When many bacteria are sequenced together, for example from a swab of a hospital sink, the reads from every organism arrive mixed in one pile. A natural way to untangle them borrows a trick from wireless receivers. First reconstruct the most abundant genome. Then remove ("cancel") every read that matches it. Then assemble what is left, which is now dominated by the next organism. Radio engineers call this successive interference cancellation (SIC): a receiver decodes the strongest overlapping signal, subtracts it, and decodes the next one from a cleaner remainder. This procedure has a known failure mode, and bacteria hit it hard. Plasmids, transposons and other mobile genetic elements are shared across hosts. If they are cancelled together with the first genome, they disappear from everyone else's assembly. Those elements are exactly what clinicians care about, because they carry the antibiotic-resistance genes.
This note proposes a different subtraction rule. Cancel only the stable chromosomal backbone of each resolved genome, never its mobile or accessory regions. Choose the cancellation order by how reliable each backbone template is, not only by how abundant the genome is. Because the templates are stable, a central server can broadcast them cheaply to field sequencers. The devices then subtract known backbones locally and upload only the residual reads, which carry the plastic, resistance-bearing content. The claim is simple to state. In plasmid-rich communities, strongest-first full-genome cancellation loses shared mobile elements, and backbone-only cancellation keeps them at little cost in contiguity or bandwidth.
Core Claim
The wireless literature makes two points that translate directly to this setting.
The first is that decoding order matters. Optimal Decoding Order and Power Allocation for Sum Throughput Maximization in Downlink NOMA Systems shows that the optimal SIC order is not fixed and shifts with operating conditions.
The second is that imperfect cancellation propagates errors down the chain. Li et al. show that when residue survives subtraction, order must be chosen to keep that residue from cascading into later decoding steps.
In read cancellation, the residue does not come from noise. It comes from shared sequence. Any read from a plasmid present in hosts A and B that is cancelled with A is an error injected into B's decoding.
Biology tells us where that shared sequence lives. Short and long-term genome stability analysis of prokaryotic genomes models chromosomes as graphs of orthologous gene families. Across 277 genomes, it finds that pathogen instability is driven by insertions and deletions of accessory genes rather than by reshuffling of the core backbone (p = 0.024). Patterns and Implications of Gene Gain and Loss in the Evolution of Prochlorococcus finds a compact core of 1,273 genes inside a non-saturating 5,736-gene pan-genome, with flexible genes clustered in genomic islands. Cartography of opportunistic pathogens and antibiotic resistance genes in a tertiary hospital environment reports hospital plasmids acting as a "plastic framework" that shuffles resistance genes across stable chromosomal lineages, and more than 60% of the plasmid and phage sequences it reconstructed were novel.
Together these motivate the design. The backbone is the part of a genome that is safe to subtract, because it is stable and host-specific. The accessory complement is the part that must never be subtracted, because it is shared and changes quickly.
The novelty is narrow and testable. Prior work does assemble-then-subtract with whole-genome templates, or it maps reads competitively against whole references. Neither restricts cancellation to an empirically identified stable backbone. Neither treats the order of cancellation as an optimization target tied to template reliability. This note does both.
Research Program
The pipeline runs in rounds.
Step 1: assemble the dominant genome. Take the residual read pool from the previous round. Assemble it with a hybrid assembler, meaning one that uses long reads to bridge repeats and short reads to correct base errors. Extract the most complete chromosome.
Step 2: label backbone and accessory regions. Split that chromosome into the two classes at gene resolution. Backbone is the gene families shared with related complete genomes and in conserved order, as in the synteny-graph compression of Brilli et al. Accessory is everything else. Any region annotated as a mobile element or flanked by insertion sequences is forced into the accessory class. Such sequences are what fractured short-read assemblies of the resistance plasmid in Genomic characterization of an emerging Enterobacteriaceae species. They are also what fractured the transposon-borne carbapenemase structures in Hu.
Step 3: cancel only backbone reads. Remove a read only if most of its length aligns confidently to backbone. Any read that overlaps a backbone–accessory boundary is retained.
