Output Formats

Output Formats

amber bin outputs

bins/bin.*.fa

One FASTA file per bin. Headers are the original contig names from the input FASTA.

>contig_001234
ACGTACGT...
>contig_005678
TTACGCGT...

bins/amber_summary.tsv

Tab-separated summary of all bins.

Column	Description
bin	Bin name (e.g. bin.001)
n_contigs	Number of contigs
total_length	Total bin length (bp)
n50	N50 contig length
n_markers	Total SCG markers found
n_unique_markers	Unique SCG markers (≤ 1 copy each)
n_duplicate_markers	SCG markers present > 1 copy
est_completeness	Estimated completeness from SCG count (%)
est_contamination	Estimated contamination from SCG duplication (%)
p_ancient_mean	Mean p_ancient across all mapped reads
ct_5p_1	Mean C→T rate at 5′ position 1
ga_3p_1	Mean G→A rate at 3′ position 1
frag_len_mean	Mean fragment length (bp)
damage_class	ancient / modern / mixed

bins/damage_per_bin.tsv

Per-bin damage profile. Same columns as the amber damage output table below.

run.abin

Binary archive produced by amber bin and consumed by amber resolve. It carries everything needed to reconstruct or aggregate a binning run without re-running the encoder.

File layout:

magic[8]       "AMBR\x01\x00\x00\n"
n_contigs[4]   uint32 — total contigs in this run
chunks...      [tag:4][size:4][data:size] repeated
END\0[4]       sentinel chunk (size = 0)

All multi-byte integers are little-endian. Unknown chunk tags are skipped by the reader, so the format is forward-compatible.

Chunks:

Tag	Data layout	Contents
NAME	offsets[n×uint32] + data_len[uint32] + packed null-terminated strings	Contig name strings. Resolve using packed_data + offsets[i]
LNGT	int32[n]	Contig lengths in bp
EMBD	dim[uint16] + float32[n×dim]	L2-normalised encoder embeddings (128 dims)
FEAT	dim[uint16] + float32[n×dim]	Full 157-dim feature vector (128 encoder + 20 aDNA + 9 CGR) used for kNN
DAMG	dim[uint16] + float32[n×dim]	Raw damage profile (20 aDNA dims: C→T positions, G→A positions, λ, fragment length, coverage)
BINL	int32[n]	Bin label per contig; −1 = unbinned
SCGM	total_markers[uint32] + n_hits[uint32] + hits[n_hits × (contig_idx:uint32 + marker_id:uint16)]	Sparse SCG marker assignments. Each hit record maps one contig to one marker ID. total_markers is the size of the marker vocabulary (e.g. 107)
END\0	—	Sentinel; signals end of file

Why a custom binary format? A single .abin file captures the full run state in one read-compatible object. amber resolve memory-maps or streams N files in parallel, builds the co-binning affinity matrix from BINL and SCGM, and uses EMBD/FEAT only if re-clustering is requested. JSON or TSV would require separate files for embeddings, markers, and labels, making resolve fragile to partial runs.

---

amber resolve outputs

consensus_bins/bin.*.fa

Consensus bins after co-binning aggregation. Same FASTA format as amber bin.

consensus_bins/resolve_summary.tsv

Includes additional columns vs the amber bin summary:

Extra column	Description
n_runs_present	Number of input runs in which at least one contig appears
mean_p_cobin	Mean co-binning probability across contig pairs in the bin
jaccard_stability	Mean Jaccard similarity of this bin across input runs

---

amber deconvolve outputs

ancient_consensus.fa / modern_consensus.fa

Reference-guided consensus sequences called from ancient-weighted and modern-weighted reads respectively. Positions with weighted depth < --min-ancient-depth / --min-modern-depth are masked with N.

deconv_stats.tsv

Per-contig EM statistics.

Column	Description
contig	Contig name
n_reads_total	Total mapped reads
n_ancient	Reads with p_ancient > 0.8
n_modern	Reads with p_ancient < 0.2
n_ambiguous	Reads in [0.2, 0.8]
pi_ancient	Estimated mixture fraction (ancient)
ct_5p_1	C→T rate at 5′ position 1
frag_len_ancient	Mean fragment length of ancient reads
frag_len_modern	Mean fragment length of modern reads

deconv_uncertainty.tsv (with --write-stats)

Per-position posterior uncertainty.

Column	Description
contig	Contig name
pos	Reference position (0-based)
depth_ancient	Weighted depth from ancient reads
depth_modern	Weighted depth from modern reads
posterior_var	Variance of the p_ancient posterior at this position
base_ancient	Called base in ancient consensus
base_modern	Called base in modern consensus
is_discordant	1 if ancient ≠ modern base call

modern_reads.bam (with --write-modern-bam)

BAM file containing all reads with p_ancient ≤ --modern-bam-threshold. Both mates of a read pair are included if either mate meets the threshold. Suitable for de novo assembly of the modern fraction.

---

amber damage output

damage_per_bin.tsv

Column	Description
bin	Bin name
n_contigs	Number of contigs
n_reads	Total mapped reads
ct_5p_1 … ct_5p_5	C→T rates at 5′ positions 1–5
ga_3p_1 … ga_3p_5	G→A rates at 3′ positions 1–5
lambda_5p	Fitted exponential decay constant at 5′
lambda_3p	Fitted exponential decay constant at 3′
p_ancient	Fraction of reads classified as ancient
frag_len_mean	Mean fragment length (bp)
frag_len_std	Std dev of fragment length (bp)
damage_class	ancient / modern / mixed

---

Validation with CheckM2

checkm2 predict -i bins/ -o bins/checkm2 -x fa --threads 16

Key output: bins/checkm2/quality_report.tsv

Column	Description
Name	Bin name
Completeness	Estimated completeness (%)
Contamination	Estimated contamination (%)
Completeness_Model_Used	Neural Network (General/Specific)

MIMAG thresholds:

• HQ (High Quality): Completeness ≥ 90%, Contamination < 5%
• MQ (Medium Quality): Completeness ≥ 50%, Contamination < 10%
• LQ (Low Quality): below MQ