes, and community complexity.
DeepMM, from Hong Kong Baptist University, takes the vision-model route instead. It converts the alignment of reads back onto their own assembly into a multi-channel image — one channel per alignment signal, positions along the contig as the spatial axis — and lets a convolutional network learn what a misassembly looks like, rather than encoding that knowledge as rules. A contrastive pretraining stage teaches the encoder to place different views of the same misassembly close together in representation space before the supervised detection head is trained.
Beyond detection, DeepMM localises the breakpoint and emits corrected contigs, which is what makes it usable as a pipeline step rather than a diagnostic. It was posted as a bioRxiv preprint in February 2025.
The encoder is a ResNet — ResNet18 by default, with ResNet50 also supported — operating on a 5 x window-length image built from the BWA alignment of reads to the assembly. The contrastive pretraining stage learns misassembly representations from unlabelled windows; the downstream head then classifies windows and localises breakpoints for correction. On the five CAMI benchmark datasets DeepMM reports the highest AUPRC among the reference-free misassembly detectors compared. Applied to a large real metagenomic dataset from a diarrhoea-predominant irritable bowel syndrome (IBS-D) cohort, correcting the assemblies with DeepMM before binning raised the count of near-complete MAGs from 905 to 1,006.
The implementation is on GitHub under the MIT licence and installs as a command-line tool
(DeepMM correct, DeepMM finetune) with BWA and samtools as external dependencies. Two
pretrained weight sets are distributed via Zenodo: a default model for a (200, 20)
insert-size profile and a CAMI-tuned model for (270, 27). Fine-tuning to a new sequencer
chemistry is supported through a documented workflow using simulated reads.
DeepMM belongs in the quality-control slot of any short-read metagenomics pipeline that produces MAGs — human gut microbiome studies, environmental surveys, and food or clinical microbiology alike. The concrete benefit is recovering more high-quality genomes from the same sequencing investment, since contigs that would otherwise be discarded or would contaminate a bin are repaired instead. It also gives assembly-method developers a reference-free way to compare assemblers on real data where no ground-truth genomes exist.
The result that matters practically is the IBS-D cohort: correcting misassemblies before binning yielded roughly 11% more near-complete MAGs, which is a larger return than most downstream parameter tuning provides. The image-plus-contrastive-learning framing also sidesteps the brittleness of hand-tuned alignment statistics that limited earlier reference-free evaluators. The evidence base is short-read data, so behaviour on long-read and hybrid assemblies is not established by these benchmarks, and the work is still a preprint awaiting peer review. Code under MIT and pretrained weights on Zenodo make it straightforward to evaluate on a new cohort.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.