Step 4: choose the order. At each round, the next genome to cancel is chosen by a score. The score combines coverage, which is the analogue of channel strength, with template reliability. Reliability is the fraction of the template's length that is backbone and the estimated rate at which its cancelled reads would also align to other candidate templates. This makes the ordering question concrete. We compare strongest-first, reliability-first and a learned score on the same communities, and ask when they disagree.
Step 5: broadcast templates to field devices. The distributed version addresses the downlink bottleneck. Downlink here means server to device. A central server keeps the catalog of resolved backbones and broadcasts it to portable sequencers running on-device base calling, as in Nanopore base calling on the edge. The devices cancel reads locally and upload only residual reads. Because backbones change slowly, the broadcast can be coded as a difference against the previous catalog. This reuses the predict-and-transmit-the-residual idea of Communication-Efficient Federated Learning via Predictive Coding, applied to sequence rather than weights.
A caution on broadcast errors. Improved Convergence Analysis and SNR Control Strategies for Federated Learning in the Presence of Noise shows that downlink corruption is identical at every receiver and therefore does not average out. The genomic analogue is serious. A lossy template that wrongly includes an accessory island would cause the same misremoval at every site. We therefore treat the backbone/accessory boundary as information that must be sent losslessly, and compress only within backbone.
Relation to the closest wireless-learning work. When NOMA Meets AIGC: Enhanced Wireless Federated Learning and Multi-Carrier NOMA-Empowered Wireless Federated Learning optimize physical-layer decoding order and power for learning workloads. In contrast, this proposal imports the logic of SIC order and imperfect cancellation into the data domain itself. It takes radio resource allocation as given.
Baselines
Naive reference: co-assembly with no cancellation. Run the same hybrid assembler on the full mixed read pool, with the same long-read and short-read inputs and the same compute budget. It cannot lose shared plasmids through subtraction, so it is the right check on whether cancellation helps at all. It wins if low-abundance chromosomes assemble adequately even without cancellation.
Strongest-first whole-genome cancellation. At each round, remove every read matching the most abundant resolved genome. This is the direct analogue of the descending-gain rule that is optimal for throughput in Zhu et al.. It is also the strongest version of the subtraction idea. Running it with an identical assembler, aligner thresholds and number of rounds isolates the effect of what is cancelled.
Competitive joint mapping, the strongest existing competitor. Map all reads simultaneously against all resolved genomes and assign each read to its best hit. This is the analogue of joint decoding rather than successive decoding. I found no reported result on Lacuna for plasmid recovery from mock bacterial communities under either competitive mapping or iterative cancellation. Baseline numbers will therefore be measured, not quoted.
Downlink comparisons. We compare uncompressed template broadcast against difference-coded broadcast. For scale, Practical Hybrid Gradient Compression for Federated Learning Systems reports a 2,911× downlink compression ratio on CIFAR-10 with ResNet-18 at 90.26% accuracy. That figure is for model weights, not sequence. It frames what downlink coding can achieve but is not a target we must hit.
Ablations. These isolate the key ingredient by changing one component at a time:
- backbone-only cancellation under strongest-first order;
- whole-genome cancellation under reliability order;
- backbone labels from synteny conservation versus labels from mobile-element annotation alone.
What counts as beating the baselines. Backbone-only cancellation must recover significantly more shared plasmids completely and correctly hosted than both cancellation baselines. It must also keep total contig count and NGA50 (a contiguity measure computed against the true genomes) within a small tolerance of the best baseline. NGA50 is the metric on which Learning to Untangle Genome Assembly with Graph Convolutional Networks reports gains: 26 contigs versus 241 for Raven on CHM13 chromosome 1. That is a single-genome task, cited only for the metric.
Why not just…?
Why not just co-assemble everything and skip cancellation? For abundant organisms this may be fine. The motivating case is the minority host. The AF18 strain in Zhang et al. was under 1% of the population yet decisive for resistance. Whether co-assembly fragments such low-coverage chromosomes inside high-coverage backgrounds is an empirical question. The abundance sweep in the experiments below answers it directly. If co-assembly matches backbone-only cancellation on minority hosts, the proposal reduces to its distributed-bandwidth benefit, and we would report that.
