OmegaPulldown

Structure-based protein–protein interaction screening with Boltz-2 and AlphaFold 3.

Analogous to AlphaPulldown (Yu et al., Bioinformatics 2023) but built on a Nextflow pipeline that runs Boltz-2 and AlphaFold 3, two next-generation predictors with native PTM support, under a single ranking schema.

Features

• Bait × candidate screening — one-vs-many or many-vs-many, from two FASTA files
• Dual backend — Boltz-2 and AlphaFold 3, selectable with --predictor boltz2 | alphafold3
• Post-translational modifications — specified via a simple CSV using CCD codes (e.g., NAG, SEP, TPO)
• MSA caching — pre-computed A3M files are reused across pairs and across runs; MMSEQS_SEARCH is skipped for proteins with cached alignments
• Unified ranking — ipTM, pTM, pDockQ, and pLDDT reported in a single TSV for either backend

Pipeline stages

1. GENERATE_PAIRS — enumerate bait × candidate combinations
2. MMSEQS_SEARCH — compute MSAs per unique protein via mmseqs2 (skipped if pre-computed)
3. Input assembly — A3M → Boltz-2 CSV or AF3 JSON with cleaned/patched MSAs
4. Structure prediction — BOLTZ_PREDICT or AF3_PREDICT (GPU)
5. PARSE_METRICS — extract confidence scores and compute pDockQ per pair
6. RANK_PREDICTIONS — aggregate and rank all pairs into a single TSV

Quick start

# Boltz-2 on a Slurm cluster
nextflow run main.nf -profile slurm \
  --baits_fasta inputs/sequences/baits.fasta \
  --candidates_fasta inputs/sequences/candidates.fasta \
  --mmseqs_db /path/to/uniref90

# AlphaFold 3
nextflow run main.nf -profile slurm \
  --predictor alphafold3 \
  --baits_fasta inputs/sequences/baits.fasta \
  --candidates_fasta inputs/sequences/candidates.fasta

# With pre-computed MSAs (skips MMSEQS_SEARCH)
nextflow run main.nf -profile slurm \
  --alignments_dir results/alignments

The -profile slurm flag is required for GPU-accelerated prediction on HPC clusters. Without it, Nextflow runs locally and predictions will fail if no GPU is available.

Profiles

Profile	Description
standard	Local execution (default)
slurm	Slurm scheduler with GPU partitions
sge	Sun Grid Engine
test	Quick test with reduced sampling steps

Post-translational modifications

Place a CSV file alongside each FASTA (e.g., bait_ptms.csv next to baits.fasta), or specify paths with --bait_ptms and --candidate_ptms:

protein_id,position,ccd
MyProtein,84,SEP
MyProtein,142,NAG

Codes follow the RCSB PDB Chemical Component Dictionary.

Sampling strategy

• Boltz-2: 5 diffusion samples per prediction. Boltz-2 uses conditional flow matching, which preserves structural diversity across samples.
• AlphaFold 3: 5 seeds × 1 diffusion sample. Within-seed samples show very narrow variance in practice, so seed-to-seed variation (independent MSA subsampling and trunk computation) better captures prediction uncertainty.

Repository structure

OmegaPulldown/
├── main.nf                 # Pipeline entry point
├── nextflow.config         # Parameters and profiles
├── modules/
│   ├── search_msa.nf       # MMSEQS_SEARCH, A3M_TO_BOLTZ_CSV
│   ├── prepare_input.nf    # GENERATE_PAIRS, MAKE_BOLTZ_YAML
│   ├── predict.nf          # BOLTZ_PREDICT
│   ├── predict_af3.nf      # MAKE_AF3_JSON, AF3_PREDICT
│   ├── rank.nf             # PARSE_METRICS, RANK_PREDICTIONS
│   └── rank_af3.nf         # PARSE_METRICS_AF3
├── bin/
│   └── make_af3_json.py    # AF3 JSON builder with A3M cleaning
├── inputs/
│   └── sequences/          # FASTA files and PTM CSVs
├── extras/
│   └── pipeline_subway.png # Pipeline schematic
└── results/                # Output (not tracked in git)

Requirements

• Nextflow ≥ 22.10
• mmseqs2 (for local MSA computation)
• Boltz-2 (conda environment)
• AlphaFold 3 (Singularity container). Model weights require a license agreement with Google DeepMind and should be stored in each user's own directory (set af3_model_dir in nextflow.config).
• Python 3.9+ with numpy

Local setup

The shipped nextflow.config contains placeholder paths for the conda environment, mmseqs2 database, and AlphaFold 3 container/databases. If you are authorized to use the Whitehead Institute computing resources, contact Troy for the pre-configured local paths.

References

• Yu D, Chojnowski G, Rosenthal M, Kosinski J. AlphaPulldown — a python package for protein–protein interaction screens using AlphaFold-Multimer. Bioinformatics 2023;39(1):btac749.
• Passaro S et al. Boltz-2: Exploring the frontiers of biomolecular structure prediction. bioRxiv 2025. DOI:10.1101/2025.06.14.659707.
• Abramson J et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 2024;630:493–500.