Home
AMBER bins metagenomic contigs from ancient DNA (aDNA) samples using damage-aware contrastive learning and quality-guided Leiden clustering. It is optimised for paleogenomic assemblies where post-mortem DNA damage, short fragment lengths, and mixed ancient/modern populations make standard binners unreliable.
Post-mortem deamination converts cytosines to uracil (read as T) at fragment termini. In standard binners this:
- Shifts apparent tetranucleotide composition of each contig → damaged and undamaged reads from the same genome look like different organisms
- Biases coverage profiles → ancient reads (50–150 bp) cover reference positions differently from modern reads (150–300 bp)
- Mixes populations → assemblies contain both ancient (damaged) and modern (undamaged) strains; binners conflate or fragment them
| Problem | AMBER solution |
|---|---|
| Composition distortion | Damage-aware InfoNCE: downweights negatives with incompatible damage signatures |
| Coverage bias | 20-dimensional aDNA feature vector captures damage, fragment length, decay |
| Mixed populations |
amber deconvolve: EM separation of ancient/modern reads before or after binning |
| Bin fragmentation | Quality-guided Leiden: SCG-penalised edges + contamination splitting + near-HQ rescue |
| Run-to-run stochasticity |
amber resolve: co-binning consensus from multiple independent runs |
- Quick Start Tutorial — full walkthrough from BAM to HQ bins
- Methods and Model — damage model math, InfoNCE derivation, Leiden quality refinement, EM deconvolution
- aDNA Features — all 157 embedding dimensions explained
- Command Reference — every flag for every subcommand
- Output Formats — file schemas for bins, abin, damage TSV, deconvolution outputs
- Benchmarks — HQ bin counts vs baselines on KapK sediment data
# Build (no GPU)
cmake -DCMAKE_BUILD_TYPE=Release .. && cmake --build . --parallel
# Bin (GPU/CPU)
amber bin --contigs contigs.fa --bam alignments.bam \
--output bins/ --threads 16
# Aggregate runs
amber resolve --runs run1/run.abin run2/run.abin run3/run.abin --output consensus/
# Deconvolve ancient/modern
amber deconvolve --contigs ref.fa --bam alignments.bam --output deconvolve/