Why not just use competitive mapping instead of successive cancellation? This is the real threat. It handles shared plasmids by refusing to commit early. Two concerns remain. It needs a complete reference catalog, while over 60% of hospital mobile elements were novel in Chng et al.. It also forces a single assignment for reads that are truly multi-host. Whether these matter is settled by measuring correctly hosted plasmid recovery under both methods at matched read pools. A hybrid in which competitive mapping is restricted to backbone templates is included as an arm. If that hybrid wins, the contribution is the backbone restriction rather than the successive order, and we would say so.
Why not just cancel strongest-first, which is the textbook optimum? It is optimal for sum throughput when interference is noise-like. Li et al. show that the rule must be justified by how residue propagates. Here the residue is structured and shared. The ordering ablation measures whether reliability-aware order adds anything beyond backbone restriction. If it adds nothing, order can simply stay strongest-first.
Proposed Experiments
Mock communities. We build communities from complete genomes with ground truth, including closed hybrid assemblies of the kind reported for AF18. We then insert the same plasmid into several hosts at controlled abundances: a 42 kb IncN resistance plasmid, plus transposon cassettes such as those described by Hu. Long reads are simulated at the roughly 92.4% median accuracy that DeepNano-coral reports for K. pneumoniae, which matches the field-device regime, and also at higher accuracy. Short reads are simulated at standard accuracy.
Real reads. Real nanopore reads from the E. coli subsets of NanoBaseLib are spiked into these mixtures. This tests robustness to real base-calling error, which NanoBaseLib places around 93–95% for current tools on R9.4 chemistry.
Primary metrics. These are:
- the fraction of shared plasmids recovered as circular contigs;
- correct host attribution of those plasmids;
- the count of mobile-element reads wrongly removed;
- per-genome contig count and NGA50.
Distributed metrics. These are downlink bytes per round, uplink residual volume, and the change in plasmid recovery when template broadcasts are deliberately corrupted.
Decisive check 1: misremoval curve. Strongest-first whole-genome cancellation should lose a growing share of shared-plasmid reads as the number of hosts carrying the plasmid rises. Backbone-only cancellation should stay flat.
Decisive check 2: corrupted broadcasts. Injecting errors into the backbone/accessory boundary of broadcast templates should cause correlated plasmid loss at all simulated sites. Injecting errors only within backbone should not. This would confirm that the boundary must be sent losslessly.
If check 1 fails, the central claim is wrong.
References
- Optimal Decoding Order and Power Allocation for Sum Throughput Maximization in Downlink NOMA Systems
- Achievable Rate Analysis of the STAR-RIS Aided NOMA Uplink in the Face of Imperfect CSI and Hardware Impairments
- Short and long-term genome stability analysis of prokaryotic genomes
- Patterns and Implications of Gene Gain and Loss in the Evolution of Prochlorococcus
- Cartography of opportunistic pathogens and antibiotic resistance genes in a tertiary hospital environment
- Genomic characterization of an emerging Enterobacteriaceae species: the first case of co-infection with a typical pathogen in a human patient
- Microbial Characteristics and Genomic Analysis of an ST11 Carbapenem-Resistant Klebsiella pneumoniae Strain Carrying blaKPC−2 Conjugative Drug-Resistant Plasmid
- Nanopore base calling on the edge
- Communication-Efficient Federated Learning via Predictive Coding
- Improved Convergence Analysis and SNR Control Strategies for Federated Learning in the Presence of Noise
- When NOMA Meets AIGC: Enhanced Wireless Federated Learning
- Multi-Carrier NOMA-Empowered Wireless Federated Learning With Optimal Power and Bandwidth Allocation
- Joint Power Control and Beamforming for Uplink Non-Orthogonal Multiple Access in 5G Millimeter-Wave Communications
- Practical Hybrid Gradient Compression for Federated Learning Systems
- Learning to Untangle Genome Assembly with Graph Convolutional Networks
- NanoBaseLib: A Multi-Task Benchmark Dataset for Nanopore Sequencing