Pipeline Architecture

BASALT is organised into three main functional modules operating in sequence:

flowchart LR
    A[Raw Reads + Assemblies] --> B[Autobinning]
    B --> C[Bin Selection]
    C --> D[Refinement<br/>DL + Contig Retrieval]
    D --> E[Reassembly]
    E --> F[Final MAGs]

Module 1: Autobinning + Bin Selection

Runs multiple binning algorithms in parallel across all assemblies, evaluates quality, and selects the optimal non-redundant binset.

Steps

Step	Script	Description
S1	S1_Autobinners_2qc_11152023.py	Run MetaBAT2, Maxbin2, CONCOCT, Semibin2 with multiple parameter sets
S1e	S1e_extra_binners.py	Run optional extra binners (MetaBinner, VAMB, LorBin)
S2	S2_BinsAbundance_PE_connections_*.py	Compute abundance profiles and paired-end connectivity
S3	S3_Bins_comparator_within_group_*.py	Compare bins within each assembly group, select best
S4	S4_Multiple_Assembly_Comparitor_*.py	Cross-assembly dereplication to remove redundant bins

Binners and Parameters

Binner	Parameter Values	Sensitivity
MetaBAT2	200, 300, 400, 500	All presets
Maxbin2	0.3, 0.5, 0.7, 0.9	more-sensitive only
CONCOCT	1–3 flexible thresholds	sensitive, more-sensitive
Semibin2	100 (default)	All presets
MetaBinner	100	Extra (-e m)
VAMB	100	Extra (-e v)
LorBin	100	Extra (-e l)

Module 2: Refinement

Removes contamination from individual bins using a deep learning ensemble, then retrieves missed contigs.

Steps

Step	Script	Description
S5	S5_Outlier_remover_DL_*.py	DL ensemble classifies each contig as Real or Contaminated; contaminated contigs are removed
S6	S6_retrieve_contigs_from_PE_contigs_*.py	Retrieve missing contigs using paired-end connectivity patterns
S7	S7_Contigs_retrieve_within_group_*.py	Intra-group contig retrieval and OLC refinement
S7lr	S7lr_finding_sr_contigs_basing_lr_and_polishing_*.py	Long-read-based contig retrieval + Racon polishing
S8	S8_OLC_new_*.py	Overlap-Layout-Consensus contig merging

Deep Learning Architecture

The S5 outlier remover uses an ensemble of 5 MLP models with the following architecture:

Input (TNF + Coverage + Coverage Change Ratio)
    → Linear(hidden_size)
    → LBR Block × 2 (Linear → BatchNorm → ReLU → residual skip)
    → Linear(num_classes=2)
    → Softmax → Real / Contaminated

Each LBR block expands to 4× hidden size and contracts back with a residual connection. The 5 models are trained independently and their predictions are combined via majority voting.

Module 3: Reassembly

Reassembles refined bins to improve genome quality and contiguity.

Steps

Step	Script	Description
S9	S9_Reassembly_*.py	Short-read reassembly using SPAdes
S9p	S9p_Hybrid_Reassembly_*.py	Hybrid reassembly using Unicycler (requires long reads)
S10	S10_OLC_new_*.py	Final Overlap-Layout-Consensus refinement

Checkpoint System

BASALT records progress after each step in Basalt_checkpoint.txt. The checkpoint file tracks:

1st autobinner done!
2nd bin selection within group done!
3rd bin selection within multiple groups done!
4th outlier removal done!
5th contig retrieval done!
6th long-read contig retrieval done!
7th intra-group contig retrieval done!
8th reassembly done!
9th final OLC done!

Use --mode continue to resume from the last completed step. Use --mode new to start fresh.

Data Feeding Workflow

The Data Feeding module (Data_feeding.py) allows users to bypass the Autobinning module entirely:

1. Rename and reindex external bins to be compatible with BASALT
2. Recompute coverage matrices using Bowtie2 mapping
3. Generate paired-end connection files
4. Run CheckM/CheckM2 quality assessment
5. Feed the prepared data directly into the Refinement module