Boltz-2 Virtual Screening

Screen a ligand library against a protein sequence with Boltz-2: co-folded complex structures plus predicted binding affinity, ranked into a shortlist.

by Temple Compute

Run Boltz-2 Virtual Screening on Horus

W-01 · Boltz-2 Virtual Screening

Screen a ligand library against a protein sequence with Boltz-2: co-folded complex structures plus predicted binding affinity, ranked into a shortlist.

Domain: Drug Discovery

Overview

Screens a ligand library against a protein sequence with Boltz-2, which co-folds each protein-ligand pair and predicts binding affinity, then ranks the ligands into a shortlist. The workflow has 3 tasks.

Pipeline

prep    target.fasta + ligands.smi ──► results/boltz_inputs.tar.gz (one YAML per ligand)
predict boltz predict              ──► results/predictions.tar.gz
rank    affinity + confidence      ──► results/top_hits.csv

Inputs / Outputs

fasta: target protein FASTA; the first record is used.
ligands: one SMILES [name] per line; the name becomes the output id.
Outputs: results/boltz_inputs.tar.gz, results/predictions.tar.gz, results/top_hits.csv.

results/top_hits.csv is sorted best first (highest binder probability, then lowest predicted value):

| Column | Meaning | |---|---| | rank | 1 = best | | ligand | name from ligands.smi | | affinity_probability_binary | Boltz-2 probability that the ligand binds | | affinity_pred_value | predicted affinity, log10(IC50 in µM); lower is tighter | | confidence_score | structure confidence of the complex |

Parameters

| id | Default | Meaning | |---|---|---| | use_msa | on | Fetch the MSA from the ColabFold server; off runs single-sequence | | recycling_steps | 3 | Boltz trunk recycling iterations (1-10) | | sampling_steps | 200 | Diffusion steps per structure (50-500) | | diffusion_samples | 1 | Structure samples per complex (1-10) | | accelerator | gpu | gpu or cpu |